1. Introduction

2. Methodology

2.1. Data Collection

The data used for this study was obtained from the cancer imaging archive database, and its access is public, indicating it can be used publicly for research purposes [8]. The data consist of contrast-enhanced CT images from a total of 53 confirmed ACC patients, resulting in 18,215 CT images, as shown in Table 1. The data spanned 2006 to 2018 and included comprehensive clinical and pathological information, including the crucial Ki-67 index. All participants met specific inclusion criteria, encompassing a histopathological confirmation of ACC, surgical resection of the tumor, and determination of the Ki-67 index from the resected tissue as part of the standard evaluation. Only patients with available pre-resection contrast-enhanced CT scans of the abdomen were included in the dataset. Notably, patients whose Ki-67 index was quantified solely in biopsied tissue samples rather than in the resected whole tumor were excluded from the study. Previous research had established that Ki-67 quantification should be based on samples collected from the entire tumor [13]. Notably, prior to the availability of this dataset, there was no publicly-accessible library designed explicitly for adrenal lesions. Consequently, this dataset fills a critical gap in the field. It can serve as a valuable training set for machine learning algorithms for diverse applications, including adrenal tumor segmentation and classification.

Because the study involves binary classification modeling, we used a second data- the kidney tumor data. The data was obtained from the training set of the 2019 Kidney and Kidney Tumor Segmentation Challenge (KiTS19) [13]. The dataset consists of CT scans from 210 patients who underwent partial or radical nephrectomy at the University of Minnesota Medical Center. The CT scans were collected during routine patient care and were heterogeneous in terms of scanner manufacturers and acquisition protocols [13].

Table 1. The demographic distribution of the data.

Table 1. The demographic distribution of the data.

Parameter	Statistic	ACC	Kidney tumor
Gender - Male	Total	22	123
Gender - Female	Total	31	87
Age	Min/Max	22/82	1/90
	Mean	53.00	58.35
	Median	54	61
	Mode	56	73
	SD	13.47421	14.38798
Number of CT images used	Total	18,215	18,215
Data collection period	Years	2006 - 2018	2010 – mid-2018

2.3. Data Augmentation

We employed several data augmentation techniques to increase the size and diversity of our training dataset and enhance the robustness of the model by exposing it to a broader range of variations and scenarios. This involves applying various transformations to the existing data, such as flipping, shifting, rotating, or zooming, to create new training samples slightly different from the original ones, as shown in Table 2. The model becomes less sensitive to small changes or noise in the input data by augmenting the data, leading to improved generalization performance [16]. Data augmentation plays a crucial role in mitigating overfitting, where the model memorizes the training data rather than learning meaningful patterns. By introducing randomness and variability through data augmentation, the model is less likely to overfit and learns more robust and generalizable representations [16].

We applied various data augmentation techniques with specific ranges or scales. We utilized horizontal and vertical flips to mirror images, width and height shifts to randomly shift images within a fraction of their total width or height, shear transformations within a maximum shear angle, zooming in or out by a specified range, rotation within a certain angle, ZCA whitening (disabled in our case), and channel shifting within a specified range. These scales or ranges provide flexibility in controlling the extent of the applied transformations during data augmentation, allowing us to tailor the augmentation process to the specific requirements of our deep learning task and the characteristics of our dataset. By adjusting these parameters, we can effectively enhance the diversity and variability of our training data, thereby improving the model’s ability to learn and generalize from the limited available data.

Table 2. Data augmentation techniques and their range.

Table 2. Data augmentation techniques and their range.

Techniques	Range/Scale
Horizontal flip	True
Vertical flip	True
Width shift range	0.3
Height shift range	0.3
Shear range	0.2
Zoom range	0.2
Rotation range	0.2
ZCA whitening	False
Channel shift range	0.2

2.5. MSHA Model

The MSHA is a novel multi-modal contextual fusion CNN model. The study adopted the MSHA model to illustrate the capability of a CNN model developed from scratch with unique modifications and peculiarity to the data in use to attain optimal performance. It conforms with the data used and the frame of the problem we intend to solve. We designed the DL network as a simple CNN model and improved it by adding layers (convolution, pooling, and dense) and hyperparameter tuning until the utmost performance was achieved. While tuning the hyperparameter, we avoided overfitting and underfitting to ensure adequate generalization of the unseen data. This is visible in the model’s performance compared with the performance of the state-of-the-art pre-trained model evaluated in this study. The model is designed to handle both image classification and object detection tasks. It contains components dedicated to feature extraction, object detection using the region proposal network (RPN), and final classification.

