Selecting the optimal activation function for training deep neural networks has always been challenging because it significantly impacts the neural network’s performance and training speed. At this point, researchers are more likely to employ RELU than other well-known activation functions. After RELU, researchers have proposed many activation functions to improve RELU. None of them was capable of surpassing RELU as their most significant rival. SWISH outperformed RELU in several challenging experiments like classification, object detection, and tracking. Replacing RELU units with SWISH, which improves performance in many tasks. This paper proposed a new activation function surpassing Google’s brain’s SWISH function, Which we named AIF. Experiments indicate that our proposed activation function outperforms SWISH in various tasks.

Computer vision methods have shown to be effective in classifying garbage into recycling categories for waste processing, existing methods are costly, imprecise, and unclear. To tackle this issue, we introduce MWaste, a mobile application that uses computer vision and deep learning techniques to classify waste materials as trash, plastic, paper, metal, glass or cardboard. Its effectiveness was tested on various neural network architectures and real world images, achieving an average precision of 92% on the test set. This app can help combat climate change by enabling efficient waste processing and reducing the generation of greenhouse gases caused by incorrect waste disposal.

Auto-labeling is one of the main challenges in 3D vehicle detection. Auto-labeled datasets can be used to identify objects in LiDAR data, which is a challenging task due to the large size of the dataset. In this work, we propose a novel methodology to generate new 3D based auto-labeling datasets with a different point of view setup than the one used in the most recognized datasets (KITTI, WAYMO, etc.). The performance of the methodology has been further demonstrated with the development of our own dataset with the auto-generated labels and tested under boundary conditions on a bridge in a fixed position. The proposed methodology is based on the YOLO model trained with the KITTI dataset. From a camera-LiDAR sensory fusion, it is intended to auto-label new datasets while maintaining the consistency of the Ground Truth. The main contribution of this work is a novel methodology to auto-label autonomous driving datasets using YOLO as the main labelling system. The performance of this approach is measured retraining the contrast models of the KITTI benchmark.

Sometimes when diabetic retinopathy (DR) is found and treated quickly, vision loss can indeed be spared. This study deploys a deep learning (DL) model that can discover all 5 stages of DR more accurately than other methods. The proposed methodology shows two cases scenarios: case 1 with image enhancement using CLAHE and ESRGAN, and case 2 without image enhancement. Augmentation techniques are then employed to produce a balanced dataset with the identical criteria for both scenarios. The generated model using DenseNet-121 on the APTOS dataset outperformed other approaches for locating the 5 stages of DR, with an accuracy of 98.7 percent for case 1 and 81.2 percent for case 2. Using CLAHE and ESRGAN was shown to improve a model's performance and ability to learn.

Deep learning technologies for skin cancer detection have been dramatically advanced based on high resolution dermoscopic images. The low-cost approach based on clinical skin images of low resolution is promising but remains technically challenging due to their undermined image quality. In this paper, we propose a coarse-to-fine efficient super resolution transformer (CF-ESRT) network to reconstruct the dermoscopy-level high resolution skin image from a low resolution clinical image. By connecting the refinement network to the original super resolution transformers and applying perceptual and gradient losses, our framework noticeably improves the finer texture details of skin lesions in the super resolution (SR) images, and is effective to elevate the perceptual quality of the SR images. Quantitative and qualitative evaluations show that our method outperforms ESRT the basis model as well as the other state-of-the-art SR models.

Pulsed thermography is a vital technique in the nondestructive evaluation field. However, its data analysis can be complex and requires skilled experts. Advances in deep learning have yielded exceptional results, including image segmentation. Therefore, many efforts have been made to apply deep learning methods to data processing for nondestructive evaluation. Despite this, there is currently no public Pulsed thermographic dataset available for evaluating various spatial-temporal deep methods of segmenting pulsed thermographic data. This article aims to provide such a dataset and assess the performance of commonly used deep learning-based instance segmentation models on it. Additionally, the impact of the number of frames and data transformations on model performance is examined. The findings suggest that suitable preprocessing methods can effectively reduce the data size without compromising the deep models’ performance.

As AI models become more and more common in business and even in our daily lives, it is important to understand what the carbon impact of these models is. Recent papers have shown that this impact can be quite great, i.e., the training of a single high-end model can result in emissions of more than 500t of CO2eq. In this paper we discuss the factors that influence the carbon footprint of AI models, explore what impact different decisions have, and show how the footprint can be reduced. We also examine the footprint of different models to give a guideline on how urgent action is for different organizations.

One of the most common forms of dementia is Alzheimer’s disease (AD), which leads to progressive mental deterioration. Unfortunately, there is no definitive diagnosis and cure that can stop the condition progressing. The diagnosis is often performed based on the clinical history and neuropsychological data, including magnetic resonance imaging (MRI). Deep neural networks (DNN) algorithms are gaining popularity for medical diagnosis, and have been used widely for the analysis of MRI data. DNNs can extract hidden features from thousands of training images automatically. However, they cannot judge how confident they are about their predictions. To use DNNs in safety-critical applications such as medical diagnosis, uncertainty quantification of DNNs predictions is crucial. For this purpose, Monte Carlo dropout (MCD) has been widely used, however, it may lead to overconfident and miss calibrated results. This paper proposes a framework in which the MCD algorithm’s hyper-parameters are optimized during training using Bayesian optimization for the first time. The conducted optimization leads to assigning high predictive entropy to erroneous predictions and making it possible to recognize risky predictions. The proposed framework is used for AD diagnosis, which has not been done before. We compare our method with some existing methods in the literature based on different uncertainty quantification criteria. The results of comprehensive experiments on the Kaggle dataset using a deep model pre-trained on the ImageNet dataset show that the proposed algorithm can quantify uncertainty much better than the existing methods.

MicroRNAs (miRNAs) are small non-coding RNAs that play a central role in the post-transcriptional regulation of biological processes. miRNAs regulate transcripts by direct binding involving the Argonaute protein family. The exact rules of binding are not known, and several in silico miRNA target prediction methods have been developed to date. Deep Learning has recently revolutionized miRNA target prediction. However, the higher predictive power comes with decreased ability to interpret increasingly complex models. Here, we present a novel interpretation technique, called attribution sequence alignment, for miRNA target site prediction models that can interpret such Deep Learning models on a two-dimensional representation of miRNA and putative target sequence. Our method produces a human readable visual representation of miRNA:target interactions and can be used as a proxy for further interpretation of biological concepts learned by the neural network. We demonstrate applications of this method in clustering of experimental data into binding classes, as well as using the method to narrow down predicted miRNA binding sites on long transcript sequences. Importantly, the presented method works with any neural network model trained on a two-dimensional representation of interactions and can be easily extended to further domains such as protein-protein interactions.

Most machine learning and deep learning algorithms can only use low-dimensional data as input, but the data that must be processed in practical applications is diverse and irregular. There are two main problems with big dynamic data. (1) The size of the embedding table grows linearly with the vocabulary size, resulting in massive memory consumption. (2) For different newly added vocabularies. To solve these two problems, this paper proposes a novel embedding algorithm that can learn attribute associations with entities based on deep and hash algorithms. Taking movie data as an example, the encoding method and the specific flow of the algorithm are presented in detail, and the effect of dynamic reuse of data models is realized. Compared with the four existing embedding algorithms which can fuse entity attribute information, the deep hash embedding algorithm proposed in this paper has obvious optimization of time and space complexity.

Convolutional neural network architectures have become increasingly complex, which has improved the performance slowly on well-known benchmark datasets in the recent years. In this research, we have analyzed the true need for such complexity. We have introduced G-Net light, a lightweight modified GoogleNet with improved filter count per layer to reduce feature overlaps and complexity. Additionally, by limiting the amount of pooling layers in the proposed architecture, we have exploited the skip connections to minimize the spatial information loss. The investigations on the proposed architecture are evaluated on three retinal vessel segmentation publicly available datasets. The proposed G-Net light outperforms other vessel segmentation architectures by reducing the number of trainable parameters..

The classification of vessel types in SAR imagery is of crucial importance for maritime applications. However, the ability to use real SAR imagery for deep learning classification is limited, due to the general lack of such data and/or the labor-intensive nature of labeling them. Simulating SAR images can overcome these limitations, allowing the generation of an infinite number of datasets. In this contribution, we present a synthetic SAR imagery dataset with ship wakes, which comprises 46080 images for ten different real vessel models. The variety of simulation parameters includes 16 ship heading directions, 6 ship velocities, 8 wind directions, 2 wind velocities, and 3 incidence angles. In addition, we extensively investigate classification performance for noise-free, noisy, and denoised ship wake scenes. We utilize the standard AlexNet architecture and employ training from scratch. To achieve the best classification performance, we conduct Bayesian optimization to determine hyperparameters. Results demonstrate that the classification of vessel types based on their SAR signatures is highly efficient, with maximum accuracies of 96.16%, 92.7%, and 93.59%, when training using noise-free, noisy, and denoised datasets respectively. Thus, we conclude that the best strategy in practical applications should be to train convolutional neural networks on denoised SAR datasets. The results show that the versatility of the SAR simulator can open up new horizons in the application of machine learning to a variety of SAR platforms.

A common spinal condition, spondylolisthesis is the presence of a relative back or forth displacement between the upper and lower vertebra due to one vertebra being oriented away from the smooth curvature of a normal spine. Aging-related illnesses such as degenerative spondylolisthesis are especially burdensome on social welfare and health-care systems in an aging society, especially radiologists and clinical physicians. Therefore, we proposed a computer aided diagnosis algorithm, named LumbarNet, for vertebral slippage detection on clinical X-ray images. Collaborating with i) a P-grade, ii) a piecewise slope detection scheme, and iii) a dynamic shift detection routine, LumbarNet was thus specialized for analyzing complex structural patterns in lumbar spine X-ray images and outcompeted other U-Net based methods. Extensive experiments on lumbar spine X-ray images in standard clinical practices showed that LumbarNet achieved a mean intersection over union value of 0.88 in vertebral region detection and an accuracy of 88.83% in vertebral slippage detection.

Recently, the use of drones/unmanned aerial vehicles (UAVs) has notably increased due to their broad commercial spread and low cost. The wide diffusion of drones increases the hazards of their misuse in illegitimate actions such as drug smuggling and terrorism. Thus, the surveillance and automated detection of drones are crucial for safeguarding restricted regions or special zones from illegal drone interventions. One of the most challenging issues in drone detection in surveillance videos is the apparent similarity of drones and birds against complex backgrounds. In this work, an automated image-based drone-detection system utilizing an advanced deep-learning-based object-detection method known as you only look once (YOLOv5) is introduced for protecting restricted regions or special zones from unlawful drone interventions. Due to the lack of sufficient data, transfer learning was utilized to pretrain the object-detection method to increase the performance. The experiments showed outstanding results, and an average precision of 94.7% was accomplished.

Scientists and astronomers have attached Scientists and astronomers have attached great importance to the task of discovering new exoplanets, even more so if they are in the habitable zone. To date, more than 4300 exoplanets have been confirmed by NASA, using various discovery techniques, including planetary transits, in addition to the use of various databases provided by space and ground-based telescopes. This article proposes the development of a deep learning system for detecting planetary transits in Kepler Telescope lightcurves. The approach is based on related work from the literature and enhanced to validation with real lightcurves. A CNN classification model is trained from a mixture of real and synthetic data, and validated only with real data and different from those used in the training stage. The best ratio of synthetic data is determined by the perform of an optimisation technique and a sensitivity analysis. The precision, accuracy and true positive rate of the best model obtained are determined and compared with other similar works. The results demonstrate that the use of synthetic data on the training stage can improve the transit detection performance on real light curves.

Convolutional Neural Networks (CNNs) have gained prominence in the research literature on image classification over the last decade. One shortcoming of CNNs, however, is their lack of generalizability and tendency to overfit when presented with small training sets. Augmentation directly confronts this problem by generating new data points providing additional information. In this paper, we investigate the performance of more than ten different sets of data augmentation methods, with two novel approaches proposed here: one based on the Discrete Wavelet Transform and the other on the Constant-Q Gabor transform. Pretrained ResNet50 networks are finetuned on each augmentation method. Combinations of these networks are evaluated and compared across three benchmark data sets of images representing diverse problems and collected by instruments that capture information at different scales: a virus data set, a bark data set, and a LIGO glitches data set. Experiments demonstrate the superiority of this approach. The best ensemble proposed in this work achieves state-of-the-art performance across all three data sets. This result shows that varying data augmentation is a feasible way for building an ensemble of classifiers for image classification (code available at https://github.com/LorisNanni).

