A lightweight deep learning method to identify different types of cervical cancer Authors: Md Humaion Kabir Mehedi, Moumita Khandaker, Shaneen Ara, Md Ashraful Alam, M F Mridha, Zeyar Aung Published: 2024 ABSTRACT AbstractCervical cancer is the second most common cancer in women’s bodies after breast cancer. Cervical cancer develops from dysplasia or cervical intraepithelial neoplasm (CIN), the early stage of the disease, and is characterized by the aberrant growth of cells in the cervix lining. It is primarily caused by Human Papillomavirus (HPV) infection, which spreads through sexual activity. This study focuses on detecting cervical cancer types efficiently using a novel lightweight deep learning model named CCanNet, which combines squeeze block, residual blocks, and skip layer connections. SipakMed, which is not only popular but also publicly available dataset, was used in this study. We conducted a comparative analysis between several transfer learning and transformer models such as VGG19, VGG16, MobileNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileViT, and Swin Transformer with the proposed CCanNet. Our proposed model outperformed other state-of-the-art models, with 98.53% accuracy and the lowest number of parameters, which is 1,274,663. In addition, accuracy, precision, recall, and the F1 score were used to evaluate the performance of the models. Finally, explainable AI (XAI) was applied to analyze the performance of CCanNet and ensure the results were trustworthy.Keywords: Cervical cancer, Cancer type identification, Deep learning, Lightweight algorithmSubject terms: Cancer imaging, Mathematics and computing, Computational science Introduction The term “Cervical cancer” is characterized by an abrupt and unusual growth of cells in the cervical lining. The cause of this disease is the human papillomavirus (HPV) attack on the body and the medium of transmission is sexual intercourse. This abnormal cell growth is called dysplasia or Cervical Intraepithelial Neoplasia (CIN) which is the first stage of cervical cancer1. Cervical cancer is primarily attributable to chronic infection with specific types of HPV, with HPV 16 and 18 accounting for about 70% of cervical cancer cases. Despite the availability of HPV vaccines, the virus still infects many people worldwide, particularly in low-income areas. Some more facts are causing cervical cancer, for instance, smoking increases the risk by two times because chemicals from cigarettes mutate cervical cells’ DNA and also compromise the immune system to eliminate the HPV virus. Immunosuppressed women, for example, those with HIV or undergoing treatment for AIDS or with other immunosuppressive disorders or taking immunosuppressive therapies are also at a higher risk because their immune systems are not strong enough to clear persistent HPV infections. Taking oral contraceptives has adverse health effects among women especially if they use it for the long term, it increases the chances of cervical cancer mainly because of hormonal effects on cervical cells. Additionally, women who have successfully carried multiple pregnancies to term face an increased risk, as the cervix becomes particularly vulnerable during pregnancy due to hormonal changes and the physical stress of childbirth. Other factors associated with increased risks include engaging in sexual activities at early ages with multiple sexual partners and economic factors including low access to adequate healthcare. Traditional treatments include surgery, chemotherapy, and radiotherapy, according to the patient’s case study2. Cervical cancer alone affects the lives of 57,000 women every year with 30,000 deaths recorded annually. Cervical cancer is the fourth most common cancer among women all over the world. In 2020, 604,000 new cases were recorded, and the death toll was 342,0003. Most of them are detected in developing and developed nations, and preventive methods including HPV vaccination and regular examination, are still hard to come by in these parts. Surveys indicate that low and middle-income countries, including Malawi in Sub-Saharan Africa, Latin America, and some parts of Asia, have experienced the highest increase in these incidences. In South Asian countries like India and Bangladesh cervical cancer still poses a great threat to women and this is attributed to poor awareness, lack of immunization, and poor cervical cancer screening services. As per the data of GLOBOCAN 2020, on the one hand, in Bangladesh, 8267 women are newly diagnosed every year with cervical cancer and 5214 women die from cervical cancer annually. On the other hand, in India 123,907 new cervical cancer cases are estimated in women annually and 77,348 women die of cervical cancer annually. This is equivalent to a mortality-to-incidence ratio of 62.4%4. On a larger scale, the World Health Organization (WHO) estimates that cervical cancer results in over $3 billion in lost productivity globally each year. Globally in 2020, there were an estimated 604,127 cervical cancer cases and 341,831 deaths, with a corresponding age-standardized incidence of 13.3 cases per 100,000 women-years (95% CI 13.3–13.3) and mortality rate of 7.2 deaths per 100,000 women-years (95% CI 7.2–7.3). Cervical cancer rate per 100,000 women ranged from 2.2 (2.9–2.4) in Iraq to 84.6 (74.8–94.3) in Eswatini. Mortality varied between 1.0 (95% CI 0.8–1.2) in Switzerland to 55.7 (95% CI 47.7–63.7) in Eswatini. Age-standardized incidence ranged from 67.9 (95% CI 65.7–70.1) per 100,000 person-years in Malawi and 65.5 (95% CI 63.0–67.9) in Zambia in Africa, 36.6 (95% CI 35.0–38.2) in Bolivia and 34.1 (95% CI 32.1–36). Such expenses in addition to stripping away female employees have a definite economic and social cost5. In the fight against cancer, early diagnostic techniques will aid in addressing the disease in its early stages and implementing preventive measures. Gardasil and Cervarix vaccines can prevent cervical cancer and HPV and it is most effective when given to girls and boys of 9 to 14 years before infection. Studies, such as “A Hybrid Deep Learning-Based Automated System for HPV Vaccination Impact Prediction”6 as noted by others Aylin et al., where robust vaccination programs are in places such as Australia and the UK, cervical cancer incidences are considerably low. Preliminary data sustained that vaccination in the ‘adjuvant’ setting following conization reduces the risk of recurrent Cervical Intraepithelial Neoplasia (CIN) by about 50–80%7–9. However, as with many other prices, the issue at hand proves extremely costly in low-income countries such as Bangladesh, where awareness is low, and the essentials of implementation; vaccination infrastructure remains severely underdeveloped. A Cost-Effective Analysis of HPV Vaccination in Low-Income Countries10 calls for a support system to overcome these barriers and enhance vaccination coverage in developing nations. As the majority of HPV-related lesions occur in the uterine cervix, hysterectomy protects from the risk of developing most HPV-related lesions11. One of the studies on cervical