We adopted a CNN architecture using the detection method of classification, which determines the output information from a single image. Our CNN architecture consists of a 2-dimensional (2-D) CNN architecture. The network comprises four convolutional and three max-pooling layers. The two layers used kernel and pool sizes of 3 × 3 and 2 × 2, respectively. A series of 4 fully connected layers with 128, 64, 32, and 16 neurons, and a final output layer, as shown in Figure 3, provided high-level reasoning with a sigmoid activation function for the classification task. The sigmoid activation function squashes the output of each neuron between the range of 0 and 1. Details regarding training are as follows: Adam, a gradient-based stochastic optimizer, was utilized with a batch size of 32 and a dropout of 25% on the convolutional and fully connected layers, respectively. We used the binary cross entropy loss to compare the predicted probabilities to the actual class output, which can be 0 or 1. Finally, we compile the model using accuracy metrics. The rectified linear activation function (ReLu) was the activation function of choice across the entire network before the final sigmoid activation function.

Our model stands out as a unique and novel approach in several aspects. Firstly, it incorporates additional features such as batch normalization layers, dropout layers, residual connections, and attention mechanisms. These additions contribute to improved training stability, regularization against overfitting, enhanced gradient flow, and the ability to capture long-range dependencies. Secondly, our model is designed to handle both image classification and object detection tasks within a single unified framework. It consists of components dedicated to feature extraction, object detection using the region proposal network (RPN), and final classification. This multi-purpose architecture offers versatility and efficiency, eliminating the need for separate models for different tasks. Moreover, the attention mechanism in our model allows for capturing dependencies across spatial locations, enabling the model to focus on relevant image regions and understand contextual relationships. This attention mechanism enhances performance in both image classification and object detection tasks. Combining convolutional layers and residual connections also empowers the model to learn complex image features and structures effectively. The convolutional layers capture hierarchical features, while residual connections facilitate gradient flow and enable the training of deeper models. Lastly, our model’s flexibility and modifiability make it a valuable tool for researchers and practitioners.

Figure 3. The basic Architecture of the Proposed Model.

Figure 3. The basic Architecture of the Proposed Model.

3. Result and Discussion

3.1. Performance Evaluation Metrics

Performance evaluation metrics are crucial in developing, testing, and deploying machine learning models, allowing for more accurate and effective AI solutions [19]. They provide a way to quantitatively measure the accuracy, precision, sensitivity, specificity, and other aspects of the model’s performance. With performance evaluation metrics, it is possible to determine how well a model performs or compare different models’ performance. Performance evaluation metrics also help improve machine learning models’ transparency and interpretability, essential for building trust in these systems [20].

In this study, the performance of the novel MSHA model was evaluated using performance evaluation metrics. The results indicate that the MSHA model was highly accurate in detecting and classifying ACC and kidney tumors, with a 97.00% and 95.00% precision, as shown in Table 3. The model’s high precision score indicates it could correctly identify ACC and kidney tumors in most cases. Also, the sensitivity score of 94.00% for ACC and 97.00% for kidney tumors show that the model could correctly identify all positive cases. This indicates a highly sensitive model with promising clinical significance in accurately detecting ACC and kidney tumors in CT images. Subsequently, the specificity score of 96.80% for ACC and 94.50% for kidney tumors show that the model could correctly identify all negative cases. This indicates that the model is highly specific and can accurately tell when the tumor is absent. Furthermore, the F1 score of 96.00% for ACC and kidney tumors indicates the model’s potential to balance precision and sensitivity efficiently. This means the model could identify the positive cases (ACC and kidney tumor) while correctly minimizing the false positives. Finally, the accuracy score of 95.65% indicates that the model could correctly classify the CT images of ACC and kidney tumors with a high degree of accuracy.

The high performance of the MSHA model in detecting and accurately classifying CT images of ACC and kidney tumors has significant clinical implications in providing a rapid and accurate diagnosis, particularly in regions with limited access to specialized medical care and facilities. The model highlights the potential of the MSHA model as a valuable tool in detecting and classifying ACC and kidney tumors using CT images.

Table 3. The Performance Evaluation Metrics for the MSHA Model.