Online social networks (OSNs) play an integral role in facilitating social interaction; however, these social networks increase antisocial behavior, such as cyberbullying, hate speech, and trolling. Aggression or hate speech that takes place through short message service (SMS) or the Internet (e.g., in social media platforms) is known as cyberbullying. Therefore, automatic detection utilizing natural language processing (NLP) is a necessary first step that helps prevent cyberbullying. This research proposes an automatic cyberbullying method to detect aggressive behavior using a consolidated deep learning model. This technique utilizes multichannel deep learning based on three models, namely, the bidirectional gated recurrent unit (BiGRU), transformer block, and convolutional neural network (CNN), to classify Twitter comments into two categories: aggressive and not aggressive. Three well-known hate speech datasets were combined to evaluate the performance of the proposed method. The proposed method achieved promising results. The accuracy of the proposed method was approximately 88%.

This paper proposes a neural network which fuses the data received from a camera system on a gantry, to detect moving objects and calculate relative position and velocity of the vehicles traveling on a freeway, this information is used to estimate the traffic flow. To estimate the traffic flow at both microscopic and macroscopic view, this paper used YOLO v4 and DeepSORT for vehicle detection and tracking, then counting the number of vehicles pass through the freeway by drawing virtual lines and hot zones, also counting the velocity of each vehicles. The information is then pass to the traffic control center, in order to monitoring and control traffic flow on freeways, and analyzing freeway conditions.

The study aimed to identify different molds that grow on various food surfaces. As a result, we conducted a case study for the detection of mold on food surfaces based on the “you only look once (YOLO) v5” principle. In this context, a dataset of 2050 food images with mold growing on their surfaces was created. The dataset was trained using the pre-trained YOLOv5 algorithm. In comparison to YOLOv3 and YOLOv4, this current YOLOv5 model had better precision, recall, and average precision (AP), which were 98.10%, 100%, and 99.60%, respectively. The YOLOv5 algorithm was used for the first time in this study to detect mold on food surfaces. In conclusion, the proposed model successfully recognizes any kind of mold present on the food surface. Using YOLOv5, we are currently conducting research to identify the specific species of the detected mold.

We confirm that energy dissipation weighting provides the most accurate approach to determining the effective hydraulic conductivity (Keff) of a binary K grid. A deep learning algorithm (UNET) can infer Keff with extremely high accuracy (R2 > 0.99). The UNET architecture could be trained to infer the energy dissipation weighting pattern from an image of the K distribution with high fidelity, although it was less accurate for cases with highly localized structures that controlled flow. Furthermore, the UNET architecture learned to infer the energy dissipation weighting even if it was not trained on this information directly. However, the weights were represented within the UNET in a way that was not immediately interpretable by a human user. This reiterates the idea that even if ML/DL algorithms are trained to make some hydrologic predictions accurately, they must be designed and trained to provide each user-required output if their results are to be used to improve our understanding of hydrologic systems most effectively.

High-throughput screening technologies continues to produce large amounts of multiomics data from different populations and cell types for various diseases, such as cancer. However, analysis of such data encounters difficulties due to cancer heterogeneity, further exacerbated by human biological complexity and genomic variability. There is a need to redefine the drug discovery development pipeline, bringing an Artificial Intelligence (AI)-powered informational view that integrates relevant biological information and explores new ways to develop effective anticancer approaches. Here, we show SynPred, an interdisciplinary approach that leverages specifically designed ensembles of AI-algorithms, links omics and biophysical traits to predict synergistic anticancer drug synergy. SynPred exhibits state-of-the-art performance metrics: accuracy – 0.85, precision – 0.77, recall – 0.75, AUROC – 0.82, and F1-score - 0.76 in an independent test set. Moreover, data interpretability was achieved by deploying the most current and robust feature importance approaches. A simple web-based application available online at http://www.moreiralab.com/resources/synpred/ was constructed to predict synergistic anticancer drug combinations requiring only the upload of the two drug SMILES to be tested, allowing easy access by non-expert researchers.

White matter hyperintensity (WMH) has been considered the primary biomarker from small-vessel cerebrovascular disease to Alzheimer’s disease (AD) and has been reported for its correlation of brain structural changes. To perform WMH related analysis with brain structure, both T1-weighted (T1w) and (Fluid Attenuated Inversion Recovery(FLAIR) are required. However, in a clinical situation, it is limited to obtain 3D T1w and FLAIR images simultaneously. Also, the most of brain segmentation technique supports 3D T1w only. Therefore, we introduced the semi-supervised learning method that can perform brain segmentation using FLAIR image only. Our method achieved a dice overlap score of 0.86 for brain tissue segmentation on FLAIR, with the relative volume difference between T1w and FLAIR segmentation under 4.8%, which is just as reliable as the segmentation done by its paired T1w image. We believe our semi-supervised learning method has a great potential to be used to other MRI sequences and provide encouragement to people who seek brain tissue segmentation from a non-T1w image.

The evolution in imaging technologies and artificial intelligence algorithms, coupled with improvements in UAV technology, has enabled the use of unmanned aircraft in a wide range of applications. The feasibility of this kind of approach for cattle monitoring has been demonstrated by several studies, but practical use is still challenging due to the particular characteristics of this application, such as the need to track mobile targets and the extensive areas that need to be covered in most cases. The objective of this study was to investigate the feasibility of using a tilted angle to increase the area covered by each image. Deep Convolutional Neural Networks (Xception architecture) were used to generate the models for the experiments, which covered aspects like ideal input dimensions, effect of the distance between animals and sensor, effect of classification error on the overall detection process, and impact of physical obstacles on the accuracy of the model. Experimental results indicate that oblique images can be successfully used under certain conditions, but some practical limitations need to be addressed in order to make this approach appealing.

The explosion of data has transformed the world since much more information is available for collection and analysis than ever before. To extract valuable information from the data in different dimensions, various deep learning models have been developed in the past years. Although these models have demonstrated their strong capability on improving products and services in various applications, training them is still a time-consuming and resource-intensive process. Presently, cloud, one of the most powerful computing infrastructures, has been used for the training. However, how to manage cloud computing resources and to perform the training efficiently is still challenging current techniques. For example, general resource scheduling approaches, such as spread priority and balanced resource schedulers, actually do not work well with deep learning workloads. Besides, the resource allocation problem on a cluster can be divide into two subproblems: (1) local resource optimization: improve resource configuration for a single machine; (2) global resource optimization: improve the cluster-wide resource allocation. In this thesis, we propose two novel container schedulers, FlowCon and SpeCon, that are designed to address these two subproblems respectively and specifically to optimize performance of short-lived deep learning applications in the cloud. FlowCon focuses on resource configuration of single-node in a cluster, as show that it efficiently improves deep learning tasks completion time and resource utilization, and reduces the completion time of a specific job by up to 42.06\% without sacrificing the overall system time. SpeCon targets on cluster-wide resource configuration that speculatively migrate slow-growing models to release resources for fast-growing ones. Based on our experiments, SpeCon improves makespan for up to 24.7\%, compared to current approaches.

The advent of deep learning has completely reshaped our world. Now, our daily life is fulfilled with many well-known applications that adopt deep learning techniques, such as self-driving cars and face recognition. Furthermore, robotics developed more forms of technology which share the same principle with face recognition, such as hand pose recognition and fingerprint recognition. Image recognition technology requires a huge database and various learning algorithms, such as convolutional neural network and recurrent neural network, that requires lots of computational power, such as CPUs and GPUs. Thus, clients could not be satisfied with the computational resource of the local machine. The cloud resource platform emerged at a historic moment. Docker containers play a significant role of microservices-based applications in the next generation. However, it could not guarantee the quality of service. From clients’ perspective, they have to balance the budget and quality of experiences (e.g. response time). The budget leans on individual business owners and the required Quality of Experience (QoE) depends on usage scenarios of different applications, for instance, an autonomous vehicle requires real-time response, but, unlocking your smartphone can tolerate delays. Plenty of on-going projects developed user-oriented optimization resource allocation to improve the quality of the service. Considering the users’ specifications, including accelerating the training process and specifying the quality of experience, this thesis proposes two differentiate containers scheduling for deep learning applications: TRADL and DQoES .

Invasive species are significant threats to global agriculture and food security being the major causes of crop loss. An operative biosecurity policy requires full automation of detection and habitat identification of the potential pests and pathogens. Unmanned Aerial Vehicles (UAVs) mounted thermal imaging cameras can observe and detect pest animals and their habitats, and estimate their population size around the clock. However, their effectiveness becomes limited due to manual detection of cryptic species in hours of captured flight videos, failure in habitat disclosure and the requirement of expensive high-resolution cameras. Therefore, the cost and efficiency trade-off often restricts the use of these systems. In this paper, we present an invasive animal species detection system that uses cost-effectiveness of consumer-level cameras while harnessing the power of transfer learning and an optimised small object detection algorithm. Our proposed optimised object detection algorithm named Optimised YOLO (OYOLO) enhances YOLO (You Only Look Once) by improving its training and structure for remote detection of elusive targets. Our system, trained on the massive data collected from New South Wales and Western Australia, can detect invasive species (rabbits, Kangaroos and pigs) in real-time with a higher probability of detection (85–100 %), compared to the manual detection. This work will enhance the visual analysis of pest species while performing well on low, medium and high-resolution thermal imagery, and equally accessible to all stakeholders and end-users in Australia via a public cloud.

Face recognition is one of the most active research fields of computer vision and pattern recognition, with many practical and commercial applications including identification, access control, forensics, and human-computer interactions. However, identifying a face in a crowd raises serious questions about individual freedoms and poses ethical issues. Significant methods, algorithms, approaches, and databases have been proposed over recent years to study constrained and unconstrained face recognition. 2D approaches reached some degree of maturity and reported very high rates of recognition. This performance is achieved in controlled environments where the acquisition parameters are controlled, such as lighting, angle of view, and distance between the camera-subject. However, if the ambient conditions (e.g., lighting) or the facial appearance (e.g., pose or facial expression) change, this performance will degrade dramatically. 3D approaches were proposed as an alternative solution to the problems mentioned above. The advantage of 3D data lies in its invariance to pose and lighting conditions, which has enhanced recognition systems efficiency. 3D data, however, is somewhat sensitive to changes in facial expressions. This review presents the history of face recognition technology, the current state-of-the-art methodologies, and future directions. We specifically concentrate on the most recent databases, 2D and 3D face recognition methods. Besides, we pay particular attention to deep learning approach as it presents the actuality in this field. Open issues are examined and potential directions for research in facial recognition are proposed in order to provide the reader with a point of reference for topics that deserve consideration.

Learning to maintain postural balance while standing requires a significant fine coordination effort between the neuromuscular system and the sensory system. It is one of the key contributing factors towards fall prevention, especially in the older population. Using artificial intelligence (AI), we can similarly teach an agent to maintain a standing posture, and thus teach the agent not to fall. In this paper, we investigate the learning progress of an AI agent and how it maintains a stable standing posture through reinforcement learning. During training, the AI agent learnt three policies. First, it learnt to maintain the Centre-of-Gravity and Zero-Moment-Point in front of the body. Then, it learnt to shift the load of the entire body on one leg while using the other leg for fine tuning the balancing action. Finally, it started to learn the coordination between the two pre-trained policies. This study shows the potentials of using deep reinforcement learning in human movement studies. The learnt AI behaviour also exhibited attempts to achieve an unplanned goal because it correlated with the set goal (e.g. walking in order to prevent falling). The failed attempts to maintain a standing posture is an interesting by-product which can enrich the fall detection and prevention research efforts.

In the past years, various intelligent machine learning and deep learning algorithms have been developed and widely applied for gearbox fault detection and diagnosis. However, the real-time application of these intelligent algorithms has been limited, mainly due to the fact that the model developed using data from one machine or one operating condition has serious diagnosis performance degradation when applied to another machine or the same machine with a different operating condition. The reason for poor model generalization is the distribution discrepancy between the training and testing data. This paper proposes to address this issue using a deep learning based cross domain adaptation approach for gearbox fault diagnosis. Labelled data from training dataset and unlabeled data from testing dataset is used to achieve the cross-domain adaptation task. A deep convolutional neural network (CNN) is used as the main architecture. Maximum mean discrepancy is used as a measure to minimize the distribution distance between the labelled training data and unlabeled testing data. The study proposes to reduce the discrepancy between the two domains in multiple layers of the designed CNN to adapt the learned representations from the training data to be applied in the testing data. The proposed approach is evaluated using experimental data from a gearbox under significant speed variation and multiple health conditions. An appropriate benchmarking with both traditional machine learning methods and other domain adaptation methods demonstrates the superiority of the proposed method.