cancer found that while CNN models performed similarly to ViT models, CNNs were less effective at handling image complexity, particularly with diverse datasets6. Another study, which applied the Swin Transformer model for brain tumor classification, was limited by its evaluation on a single MRI dataset, affecting its generalizability10. One more study where a modified Swin Transformer for skin cancer diagnosis, despite improvements, had an accuracy of 89.36%, struggling with the diverse characteristics of skin cancer images12. A further study that utilized an enhanced EfficientNetv2 model for brain tumor classification achieved high accuracy but faced limitations due to its reliance on a single dataset, affecting its real-world applicability13. Behind every successful work, there is some motivation. Studying previous work in this field also motivates us. The recent deaths of individuals undergoing treatment have deeply concerned us. Our paper improves the previous works by offering a solution to their shortcomings and by presenting a CCanNet lightweight model appropriate for cervical cancer detection. This study makes two significant contributions. Our study extends prior deep learning models and proposes CCanNet, which outperforms more standard structures such as VGG16, VGG19, MobileNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileVit, and Swin Transformer. This paper contributes a model with fewer parameters but higher accuracy to the automated cancer detection literature, which suggests the introduction of a new option. Then, for automated and early detection, we developed a lightweight model named CCanNet. This model was ideal for deployment because it was computationally less intensive and could operate in setups with limited resources, making its implementation feasible in developing countries or local clinics. This model would also perform well on mobile devices and enable real-time analysis, reducing infrastructure costs. Early and accurate detection of cervical cancer lessens the mortality rate. If diagnostic tools were more accessible to the mass population, it would reduce the disparity in healthcare that provides quality treatment services to the underprivileged. This model has the potential to foster innovation in deep learning and medical imaging research, leading to the advancement of new technologies and applications. The most significant contribution to our suggested model is establishing a lightweight network in the spectrum of deep learning. Therefore, parallel studies on the existing state-of-the-art models have been described to show the comparison and improved features in this model. In addition, the process includes the application of data augmentation and different preprocessing techniques to enhance performance. Accuracy, precision, recall, F1 score, and confusion matrix are on show to conduct an evaluation. Moreover, GradCam and guided GradCam are applied to analyze the model’s performance. Thus, it provides a visual explanation of the affected area in the image data, which elevates an accurate prediction. The major contributions of this paper are: We introduce an improved lightweight deep learning model with only 1.27M parameters named CCanNe. The model combines residual blocks, squeeze blocks and skip connections. We introduce an improved lightweight deep learning model with only 1.27M parameters named CCanNe. The model combines residual blocks, squeeze blocks and skip connections. Compare the proposed model with the other state-of-the-art models such as VGG16, VGG19, MobleNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileViT and Swin Transformer, evaluating accuracy, precision, recall, F1 score, and confusion matrix. Compare the proposed model with the other state-of-the-art models such as VGG16, VGG19, MobleNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileViT and Swin Transformer, evaluating accuracy, precision, recall, F1 score, and confusion matrix. The data augmentation method has been used to expand data, and different preprocessing techniques have been applied to enhance performance. The data augmentation method has been used to expand data, and different preprocessing techniques have been applied to enhance performance. GradCam and guided GradCam have been applied to analyze the model’s performance and make the model more trustworthy. GradCam and guided GradCam have been applied to analyze the model’s performance and make the model more trustworthy. The “Introduction” section provides a brief explanation of the onset of cervical cancer, basic treatments, and the use of AI in automated detection. The section “Related works” talks about the most up-to-date model for solving the problem more accurately. This includes the application of machine learning and deep learning in various fields such as medical diagnostics and cervical cancer classification. Subsequently, the “Methodology” section, our study introduces a novel lightweight deep-learning model. Furthermore, the dataset, preprocessing, models, and evaluation metrics are also addressed. The “Results and discussions” section provides a comparison of the CCanNet model’s performance with that of extant models, as well as an explanation of the role of XAI in improving the model’s performance. Furthermore, the “Strengths and limitations” segment explains the major strengths and limitations of the proposed model. Lastly, the information in the “Conclusion and future works” section elucidates the rationale behind the optimal application of this study and identifies potential areas for continued research. Related works In recent years, researchers have explored and exploited several models namely, machine learning (ML), and deep learning (DL). In addition, there adaptive learning-based methods are available for elucidating crack image datasets. In the upcoming years, this has not only dramatically escalated the development but also the enactment of ML models. Multiple investigations have been performed on the cataloging of cervical cancers. The alarming statistics of cervical cancer rates globally have pushed researchers to perform more work in the identification, treatment, and diagnosis of this disease. As a result, much research has been conducted in the recent past on image data sets using deep learning methodologies. By analyzing medical images and databases, it has been understood that technologies have greatly impacted the efficiency and accuracy of identifying several types of disorders. For instance, machine learning and deep learning techniques have had a positive impact on Pap smear analysis for cervical cancer screening to overcome frequent challenges such as unauthorized data and image quality problems. Machine learning (ML) and deep learning (DL) models are essential for advancing medical diagnostics. Pap smear analysis (cervical cancer detection) was performed using ML and DL methodologies while labeling common snags such as data insufficiency and image quality issues14. Cervical cancer, for the most part, is caused by human papillomavirus (HPV), and is a major health concern