1. Introduction

• Proposed a novel deep learning framework combining NFE and AutoInt models for automatic cervical cancer image classification.
• Utilized the pre-trained VGG16 model for high-quality feature extraction, enhancing feature extraction efficiency.
• Conducted an extensive evaluation of multiple machine learning classifiers (KNN, LGBM, ECT, etc.), achieving near-perfect classification accuracy.
• Optimized time efficiency, with models like LDA and ECT performing well in time-sensitive tasks.
• Employed a robust, well-labeled dataset and K-fold cross-validation to ensure model reliability and generalizability.

2. Literature Review

Despite the notable progress highlighted in the literature, it's important to acknowledge certain limitations within the existing studies. One recurring challenge is the availability and standardization of datasets. Many studies rely on specific datasets, such as SIPaKMeD or Herlev, which cannot fully capture the diversity and variability present in real-world clinical scenarios. Limited access to large, diverse datasets can impact the generalizability of the proposed models. A notable limitation in several studies is the relatively small sample sizes, which can impact the statistical robustness of the findings. Limited access to extensive datasets can constrain the ability to fully explore the nuances of cervical cancer variations and subtypes. Another common limitation lies in the interpretability of deep learning models. While these models often demonstrate impressive accuracy, understanding the underlying decision-making processes remains a challenge.

Table 1. Comparison of Machine Learning and Deep Learning Models for Cervical Cell Image Classification – Accuracy Rates Across Various Studies.

Table 1. Comparison of Machine Learning and Deep Learning Models for Cervical Cell Image Classification – Accuracy Rates Across Various Studies.

Study	Model Used	Accuracy (%)	Dataset
Study [17]	ResNet-50, XGB, SVM, RF	97 (ResNet-50), 82 (XGB), 84 (SVM), 79 (RF)	Cervicography images (4119)
Study [18]	Pre-trained DL models	82	Cervigram images
Study [19]	ResNet101, SVM	100 (SVM, normal vs abnormal), 97.3 (overall)	Pap smear images
Study [20]	ResNet-152	94.89	SIPaKMeD pap-smear images
Study [21]	Transfer Learning	85	Histopathology images (59 WSIs)
Study [22]	DL-based hybrid deep feature fusion	99.85 (2-class), 99.14 (5-class)	SIPAKMED and Herlev datasets
Study [23]	CNN (AlexNet with padding)	87.32 (AlexNet with padding)	SIPaKMeD image dataset
Study [24]	Hybrid model (Darknet53 + Mobilenetv2)	98.9 (SVM)	Pap smear images
Study [25]	CNN (4-layer simplified)	91.13	SIPaKMeD dataset (5 cell types)
Current Study	KNN, LGBM, ECT, RF	99.96 (KNN), 99.92 (LGBM), 99.88 (ECT)	Public dataset (25,000 images)

3. Materials and Methods

3.2 Dataset:

The dataset utilized in this study, retrieved from [28], comprises a substantial collection of 25,000 images. These images were meticulously captured using a Charge-Coupled Device (CCD) camera, meticulously modified to seamlessly integrate with an optical microscope. This tailored approach ensures the acquisition of precise depictions of squamous cells, contributing to the dataset's robustness and relevance to cervical cancer analysis. Categorized into five distinct groups—Dyskeratotic, Koilocytotic, Parabasal, Metaplastic, and Superficial-Intermediate—the dataset captures a comprehensive spectrum of morphological characteristics inherent to cervical cells. Dyskeratotic cells, for instance, exhibit early and aberrant keratinization, distinguishing them by specific visual features. Koilocytotic cells, on the other hand, are characterized by vesicular nuclei, often observed in binucleated or multinucleated arrangements.

To further enhance the dataset, we have collected an additional 1,000 images per class from Sheikh Zayed Medical College, as shown in Figure 2, along with sample images from our own data collection in Figure 3 and from a Kaggle dataset in Figure 5. This augmentation strengthens the dataset's diversity and improves its capacity for accurate classification and analysis.