In this paper, our goal is to perform a virtual restoration of an ancient coin from its image. The present work is the first one to propose this problem, and it is motivated by two key promising applications. The first of these emerges from the recently recognised dependence of automatic image based coin type matching on the condition of the imaged coins; the algorithm introduced herein could be used as a pre-processing step, aimed at overcoming the aforementioned weakness. The second application concerns the utility both to professional and hobby numismatists of being able to visualise and study an ancient coin in a state closer to its original (minted) appearance. To address the conceptual problem at hand, we introduce a framework which comprises a deep learning based method using Generative Adversarial Networks, capable of learning the range of appearance variation of different semantic elements artistically depicted on coins, and a complementary algorithm used to collect, correctly label, and prepare for processing a large numbers of images (here 100,000) of ancient coins needed to facilitate the training of the aforementioned learning method. Empirical evaluation performed on a withheld subset of the data demonstrates extremely promising performance of the proposed methodology and shows that our algorithm correctly learns the spectra of appearance variation across different semantic elements, and despite the enormous variability present reconstructs the missing (damaged) detail while matching the surrounding semantic content and artistic style.

Speech signals are degraded in real-life environments, product of background noise or other factors. The processing of such signals for voice recognition and voice analysis systems presents important challenges. One of the conditions that make adverse quality difficult to handle in those systems is reverberation, produced by sound wave reflections that travel from the source to the microphone in multiple directions.To enhance signals in such adverse conditions, several deep learning-based methods have been proposed and proven to be effective. Recently, recurrent neural networks, especially those with long and short-term memory (LSTM), have presented surprising results in tasks related to time-dependent processing of signals, such as speech. One of the most challenging aspects of LSTM networks is the high computational cost of the training procedure, which has limited extended experimentation in several cases. In this work, we present a proposal to evaluate the hybrid models of neural networks to learn different reverberation conditions without any previous information. The results show that some combination of LSTM and perceptron layers produce good results in comparison to those from pure LSTM networks, given a fixed number of layers. The evaluation has been made based on quality measurements of the signal's spectrum, training time of the networks and statistical validation of results. Results help to affirm the fact that hybrid networks represent an important solution for speech signal enhancement, with advantages in efficiency, but without a significan drop in quality.

With the increase of the Artificial Neural Network (ANN), machine learning has taken a forceful twist in recent times. One of the most spectacular kinds of ANN design is the Convolutional Neural Network (CNN). The Convolutional Neural Network (CNN) is a technology that mixes artificial neural networks and up to date deep learning strategies. In deep learning, Convolutional Neural Network is at the center of spectacular advances. This artificial neural network has been applied to several image recognition tasks for decades and attracted the eye of the researchers of the many countries in recent years as the CNN has shown promising performances in several computer vision and machine learning tasks. This paper describes the underlying architecture and various applications of Convolutional Neural Network.

To generate a probability density function (PDF) for fitting probability distributions of real data, this study proposes a deep learning method which consists of two stages: (1) a training stage for estimating the cumulative distribution function (CDF) and (2) a performing stage for predicting the corresponding PDF. The CDFs of common probability distributions can be adopted as activation functions in the hidden layers of the proposed deep learning model for learning actual cumulative probabilities, and the differential equation of trained deep learning model can be used to estimate the PDF. To evaluate the proposed method, numerical experiments with single and mixed distributions are performed. The experimental results show that the values of both CDF and PDF can be precisely estimated by the proposed method.

Manipulating cloth-like deformable objects (CDOs) is a long-standing problem in the robotics community. CDOs are flexible (non-rigid) objects that do not show a detectable level of com- pression strength while two points on the article are pushed towards each otherand include objects such as ropes (1D), fabrics (2D) and bags (3D). In general, CDOs’ many degrees of freedom (DoF) introduce severe self-occlusion and complex state-action dynamics as significant obstacles for perception and manipulation systems. These challenges exacerbate existing issues of modern robotic control methods such as imitation learning (IL) and reinforcement learning (RL). This review focuses on the application details of data-driven control methods on four major task families in this domain: cloth-shaping, rope manipulation, dressing and bag manip- ulation. Furthermore, we identify specific inductive biases in these four domains that present challenges for more general IL and RL algorithms, and summarise the future direction for the development of the field.

The COVID-19 pandemic situation has created even more difficulties in the quick identification and screening of the COVID-19 patients for the medical specialists. Therefore, a significant study is necessary for detecting COVID-19 cases using an automated diagnosis method, which can aid in controlling the spreading of the virus. In this paper, the study suggests a Deep Convolutional Neural Network-based multi-classification approach (COV-MCNet) using eight different pre-trained architectures such as VGG16, VGG19, ResNet50V2, DenseNet201, InceptionV3, MobileNet, InceptionResNetV2, Xception which are trained and tested on the X-ray images of COVID-19, Normal, Viral Pneumonia, and Bacterial Pneumonia. The results from 3-class (Normal vs. COVID-19 vs. Viral Pneumonia) showed that only the ResNet50V2 model provides the highest classification performance (accuracy: 95.83%, precision: 96.12%, recall: 96.11%, F1-score: 96.11%, specificity: 97.84%) compared to rest of the models. The results from 4-class (Normal vs. COVID-19 vs. Viral Pneumonia vs. Bacterial Pneumonia) demonstrated that the pre-trained model DenseNet201 provides the highest classification performance (accuracy: 92.54%, precision: 93.05%, recall: 92.81%, F1-score: 92.83%, specificity: 97.47%). Notably, the ResNet50V2 (3-class) and DenseNet201 (4-class) models in the proposed COV-MCNet framework showed higher accuracy compared to the rest six models. This indicates that the designed system can produce promising results to detect the COVID-19 cases on the availability of more data. The proposed multi-classification network (COV-MCNet) significantly speeds up the existing radiology-based method, which will be helpful to the medical community and clinical specialists for early diagnosis of the COVID-19 cases during this pandemic.

The success of deep learning (DL) in offline brain-computer interfaces (BCIs) has not yet translated into efficient online applications. This is due to three primary challenges. First, most DL solutions are designed for offline decoding, making the transition to online decoding complex and unclear. Second, the use of sliding windows in online decoding increases the computational complexity of DL training. Third, DL models typically require large amounts of training data, which are often scarce in BCI applications. To address these issues and enable real-time decoding, even across different subjects without calibration data, we first introduce a novel approach called real-time adaptive pooling (RAP). RAP modifies the pooling layers of existing offline DL models towards the online decoding requirements. Additionally, it significantly reduces the computational demand during training by jointly decoding consecutive windows. To reduce the amount of training data required, our approach leverages different levels of domain adaptation. We show how different settings enable different adaptation solutions. The results demonstrate that our approach is both powerful and can be calibration-free, providing a robust and practical solution for real-time BCI applications. These findings pave the way for the development of co-adaptive and highly efficient DL-based online BCI systems.

Information dissemination and preservation are crucial for societal progress, especially in the technological age. While technology fosters knowledge sharing, it also risks spreading misinformation. Audio deep fakes, convincingly fabricated audio using Artificial intelligence (AI), exacerbate this issue. We present Sonic Sleuth, an AI model for detecting audio deepfakes. Our approach leverages Deep Learning (DL) for a robust detection model. Meticulous data preprocessing and rigorous experimentation with various models led to the implementation of the most effective solution with a custom CNN model. Our comprehensive testing resulted in a highly accurate model (98.27% accuracy, 0.016 EER) trained on a substantial dataset of real and synthetic audio. In addition to an 84.92% accuracy and 0.085 EER on an external dataset, these results demonstrate Sonic Sleuth’s potential as a powerful tool against audio misinformation.

Oral drug absorption is the primary route for drug administration. However, this process hinges on multiple factors, including the drug’s physicochemical properties, formulation characteristics, and gastrointestinal physiology. Given its intricacy and the exorbitant costs associated with experimentation, the trial-and-error method proves prohibitively expensive. Theoretical models have emerged as a cost-effective alternative by assimilating data from diverse experiments and theoretical considerations. These models fall into three categories: data-driven models, encompassing classical pharmacokinetics, quantitative-structure models (QSAR), and machine/deep learning. Mechanism-based models include quasi-equilibrium, steady-state, and physiologically based pharmacokinetics, alongside first principles such as molecular dynamics and continuum models. This review provides an overview of recent modeling endeavors across these categories, evaluating their respective advantages and limitations. Additionally, a primer on partial differential equations and their numerical solutions is included in the supplementary materials, recognizing their utility in modeling physiological systems despite their mathematical complexity limiting their widespread application in this field.

The number of connected IoT devices is increasing significantly due to their many benefits, including automation, improved efficiency and quality of life, and waste reducing. However, these devices have several vulnerabilities that have led to the rapid growth in the number of attacks. Therefore, several machine learning-based Intrusion Detection Systems (IDS) tools have been developed to detect intrusions and suspicious activity to and from a host (HIDS - Host IDS) or, in general, within the traffic of a network (NIDS - Network IDS). The proposed work performs a comparative analysis and an ablative study among recent machine learning-based NIDS to develop a benchmark of the different proposed strategies. The proposed work compares both shallow learning algorithms, such as Decision Trees, Random Forest, Naïve Bayes, Logistic Regression, XGBoost, and Support Vector Machine, and deep learning algorithms, such as DNN, CNN, and LSTM, whose approach is relatively new in the literature. The algorithms are evaluated on the KDD-99, NSL-KDD, UNSW-NB15, and IoT-23 datasets. The results show that the NIDS tools based on deep learning approaches achieve better performance in detecting network anomalies than shallow learning approaches.

In computer vision, human action recognition is a hot topic, popularized by the development of deep learning. Current models have achieved high accuracy results on public datasets. Despite this success, they require significant computational resources for training. Given that transfer learning based techniques allow reusing what other models have already learned and training models with less computational resources, in this work we propose using a transfer learning based approach for action recognition in videos. We describe a methodology for human action recognition using transfer learning techniques in a custom dataset. The proposed method consists of four stages: 1) human detection and tracking, 2) video preprocessing, 3) feature extraction (using pretrained models with ImageNet), and 4) action recognition using a two-stream model consisting of TCNs, LSTMs, and CNNs layers. The custom dataset is imbalanced with 189, 390, 490, 854, and 890 videos per class, respectively. For feature extraction, we analyzed the performance of seven pretrained models: Inception-v3, MobileNet-v2, MobileNet-v3-L, VGG-16, VGG-19, Xception, and ConvNeXt-L. We show that the best results were obtained with the last one. Finally, using pretrained models for feature extraction allowed training in a PC with a single GPU with an accuracy of 94.9%.

3D reality capturing has demonstrated increased efficiency and consistently accurate outcomes in architectural digitisation. Nevertheless, despite advancements in data collecting, 3D reality capturing still lacks full automation, especially in the post-processing and modelling phase. Artificial intelligence (AI) has been a significant focus, especially in computer vision, and tasks such as image classification and object recognition might be beneficial to the digitisation process and its subsequent utilisation. This article aims to examine the potential outcomes of integrating AI technology into the field of 3D reality-capturing, with a particular focus on its use in architectural scenarios. The main methods used for data collection are laser scanning (static or mobile) and photogrammetry. As a result, image data including RGB-D data (files containing both RGB colours and Depth information) and point clouds have become the most common raw datasets available for object mapping. The study comprehensively analyses the current use of 2D and 3D deep learning techniques in documentation tasks, particularly downstream applications. It also highlights the ongoing research efforts in developing real-time applications with the ultimate objective of achieving generalisation and improved accuracy.