for women. Successful treatment depends on early detection. Traditional methods such as the Pap test, colonoscopy, and HPV have been tested using advanced ML and DL models to enhance diagnostic accuracy. Kalbhor et al.15 and Devi and Thirumurugan proposed the successful classification of cervical cancer using multiple ML techniques. However, these techniques achieve high accuracy rates. Recent studies have employed septet mechanisms either trait engineering to improve classification performance or achieve notable accuracy improvements. Deep learning models have some rare abilities in medical image analysis. First, the multi-axis vision transformer framework uses ConvNeXtv2. Second, GRN-based MLP blocks attained 99.02% accuracy on the SIPaKMeD dataset16. Other noteworthy models included ResNet50V2, InceptionV3, and Xception, with ResNet50V2 displaying the highest accuracy for cervical smear images. progress urges that deep learning models to be replaced more incisively. Moreover, diagnostic procedures, enabling earlier and more accurate disease detection are easily improved through DL. Furthermore, cervical cancer, ML, and DL models are widely used to diagnose other categories of cancers and medical conditions, such as breast cancer, lung cancer, and colorectal cancer. In addition, it is instrumental in Alzheimer’s diagnosis, medical imaging, and pathology image segmentation. Advanced diagnostic tools, for instance, EEG encoding, CT imaging, and surgical navigation, are also being developed using these technologies. There are several feature reduction and classification techniques. Efficient feature depletion and classification are crucial for handling medical datasets; however, they often face problems owing to issues such as class imbalance and missing data. Several studies have attempted in different manners to overcome these challenges. In addition, Tanimu et al.17 not only used the decision tree (DT) classification algorithm but also LASSO for reducing features, addressing dataset challenges with SMOTETomek to attain exorbitant accuracy. An excellent classification performance was achieved by Abdoh et al.18 through the use of random forest (RF) with SMOTE along with additional feature reduction approaches. Adaptive learning-based methods are based on improving the model performance by addressing particular provocations such as class imbalance and incomplete datasets. For example, to generate synthetic data points for underrepresented classes, the ADASYN algorithm was used to stabilize the dataset and increase the detection accuracy of cervical cancer (Pap smear images). In this paper19, the commonly occurring obstacles in PAP smear such as insufficient data sufficiency not being enough, images containing varieties, and lack of clarity in quality have been minimized by providing an advanced framework multi-axis vision transformer. By integrating ADASYN with support vector machine (SVM) and convolutional neural network (CNN), researchers achieved an accuracy of 99.99%. Combined with ensemble learning techniques namely extreme gradient boosting, random forest, and extra tree classifier, this method acquires not only a high accuracy of 99.41% in guessing cervical cancer, but also shows the value of adaptive learning in handling incomplete and imbalanced datasets. Transfer learning-based techniques support the knowledge obtained from initially learned models to improve the performance of related tasks when data is shortened. Moreover, pre-trained deep convolutional neural network (CNN) models were used to classify Pap smear images into seven groups in the setting of cervical cancer detection, and DenseNet-201 (a pre-trained CNN model) was employed for this multi-class classification task, displaying higher-ranking accuracy and performance. In this paper20, by introducing a multi-axis vision transformer, the authors address common barriers in PAP smear analysis such as scant data and poor image quality. This framework featured ConvNeXtv2 and GRN-based MLP blocks and achieved 99.02% accuracy using the SIPaKMeD dataset. Another study21 elucidates that an ensemble of machine learning algorithms such as random forest and decision tree classifiers, achieves a 98.01% accuracy in trial detection. For fragmented datasets, a model using not only ADASYN with SVM, but also CNN features ensured 99.99% accuracy22. Moreover, using SVM and ResNet101, a multi-class procedure achieved 100% accuracy28. In Table 1, the study24 elucidates that the matter of class imbalance in datasets such as Light-Weighted Naïve Bayes, Logistic classifiers, and Random Forest offer optimal solutions. One more study25 identifies diagnostic errors in cervical cancer detection using the EfficientNetB0 model, achieving 96.84% accuracy in classifying cervix types. In addition, Gaussian Fitting Energy models were applied to the Herlev PAP smear dataset with an SVM classifier reaching 95% accuracy26. In contrast, Random Forest classifiers often achieve high accuracy but are ineffective for large datasets. To address this, the Boruta analysis with SVM is suggested as a more efficient approach27. Table 1Summary of related works for cervical cancer classification.Ref.DatasetClassesMethodsAccuracy20SIPaKMeD and LBC3 and 4MultiAxis Vision Transformer99.02% (SIPaKMeD)21Dataset from Universitario Hospital31 attributesRandom Forest & Decision Tree98.01% Random Forest22Dataset from Universitario Hospital36 characteristicsADASYN characteristics along with SVM & CNN99.99%23Herlev PAP Smear2SVM classifier and ResNet101100% SVM24Specialized Medical Centers in Jordan3LWNB, & Random ForestTP rate 0.899 (LWNB)25Tercha General Hospital3EfficientNetb096.84%26Herlev pap-smear image Dataset7 converted to 2Support Vector Machines, Gaussian Fitting Energy95%27UCI Dataset2Boruta analysis with SVMPrecision 0.91214Herlev dataset2Pap smear, CAD (Machine Learning algorithms)Not Specified15SIPaKMeD Dataset5AlexNet, ResNet-101, ResNet-152, InceptionV398.08%16SIPaKMeD, Mendeley5CerCancdotNet, using MobileNet, DarkNet-19, ResNet-18,97.7% (SIPaKMeD)transfer learning17Not Specified2Decision Tree, LASSO, and SMOTETomek98.72%18UCI2Random Forest, Principal Component Analysis (PCA), SMOTENot Specified Summary of related works for cervical cancer classification. Methodology In this section, we discuss the working procedures of our project as shown in Fig. 1, including the dataset, preprocessing, models, and evaluation metrics. Workflow of our research from dataset to cervical cancer prediction. Dataset description We used a popular, publicly available dataset named SipakMed (cipakmed)29 that was taken from Kaggle. In the dataset, there were 3 stages of cells: normal, benign, and abnormal, in a total of 966 images. As in a single image with multiple types of cells, the author cropped the images to create 5 types of cells, which fell under the 3 stages of cells. Here superficial-intermediate and parabasal fell under normal, metaplastic fell in benign, and lastly, dyskeratotic (diskeratotic) and koilocytotic fell under abnormal cells. There was a total of 4049 cells. In Fig. 2, each