Figure 4 serves as a visual representation, providing an illustrative glimpse into the diversity of sample images sourced from the dataset. This diverse and well-defined dataset, characterized by distinct morphological categories, forms the foundation for robust model training and evaluation in the subsequent stages of the research.

Figure 5. Samples images from the dataset.

Figure 5. Samples images from the dataset.

3.3 NFE Feature Extraction

The Neural Feature Extractor Model (NFEModel) is a customized neural network that leverages the architecture of the famous VGG16 model, with its focus on extracting features from images. This model modifies the VGG16 convolutional network, which is well-known for its high performance in classifying images, to focus on tasks that include analyzing and identifying complex patterns in images, such as medical imaging for cervical cancer. The NFEModel gains an advantage by utilizing the pre-trained weights obtained from the ImageNet database, which provides it with a strong and comprehensive knowledge of various visual content. It excludes the uppermost layer of the original VGG16 model to provide customization based on unique project requirements [29]. After applying these changed layers to the images, the model utilizes a Global Average Pooling (GAP) layer. This layer reduces the large amount of data from the previous convolutional layers by averaging down the spatial information, making it more manageable [30]. The simplified collection of features is now prepared for additional analysis or classification in following phases of the model, making the NFEModel an essential initial step in intricate image-based diagnosis and classification systems.

The model architecture described is designed for efficient feature extraction from images, using a pre-trained VGG16 model as its backbone. The VGG16 model, originally trained on the ImageNet dataset, has been shown to be highly effective for recognizing and analyzing common visual patterns in large-scale image datasets. In this model, the top layers responsible for classification are excluded (include_top=False), ensuring that the focus remains on the lower convolutional layers for feature extraction. The input to the model consists of images resized to 512×512×3, representing the height, width, and three-color channels (RGB).

The feature extraction process proceeds with a Global Average Pooling (GAP) layer, which significantly reduces the dimensionality of the feature maps produced by VGG16. Mathematically, for a 3D tensor x∈R^h×w×d representing the feature maps, where ℎ and w are the spatial dimensions and d is the depth, the GAP layer computes the average of all values in each feature map. This can be expressed as:

$$yk={\displaystyle \frac{1}{h\times x}} \sum _{i=0}^h\sum _{j=1}^w{x}_{ij}k$$

(1)

This operation compresses spatial information, creating a lower-dimensional but representative feature vector that retains the most essential information from the input image [31].

Following the GAP layer, a Dense (fully connected) layer with 256 neurons is applied. The Dense layer performs the following transformation on the input y:

$$z=f\left(w_y+b\right)$$

(2)

where W∈R^256×d is the weight matrix, b∈R²⁵⁶ is the bias vector, and f is the ReLU (Rectified Linear Unit) activation function. The ReLU function, defined as f(x)=max(0,x), introduces non-linearity, allowing the network to learn complex relationships between the input features [32].

To mitigate overfitting, the model includes a Dropout layer with a dropout rate of p=0.5. During training, dropout randomly sets a fraction p of the neurons to zero, as defined by:

$${\widehat{z}}_i=\left\{\begin{array}{c}{z}_i with probability (1-p) \\ 0 with probability p\end{array}\right.$$

(3)

This technique helps the model generalize better by preventing it from relying too heavily on specific neurons or patterns in the data [33].

The final layer is another Dense layer with 128 neurons, which further refines the extracted features. The mathematical operation is similar to the earlier Dense layer, with the final output represented as:

$$\widehat{y}=f\left({\acute{w}}_{\widehat{z}}+\acute{b}\right)$$

(4)

where W′∈R^128×256 is the weight matrix, and b′∈R¹²⁸ is the bias term.

The data processing pipeline includes an Image Data Generator, which loads and preprocesses the images in batches. The images are resized to 512×512, and the preprocess input function normalizes pixel values for compatibility with VGG16, ensuring the images are in a format the model can process efficiently [29].

Figure 6. Architecture of the NFE Model.