Ensuring precise segmentation of colorectal polyps holds critical importance in the early diagnosis and treatment of colorectal cancer. Nevertheless, existing deep learning-based segmentation methods are fully-supervised, requiring extensive precise manual pixel-level annotation data, which leads to high annotation costs. Additionally, it remains challenging to train large-scale segmentation models when confronted with limited colonoscopy data. To address these issues, we introduce the general segmentation foundation model—Segment Anything Model (SAM) into the field of medical image segmentation. Fine-tuning the foundation model is an effective approach to tackle sample scarcity. However, current SAM fine-tuning techniques still rely on precise annotations. To overcome this limitation, we propose WSPoly-SAM, a novel weakly-supervised approach for colonoscopy polyp segmentation. WSPoly-SAM utilizes weak annotations to guide SAM in generating segmentation masks, which are then treated as pseudo-labels to guide the fine-tuning of SAM, thereby reducing the dependence on precise annotations data. To improve the reliability and accuracy of pseudo-labels, we have designed a series of enhancement strategies to improve the quality of pseudo-labels and mitigate the negative impact of low-quality pseudo-labels. Experimental results on five medical image datasets demonstrate that WSPoly-SAM outperforms current fully-supervised mainstream polyp segmentation networks on the Kvasir-SEG, ColonDB, CVC-300, and ETIS datasets. Furthermore, by using different amounts of training data in weakly-supervised and fully-supervised experiments, it is found that weakly-supervised fine-tuning can save 70% to 73% of annotation time costs compared to fully-supervised fine-tuning. This study provides a new perspective on the combination of weakly-supervised learning and SAM models, significantly reducing annotation time and offering insights for further development in the field of colonoscopy polyp segmentation.

Classical Arabic poetry is a form of Arabic literature, and almost every piece of Arabic poetry follows one of the 16 meters. The rhythmic framework of the poems or the verses is called meters. The deep learning (DL) model was implemented using TensorFlow in this work using a large dataset of Arabic poetry. The character level encoding was used to convert text to integers for classifying the full-verse and half-verse data. The study evaluates the data without removing diacritics from the text. The train-test-split method with a 70-15-15 split was employed for classification. 15% of the total data was considered unseen test data for all models. The work was conducted with multiple deep learning models, including the Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU), and Bidirectional LSTM (Bi-LSTM). The Bi-LSTM model shows the best accuracy of all the models specified, with 97.53% for full-verse and 95.23% for half-verse without removing diacritic data.

Proteins need to be located in appropriate spatiotemporal contexts to carry out their diverse biological functions. Mislocalized proteins may lead to a broad range of diseases, such as cancer and Alzheimer’s disease. Knowing where a target protein resides within a cell will give insights into tailored drug design for a disease. As the gold validation standard, the conventional wet lab uses fluorescent microscopy imaging, immunoelectron microscopy, and fluorescent biomarker tags for protein subcellular location identification. However, the booming era of proteomics and high-throughput sequencing generates tons of newly discovered proteins, making protein subcel-lular localization by wet-lab experiments a mission impossible. To tackle this concern, in the past decades, artificial intelligence (AI) and machine learning (ML), especially deep learning methods, have made significant progress in this research area. In this article, we review the latest advances in AI-based method development in three typical types of approaches, including sequence-based, knowledge-based, and image-based methods. We also elaborately discuss existing challenges and future directions in AI-based method development in this research field.

This paper analyses the application of deep learning techniques for predicting wave overtopping events in port environments using sea state and weather forecasts as inputs. The study was conducted in the outer port of Punta Langosteira, A Coruña, Spain. A video recording infrastructure was installed to monitor overtopping events from 2015 to 2022, identifying 3709 overtopping events. The data collected was merged with actual and predicted data for the sea state and weather conditions during the overtopping events, creating three datasets. We used these datasets to create several machine learning models to predict whether an overtopping event would occur based on sea state and weather conditions. The final models achieved a high accuracy level during the training and testing stages: 0.81, 0.73, and 0.84 average accuracy during training and 0.67, 0.48, and 0.86 average accuracy during testing, respectively. The results of this study have significant implications for port safety and efficiency, as wave overtopping events can cause disruptions and potential damage. Using deep learning techniques for overtopping prediction can help port managers take preventative measures and optimize operations, ultimately improving safety and helping to minimize the economic impact that overtopping events have on the port's activities.

Yield calculation is an important link in modern precision agriculture, which is an effective means to improve breeding efficiency, and adjust planting and marketing plans. With the continuous progress of artificial intelligence and sensing technology, yield calculation schemes based on image processing technology have many advantages such as high accuracy, low cost, and non-destructive calculation, and have been favored by a large number of researchers. This article reviews the research progress of crop yield calculation based on remote sensing images and visible light images, describes the technical characteristics and applicable objects of different schemes, and focuses on detailed explanations of data acquisition, independent variable screening, algorithm selection and optimization. Common issues are also discussed and summarized. Finally, solutions are proposed for the main problems that have arisen so far, and future research directions are predicted, with the aim of achieving more progress and wider popularization of yield calculation solutions based on image technology.

This study presents a novel approach for automating the body condition scoring of dairy cows, leveraging advancements in 3D imaging technology and deep neural networks. The primary objective was to design and implement a system capable of accurately and efficiently assessing the body condition of dairy cows, a critical metric in livestock management for optimizing health and productivity. To achieve this, a 3D camera was employed to capture detailed point cloud data, reconstructing the three-dimensional morphology of individual cows. The obtained data were then fed into a deep neural network, specifically tailored for the task of ranking body condition. The neural network was trained on a diverse dataset of annotated body condition score representing varying degrees of body condition, ensuring robust performance across different physiological states. The results demonstrate the efficacy of the proposed system in automatically and objectively scoring the body condition of dairy cows. The automated process not only expedites the assessment but also reduces the subjectivity associated with manual scoring methods. This innovative approach holds promise for improving the efficiency of dairy farm management by providing timely and accurate body condition assessments. The integration of 3D imaging and deep learning techniques paves the way for future advancements in precision livestock farming, contributing to enhanced animal welfare, optimized production, and sustainable agriculture practices.

This paper investigates the adaptability of four state-of-the-art Artificial Intelligence (AI) models to the Australian mammographic context through transfer learning, explores the impact of image enhancement on model performance and analyses the relationship between AI outputs and histopathological features for clinical relevance and accuracy assessment. A total of 1712 screening mammograms (n=856 cancer cases and n=856 matched normal cases) were used in this study. The 856 cases with cancer lesions were annotated by two expert radiologists and the level of concordance between their annotations was used to establish image subsets. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of Globally aware Multiple Instance Classifier (GMIC), Global-Local Activation Maps (GLAM), I&H and End2End AI models, both in the pre-trained and transfer learning modes, with and without applying the Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithm. The four AI models with and without transfer learning in the high-concordance subset outperformed those in the entire dataset. Applying the CLAHE algorithm to mammograms improved the performance of the AI models. In the high-concordance subset with transfer learning and CLAHE algorithm applied, the AUC of the GMIC model was highest (0.912), followed by GLAM model (0.909), I&H (0.893) and End2End (0.875). There were significant differences (P<0.05) in the performances of the four AI models between high-concordance subset and entire dataset. The AI models demonstrated significant differences in malignancy probability concerning different tumour size categories in mammograms. The performance of AI models was affected by several factors such as concordance classification, image enhancement and transfer learning. Mammograms with strong concordance of radiologists’ annotations, applying image enhancement and transfer learning could enhance the accuracy of AI models.

Radiator reliability is crucial in environments characterized by high temperatures and friction, where prompt interventions are often required to prevent system failures. This study introduces a proactive approach to radiator fault diagnosis, leveraging the integration of Gaussian Mixture Model (GMM) and Long-Short Term Memory (LSTM) autoencoders. Vibration signals from radiators were systematically collected through randomized durability vibration bench tests, resulting in four operating states—two normal, one unknown, and one faulty. Time-domain statistical features of these signals were extracted and subjected to Principal Component Analysis (PCA) to facilitate efficient data interpretation. Subsequently, this study discusses the comparative effectiveness of GMM and LSTM in fault detection. GMMs are deployed for initial fault classification, leveraging their clustering capabilities, while LSTM autoencoders excel in capturing time-dependent sequences, facilitating advanced anomaly detection for previously unencountered faults. This alignment offers a potent and adaptable solution for radiator fault diagnosis, particularly in challenging high-temperature or high-friction environments. Consequently, the proposed methodology not only provides a robust framework for early-stage fault diagnosis but also effectively balances diagnostic capabilities during operation. Additionally, this study presents the foundation for advancing reliability life assessment in accelerated life testing, achieved through dynamic threshold adjustments using both the absolute log-likelihood distribution of the GMM and the reconstruction error distribution of the LSTM autoencoder model.

Radiological imaging is an essential component of a swallowing assessment. Artificial intelligence (AI), and especially deep learning (DL) models, have enhanced the efficiency and efficacy through which imaging is interpreted, and subsequently have important implications for swallow diagnostics and intervention planning. However, the application of AI for the interpretation of videofluoroscopic swallowing studies (VFSS) is still emerging. This review showcases recent literature in the use of AI to interpret VFSS and highlights clinical implications for speech pathologists (SPs). With a surge in AI research, there have been advances made in dysphagia assessment. Several studies have demonstrated successful implementation of DL algorithms to analyze VFSS. Notably, convolutional neural networks (CNNs) have been used to detect pertinent aspects of the swallowing process with high levels of precision. DL algorithms have the potential to streamline VFSS interpretation, improve efficiency and accuracy, and enable precise interpretation of instrumental dysphagia evaluation, which is especially advantageous when access to skilled clinicians is not ubiquitous. By enhancing precision, speed, and depth of VFSS interpretation, SPs can obtain a more comprehensive understanding of swallow physiology and deliver targeted and timely intervention that is tailored towards the individual. This has practical applications for both clinical practice and dysphagia research. As this research area grows and AI technologies progress, the application of DL in the field of VFSS interpretation is clinically beneficial and has the potential to transform dysphagia assessment and management. With broader validation and inter-disciplinary collaborations, AI-augmented VFSS interpretation is likely to transform swallow evaluation and ultimately improve outcomes for individuals with dysphagia.

This study explores the Edaravone solubility space encompassing both neat and binary dissolution media. The efforts were made toward revealing the concentration limits inherently associated with common pure and mixed solvents. For this purpose, the published solubility data of the title drug were scrupulously inspected and cured. This enabled to make the dataset consistent and coherent. However, the lack of some important types of solvents in the collection called for an extension of the available pool of Edaravone solubility data. Hence, new measurements were performed for collecting Edaravone solubility in polar non-protic and polar di-protic media. Such an extended set of data was used for tuning the parameters of regressors models and formulating the ensemble for predicting new data. In both phases, namely the model training and ensemble formulation, close attention was paid not only to minimizing the deviation of computed values from the experimental ones but also to ensuring the high predictive power and computing solubility for new systems. Furthermore, the environmental friendliness characteristics determined based on the common green solvent selection criteria was included in the analysis. Our applied protocol led to the conclusion that the solubility space defined by ordinary solvents is limited, and it is unlikely to find solvents that are more suited for Edaravone dissolution than those depicted in this manuscript. The theoretical framework presented in this study provides a precise guide for conducting experiments, saving time and resources in the pursuit of new findings.

The makeup of human microbiota has been linked to a number of autoimmune disorders. Recent developments in whole metagenome sequencing and 16S rRNA sequencing technology have considerably aided research into the microbiome and its relationship to disease. Due to the inherent high dimensionality and complexity of data generated by high-throughput platforms, conventional bioinformatics techniques could only provide an inadequate explanation for the most relevant changes and seldom provide correct predictions. Machine learning, on the other hand, is a subset of artificial intelligence applications that enable the untangling of high-dimensional systems and intricate knots in correlation by learning complex patterns and improving automatically from training data without being explicitly programmed. Machine learning is increasingly being utilized to research the influence of microbes on the onset of illness and other clinical features since computer power has increased dramatically in the last few decades. In this review paper, we focused on emerging methodological approaches of supervised machine learning algorithms for identification of autoimmune disorders utilizing metagenomics data, as well as the potential benefits and limitations of machine learning models in clinical applications.

Transcutaneous injection laryngoplasty is a well-known procedure for treating paralyzed vocal fold by injecting augmentation material into it. Hence, vocal fold localization plays a vital role in the preoperative planning as the fold location is required to determine the optimal injection route. In this communication, we propose a mirror environment based reinforcement learning (RL) algorithm for localizing the right and left vocal folds in preoperative neck CT. RL-based methods commonly showed noteworthy outcome in general anatomic landmark localization problem in the recent years. However, such methods suggest training individual agent for localizing each fold, though the right and left vocal folds are located in close proximity and have high feature-similarity. Utilizing the lateral symmetry between the right and left vocal folds, the proposed mirror environment allows for a single agent for localizing both the folds by treating the left fold as a flipped version of the right fold. Thus, localization of both folds can be trained using a single training session which utilizes the inter-fold correlation and avoids redundant feature learning. Experiment with 120 CT volumes showed improved localization performance and training efficiency of the proposed method compared with the standard RL method.