Table 3. The Performance Evaluation Metrics for the MSHA Model.

	Precision %	Sensitivity %	Specificity %	F1 Score %	Accuracy %
ACC	97.00	94.00	96.80	96.00	95.65
Kidney tumor	95.00	97.00	94.50	96.00

3.3. Learning Curve

Model accuracy and model loss learning curves are important tools for evaluating the performance of deep learning models. They provide information about the accuracy and loss of a model over time during the training process, which can help identify potential issues with the model’s performance and guide improvements to the model [5]. The model accuracy learning curve shows the model’s accuracy on the training and validation datasets over time. It can reveal whether the model is overfitting or underfitting the training data [22]. An overfit model will have high accuracy on the training data but low accuracy on the validation data, indicating that it needs to generalize better to new and unseen data [23]. An underfit model will have low accuracy on both the training and validation data, indicating that it needs to learn the patterns in the data better [24]. Monitoring the model accuracy learning curve makes it possible to identify the optimal number of epochs to train the model and ensure that it is not overfitting or underfitting.

The MSHA model produced a good fit and could learn the underlying patterns in the data without overfitting or underfitting. The increased training accuracy over the epochs indicates that the model is learning and improving. As shown in Figure 5, the training started slowly at epoch 0 and maintained an upward and steady increase to produce an approximate training accuracy of 92.7% between epochs 44-50. As important as the training accuracy, it is important to evaluate the model’s performance on the validation data, which represents new, unseen data, to ensure that the model is balanced with the training data. Similar to the training accuracy, the validation accuracy starts slowly and experiences a slight irregularity. However, it steadily increases until it plateaus at a validation accuracy of 94.7% at epoch 50, indicating that further training may not improve the model’s performance on new data.

The model loss learning curve shows the change in the loss function of a model as it trains over multiple epochs. The training loss decreases over time as the model learns to fit the data better, which is a good sign. The training loss started at 0.6919 at epoch 1 and gradually decreased uniformly to 0.1805 at epoch 50. This indicates that the model improves its ability to predict the correct output with less error. On the other hand, the validation loss steadily decreased from epochs 1 to 11 with consistent fluctuation between epochs 14-43 before stabilizing at epoch 45 and maintaining a uniform decrease. The validation loss measures the difference between the predicted and actual outputs on a data set that the model has not seen during training. Thus, it estimates the model’s performance on new data. A good fit is achieved when the model’s low training and validation losses have stabilized over several epochs. Low training and high validation loss indicate overfitting, while high values of both losses may suggest underfitting [5]. Therefore, analyzing the model accuracy and model loss learning curves is crucial in evaluating a model’s performance and deciding how to improve it.

Figure 5. The Model Accuracy and Loss of the MSHA Model.

Figure 5. The Model Accuracy and Loss of the MSHA Model.

3.5. Comparative Evaluation with State-of-Art Transfer Learning Techniques

Comparing our model with other state-of-the-art transfer learning techniques such as ResNet50, VGG16, VGG19, and InceptonV3 provides a benchmark for evaluation, assesses generalizability, contributes to advancements in the field, and aids decision-making for practical applications. Furthermore, these models are well-known and widely used transfer learning techniques in computer vision with remarkable performance and significant contributions to image recognition tasks. By comparing the MSHA model with these well-established models, we can effectively assess its performance, competitiveness, and potential superiority. This comparison will help position the novel MSHA model within the context of existing state-of-the-art approaches and establish its credibility and relevance in computer vision.

Similar to the implementation of the MSHA model, all the state-of-art transfer learning techniques used were implemented on the Jupyter Notebook. The last layers were frozen during the fine-tuning process to preserve learned representations and prevent them from being modified or overwritten, allowing the model to focus on adapting its parameters to the new task at hand and reducing the number of parameters that need to be updated, making the training process more efficient and faster. Furthermore, three dense layers with the ReLU activation function were adopted before the final output layer with the sigmoid activation function. Finally, the model was trained using 50 epochs while callbacks were adopted to prevent overfitting.