image represented their class. Sample images from the dataset 1 for each class. Dataset preprocessing In the dataset, all the images were not the same in size, so we resized them to 128 × 128. We normalized the image by dividing 1.0/255 and rescaling the image range from 0 to 1. Different data augmentation techniques have also been applied, such as width shift, height shift, and horizontal flips. Deep learning models We applied four transfer learning models to the proposed dataset. To train the models we used pretrained weights of the models and then added one global average pooling, two dense with 512 and 256 neuron. Subsequently, a dropout layer with a value of 0.3. Finally, an output layer with number of classes was predicted. VGG19 &VGG16: VGG16 and VGG19 are considered to be among the30 most well-known deep architectures in the world for image categorization. It is widely acknowledged that these models are among the most advanced models available for the exact outcome they obtained. The dimensions of the input picture have been set to 224 x 224 for both the VGG16 and the VGG19 samples. The primary strategy that VGG16 employs is to maintain a straightforward model while simultaneously constructing deep blocks. VGG19 is comprised of 19 layers and has potential for pattern recognition that is more adaptable. VGG19 &VGG16: VGG16 and VGG19 are considered to be among the30 most well-known deep architectures in the world for image categorization. It is widely acknowledged that these models are among the most advanced models available for the exact outcome they obtained. The dimensions of the input picture have been set to 224 x 224 for both the VGG16 and the VGG19 samples. The primary strategy that VGG16 employs is to maintain a straightforward model while simultaneously constructing deep blocks. VGG19 is comprised of 19 layers and has potential for pattern recognition that is more adaptable. MobileNetV2: MobileNetV2 is a CNN architecture31 for the efficient execution of mobile devices. It uses inverted bottleneck blocks and residual connections, which results in being around 35% quicker than MobileNetV1. It introduces non-linearity at the intermediate expansion layer through the use of lightweight depthwise convolutions. The model consists of an initial convolution layer with 32 filters, 19 bottleneck layers, requires 300 million multiply-adds for processing, and contains 3.4 million parameters. There exist two kinds of blocks: residual blocks with a stride of one and blocks with a stride of 2 for downsizing. MobileNetV2: MobileNetV2 is a CNN architecture31 for the efficient execution of mobile devices. It uses inverted bottleneck blocks and residual connections, which results in being around 35% quicker than MobileNetV1. It introduces non-linearity at the intermediate expansion layer through the use of lightweight depthwise convolutions. The model consists of an initial convolution layer with 32 filters, 19 bottleneck layers, requires 300 million multiply-adds for processing, and contains 3.4 million parameters. There exist two kinds of blocks: residual blocks with a stride of one and blocks with a stride of 2 for downsizing. AlexNet: AlexNet is a deepening development of CNNs32 and is used for tasks of picture identification and classification. It won first prize in the ImageNet Large Scale Visual Recognition Challenge in 2012 and was a milestone in deep learning. Although both Alexnet and LeNet have the same design of architecture, Alexnet has much deeper design and a large number of filters per layer. There are eight layers, five of them are convolutional, two are fully connected hidden layers, and one is a fully connected output layer. Alexnet has been introduced with the groundbreaking use of convolutional layers and rectified linear units, introducing the proper basis for modern deep learning models that speed up the progress of applications in computer vision and pattern recognition. AlexNet: AlexNet is a deepening development of CNNs32 and is used for tasks of picture identification and classification. It won first prize in the ImageNet Large Scale Visual Recognition Challenge in 2012 and was a milestone in deep learning. Although both Alexnet and LeNet have the same design of architecture, Alexnet has much deeper design and a large number of filters per layer. There are eight layers, five of them are convolutional, two are fully connected hidden layers, and one is a fully connected output layer. Alexnet has been introduced with the groundbreaking use of convolutional layers and rectified linear units, introducing the proper basis for modern deep learning models that speed up the progress of applications in computer vision and pattern recognition. Proposed CCanNet: Behind the building a new model, we had two main goals: one was making an efficient model and other was getting a better performance. We used keras framework for building the model. The main model was 54 layer including input and output layer. The model had conv2d, batchnormalization, maxpool, globalaveragepool, activation, dense and dropout layer. In addition, squeeze block33 was added to reduce the number of parameter without dropping the performance of the model. And this fulfill one of the main goal. One of the problem is, in traditional deep learning models when the number of layer increases , the performance could be in peaks or declines. A residual block34 can solve this degradation problem. So, we added two residual block with skip connection. In Fig. 3 shows the architecture of the proposed CCanNet model.Residual BlockOne of the most commonly used deep learning models for image recognition is the ResNet (residual network), which relies on the residual block as one of its basic components. It becomes much easier to train multi-layer networks using the idea of residual blocks, which helps to alleviate the issue of vanishing gradients in very deep neural networks. This architecture uses skip connections to group up layers of the network into blocks, and for each block, data goes through and around. Within the block, the layers will pass the data forward normally, but between blocks, there will be a connection by combining the input with the output from the block, giving two different paths for data to follow, one going through the block and another going around. Identity mapping is performed by the skip connection, which ensures the gradient is not degraded by eliminating the effect of zero learning layers with close to zero weight, passing the input directly to the next layer35. In our model, we use residual blocks to create skip connections and alternate paths for data flow. The residual blocks provide better accuracy and efficiency in deep network training by ensuring that the input data towards the output layers and the updates are more meaningful36. 