Deep learning techniques, such as convolutional neural networks (CNN), generative adversarial networks (GAN), and graph neural networks (GNN), have over the past decade changed the ac-curacy of prediction in many diverse fields. In recent years, the application of deep learning tech-niques in computer vision tasks in pathology demonstrated extraordinary potential in assisting clinicians, automating diagnosis, and reducing costs for patients. Formerly unknown pathologi-cal evidence, such as morphological features related to specific biomarkers, copy number varia-tions, and other molecular features, were also able to be captured by deep learning models. In this paper, we review popular deep learning methods and some recent publications about their appli-cations in pathology.

Data sets for medical images are generally imbalanced and limited in sample size because of high data collection costs, time-consuming annotations, and patient privacy concerns. The training of deep neural network classification models on these data sets to improve the generalization ability does not produce the desired results for classifying the medical condition accurately and often overfit the data on the majority of class samples. To address the issue, we propose a framework for improving the classification performance metrics of deep neural network classification models using transfer learning: pre-trained models, such as Xception, InceptionResNet, DenseNet along with the Generative Adversarial Network (GAN) – based data augmentation. Then, we trained the network by combining traditional data augmentation techniques, such as randomly flipping the image left to right and GAN-based data augmentation, and then fine-tuned the hyper-parameters of the transfer learning models, such as the learning rate, batch size, and the number of epochs. With these configurations, the Xception model outperformed all other pre-trained models achieving a test accuracy of 98.7%, the precision of 99%, recall of 99.3%, f1-score of 99.1%, receiver operating characteristic (ROC) - area under the curve (AUC) of 98.2%.

This is an article or technical note which is intended to provides an insight journey of Machine Learning Systems (MLS) testing, its evolution, current paradigm and future work. Machine Learning Models, used in critical applications such as healthcare industry[1], Automobile [2], [3] and Air Traffic control, Share Trading etc., and failure of ML Model can lead to severe consequences in terms of loss of life or property. To remediate this, developers, scientists, and ML community around the world, must build a highly reliable test architecture for critical ML application. At the very foundation layer, any test model must satisfy the core testing attributes such as test properties and its components. This attribute comes from the software engineering [5], [6], but the same cannot be applied in as-is form to the ML testing and we will tell you “why”.

Anomaly occurrences in hydraulic machinery may lead to massive systems shut down, jeopardizing the safety of the machinery and its surrounding human operator(s) and environment, and the severe economic implications succeeding the faults and their associated damage. Hydraulics are mostly placed in ruthless environments, where they are consistently vulnerable to many faults. Hence, not only the machines and their components are prone to anomalies, but also the sensors attached to them, which monitor and report their health and behavioral changes. In this work, a comprehensive applicational analysis of anomalies in hydraulic systems extracted from a hydraulic test rig is thoroughly achieved. Firstly, we provided a combination of a new architecture of LSTM autoencoders and supervised machine and deep learning methodologies to perform two separate stages of fault detection and diagnosis. The two phases are condensed by: the detection phase using the LSTM autoencoder. Followed by the fault diagnosis phase represented by the classification schema. The previously mentioned framework is applied to both component and sensor faults in hydraulic systems, deployed in the form of two in-depth applicational experiments. Moreover, a thorough literature review of the past decade related work for the two stages separately is successfully conducted in this paper.

Tail-biting convolutional codes extend the classical zero-termination convolutional codes: Both encoding schemes force the equality of start and end states, but under the tail-biting each state is a valid termination. This paper proposes a machine-learning approach to improve the state-of-the-art decoding of tail-biting codes, focusing on the widely employed short length regime as in the LTE standard. This standard also includes a CRC code. First, we parameterize the circular Viterbi algorithm, a baseline decoder that exploits the circular nature of the underlying trellis. An ensemble combines multiple such weighted decoders, each decoder specializes in decoding words from a specific region of the channel words' distribution. A region corresponds to a subset of termination states; the ensemble covers the entire states space. A non-learnable gating satisfies two goals: it filters easily decoded words and mitigates the overhead of executing multiple weighted decoders. The CRC criterion is employed to choose only a subset of experts for decoding purpose. Our method achieves FER improvement of up to 0.75dB over the CVA in the waterfall region for multiple code lengths, adding negligible computational complexity compared to the circular Viterbi algorithm in high SNRs.

Radio Frequency Interference (RFI) detection and characterization play a critical role to in ensuring the security of all wireless communication networks. Advances in Machine Learning (ML) have led to the deployment of many robust techniques dealing with various types of RFI. To sidestep an unavoidable complicated feature extraction step in ML, this paper proposes an efficient end-to-end method using the latest advances in deep learning to extract the appropriate features of the RFI signal. Moreover, this study utilizes the benefits of transfer learning to determine both the type of received RFI signals and their modulation types. To this end, the scalogram of the received signals is used as the input of the pre-trained convolutional neural networks (CNN), followed by a fully-connected classifier. This study considers a digital video stream as the signal of interest (SoI), transmitted in a real-time satellite-to-ground communication using DVB-S2 standards. To create the RFI dataset, the SoI is combined with three well-known jammers namely, continuous-wave interference (CWI), multi- continuous-wave interference (MCWI), and chirp interference (CI). This study investigated four well-known pre-trained CNN architectures, namely, AlexNet, VGG-16, GoogleNet, and ResNet-18, for the feature extraction to recognize the visual RFI patterns directly from pixel images with minimal preprocessing. Moreover, the robustness of the proposed classifiers is evaluated by the data generated at different signal to noise ratios (SNR).

For the generation of representative volume elements a statistical description of the relevant parameters is necessary. These parameters usually describe the geometric structure of a single grain. Commonly, parameters like area, aspect ratio and slope of the grain relative to the rolling direction are applied. However, usually simple distribution functions like log normal or gamma distribution are used. Yet, these do not take the interdependencies between the microstructural parameters into account. To fully describe any metallic microstructure though, these interdependencies between the singular parameters need to be accounted for. To accomplish this representation, a machine learning approach was applied in this study. By implementing a Wasserstein generative adversarial network, the distribution, as well as the interdependencies could accurately be described. A validation scheme was applied to verify the excellent match between microstructure input data and synthetically generated output data.

An increase in the use of smartphones has laid to the use of the internet and social media platforms. The most commonly used social media platforms are Twitter, Facebook, WhatsApp and Instagram. People are sharing their personal experiences, reviews, feedbacks on the web. The information which is available on the web is unstructured and enormous. Hence, there is a huge scope of research on understanding the sentiment of the data available on the web. Sentiment Analysis (SA) can be carried out on the reviews, feedbacks, discussions available on the web. There has been extensive research carried out on SA in the English language, but data on the web also contains different other languages which should be analyzed. This paper aims to analyze, review and discuss the approaches, algorithms, challenges faced by the researchers while carrying out the SA on Indigenous languages.

The upsurge of Generative Adversarial Networks (GANs) in the previous five years has led to advancements in unsupervised data manipulation, sourced feature translation, and precise input-output synthesis through a competitive optimization of the discriminator and generator networks. More specifically, the recent rise of cycle-consistent GANs enables style transfers from a discrete source (input A) to target domain (input B) by preprocessing object features for a multi-discriminative adversarial network. Traditionally, cyclical adversarial networks have been exploited for unpaired image-to-image translation and domain adaptation by determining mapped relationships between an input A graphic and an input B graphic. However, this integral mechanism of domain adaptation can be applied to the complex acoustical features of human speech. Although well-established datasets, such as the 2018 Voice Conversion Challenge repository, paved way for female-male voice transformation, cycle-GANs have rarely been re-engineered for voices outside the datasets. More critically, cycle-GANs have massive potential to extract surface-level and hidden feature to distort an input A source into a texturally unrelated target voice. By preprocessing, compressing, and packaging unique acoustical voice properties, CycleGANs can learn to decompose speech signals and implement new translation models while preserving emotion, the intent of words, rhythm, and accents. Due to the potential of CycleGAN’s autoencoder in realistic unsupervised voice-voice conversion/feature adaptation, the researchers raise the ethical implications of controlling source input A to manipulate target voice B, particularly in cases of defamation and sabotage of target B’s words. This paper analyzes the potential of cycle-consistent GANs in deceptive voice-voice conversion by manipulating interview excerpts of political candidates.

: Cybersecurity has become a critical area in the digital field in recent years. The expansion of networks has revolutionised the way network structures are organised and managed. However, with increased connectivity and the growing complexity of modern networks, the threat of cyber-attacks has become more intense. As technology continues to advance, it brings both opportunities and challenges. One of the major challenges is the need to secure networks and sensitive data from various malicious activities. Traditional networks have evolved to include Software Defined Network (SDN), which offers a more flexible and programmable framework. Researchers should focus on detecting attacks in SDN because SDN networks are becoming more popular and attractive targets for clients due to their programmability and dynamic nature. The centralised controller, known as the backbone of an SDN, becomes a single point of failure and a potential target for attackers if not properly secured. Researchers need to emphasise the detection of attacks in SDN in order to mitigate these risks. By understanding the potential vulnerabilities and attack methods specific to SDN, researchers can develop effective detection approaches and propose countermeasures. This methodology helps protect the network from potential threats and minimises the impact of successful attacks. Therefore, this survey paper provides a review of Intrusion Detection Systems (IDSs) in Software-Defined Networks (SDN) to provide a thorough understanding of SDN security issues.

The transportation industry is rapidly transitioning from Internal Combustion Engine (ICE) based vehicles to Electric Vehicles (EVs) to promote clean energy. However, large-scale adoption of EVs can compromise the reliability of the power grids by introducing large uncertainty in the demand. Demand response with a controlled charge scheduling strategy for EVs can mitigate such issues. In this paper, a deep reinforcement learning- based charge scheduling strategy is developed for individual EVs by considering user’s dynamic driving behavior and charging preferences. The temporal dynamics of user’s anxiety about charging the EV battery is rigorously addressed. A dynamic weight allocation technique is applied to continuously tune user’s priority for charging and cost-saving with respect to charging duration. The sequential charging control problem is formulated as a Markov decision process, and an episodic approach to the deep deterministic policy gradient (DDPG) algorithm with target policy smoothing and delayed policy update techniques is applied to develop the optimal charge scheduling strategy. A real-world dataset that captures user’s driving behavior, such as arrival time, departure time, and charging duration, is utilized in this study. The extensive simulation results reveal the effectiveness of the proposed algorithm in minimizing energy cost while satisfying user’s charging requirements.

Early detection of cervical cancer is the need of the hour to stop the fatality from this disease. There have been various CAD approaches in the past that promise to detect cervical cancer in an early stage, each of them having some constraints.This study proposes a novel ensemble deep learning model, VLR (variable learning rate), aimed at enhancing the accuracy of cervical cancer classification. The model architecture integrates VGG16, Logistic Regression, and ResNet50, combining their strengths in an offbeat ensemble design. VLR learns in two ways: firstly, by dynamic weights from three base models, each of them trained separately; secondly, by attention mechanisms used in the dense layer of the base models. Hyperparameter tuning is applied to further reduce loss, fine tune the model’s performance, and maximize classification accuracy. We performed K-fold cross-validation on VLR to evaluate any improvements in metric values resulting from hyperparameter fine tuning. We have also validated our model on images captured in three different solutions and on a secondary dataset. Our proposed VLR model outperformed existing methods in cervical cancer classification, achieving a remarkable training accuracy of 99.95% and a testing accuracy of 99.89%.

Brain-Computer Interfaces (BCIs) are essential in advancing medical diagnosis and treatment by providing non-invasive tools to assess neurological states. Among these, Motor Imagery (MI), where patients mentally simulate motor tasks without physical movement, has proven to be an effective paradigm for diagnosing and monitoring neurological conditions. Electroencephalography (EEG) is widely used for MI data collection due to its high temporal resolution, cost-effectiveness, and portability. However, EEG signals can be noisy from a number of sources, including physiological artifacts and electromagnetic interference. They can also vary from person to person, which makes it harder to extract features and understand the signals. Additionally, this variability, influenced by genetic and cognitive factors, presents challenges for developing subject-independent solutions. To address these limitations, this paper presents a Multimodal and Explainable Deep Learning (MEDL) approach for MI-EEG classification and physiological interpretability. Our approach involves: i) evaluating different deep learning (DL) models for subject-dependent MI-EEG discrimination; ii) employing Class Activation Mapping (CAM) to visualize relevant MI-EEG features; and iii) utilizing a Questionnaire-MI Performance Canonical Correlation Analysis (QMIP-CCA) to provide multidomain interpretability. On the GIGAScience MI dataset, experiments show that shallow neural networks are good at classifying MI-EEG data, while the CAM-based method finds spatio-frequency patterns. Moreover, the QMIP-CCA framework successfully correlates physiological data with MI-EEG performance, offering an enhanced, interpretable solution for BCIs.