The ResNet50 is a convolutional neural network (CNN) architecture known for its deep residual learning framework. It addresses the problem of vanishing gradients in very deep networks, allowing for the training of extremely deep models. It has been successful in various image classification challenges and is renowned for its ability to capture intricate features from images [28]. The VGG16 and VGG19 are deep CNN architectures developed by the Visual Geometry Group (VGG) at the University of Oxford. These models are characterized by their uniform architecture, consisting of multiple stacked convolutional and fully connected layers. VGG16 and VGG19 are known for their excellent performance on large-scale image classification tasks, exhibiting high accuracy due to their deep and fine-grained feature extraction capabilities [29]. The InceptionV3, also known as GoogleNet, introduced the concept of inception modules, efficiently capturing multi-scale features by employing parallel convolutions at different spatial resolutions. This architecture reduces the computational complexity while maintaining high accuracy. InceptionV3 has been widely used in various image recognition tasks and has demonstrated excellent object recognition and localization performance [30].

The comparative analysis of the novel MSHA model with state-of-the-art models reveals significant variation in performance and provides differing clinical implications. The novel MSHA model significantly outperformed other models with improved performance. The MSHA model outperforms ResNet50, VGG16, VGG19, and InceptionV3 in terms of accuracy, precision, sensitivity, specificity, F1 score, AUC, and loss. It achieves an accuracy of 96.65%, significantly higher than the other models, as shown in Table 4. The MSHA model also demonstrates higher precision, sensitivity, and specificity than ResNet50, VGG16, VGG19, and InceptionV3. Its F1 score of 96.0% indicates a superior balance between precision and sensitivity. Additionally, the MSHA model achieves an AUC value of 0.99, reflecting excellent discriminative ability, and has a lower loss value of 0.108, indicating better optimization and fewer errors.

The improved accuracy of the MSHA model holds significant clinical significance. With a 96.65% accuracy in correctly classifying CT images of ACC and kidney tumors, the MSHA model provides reliable and precise results. This high level of accuracy can greatly benefit oncologists and healthcare professionals involved in diagnosing and treating these types of tumors. It reduces the chances of misclassification, enabling early detection and appropriate intervention and improving patient outcomes.

By outperforming other models across various metrics, the MSHA model offers a more robust and accurate tool for assisting oncologists in making critical decisions. Its higher precision, sensitivity, and specificity values ensure better identification of positive cases and accurate exclusion of negative cases. The model’s superior F1 score indicates a well-balanced trade-off between precision and sensitivity, striking an optimal equilibrium in tumor classification. Moreover, the high AUC value of 0.99 signifies its excellent discriminative ability, distinguishing between ACC and kidney tumors with high confidence. The MSHA model’s lower loss value demonstrates effective error minimization and optimization, enhancing its overall performance.

The superior performance of the novel MSHA model can be attributed to several factors, such as its unique architectural design, effective training strategy, and better capability to learn and represent the relevant features in the skin lesion images. The incorporation of specific design choices, such as the Mixed-Scale Dense Convolution Layer, Self-Attention Mechanism, Hierarchical Feature Fusion, and Attention-Based Contextual Information, enabled the MSHA model to capture and extract relevant features more effectively for skin lesion classification. The MSHA model’s architecture seems better suited to learning and representing the intricate patterns and structures in the skin lesion images associated with ACC and kidney tumors. Also, the MSHA model was trained using an optimized configuration and effective training strategies, such as carefully selecting hyperparameters such as learning rate, batch size, and regularization techniques. These configurations will facilitate faster convergence and help the model find a more optimal solution.

Despite ResNet50 being a sophisticated model known for its deep architecture and skip connections, it exhibited the least performance compared to other models. The lower precision and F1 score of 66.0% suggests that ResNet50 had a higher rate of false positives and false negatives, resulting in suboptimal predictions. This could result from the unique data characteristic of the data used and the architectural complexity of the ResNet50 model, which may not be ideal for detecting CT images associated with ACC and kidney tumors. Also, the VGG19 model, designed as a more sophisticated version of VGG16, achieved slightly lower performance with a sensitivity, specificity, F1 score, TP, AUC, and accuracy lower than the VGG16 model. This difference can also be attributed to factors such as increased model complexity in the case of the VGG19 model leading to insufficient representation of the specific features relevant to the classification task.

Table 4. The Comparative Analysis of the Novel MSHA and Other State-of-the-art Models.

Table 4. The Comparative Analysis of the Novel MSHA and Other State-of-the-art Models.