Figure 6. Architecture of the NFE Model.

The use of powers of 2 for layer sizes is justified by its computational efficiency, particularly in GPU-based frameworks like TensorFlow and PyTorch. These sizes align better with binary systems, leading to marginally improved memory usage and performance. Additionally, the powers of 2 provide a convenient scaling factor for hyperparameter tuning, simplifying experimentation and reducing the search space for model optimization. While not strictly necessary, this approach strikes a balance between performance and practical ease of use.

3.3.1 AutoInt Features Extraction

The AutoInt model is a neural network structure specifically created to understand and represent the complex connections between features in datasets with a large number of dimensions. AutoInt's strength lies in its ability to automatically learn feature interactions through self-attention mechanisms, which are well-suited for capturing complex relationships between input features, especially in high-dimensional datasets like medical imaging. Unlike traditional models that rely on manual feature engineering, AutoInt identifies important feature interactions at multiple levels, focusing on the most relevant combinations. This leads to improved generalization and prediction performance. We will expand on this theoretical foundation to better explain how AutoInt enhances performance, particularly in tasks involving intricate patterns such as medical image classification. AutoInt utilizes attention processes to automatically identify and represent interactions at multiple levels, instead of relying on manually designed feature interactions as traditional systems do. The fundamental concept underlying AutoInt is to employ self-attention layers, similar to those present in Transformer models, to acquire knowledge about the significance of interactions between pairs of features without relying on explicit manual feature engineering [34]. This strategy allows the model to concentrate on the most pertinent feature combinations, which could enhance prediction performance in tasks such as classification, regression, and recommendation systems.

AutoIntModel, is designed for automatic feature interaction learning through a series of dense (fully connected) layers. It processes input data in a hierarchical fashion, gradually reducing the dimensionality while applying non-linear transformations, thereby enabling the model to capture complex relationships between input features. The use of dense layers in this model is fundamental for transforming and combining features, which is a critical step in many machine learning tasks, such as classification and regression.

The first layer of the model, referred to as Dense Layer 1, consists of 128 units (neurons) and utilizes the Rectified Linear Unit (ReLU) activation function. The dense layer can be mathematically described as:

$${\mathrm{z}}_1\mathrm{}=\mathrm{f}\left({\mathrm{W}}_{1\cdot \mathrm{x}}+{\mathrm{b}}_1\mathrm{}\right)$$

(5)

where W1∈R ^128×d is the weight matrix, b1∈R¹²⁸ is the bias vector, x∈R^d is the input vector of dimensionality d, and f represents the ReLU activation function, defined as f(z)=max (0, z). This operation applies a linear transformation to the input data, followed by the ReLU activation, which introduces non-linearity into the model. The non-linearity enables the model to learn complex feature interactions by allowing certain neurons to become inactive (outputting zero), depending on the value of their inputs [32].

Following Dense Layer 1, the model passes the output to Dense Layer 2, which consists of 64 units and similarly applies the ReLU activation function. This layer can be described mathematically as:

$${\mathrm{z}}_2\mathrm{}=\mathrm{f}\left({\mathrm{W}}_{2\cdot {\mathrm{z}}_1}+{\mathrm{b}}_2\mathrm{}\right)$$

(6)

where W₂∈R^64×128 and b2 ∈R⁶⁴ are the weight matrix and bias vector, respectively. Dense Layer 2 further refines the feature interactions learned in the first layer by applying another set of transformations and non-linear activations. This hierarchical structure, where each dense layer learns progressively more abstract representations of the input, is a standard approach in deep learning architectures for modeling complex patterns in data [35].

The final layer of the model, called the Output Layer, consists of 32 units and also utilizes the ReLU activation function. The transformation applied in the output layer is given by:

$${\mathrm{z}}_3\mathrm{}=\mathrm{f}\left({\mathrm{W}}_{3\cdot {\mathrm{z}}_2}+{\mathrm{b}}_3\mathrm{}\right)$$

(7)

where W₃∈R^32×64 and b₃∈R³². The output of this layer is a 32-dimensional vector, representing the final feature interactions learned by the model. These interactions are the result of the sequential transformations and activations applied through the dense layers. Depending on the specific task, this output could be used for further processing, such as classification or regression, in subsequent layers or models.