This paper considers the question of what kind of knowledge is produced by deep learning. Ryle’s concept of knowledge how is examined and is contrasted with knowledge with a rationale. It is then argued that deep neural networks do produce knowledge how, but, because of their opacity, they do not in general, though there may be some special cases to the contrary, produce knowledge with a rationale. It is concluded that the distinction between knowledge how and knowledge with a rationale is a useful one for judging whether a particular application of deep learning AI is appropriate.

Heart rate and blood pressure, along with body temperature and respiratory rate, are two of the four parameters used to assess the state of bodily functions, as both measures are indicators of heart activity. Therefore, these measurements become a basic procedure in patient diagnosis, necessitating techniques that can obtain them accurately and reliably, while being as non-intrusive as possible for the patient. Using video data, these two vital signs can be determined at any time and place. To achieve this, artificial intelligence techniques such as machine learning and deep learning 6 can be applied to overcome the main limitations of conventional remote photoplethysmography (rPPG). This document explains the technical and methodological aspects involved, including the acquisition and preprocessing of video data, the extraction of the rPPG signal, and its subsequent processing, as well as the artificial intelligence model for heart rate estimation and the mathematical modeling for blood pressure estimation. We compared the results for heart rate and blood pressure extracted from videos recorded by a basic smartphone front-camera to a blood pressure monitor and achieved results with reasonable accuracy and correlation. These results are part of the AIVA project, grant agreement 2021/C005/00145071. This project has been funded by Red.es through the Spanish 2021 call for grants aimed at research and development projects in artificial intelligence and other digital technologies.

Detecting people and technical objects in various situations, such as natural disasters and warfare, is critical to search and rescue operations and the safety of the civilian. A fast and accurate detection for people and equipment can significantly increase the effectiveness of search and rescue missions and provide timely assistance to people. Computer vision and deep learning technologies play a key role in detecting the required objects due to their ability to analyse big volumes of visual data in real time. The performance of the neural networks such as YOLOv4-v8, Faster R-CNN, SSD, and EfficientDet has been analysed using COCO2017, SARD, SeaDronesSee, and VisDrone2019 datasets. Main criteria for comparison were mAP, Precision, Recall, F1-Score, and the ability of the neural network to work in real time. The most important metrics for evaluating the efficiency and performance of models for a given task are accuracy (mAP), F1-Score, and processing speed (FPS). These metrics allow us to evaluate both the accuracy of object recognition and the ability to use the models in real-world environments where high processing speed is important. Although different neural networks perform better on certain types of metrics, YOLO outperforms them on all metrics, showing the best results of mAP-0.88, F1 - 0.88, and FPS - 48, so the focus was on these models.

The growing use of drones in precision agriculture highlights the need for enhanced operational efficiency. Despite the ability of computer vision based on deep learning has made remarkable progress in the past ten years, when it comes to segmentation task on UAVs, there is always a conflict between the demand of high precision and low inference latency. Due to such a dilemma, we propose the IRB-YOLO, an efficiency model based on Inverted Residual Block, devoting to provide constructive strategies in real-time detection tasks of UAV camera. The working details of this paper are as follows: (1) This paper innovates with a IR-Block(Inverted Residual Block), integrated into a refined YOLOv8-seg structure to create IRB-YOLO. This model specializes in pixel-level classification of UAV-acquired RGB images, facilitating the creation of exact maps to guide agricultural strategies. (2)When it comes to the experiments on a Vatica dataset with any other light-weight segmentation model, IRB-YOLO achieve at least a 3.3% increase in mAP. Further validation using a diverse species dataset confirms its robust generalization. (3)Without overloading the complex attention mechanism and deeper and deeper network, a stem that incorporates efficient feature extraction components, inverted residual block, can still possess outstanding modeling capabilities. IRB-YOLO builds a bridge between academic research and edge deployment of drones, making it applicable in real-world scenarios.

Crawling robots are the focus of intelligent inspection research, and the main feature of this type of robot is the flexibility of in-plane attitude adjustment. The crawling robot HIT_Spibot is a new type of steam generator heat transfer tube inspection robot with a unique mobility capability different from traditional quadrupedal robots. This paper introduces a hierachical motion planning approach for HIT_Spibot, aiming to achieve efficient and agile maneuverability. The proposed method integrates three distinct planners to handle complex motion tasks: a nonlinear optimization-based base motion planner, a TOPSIS-based base orientation planner, and a Mask-D3QN (MD3QN) algorithm-based gait motion planner. Initially, the robot's base and foot workspace are delineated through envelope analysis, followed by trajectory computation using Larangian methods. Subsequently, the TOPSIS algorithm is employed to establish an evaluation framework conducive to foundational turning planning. Finally, the MD3QN algorithm trains foot points to facilitate robot movement along predefined paths. Experimental results demonstrate the method's adaptability across diverse tube structures, showcasing robust performance even in environments with random obstacles. Compared to the D3QN algorithm, MD3QN achieves a 100% success rate, enhances average overall scores by 6.27%, reduces average stride lengths by 39.04%, and attains a stability rate of 58.02%. These results not only validate the effectiveness and practicality of the method but also showcase the significant potential of HIT_Spibot in the field of industrial inspection.

Increasing concerns on animal welfare in commercial pig industry include aggression between pigs as it affects their health and growth. Early detection of aggressive behaviors is essential for optimizing their living environment. A major challenge for detection is that these behaviors are observed occasionally in normal conditions. Under this circumstance, a limited amount of aggressive behavior data will lead to class imbalance issue, making it difficult to develop an effective classification model for the detection of aggressive behaviors. To overcome this problem, we approach the aggressive behavior detection problem as a case of anomaly detection rather than classification and propose a model based on anomaly detection, which can better handle unbalanced class distribution and effectively detect the infrequent, aggressive episodes of pigs. The model consists of a convolutional neural network (CNN) and a variational long short-term memory (LSTM) autoencoder. Additionally, we adopted a training method similar to weakly supervised anomaly detection and included a few aggressive behavior data in the training set for prior learning. To effectively utilize the aggressive behavior data, we created Reconstruction Loss Inversion, a novel objective function, to train the autoencoder-based model to increase the reconstruction error of aggressive behaviors by inverting the loss function. Our anomaly detection approach significantly outperforms traditional classification-based methods, effectively identifying aggressive behaviors in a natural farming environment. This method offers a robust solution for detecting aggressive animal behaviors and contributes to improving their welfare.

Large-scale deep learning models have produced significant results in underwater visual image target detection. However, deploying these models on underwater embedded devices poses a series of issues and challenges. Firstly, underwater devices have limited communication capabilities and carry relatively short supply of energy, resulting in constrained computing power for edge devices. Secondly, the large scale of deep neural networks creates a conflict in power demands with unmanned underwater equipment. Thirdly, insufficient underwater light and turbid water quality will lead to degradation problems in the images acquired by the device. In this paper, we build an offline system to complete the task of underwater scene target recognition. This paper introduces several innovations: 1) The new YOLO-TN model is developed by compressing the target recognition model based on YOLO-V5. The model use a model compression algorithm replace the YOLO-V5’s backbone network with a lightweight network, then perform parameter pruning on the model detection head, while reduce the size of the input image. The model uses a Teacher-guided Neural Architecture Search method based on YOLO-V5, named as YOLO-TN. 2)A method for constructing and processing real underwater environment data sets is designed. The datasets are all captured using underwater unmanned vehicles in real underwater environments, addressing issues present in existing underwater datasets, such as unbalanced target volume and single-image environment. Experimental results demonstrate that the YOLO-TN model significantly reduces the model size while maintaining the original recognition accuracy. Furthermore, it achieves 28.6 FPS on embedded devices, which is 12 times that of the original model.

Gastrointestinal (GI) tract disorders, ranging from benign polyps to aggressive forms of cancer, pose significant health challenges globally. Early detection and precise classification of these conditions are crucial for effective treatment and improving patient survival rates. This study employs the Hyper-Kvasir dataset, a comprehensive collection of endoscopic images, to develop deep learning models that harness the power of the YOLO (You Only Look Once) architecture for real-time detection and classification of GI abnormalities. The focus is on overcoming inherent challenges such as class imbalance and limited annotated data availability. Advanced machine learning strategies, including data augmentation and semi-supervised learning, are utilized to enhance the model's performance. Our experiments demonstrate notable improvements in the detection of pre-cancerous lesions and other GI abnormalities, confirming the potential of integrating AI into endoscopic practices to support clinicians, reduce diagnostic errors, and contribute to more accurate and timely diagnoses. The implications of these findings are significant, offering a pathway to more reliable diagnostic processes and ultimately, better patient management in gastroenterology.

This research aims to explore the integration method of deep learning and large language models in speech recognition to improve the system’s recognition accuracy and ability to handle complex contexts. Deep neural network (DNN), convolutional neural network (CNN), long short-term memory network (LSTM) and Transformer-based large language model are used to build an integrated acoustic and language model framework. Experiments on TIMIT, LibriSpeech and Common Voice datasets show that the ensemble model shows significant improvements in both word error rate (WER) and real-time factor (RTF) compared to traditional models. Especially in terms of adaptability to multiple languages and accent changes, the model shows superior performance. The conclusion shows that through technology integration, the performance of the speech recognition system in complex environments can be effectively improved, providing a new direction for the future development of speech recognition technology.

Pakistan is an agricultural country, and an exporter of crops in many countries. The crop production is an important factor to earn revenue for the country. But the agriculture of Pakistan is still not revolutionized and no modern system has been deployed commonly. Whereas the farmers face difficulty in growing some crops because those crops needs more care and calculated steps. Moreover the farmers face difficulty at different stages of the crop growth like to check fertility of soil and if the soil is non-fertile then which chemical property is lacking. The objective of this research is to provide digitize the agriculture of Pakistan. An emerging technology IoT and Deep learning based system will be developed. The IoT part contains some devices like sensors,gateways and communication technologies whereas the software part consists of Deep learning models that can make predictions about the fertility of the soil. The motive of this system is to provide best results in term of accuracy. Whereas the dataset used is publicly available on kaggle but it was of non-fertile values, so data augmentation is also performed to generate data for fertile values. We have implemented LSTM,RNN,CNN among them LSTM performed best for the prediction of soil fertility.

The rapidly evolving domain of multimodal sentiment analysis is crucial for unraveling the intricate layers of emotional expression in social media content, customer service interactions, and personal vlogs. This research introduces a cutting-edge Advanced Cross-Modal Gating (ACMG) framework that significantly enhances the precision of sentiment analysis by refining the interplay among textual, auditory, and visual modalities. Our approach addresses three foundational aspects of sentiment analysis: (1) Advanced learning of cross-modal interactions, which focuses on extracting and synthesizing sentiment from varied modal inputs, thus providing a holistic view of expressed emotions; (2) Mastery over the temporal dynamics of multimodal data, enabling the model to maintain context and sentiment continuity over extended interactions; and (3) Deployment of a novel fusion strategy that not only integrates unimodal and cross-modal cues but also dynamically adjusts the influence of each modality based on its contextual relevance to the sentiment being expressed. The exploration of these dimensions reveals that the nuanced modeling of cross-modal interactions is crucial for enhancing model responsiveness and accuracy. By applying the ACMG model to two highly regarded datasets—CMU Multimodal Opinion Level Sentiment Intensity (CMU-MOSI) and CMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI)—we achieve groundbreaking accuracies of 83.9\% and 81.1\%, respectively. These results represent significant improvements of 1.6\% and 1.34\% over the current state-of-the-art, showcasing the superior performance and potential of our approach in navigating the complexities of multimodal sentiment analysis.