1 Here in the Eq. (1),,,x as input, denotes the first convolutional layer with weights W1, BN denotes batch normalization, ReLU denotes the Rectified Linear Unit activation function and denotes the 1 × 1 convolution for the shortcut connection when the dimensions do not match.Squeeze BlockThe squeeze block works only for channel-wise learning by predicting the importance of individual channels or feature maps. The importance shows the impact on the output of a specific input. According to that prediction, the channels are then modified, and only important information is kept through this process by cancelling others. Adaptive feature recalibration is conducted in the scaling of channels37. In our model, the squeeze block is implied, divided into two branches, one passing the input and the other computing modulation weight for each input. This leads to each channel having a calculated modulation weight. In the squeeze phase, a scalar descriptor is calculated for each channel of input x. The descriptor uses global average pooling, dense layers, reshaping and multiplying. Squeeze output shapes as where the parameters are height, width, and number of channels38. After the completion of the squeeze phase, these descriptors are inserted into the fully connected layer, where the dimension is reduced by the reduction ratio r. This block reduces the number of parameters without affecting the performance. 2 Here in the Eq. (2),Global average poolingThe input passes a global average pooling (GAP) operation, which calculates the average value of each feature in the single feature map. This average value is then returned as the output node. GAP divides input feature map in non-overlapping regions in matrix form39. In our applied model global average p they put this pulling operation when there was pooling Layer reduces the dimensions of input layer in huge amount also lessening the number of parameters which lead to deduction in overfitting. Data is passed from Max Pool layer after filtrating the maximum valued feature to the GAP layer taking the average value.Why do we use global average pooling layer, not flatten layer?There is a layer named flatten layer used in the fully connected neural network where every node is densely connected to each other, pooled feature maps are passed on input nodes. These pooled feature maps are 2 dimensional, which is transformed into 1 dimensional through flattening. We have not used flatten layer in our model, instead global average pooling Layer has been implemented. GAP and flattening both are used to deduct the dimensions of feature maps in CNN before putting as input in a fully connected network40. But GAP is considered as a better technique for various reasons. In GAP the number of parameters after taking the average out of convolutional layer decreases. As it generates a vector of only considering parameters of D. Whereas flattening generates a vector of ( which feeds on a large number of parameters depending on Hand Wvalues in the fully connected network. Flattening may lead to overfitting by passing on many parameters and not having a large enough dataset. In GAP the risk is reduced of overfitting due to smaller number of parameters. There is a specific input size necessary for the fully connected layers in terms of flattening. But in case of GAP the input size can vary since each feature map is transformed into a single value, which gives more diversity in input sizes. GAP attains a low dimensional structure that makes the architecture simpler and easily manageable whereas flattening makes the interpretation more difficult41.The mathematical representation of the proposed model is given here step by step including number of filter with filter size, pooling size, activation function, dropout rate and number of neurons in the dense or fully connected layer. 3 Input size 128 × 128 × 3, 3 defines RGB channel in the Eq. (3). 4 Activation Function “ReLU”, number of filter 32 and size (3 × 3), keeping the padding same in the Eq. (4). 5 where GAP is the global average pooling operation, and are weight matrices, and represents the sigmoid activation in the Eq. (5). 6 Activation Function “ReLU”, number of filter for each convolution layer 64 and size (3 × 3) in the Eq. (6). 78 Activation Function “ReLU”, number of filter for each convolution layer 128 and size (3 × 3) in the Eq. (8). 910 Activation Function “ReLU”, number of filter for each convolution layer 256 and size (3 × 3) in the Eq. (10). 1112 All the maxpooling layer used pool size (2 × 2) shown in the Eqs. (7), (9) and (11). In addition, global average pooling in the Eq. (12) 13 Dense layer with 64 neurons and dropout rate 30% in the Eq. (13). 14 Dense layer with 32 neurons and dropout rate 30% in the Eq. (14). 15 Final or output layer with “Softmax” activation function and neurons= number of classes in the dataset. Here, Y denotes output in the Eq. (15).The hyperparameter configuration used in the proposed CCanNet Model Table 2 presents the hyperparameters used in classifying cervical cytology images. Here, the image size of 128 × 128 has been taken, and 100 epochs have been used. The learning rate of the proposed CCanNet model is 0.001, and the batch size used is 20. As the optimizer, Adamax is used, and for the execution environment, a GPU is used to train the model. Furthermore, the training-validation-test split is 8:1:1. Proposed CCanNet: Behind the building a new model, we had two main goals: one was making an efficient model and other was getting a better performance. We used keras framework for building the model. The main model was 54 layer including input and output layer. The model had conv2d, batchnormalization, maxpool, globalaveragepool, activation, dense and dropout layer. In addition, squeeze block33 was added to reduce the number of parameter without dropping the performance of the model. And this fulfill one of the main goal. One of the problem is, in traditional deep learning models when the number of layer increases , the performance could be in peaks or declines. A residual block34 can solve this degradation problem. So, we added two residual block with skip connection. In Fig. 3 shows the architecture of the proposed CCanNet model. Residual Block One of the most commonly used deep learning models for image recognition is the ResNet (residual network), which relies on the residual block as one of its basic components. It becomes much easier to train multi-layer networks using the idea of residual blocks, which helps to alleviate the issue of vanishing gradients in very deep neural networks. This architecture uses skip connections to group up layers of the network into blocks, and for each block, data goes through and around. Within the block, the layers will pass the data forward normally, but between blocks, there will be a connection by combining the input with the output from the block, giving two different paths for data to follow, one going through the block and another going around. Identity mapping is performed by the skip connection, which ensures the gradient is not degraded by eliminating the effect of zero learning layers with close to zero weight, passing the input directly to the next layer35. In our model, we use residual blocks to create skip connections and alternate paths for data flow. The residual blocks provide better accuracy and efficiency in deep network training by ensuring that the input data towards the output layers and the updates are more meaningful36. 