The use of dense layers for automatic feature interaction learning has been widely studied and found to be effective in capturing intricate patterns in data [36]. The ReLU activation function is particularly advantageous in such models due to its computational simplicity and ability to prevent gradient vanishing during backpropagation, a common problem in deep neural networks. The dense structure, combined with ReLU activations, facilitates efficient and scalable learning from high-dimensional input data.

Figure 7. Architecture of the AutoInt Model.

Figure 7. Architecture of the AutoInt Model.

4. Results

NFE Model Features Results:

AutoInt model features results:

Combined Features Results:

5. K-Fold Validations

6. PREDICTION TIME

8. Conclusion

References

1. P. A. Cohen, A. Jhingran, A. Oaknin, and L. Denny, "Cervical cancer," The Lancet, vol. 393, no. 10167, pp. 169-182, 2019.
2. S. E. Waggoner, "Cervical cancer," The lancet, vol. 361, no. 9376, pp. 2217-2225, 2003.
3. K. Canfell et al., "Mortality impact of achieving WHO cervical cancer elimination targets: a comparative modelling analysis in 78 low-income and lower-middle-income countries," The Lancet, vol. 395, no. 10224, pp. 591-603, 2020. [CrossRef]
4. R. L. Siegel, K. D. Miller, and A. Jemal, "Cancer statistics, 2019," CA: a cancer journal for clinicians, vol. 69, no. 1, pp. 7-34, 2019. [CrossRef]
5. A. Stafl, "Cervicography: a new method for cervical cancer detection," American Journal of Obstetrics and Gynecology, vol. 139, no. 7, pp. 815-821, 1981. [CrossRef]
6. M. F. Janicek and H. E. Averette, "Cervical cancer: prevention, diagnosis, and therapeutics," CA: a cancer journal for clinicians, vol. 51, no. 2, pp. 92-114, 2001. [CrossRef]
7. J. S. Mandelblatt et al., "Costs and benefits of different strategies to screen for cervical cancer in less-developed countries," Journal of the National Cancer Institute, vol. 94, no. 19, pp. 1469-1483, 2002. [CrossRef]
8. J. R. Thomas C Wright, "Cervical cancer screening in the 21st century: is it time to retire the PAP smear?," Clinical obstetrics and gynecology, vol. 50, no. 2, pp. 313-323, 2007. [CrossRef]
9. C. Johnson-Greene. "Don’t Wait for Symptoms of Cervical Cancer to Appear." Henderson. https://universityhealthnews.com/daily/cancer/dont-wait-for-symptoms-of-cervical-cancer-to-appear/ (accessed 7-12-2023.
10. M. Ottaviano and P. La Torre, "Examination of the cervix with the naked eye using acetic acid test," American journal of obstetrics and gynecology, vol. 143, no. 2, pp. 139-142, 1982. [CrossRef]
11. W. Small Jr et al., "Cervical cancer: a global health crisis," Cancer, vol. 123, no. 13, pp. 2404-2412, 2017. [CrossRef]
12. M. Schiffman and P. E. Castle, "The promise of global cervical-cancer prevention," New England Journal of Medicine, vol. 353, no. 20, pp. 2101-2104, 2005. [CrossRef]
13. S. M. Ismail et al., "Observer variation in histopathological diagnosis and grading of cervical intraepithelial neoplasia," British Medical Journal, vol. 298, no. 6675, pp. 707-710, 1989. [CrossRef]
14. D. C. Sigler and J. W. Howe, "Inter-and intra-examiner reliability of the upper cervical X-ray marking system," Journal of Manipulative and Physiological Therapeutics, vol. 8, no. 2, pp. 75-80, 1985.
15. R. Lozano, "Comparison of computer-assisted and manual screening of cervical cytology," Gynecologic oncology, vol. 104, no. 1, pp. 134-138, 2007. [CrossRef]