Models	Precision %	Sensitivity %	Specificity %	F1 Score %	TP	FP	TN	FN	AUC	Loss	Accuracy %
MSHA	96.0	96.0	96.0	96.0	3444	204	3525	113	0.99	0.108	96.65
ResNet50	66.0	66.0	66.5	66.0	2475	1173	2335	1303	0.72	0.615	66.02
VGG16	81.0	81.0	81.0	81.0	2872	776	3004	634	0.90	0.424	80.65
VGG19	81.0	80.0	80.0	80.0	2667	981	3181	457	0.89	0.424	80.26
InceptionV3	72.0	72.0	72.0	72.0	2900	748	2302	1336	0.79	0.592	71.40

4. Conclusions

References

1. Farina, E., Nabhen, J. J., Dacoregio, M. I., Batalini, F., & Moraes, F. Y. (2022). An overview of artificial intelligence in oncology. Future Science OA, 8(4). [CrossRef]
2. Talukder, Md. A. (2022). Machine Learning-Based Lung and Colon Cancer Detection Using Deep Feature Extraction and Ensemble Learning. [CrossRef]
3. Nandini, B. (2021). Detection of skin cancer using inception V3 and inception V4 Convolutional Neural Network (CNN) for Accuracy Improvement. Revista Gestão Inovação e Tecnologias, 11(4), 1138–1148. [CrossRef]
4. Reddy, Y. V., Chandana, G., Redddy, G. C., Kumar, A., Kumar, S., & Ahmed, Dr. S. (2023). Lung cancer detection using Yolo CNN algorithm. International Journal of Research Publication and Reviews, 4(5), 5297–5300. [CrossRef]
5. Uzun Ozsahin, D., Onakpojeruo, E. P., Uzun, B., Mustapha, M. T., & Ozsahin, I. (2023). Mathematical assessment of machine learning models used for brain tumor diagnosis. Diagnostics, 13(4), 618. [CrossRef]
6. Bi, W. L., Hosny, A., Schabath, M. B., Giger, M. L., Birkbak, N. J., Mehrtash, A., Allison, T., Arnaout, O., Abbosh, C., Dunn, I. F., Mak, R. H., Tamimi, R. M., Tempany, C. M., Swanton, C., Hoffmann, U., Schwartz, L. H., Gillies, R. J., Huang, R. Y., & Aerts, H. J. (2019). Artificial Intelligence in cancer imaging: Clinical challenges and applications. CA: A Cancer Journal for Clinicians. [CrossRef]
7. Shi, Y., Qin, Y., Zheng, Z., Wang, P., & Liu, J. (2023). Risk factor analysis and multiple predictive machine learning models for mortality in COVID-19: A Multicenter and multi-ethnic cohort study. The Journal of Emergency Medicine. [CrossRef]
8. Nolan, T., & Kirby, J. (2023, June 9). Voxel-level segmentation of pathologically-proven adrenocortical carcinoma with Ki-67 expression (adrenal-ACC-KI67-SEG). Voxel-level segmentation of pathologically-proven Adrenocortical carcinoma with Ki-67 expression (Adrenal-ACC-Ki67-Seg) - The Cancer Imaging Archive (TCIA) Public Access - Cancer Imaging Archive Wiki. https://wiki.cancerimagingarchive.net/pages/viewpage.action?pageId=93257945.
9. Singh, Y., Kelm, Z. S., Faghani, S., Erickson, D., Yalon, T., Bancos, I., & Erickson, B. J. (2023). Deep learning approach for differentiating indeterminate adrenal masses using CT imaging. Abdominal Radiology. [CrossRef]
10. Avanzo, M., Wei, L., Stancanello, J., Vallières, M., Rao, A., Morin, O., Mattonen, S. A., & El Naqa, I. (2020). Machine and deep learning methods for radiomics. Medical Physics, 47(5). [CrossRef]
11. Levy, L., & Tsaltas, J. (2021). Recent advances in benign gynecological laparoscopic surgery. Faculty Reviews, 10. [CrossRef]
12. Wei, J., Zhu, R., Zhang, H., Li, P., Okasha, A., & Muttar, A. K. H. (2021). Application of PET/CT image under convolutional neural network model in postoperative pneumonia virus infection monitoring of patients with non-small cell lung cancer. Results in Physics, 26, 104385. [CrossRef]
13. Heller, N., Sathianathen, N., Kalapara, A., Walczak, E., Moore, K., Kaluzniak, H., ... & Weight, C. (2019). The kits19 challenge data: 300 kidney tumor cases with clinical context, ct semantic segmentations, and surgical outcomes. arXiv preprint arXiv:1904.00445.