The aim of the present study was to predict the response of non-small cell lung cancer (NSCLC) patients to immune checkpoint inhibitors (ICIs) by leveraging computed tomography (CT) images using deep learning techniques. Retrospectively, 624 sequential CT images were gathered from 156 patients at Jiangsu Province Hospital, along with their clinical data. The dataset was subsequently partitioned into three groups: training (n=547), validation (n=64), and test (n=64). Moreover, an external validation cohort included 37 CT images from patients at Nanjing Pukou Peoples' Hospital, accompanied by comprehensive clinical data. An advanced Video Vision Transformer (ViViT) model incorporating global self-attention was utilized to analyse patients treated with ICIs and predict their response. The ViViT model's efficacy was evaluated using a confusion matrix and a receiver operating characteristic curve (ROC). Notably, the ViViT model demonstrated predictive prowess for ICIs response, yielding respective areas under the receiver operating characteristic curve (AUC) of 0.74 (95% CI: 0.69-0.78), 0.74 (95% CI: 0.61-0.86), 0.76 (95% CI: 0.62-0.88), and 0.69 (95% CI: 0.5-0.87) in the training, validation, test, and external validation cohorts. The present study illustrates how a deep learning model can provide a non-invasive means to predict clinical outcomes in NSCLC patients undergoing ICIs, potentially transforming personalized treatment approaches for individuals with NSCLC.

This study evaluates the effectiveness of Darknet53 and VGG16 pretrained models to develop a hybrid ensemble model named HybridPlantNet23 deep learning model for effective plant disease classification, using the publicly available dataset ‘A Dataset of Leaf Images’ from Mendeley webpage. The preprocessing of the dataset involved resizing the images to the respective requirements of the pretrained models to develop a combined datastore which served as the input. All experiments were conducted using MATLAB R2023b deep network designer interface. Model performances were evaluated based on the conventional accuracy, precision, recall, and F1-score. VGG16 achieved the highest accuracy (97.25%), precision (96.30%), recall (96.96%), and F1-score (96.55%), outperforming both Darknet53 and the ensemble HybridPlantNet23 model. Despite the advanced architecture of the ensemble model, its performance (accuracy: 95.37%, precision: 93.56%, recall: 94.20%, F1-score: 93.74%) was not significantly superior, thus highlighting scientific reasons behind this development. Recommendations for future works were provided accordingly.

Microscopic and ultra-microscopic vascular sutures are indispensable in the surgical procedures such as arm transplantation and finger reattachment. The state of the blood vessels after suturing, such as vascular patency, narrow, and blocked, determines the success rate of the operation. If we can grasp the golden window period after blood vessel suture and before muscle tissue suture to achieve accurate and objective assessment of blood vessel status, it can not only reduce medical costs but also improve social benefits. Doppler optical coherence tomography enables high-speed, high-resolution imaging of biological tissues, especially microscopic and ultra-microscopic blood vessels. Using Doppler optical coherence tomography to image the sutured blood vessels, not only the three-dimensional structure of the blood vessels, but also the blood flow information can be obtained. By extracting the contour of blood vessel wall and the contour of blood flow area, the three-dimensional shape of blood vessel can be reconstructed in three dimensions, which provides parameter support for the assessment of blood vessel status. In this work, we propose a neural network-based multi-classification deep learning model that can simultaneously and automatically extract blood vessel boundaries from Doppler OCT vessel intensity images and the contours of blood flow regions from corresponding Doppler OCT vessel phase images. Compared with the traditional random walk segmentation algorithm and cascade neural network method, the proposed model has better performance. This method is easier to realize system integration and has great potential for clinical evaluation. It is expected to be applied to the evaluation of microscopic and ultra-microscopic vascular status in microvascular anastomosis.

Prochlorococcus, a cyanobacteria genus of the smallest and most abundant oceanic phototrophs, encompasses ecotype strains adapted to high-light (HL) and low-light (LL) niches. To elucidate the adaptive evolution of this genus, we analyzed 40 Prochlorococcus marinus ORFeomes, including two cornerstone strains, MED4 and NATL1A. Employing deep learning with robust statistical methods, we detected new protein family distributions in the strains and identified key genes differentiating HL and LL strains. HL strains harbor genes (ABC-2 transporters) related to stress resistance, such as DNA repair and RNA processing, while LL strains exhibit unique chlorophyll adaptations (ion transport proteins, HEAT repeats). We report variable, depth-dependent endogenous viral elements in the 40 strains. We constructed the ORFeomes of MED4 and NATL1A covering 99 % of the annotated protein-coding sequences of the two species, totaling 3976 cloned, sequence-verified ORFs. These comparative genomics analyses, paired with MED4 and NATL1A ORFeomes, will facilitate future genotype-to-phenotype mappings and systems biology exploration of Prochlorococcus ecology.

Underwater visual detection technology plays a pivotal role in fields such as marine exploration. With the increasing demand for underwater monitoring, the quest for efficient and reliable methods for underwater target recognition has become particularly significant. To address this requirement, this study developed an innovative underwater object detection architecture based on YOLOv8, named YOLOv8-MU, aimed at significantly enhancing detection accuracy.By integrating the LarK module proposed in UniRepLKNet to optimize the backbone network, YOLOv8-MU aims to achieve a larger receptive field without increasing the model’s depth. Further, this research introduces C2fSTR, an innovative method that combines Swin Transformer with the C2f module. Additionally, we have incorporated the SPPFCSPC_EMA module, which combines Cross-Stage Partial Fast Spatial Pyramid Pooling (SPPFCSPC) with attention mechanisms, significantly improving the detection accuracy and robustness of various biological targets. Moreover, by introducing a fusion block based on DAMO-YOLO into the neck of the model, we further enhanced the model’s capability in multi-scale feature extraction. Finally, the adoption of the MPDIoU loss function, designed around vertex distance, effectively tackles the challenges of localization accuracy and boundary clarity in underwater organism detection. Experimental results on the URPC2019 dataset demonstrate that the YOLOv8-MU model achieved an mAP@0.5 of 78.4%, marking improvements of 5.6%, 1.1%, and 4.0% over YOLOv5s, YOLOv7, and YOLOv8n respectively, indicating the leading performance (SOTA) of this method. On the other hand, further evaluation on the URPC2020 dataset confirmed the generalization capability of the YOLOv8-MU architecture, with its mAP@0.5 reaching 80.4%, surpassing various models including YOLOv5x and YOLOv8n, showcasing the wide applicability of our proposed improved model architecture.

Stagnation at local optima represents a significant challenge in bio–inspired optimization algorithms, often leading to suboptimal solutions. This paper addresses this issue by proposing a hybrid model that combines the Orca Predator Algorithm with Deep Q–Learning. Orca Predator Algorithm is an optimization technique that mimics the hunting behavior of orcas. It solves complex optimization problems by exploring and exploiting search spaces efficiently. Deep Q–Learning is a reinforcement learning technique that combines Q–Learning with deep neural networks. This integration aims to turn the stagnation problem into an opportunity for more focused and effective exploitation, enhancing the optimization technique’s performance and accuracy. The proposed hybrid model leverages the biomimetic strengths of Orca Predator Algorithm to identify promising regions nearby in the search space, complemented by the fine–tuning capabilities of Deep Q–Learning to navigate these areas precisely. The practical application of this approach is evaluated using the high–dimensional Heartbeat Categorization Dataset, focusing on the feature selection problem. This dataset, comprising complex electrocardiogram signals, provided a robust platform for testing the feature selection capabilities of our hybrid model. Our experimental results are encouraging, showcasing the hybrid strategy capability to identify relevant features without significantly compromising the performance metrics of machine learning models. This analysis was performed by comparing the improved method of Orca Predator Algorithm against its native version and a set of state–of–the–art algorithms.

X-ray Fluorescence Computed Tomography (XFCT) is an emerging non-invasive imaging technique providing molecular-level data, gaining attention for high-resolution imaging. However, increased sensitivity with benchtop X-ray sources raises radiation exposure. Artificial Intelligence (AI), particularly deep learning (DL), has revolutionized medical imaging by delivering high-quality images in the presence of noise. In XFCT, traditional methods rely on complex algorithms for noise reduction, but AI holds promise in addressing high-dose concerns. We present an optimized SCUNet model for noise reduction in low-concentration XRF images. Compared to higher-dose images, our method’s effectiveness is evaluated. While various denoising techniques exist for X-ray and CT, few address to XFCT. The DL model is trained and assessed using the augmented data, focusing on background noise reduction. Image quality is measured using PSNR and SSIM, comparing outcomes with 100% X-ray dose images. Results show the proposed algorithm achieves high-quality images from low-dose and low-contrast agents, with a maximum PSNR of 39.05 and SSIM of 0.86. The model outperforms BM3D, NLM, DnCNN, and SCUNet in both visual inspection and quantitative analysis, particularly in high-noise scenarios. This indicates the potential of AI, specifically the SCUNet model, in significantly improving XFCT imaging by mitigating the trade-off between sensitivity and radiation exposure.

Despite UAV multi-target detection exhibits considerable developmental potential worldwide, it suffers distinct challenges compared with traditional tasks in this field. These challenges include insufficient feature extraction capabilities for small targets, limited capabilities for multi-dimensional feature fusion, as well as constraints on hardware computation parameters. Especially in mission scenarios such as disaster detection, these challenges will be further amplified. Consequently, this paper introduces SSE-YOLO, an innovative YOLO framework algorithm specially designed to address these challenges. To enhance the model's feature extraction capability, we employ the SPDConv module to replace the original Conv in the backbone section, utilizing depth-separable convolution instead of traditional convolution pooling. Concurrently, we eliminate the SPPF module at the bottom and address a new Separate Kernel Attention Pyramid Pooling (SKAPP) module, substantially enhancing the feature fusion capability at the model's core. Moreover, to address the challenge of multi-dimensional feature fusion, we replace the Concat module of the neck and head with E-BiFPN, transmitting feature information from the backbone to the lower network through four CBS blocks, which effectively resolves the issue of lost contextual information in the model. Meanwhile, SSE-YOLO undergoes ablation experiments on the VisDrone2019 dataset to evaluate its effectiveness against alternative methods, and experimental results illustrate the model's exceptional precision in detecting UAV targets. In comparison to models with comparable experimental accuracy, SSE-YOLO requires remarkably fewer parameters. On the VisDrone2019-DET-test-dev dataset, SSE-YOLO enhances mAP by 17.3%, with a 42.5% reduction in the parameter amounts. Therefore, the proposed method effectively tackles the challenge of reconciling low parameters and high accuracy, presenting a novel pathway for deep learning-based UAV multi-target detection.

Solar energy adoption worldwide has expanded exponentially due to a surge in international interest in producing clean energy and the declining cost of solar power plants and their technology. It is anticipated that by 2050, solar will have surpassed fossil fuels to become the primary source of energy. However, one of the main challenges associated with solar energy production is the instability of photovoltaic (PV) power generation because of weather changes. Short-term forecasting of the power output of photovoltaic systems is essential for efficient management of the power grid and energy markets. This paper aims to evaluate the ability of deep learning (DL) models to provide accurate forecasting of hourly global horizontal irradiance (GHI) using different sets of features, including weather and aerosol variables along with solar radiation components. The results show that the best forecast skills are achieved by the long short-term memory autoencoder (LSTM-AE) model.

To support pathologists in breast tumor diagnosis, deep learning plays a crucial role in the development of histological whole slide image (WSI) classification methods. However, automatic classification is challenging due to the high-resolution data and the scarcity of representative training data. To tackle these limitations, we propose a deep learning-based breast tumor gigapixel histological image multi-classifier integrated with a high-resolution data augmentation model to process the entire slide by exploring its local and global information and generating its different synthetic versions. The key idea is to perform classification and augmentation in feature latent space, reducing the computational cost while preserving the class label of the input. We adopt a deep learning-based multi-classification method and evaluate the contribution given by a Conditional Generative Adversarial Network-based data augmentation model on the classifier’s performance for three tumor classes in the BRIGHT Challenge dataset. The proposed method has allowed us to achieve an average F1 equal to 69.5, considering only the WSI dataset of the Challenge. The results are comparable to those obtained by the Challenge winning method (71.6), also trained on the annotated tumor region dataset of the Challenge.

A high-resolution catalog for the 2014 Ms6.5 Ludian aftershocks was constructed based on the deep learning phase-picking model (CERP) and the seismic-phase association technology (PALM). A specific training strategy, which combines the advantages of the conventional short-long window average energy ratio algorithm (STA/LTA) and AI algorithm, is employed to retrain the CERP model. The P & S phases were accurately detected and picked on continuous seismic waveforms by the retained AI model. Hypoinverse and HypoDD were utilized for precise location of 3286 events. Compared to the previous results, our new catalog exhibits superior performances in terms of location accuracy and the number of aftershock events, thereby enabling a more de-tailed depiction of the deep-seated tectonic features. According to the distribution of aftershocks, it can be inferred that, (1) the seismogenic fault of Ludian earthquake is the NW-trending Baogunao-Xiaohe Fault. (2) the Ludian aftershocks interconnected the discontinuous NW-trending Baogunao-Xiaohe Fault, and intersected with the Zhaotong-Ludian Fault. (3) It suggests that the NE-trending Zhaotong-Ludian Fault may have been intersected by the NW-trending Baogunao -Xiaohe Fault, indicating that the Baogunao-Xiaohe Fault is likely a rela-tively young Neogene fault.