1 Here in the Eq. (1), , , x as input, denotes the first convolutional layer with weights W1, BN denotes batch normalization, ReLU denotes the Rectified Linear Unit activation function and denotes the 1 × 1 convolution for the shortcut connection when the dimensions do not match. Squeeze Block The squeeze block works only for channel-wise learning by predicting the importance of individual channels or feature maps. The importance shows the impact on the output of a specific input. According to that prediction, the channels are then modified, and only important information is kept through this process by cancelling others. Adaptive feature recalibration is conducted in the scaling of channels37. In our model, the squeeze block is implied, divided into two branches, one passing the input and the other computing modulation weight for each input. This leads to each channel having a calculated modulation weight. In the squeeze phase, a scalar descriptor is calculated for each channel of input x. The descriptor uses global average pooling, dense layers, reshaping and multiplying. Squeeze output shapes as where the parameters are height, width, and number of channels38. After the completion of the squeeze phase, these descriptors are inserted into the fully connected layer, where the dimension is reduced by the reduction ratio r. This block reduces the number of parameters without affecting the performance. 2 Here in the Eq. (2), Global average pooling The input passes a global average pooling (GAP) operation, which calculates the average value of each feature in the single feature map. This average value is then returned as the output node. GAP divides input feature map in non-overlapping regions in matrix form39. In our applied model global average p they put this pulling operation when there was pooling Layer reduces the dimensions of input layer in huge amount also lessening the number of parameters which lead to deduction in overfitting. Data is passed from Max Pool layer after filtrating the maximum valued feature to the GAP layer taking the average value. Why do we use global average pooling layer, not flatten layer? There is a layer named flatten layer used in the fully connected neural network where every node is densely connected to each other, pooled feature maps are passed on input nodes. These pooled feature maps are 2 dimensional, which is transformed into 1 dimensional through flattening. We have not used flatten layer in our model, instead global average pooling Layer has been implemented. GAP and flattening both are used to deduct the dimensions of feature maps in CNN before putting as input in a fully connected network40. But GAP is considered as a better technique for various reasons. In GAP the number of parameters after taking the average out of convolutional layer decreases. As it generates a vector of only considering parameters of D. Whereas flattening generates a vector of ( which feeds on a large number of parameters depending on Hand Wvalues in the fully connected network. Flattening may lead to overfitting by passing on many parameters and not having a large enough dataset. In GAP the risk is reduced of overfitting due to smaller number of parameters. There is a specific input size necessary for the fully connected layers in terms of flattening. But in case of GAP the input size can vary since each feature map is transformed into a single value, which gives more diversity in input sizes. GAP attains a low dimensional structure that makes the architecture simpler and easily manageable whereas flattening makes the interpretation more difficult41. The mathematical representation of the proposed model is given here step by step including number of filter with filter size, pooling size, activation function, dropout rate and number of neurons in the dense or fully connected layer. 3 Input size 128 × 128 × 3, 3 defines RGB channel in the Eq. (3). 4 Activation Function “ReLU”, number of filter 32 and size (3 × 3), keeping the padding same in the Eq. (4). 5 where GAP is the global average pooling operation, and are weight matrices, and represents the sigmoid activation in the Eq. (5). 6 Activation Function “ReLU”, number of filter for each convolution layer 64 and size (3 × 3) in the Eq. (6). 78 Activation Function “ReLU”, number of filter for each convolution layer 128 and size (3 × 3) in the Eq. (8). 910 Activation Function “ReLU”, number of filter for each convolution layer 256 and size (3 × 3) in the Eq. (10). 1112 All the maxpooling layer used pool size (2 × 2) shown in the Eqs. (7), (9) and (11). In addition, global average pooling in the Eq. (12) 13 Dense layer with 64 neurons and dropout rate 30% in the Eq. (13). 14 Dense layer with 32 neurons and dropout rate 30% in the Eq. (14). 15 Final or output layer with “Softmax” activation function and neurons= number of classes in the dataset. Here, Y denotes output in the Eq. (15). The hyperparameter configuration used in the proposed CCanNet Model Table 2 presents the hyperparameters used in classifying cervical cytology images. Here, the image size of 128 × 128 has been taken, and 100 epochs have been used. The learning rate of the proposed CCanNet model is 0.001, and the batch size used is 20. As the optimizer, Adamax is used, and for the execution environment, a GPU is used to train the model. Furthermore, the training-validation-test split is 8:1:1. Architectural view of the proposed CCanNet model, where each unique color blocks define different layers in the model. Table 2Hyperparameters configuration used in the proposed CCanNet model.HyperparametersValuesImage size128 × 128No. of epoch100Learning rate0.001Batch size20OptimizerAdamaxExecution EnvironmentGPUTraining-validation-test split8:1:1 Hyperparameters configuration used in the proposed CCanNet model. We used different data split values to find the best possible results and found 8:1:1 as optimum. Here, 80% for training the model, 10% to validating the model and rest of the 10% to testing the model’s performance. The data distribution after splitting is shown in Table 3. Table 3Data distribution among training (Train), validation (val) and testing (test).LabelsClassesShortTrainValTest1dyskeratoticDysk65081822koilocytoticKoil66082833metaplasticMeta63479804parabasalPara62979795superficial-intermediateSuper6638384 Data distribution among training (Train), validation (val) and testing (test). Evaluation metrics In this study, we utilized four key evaluation metrics42 - accuracy, precision, recall, and F1 score shown in the Eqs. (16), (17), (18) and (19). Each metric provides unique insights into the models’ capabilities in classifying cervical cancer types. Here and are the number of true positives, true negatives, false positives and false negatives respectively. A true positive refers to the situation when the model accurately predicts the positive class. A true negative occurs when the model accurately predicts the negative class. A false positive