16. A. H. Shahid and M. P. Singh, "A deep learning approach for prediction of Parkinson’s disease progression," Biomedical Engineering Letters, vol. 10, pp. 227-239, 2020. [CrossRef]
17. Y. R. Park, Y. J. Kim, W. Ju, K. Nam, S. Kim, and K. G. Kim, "Comparison of machine and deep learning for the classification of cervical cancer based on cervicography images," Scientific Reports, vol. 11, no. 1, p. 16143, 2021. [CrossRef]
18. Z. Alyafeai and L. Ghouti, "A fully-automated deep learning pipeline for cervical cancer classification," Expert Systems with Applications, vol. 141, p. 112951, 2020. [CrossRef]
19. H. Alquran et al., "Cervical cancer classification using combined machine learning and deep learning approach," Comput. Mater. Contin, vol. 72, no. 3, pp. 5117-5134, 2022. [CrossRef]
20. A. Tripathi, A. Arora, and A. Bhan, "Classification of cervical cancer using Deep Learning Algorithm," 2021: IEEE, pp. 1210-1218.
21. T. Majeed et al., "Transfer Learning approach for Classification of Cervical Cancer based on Histopathological Images," 2023: IEEE, pp. 1-5.
22. M. M. Rahaman et al., "DeepCervix: A deep learning-based framework for the classification of cervical cells using hybrid deep feature fusion techniques," Computers in Biology and Medicine, vol. 136, p. 104649, 2021. [CrossRef]
23. T. Haryanto, I. S. Sitanggang, M. A. Agmalaro, and R. Rulaningtyas, "The utilization of padding scheme on convolutional neural network for cervical cell images classification," 2020: IEEE, pp. 34-38.
24. H. Bingol, "NCA-based hybrid convolutional neural network model for classification of cervical cancer on gauss-enhanced pap-smear images," International Journal of Imaging Systems and Technology, vol. 32, no. 6, pp. 1978-1989, 2022. [CrossRef]
25. S. Alsubai et al., "Privacy Preserved Cervical Cancer Detection Using Convolutional Neural Networks Applied to Pap Smear Images," Computational and Mathematical Methods in Medicine, vol. 2023, 2023. [CrossRef]
26. "GPU Architecture." https://colab.research.google.com/github/d2l-ai/d2l-tvm-colab/blob/master/chapter_gpu_schedules/arch.ipynb (accessed 10, December 2023).
27. "TensorFlow." https://www.tensorflow.org/ (accessed 11 November, 2023).
28. A. B. ARJUN BASANDRAI. "Medical Scan Classification Dataset." kaggle. https://www.kaggle.com/datasets/arjunbasandrai/medical-scan-classification-dataset (accessed 09, December, 2023).
29. arXiv:1409.1556, 2014.K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014.
30. A. Gulli and S. Pal, Deep learning with Keras. Packt Publishing Ltd, 2017.
31. arXiv:1312.4400, 2013.M. Lin, "Network in network," arXiv preprint arXiv:1312.4400, 2013.
32. V. Nair and G. E. Hinton, "Rectified linear units improve restricted boltzmann machines," in Proceedings of the 27th international conference on machine learning (ICML-10), 2010, pp. 807-814.
33. N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov, "Dropout: a simple way to prevent neural networks from overfitting," The journal of machine learning research, vol. 15, no. 1, pp. 1929-1958, 2014.
34. W. Song et al., "Autoint: Automatic feature interaction learning via self-attentive neural networks," in Proceedings of the 28th ACM international conference on information and knowledge management, 2019, pp. 1161-1170.
35. I. Goodfellow, "Deep learning," ed: MIT press, 2016.
36. R. Wang, B. Fu, G. Fu, and M. Wang, "Deep & cross network for ad click predictions," in Proceedings of the ADKDD'17, 2017, pp. 1-7.