This paper introduces an innovative approach to lipreading, leveraging a web application designed to generate subtitles for videos where the speaker's mouth is visible and a comprehensive literature review that precedes the discussion, encompassing a thorough examination of various lipreading methods employed over the past decade. Our method employs a powerful deep learning model, featuring a 3D-convolution network and bidirectional LSTM, enabling accurate sentence-level predictions based solely on visual lip movements. With an impressive accuracy of 97%, our model is trained using pre-segmented lips regions, transformed into animated GIFs for effective pre-training. This work stands as a significant contribution to the evolving landscape of lipreading research, offering a practical and accurate solution for real-world applications.

With the development of remote sensing satellite technology for Earth observation, remote sensing stereo images have been used for three-dimensional reconstruction in various fields, such as urban planning and construction. However, remote sensing images often include noise, occluded regions, weakly textured areas and repeated textures, which can lead to reduced accuracy in stereo matching and affect the quality of the 3D reconstruction results. To reduce the impact of complex scenes in remote sensing images on stereo matching and to ensure both speed and accuracy, we propose a new end-to-end stereo matching network based on convolutional neural networks (CNNs). The proposed stereo matching network can learn features at different scales from the original images and construct cost volumes with varying scales to obtain richer scale information. Additionally, when constructing the cost volume, we introduce negative disparity to adapt to the common occurrence of both negative and nonnegative disparities in remote sensing stereo image pairs. For cost aggregation, we employ a 3D convolution-based encoder-decoder structure that allows the network to adaptively aggregate information. Before feature aggregation, we also introduce an attention module to retain more valuable feature information, enhance feature representation, and obtain a higher-quality disparity map. By training on the publicly available US3D dataset, we obtain the accuracy that 1.115 pixel in end-point error (EPE) and 5.32% in the error pixel ratio (D1) on the test dataset, and the inference speed is 92 ms. Comparing our model with existing state-of-the-art models, we achieve higher accuracy, and the network is beneficial for the three-dimensional reconstruction of remote sensing images.

Tropical cyclones (TC) are dangerous weather events and accurate monitoring and forecasting can provide significant early warning to reduce loss of life and property. However, the study of tropical cyclone intensity remains challenging, both in terms of theory and forecasting. ERA5 reanalysis is benchmark data set for tropical cyclone studies, yet the maximum wind speed error is very large (68 kts) and still 19 kts after simple linear correction even in the better sampled North Atlantic. Here, we develop an adaptive learning approach to correct the intensity in the ERA5 reanalysis, by optimising the inputs to overcome the problems because of the poor data quality and updating the features to improve the generalisability of the deep learning-based model. Specifically, we use TC knowledge to increase the representativeness of the inputs so that the general features can be learned with deep neural networks in the sample space, and then use domain adaptation to update the general features from the known domain with historical storms to the specific features for the unknown domain of new storms. This approach can reduce the error to only 6 kts which is within the uncertainty of the best track data in IBTrACS in the North Atlantic. The method may have wide applicability, such as extending it to the correction of intensity estimation from satellite imagery and intensity prediction from dynamical models.

Oral Squamous Cell Carcinoma is one among the most common cancer and early detection is the main key to avoid deaths. Automated diagnostic tools that process the histopathological images of a patient to detect abnormal oral lesions will be very much useful for the clinicians. A deep learning framework have been designed with an intermediate layer between feature extraction layers and classification layers for classifying the histopathological images into two categories, namely normal and oral squamous cell carcinoma. The intermediate layer is constructed using the proposed Swarm Intelligence technique called Modified Gorilla Troops Optimizer. Various optimization algorithms are implemented in literature for optimal parameter identification, weights updating and feature selection in deep learning models, but this work focuses on usage of optimization algorithm as intermediate layer that transforms the extracted features into the features that are more suitable for classification. Three datasets totally comprising 2784 normal and 3632 oral squamous cell carcinoma subjects are considered in this work. Three popular CNN architectures namely InceptionV2, MobileNetV3, and EfficientNetB3 are investigated as feature extraction layers. Two fully connected Neural Network layers along with batch normalization and dropout are used as classification layers. Among the investigated feature extraction models, MobileNetV3 performs well in all the three datasets with the highest accuracy of 0.89. Usage of the proposed Modified Gorilla Troops Optimizer as an intermediate layer boosts this accuracy to 0.95.

Real-time tracking is one of the most challenging problems in computer vision. Most Transformer-based trackers usually require expensive computational and storage power, which leads to the inability of these robust trackers to achieve real-time performance in resource-constrained devices. In this work, we propose a lightweight tracker, AnteaTrack. To localize the target more precisely, we develop a scaling-invariant max-filtering operator employing a sliding window combined with local max-pooling, which filters out the suspected target from the feature and performs an augmented representation while suppressing the background. For a more tight bounding box, we employ Pixel-Shuffle to increase the resolution of the feature map and get a more fine-grained representation of the target. In addition, AnteaTrack can run in real-time at 47 frames per second(FPS) on the CPU. We tested AnteaTrack on 5 datasets, and a large number of experiments have shown that AnteaTrack provides the most efficient solution compared to the same type of CPU real-time trackers. The code will be available at https://github.com/cnchange/AnteaTrack.

Background: External beam radiation therapy requires a sophisticated and laborious planning procedure. To improve the efficiency and quality of the planning procedure, machine learning predictions of the dose distributions have been introduced to speed up the planning procedure and to serve as quality assurance. The most recent dose prediction models are based on deep learning U-Nets that give good approximations of the dose in 3D almost instantly. It is our purpose to train a 3D dose prediction for glioblastoma VMAT treatment and test its robustness and sensitivity for the purpose of quality assurance of automatic contouring. Methods: From a cohort of 125 glioblastoma (GBM) patients, VMAT plans were created according to a clinical protocol. The initial model was trained on a concatenated 3D U-Net. A total of 60 cases were used for training, 15 for validation and 20 for testing. The prediction model was tested for sensitivity to dose changes according to realistic contour variations. Additionally, the model was tested for robustness by exposing it to a worst-case test set containing out-of-distribution cases.Results: The initially trained prediction model had a dose score of 0.94 Gy and a mean DVH score for all structures of 1.95 Gy. In terms of sensitivity, the model was able to predict the dose changes that occurred due to the contour variations with a mean error of 1.38 Gy.Conclusions: We obtained a 3D VMAT dose prediction model for GBM with limited data, providing good sensitivity to realistic contour variations. We tested and improved the model's robustness, by targeted updating the training set, making it a useful technique for dose awareness in the contouring evaluation and quality assurance.

Due to the complex underground environment, pumping machines are prone to produce numerous failures. The indicator diagrams of faults are similar in a certain degree, which produces indistinguishable samples. As the samples increases, manual diagnosis becomes difficult, which decreases the accuracy of fault diagnosis. For accurately and quickly judging the fault type, we propose an improved adaptive activation function and apply it to five types of neural networks. The adaptive activation function improves the negative semi-axis slope of the ReLU activation function by combining the gated channel conversion unit to improve the performance of the deep learning model. The proposed adaptive activation function is compared with the traditional activation function through the fault diagnosis data set and the public data set. The results show that the activation function has better nonlinearity, can improve the generalization performance of deep learning model, the accuracy of fault diagnosis. In addition, the proposed adaptive activation function can also be well embedded in other neural networks.

Near Earth Asteroids represent potential threats to human life because their trajectories may bring them in the proximity of the Earth. Monitoring these objects could help predict future impact events, but such efforts are hindered by the large numbers of objects that pass through the Earth’s vicinity. Additionally, there is also the problem of distinguishing asteroids from other objects in the night sky, which implies sifting through large sets of telescope image data. Within this context, we believe that employing machine learning techniques could greatly improve the detection process by sorting out the most likely asteroid candidates to be reviewed by human experts. At the moment, the use of machine learning techniques is still limited in the field of astronomy and the main goal of the present paper is to study the effectiveness of deep CNNs for the classification of astronomical objects, asteroids in this particular case, by comparing some of the well-known deep convolutional neural networks, including InceptionV3, Xception, InceptionResNetV2 and ResNet152V2. We have applied transfer learning and fine-tuning on these pre-existing deep convolutional networks and from the results that we have obtained one can see the potential of using deep convolutional neural networks in the process of asteroid classification. The InceptionV3 model has the best results in the asteroid class, meaning that by using it, we loose the least number of valid asteroids.

This paper proposes a multimodal fusion 3D target detection algorithm based on the attention mechanism to improve the performance of 3D target detection. The algorithm utilizes point cloud data and information from camera. For image feature extraction, the ResNet50+FPN architecture extracts features at four levels. Point cloud feature extraction employs the voxel method and FCN to extract point and voxel features. The fusion of image and point cloud features is achieved through regional point fusion and regional voxel fusion methods. After information fusion, the Coordinate attention mechanism and SimAM attention mechanism extract fusion features at a deep level. The algorithm's performance is evaluated using the DAIR-V2X dataset. The results show that compared to the Part-A2 algorithm, the proposed algorithm improves the mAP value by 7.9% in BEV view and 7.8% in 3D view at IOU=0.5 (cars) and IOU=0.25 (pedestrians and cyclist). At IOU=0.7 (cars) and IOU=0.5 (pedestrians and cyclist), the mAP value of the SECOND algorithm is improved by 5.4% in the BEV view and 4.3% in the 3D view, compared to other comparison algorithms.

Extracting railway tracks is crucial for creating electronic railway maps. Traditional methods require significant manual labor and resources, while existing neural networks have limitations in efficiency and precision. To address these challenges, a railway track extraction method using an improved DeepLabV3+ model is proposed, which incorporates several key enhancements. Firstly, the encoder part of the method utilizes the lightweight network MobileNetV3 as the backbone extraction network for DeepLabV3+. Secondly, the decoder part adopts the lightweight universal upsampling operator CARAFE for upsampling. Lastly, to address any potential extraction errors, morphological algorithms are applied to optimize the extraction results. Additionally, a dedicated railway track segmentation dataset is created to train and evaluate the proposed method. The experimental results demonstrate that the model achieves impressive performance on the railway track segmentation dataset and DeepGlobe dataset. The MIoU scores are 88.93% and 84.72%, with Recall values of 89.02% and 86.96%. Moreover, the overall accuracy stands at 97.69% and 94.84%. The algorithm's operation time is about 5% lower in comparison to the original network. Furthermore, the morphological algorithm effectively eliminates errors like holes and spots. These findings indicate the model's accuracy, efficiency, and the enhancement brought by the morphological algorithm in error elimination.

The creation of crop-type maps from satellite data has proven challenging, often impeded by a lack of accurate in-situ data. This paper aims to demonstrate a method for crop-type (ie. Maize, Wheat and Other) recognition based on Convolutional Neural Networks using a bottom-up approach. We trained the model with a highly accurate dataset of crowdsourced labelled street-level imagery. Classification results achieved an AUC of 0.87 for wheat, 0.85 for maize and 0.73 for other. Given that wheat and maize are the two most common food crops globally, combined with an ever-increasing amount of available street-level imagery, this approach could help address the need for improved crop-type monitoring globally. Challenges remain in addressing the noisy aspect of street-level imagery (ie. buildings, hedgerows, automobiles, etc.), where a variety of different objects tend to restrict the view and confound the algorithms

In this study, we present a statistical forecasting framework and assess its efficacy using a range of established machine learning algorithms for predicting Particulate Matter (PM) concentrations in the Arctic, specifically in Pallas (FI), Reykjavik (IS), and Tromso (NO). Our framework leverages historical ground measures and 24-hour predictions from 9 models provided by the Copernicus Atmosphere Monitoring Service (CAMS) to provide PM predictions for the following 24 hours. Furthermore, we compare the performance of various memory cells based on artificial neural networks (ANN), including recurrent neural networks (RNNs), gated recurrent units (GRUs), long short-term memory networks (LSTMs), echo state networks (ESNs), and windowed multi-layer perceptrons (MLPs), which are commonly employed in time series forecasting tasks. Irrespective of the chosen memory cell type, our results demonstrate that the proposed framework consistently outperforms the CAM models in terms of mean squared error (MSE), with average improvements ranging from 25% to 40%. Additionally, we investigate the impact of outliers on the overall model performance.