14. Subashini, M.; Sahoo, S.; Sunil, V.; Easwaran, S. A Non-Invasive Methodology for the Grade Identification of Astrocytoma Using Image Processing and Artificial Intelligence Techniques. Expert Syst. Appl. 2016, 43, 186–196. [CrossRef]
15. Wang, S.; Hamian, M. Skin Cancer Detection Based on Extreme Learning Machine and a Developed Version of Thermal Exchange Optimization. Comput. Intell. Neurosci. 2021, 2021, 9528664. [CrossRef]
16. Shorten, C.; Khoshgoftaar, T.M. A survey on image data augmentation for Deep Learning. J. Big Data 2019, 6, 60. [CrossRef]
17. Brownlee, J. (2020, August 25). Use early stopping to halt the training of neural networks at the Right Time. MachineLearningMastery.com. https://machinelearningmastery.com/how-to-stop-training-deep-neural-networks-at-the-right-time-using-early-stopping/. (accessed on 28 April 2023).
18. 18. Tensorflow callbacks - how to monitor neural network training like a pro. Better Data Science. (n.d.). https://betterdatascience.com/tensorflow-callbacks-how-to-monitor-neural-network-training/#:~:text=If%20a%20value%20of%20the%20evaluation%20metric%20doesn%E2%80%99t,old%20learning%20rate%20multiplied%20by%20a%20user-defined%20factor. (accessed on 28 April 2023).
19. Mustapha, M. T.; Ozsahin, D. U.; Ozsahin, I.; Uzun, B. Breast cancer screening based on supervised learning and multi-criteria decision-making. Diagnostics, 2022. 12(6), 1326. [CrossRef]
20. Uzun Ozsahin, D.; Mustapha, M. T.; Bartholomew Duwa, B.; Ozsahin, I. Evaluating the performance of deep learning frameworks for malaria parasite detection using microscopic images of peripheral blood smears. Diagnostics, 2022. 12(11), 2702. [CrossRef]
21. Seyer Cagatan, A.; Taiwo Mustapha, M.; Bagkur, C.; Sanlidag, T.; Ozsahin, D. U. An alternative diagnostic method for C. Neoformans: Preliminary results of deep-learning based detection model. Diagnostics, 2022. 13(1), 81.
22. Muralidhar, K. S. V. Learning curve to identify overfitting and underfitting in machine learning. Medium. Available online: https://towardsdatascience.com/learning-curve-to-identify-overfitting-underfitting-problems-133177f38df5. (accessed on 21 February, 2023).
23. Montesinos López, O. A.; Montesinos López, A.; Crossa, J. Overfitting, model tuning, and evaluation of prediction performance. Multivariate Statistical Machine Learning Methods for Genomic Prediction, 2022, 109–139.
24. Wolfe, C. R. Using transformers for computer vision. Medium. Available online: https://towardsdatascience.com/using-transformers-for-computer-vision-6f764c5a078b. (accessed on 25 April, 2023).
25. Trevisan, V. (2022, March 25). Interpreting ROC curve and ROC AUC for Classification Evaluation. Medium. https://towardsdatascience.com/interpreting-roc-curve-and-roc-auc-for-classification-evaluation-28ec3983f077.
26. Brownlee, J. (2020, September 15). ROC curves and precision-recall curves for imbalanced classification. MachineLearningMastery.com. https://machinelearningmastery.com/roc-curves-and-precision-recall-curves-for-imbalanced-classification/.(accessed on 28 April 2023).
27. Sabin, J. A. (2022). Tackling implicit bias in health care. New England Journal of Medicine, 387(2), 105–107. https://doi.org/10.1056/nejmp2201180. (accessed on 28 April 2023). [CrossRef]
28. Mukherjee, S. (2022, August 18). The annotated resnet-50. Medium. https://towardsdatascience.com/the-annotated-resnet-50-a6c536034758. (accessed on 28 April 2023).
29. Understanding VGG16: Concepts, architecture, and performance. Datagen. (2023, May 22). https://datagen.tech/guides/computer-vision/vgg16/.(accessed on 28 April 2023).
30. Alake, R. (2020, December 22). Deep learning: Understand the inception module. Medium. https://towardsdatascience.com/deep-learning-understand-the-inception-module-56146866e652. (accessed on 28 April 2023).