occurs when the model inaccurately predicts the positive class. A false negative occurs when the model makes an inaccurate prediction of the negative class. Results and discussions This section will illustrate the results of the four transfer learning methods and the proposed CCanNet models. A popular dataset named as “Sipakmed” was used to train and test the models. Each model was trained with the same number of epochs (100), learning rate (lr=0.001), activation function (adamax) and loss function (“categorical_crossentropy”). In addition, the models were trained in a local environment with the support of RTX4070 gpu and Jupyter notebook. As shown in Fig. 4, both the training and validation accuracies increased as the number of epochs increased. This indicates that the models learned to classify correctly. In the VGG16 model, the training accuracy was the highest of the four transfer learning models. The training accuracy of VGG16 is similar to that of VGG19. However, the training accuracy of MobileNetV2 was lower than that of VGG19 and VGG16. The training accuracy of AlexNet was the lowest compared with that of the four transfer learning models. Accuracy vs. Epochs graph of the popular transfer learning models. In Fig. 5, the training and validation losses vs. epochs clearly show that the training loss has been continuously decreased after around 15 epochs, but the validation loss sometimes fluctuated. The VGG19 and the VGG16 model’s validation losses increased, followed by training losses after approximately 20 epochs. In contrast, AlexNet had the largest spike in validation loss, but at 100 epochs, it reached a steady point. MobileNetV2’s training and validation losses decreased in a parallel manner after approximately 30 epochs. The importance of the validation accuracy should be considered when choosing a model. If the dataset is large, then VGG19 or VGG16 is better, and vice versa. Loss vs. Epochs graph of the popular transfer learning models. The confusion matrix of popular transfer learning models’ is shown in Fig. 6. The off-diagonal cells show wrong classifications, while the diagonal cells show correct classification. Firstly, in VGG19, the model correctly estimated 82 of 82 for label 1 (dyskeratotic). The model could not correctly guess 7 out of 83 for label 2 (koilocytotic). For label 3 (metaplastic), the model correctly predicted 77 out of 80. The label 4 (parabasal), could not predict 1 out of 79, and the label 5 (superficial-intermediate), correctly predicted 2 out of 84. Secondly, in VGG19, the model incorrectly anticipated 4 out of 82 for label 1 (dyskeratotic). The model is incorrectly 7 out of 83 for label 2 (koilocytotic). For label 3 (metaplastic), anticipated correctly 76 out of 80. The label 4 (parabasal), anticipated 79 out of 79, and the label 5 (Superficial-Intermediate) predicted 3 wrongly out of 84. Thirdly, in MobileNetV2, the model correctly anticipated 80 out of 82 (dyskeratotic). The model anticipated incorrectly 11 out of 83 (koilocytotic). For metaplastic, anticipated correctly 79 out of 80. The parabasal anticipated 78 out of 79, and the Superficial-Intermediate predicted 3 incorrectly out of 84 data. Finally, in AlexNet, the model wrongly anticipated 4 out of 82 in dyskeratotic, 81 out of 83 in koilocytotic, 77 out of 80 in metaplastic, 75 out of 79 in parabasal, and 80 out of 84 in superficial-intermediate. Confusion matrix of the popular transfer learning models. Here Fig. 8, demonstrated the training and validation accuracies vs. epochs for the proposed model. The training accuracy gradually increased from 63 to 99.36% after 100 epochs. On the other hand, validation accuracy had its ups and downs. At 21 epochs it reached lowest point after the initial point, however with an increase in epochs the value also improved and reached its peak point in the last epochs. Accuracy vs. Epochs graph of the proposed CCanNet model. In the Fig. 7, refers to the training and validation loss graphs of the proposed CCanNet model. Loss vs. Epochs graph of the proposed CCanNet model. In Fig. 9, in the proposed CCanNet model, the model correctly estimated 82 out of 82 for label 1. The model correctly anticipated 81 out of 83 for label 2 and incorrectly predicted 2. In the 3rd label, 78 out of 80 correctly predicated and could not predict correctly approximate 2. The fourth label correctly predicted 78 out of 79 but incorrectly predicted only 1, and the fifth label, or Superficial-Intermediate classes, predicted 83 out of 84 points correctly but could not correctly predict 1. Confusion matrix of the proposed CCanNet model. Discussions In Table 4, the training accuracy, test accuracy, and parameters of the transfer learning models (VGG19, VGG16, MobileNetV2, AlexNet), transformer models (ConvNeXT, DeiT_tiny, MobileViT, Swin Transformer) and proposed CCanNet are compared. The model parameters are VGG19 (20.42 million), VGG16 (15.11 million), MobileNetV2 (3.05 million), and AlexNet (2.74 million), and the proposed CCanNet (1.27 million). The test accuracies of MobileNetV2 is 95.59%, AlexNet is 95.83%, VGG16 is 95.59%, and VGG19 is 95.34%, where the highest accuracy for the proposed CCanNet 98.53%. Among the transformer models, the train and test model accuracy varies; ConvNeXT has the highest train accuracy of 99.85% with 27.8 million parameters and a comparatively low test accuracy of 89.46%, while MobileViT (4.94 million) & Swin Transformer (27.5 million) have good test accuracy of 96.32% & 96.57%, respectively. This indicates a better balance between our model’s complexity and accuracy. The proposed CCanNet’s performance clearly shows its potential as a better model in terms of accuracy and the use of parameters, making it a better option for a variety of applications, even if more validation is required in different datasets to prove its capability. Table 4Train and test accuracy with #paramerters of the transfer learning, transformer models and proposed CCanNet model.ModelsTrain accuracyTest accuracy# ParametersVGG1999.5495.3420,419,653VGG1699.9795.5915,109,957MobileNetV299.7295.593,046,469AlexNet98.8695.832,738,501ConvNeXT99.8589.4627,817,349DeiT_tiny98.3394.125,487,365MobileViT98.9896.324,940,837Swin Transformer99.3596.5727,499,877Proposed CCanNet99.3698.531,274,663 Train and test accuracy with #paramerters of the transfer learning, transformer models and proposed CCanNet model. The classification report for transfer learning, transformer models, and the proposed CCanNet model is shown in Table 5. It shows how the F1 score, precision, and recall of each model (VGG19, VGG16, MobileNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileViT, Swin Transformer and CCanNet) compares for each label (dyskeratotic, koilocytotic, metaplastic, parabasal, and superficial-intermediate). For F1 score, it is showed that CCanNet outperforms all other compared models in the labels dyskeratotic, koilocytotic and metaplastic having the values 0.99, 0.97 and 0.98 resepectively. Although in F1 score we witness CCanNet having the highest value for the above mentioned three labels, in the other two labels parabasal and superficial-intermediate, the highest scoring models are MobileViT and VGG16 with the value of 1 for parabasal and 0.99 for MobileViT respectively. If we look at the next performance metrics recall, CCanNet has to best score 1.00 in the dyskeratotic class, second best in koilocytotic, metaplastic and superficial-intermediate classes with the scores of 0.97, 0.97 and 0.99 respectively and at the third position in parabasal class with a score of 0.98. In koilocytotic class the best score is given by AlexNet model with the value of 0.98. Metaplastic class has two best performing models VGG19 and MobileNetV2 both with the score of 0.99. Parabasal class has few outshining models like VGG16, MobileViT scoring the highest value 1.00; VGG19, MobileNetV2, DeiT_tiny, Swin Transformer scoring the second highest value 0.99. Lastly on the superficial-intermediate class DeiT_tiny model is seen to give the best result 1.00. For precision, CCanNet is the best model in both metaplastic and parabasal classes giving the perfect 1.00 score. Dyskeratotic and koilocytotic classes have CCanNet as the second best performing model with scores 0.98, 0.97 and superficial-intermediate class as the third best scorer at 0.96. CCanNet shares the best scoring position with VGG19, VGG16, AlexNet, MobileViT in parabasal class, whereas AlexNet, MobileViT beat CCanNet in dyskeratotic class with 0.99 and MobileNetV2 being the best model in koilocytotic class with 0.99. In general, CCanNet achieves similar precision, recall, and F1 score as the best transformer-based model, but requires fewer parameters for its operation, demonstrating its superiority over transformer models in terms of efficiency. Therefore, CCanNet is a tremendous performer in classifying cervical cytology images, and it also excels in identifying specific cell types. Table 5Classification report of the transfer learning, transformer models and proposed CCanNet model.ModelsPerformance metricsClassesDyskKoilMetaParaSuperVGG19Precision0.950.970.871.000.99Recall0.980.860.990.990.96F1 score0.960.910.920.990.98VGG16Precision0.950.940.931.000.96Recall0.950.920.951.000.96F1 score0.950.930.941.000.96MobileNetV2Precision0.940.990.890.990.99Recall0.980.870.990.990.96F1 score0.960.920.930.990.98AlexNetPrecision0.990.900.961.000.95Recall0.950.980.960.950.95F1 score0.970.940.960.970.95ConvNeXTPrecision0.930.840.820.990.90Recall0.940.770.930.940.90F1 score0.930.810.870.960.90DeiT_tinyPrecision0.930.970.910.980.92Recall0.960.840.910.991.00F1 score0.950.900.910.980.96MobileViTPrecision0.990.910.951.000.99Recall0.940.940.951.000.99F1 score0.960.920.951.000.99Swin TransformerPrecision0.980.950.930.990.98Recall0.980.950.950.990.96F1 score0.980.950.940.990.97CCanNetPrecision0.980.971.001.000.96Recall1.000.970.970.980.99F1 score0.990.970.980.990.97 Classification report of the transfer learning, transformer models and proposed CCanNet model. In Table 6, there was a comparison shown that, our proposed model performed better than other state-of-the-art models with a health margin. The researchers45 achieved 99.75% accuracy with their Inception-Ensemble model which consist of Xception (22.9M), InceptionV3 (23.9M) and InceptionResnetV2 (55.9M). However, our proposed CCanNet (1.27M) had fewer parameters than their baseline models. The computational cost was lower for our model than others. In addition, only15,21 went closer to ours but21 used tabular data where15 worked with 2 labels. Table 6Classification accuracy comparison of the proposed and state-of-the-art models in different datasets.Ref.DatasetLabelsModelsTest accuracy21Dataset from Universitario Hospital31 attributesRandom forest98.0125Tercha general hospital3EfficientNetb096.8443SIPaKMeD5CNN91.1315SIPaKMeD2RestNet15298.0816SIPaKMeD5CercanNet97.7044SIPaKMeD5ResNet5096.0145SIPaKMeD5Inception-ensemble99.75OurSIPaKMeD5CCanNet98.53 Classification accuracy comparison of the proposed and state-of-the-art models in different datasets. Explainable AI implementation (XAI) XAI is an important tool for image-based application. It can visualize which features the model took the decisions, show which features are important and which are not, and identify possible errors in the model. In addition, it helps interact with the used by understanding model’s prediction. We applied both GradCam and guided GradCam, however, guided gradCam performed better in identifying the important features of the model’s prediction. 1. Forward pass: Compute the feature maps from the final convolutional layer. 2. Backward pass: Compute the gradients of the score for class with respect to feature maps: where is the score for class and is the -th feature map. 3. Weight computation: Compute the importance weights by global average pooling over the gradients: where is the number of pixels in the feature map, and and are spatial locations. 4. Heatmap generation: Compute the weighted sum of the feature maps followed by a ReLU: Both GradCam and guided GradCam perform the same mathematical operation however, guided gradCam need 1 more extra operation. 5. Combination: Multiply the GradCAM heatmap with the Guided Backpropagation gradient element-wise: where denotes element-wise multiplication. GradCAM but also guided GradCAM are techniques used to predict which parts of an image a CNN focuses on which area should be extracted and make a prediction. Depending on the final convolutional layer feature maps, GradCAM applied gradients to create a heatmap elucidating important areas. The influential (and colorful) parts of the original image are highlighted using gradients; Guided GradCAM amplifies this by producing a more concentrated heatmap. In addition, these methods were used to identify distinct regions for classification as shown in Fig. 10. The heatmap highlights the spotted area on the cervical region to identify cancer, suggesting its impact on the algorithm’s prediction. Performance of GradCam and guided GradCam on different classes of the dataset. Strengths and limitations Our study has a number of benefits, in terms of model design, better accuracy, higher recall that is especially helpful in the medical context, and usage of the XAI approach which in addition to making our work highly effective, makes it more transparent and believable. Lightweight: We build a lightweight deep learning model that combines squeeze block and residual blocks. The proposed model has only 1.27 million parameters where nearest contender MobileViT has almost four times higher parameters. Lightweight: We build a lightweight deep learning model that combines squeeze block and residual blocks. The proposed model has only 1.27 million parameters where nearest contender MobileViT has almost four times higher parameters. High Recall: In medical field getting high recall value is very important or keeping false negative value low. Our proposed model got higher recall value for all the 5 classes and the value is over 0.97. High Recall: In medical field getting high recall value is very important or keeping false negative value low. Our proposed model got higher recall value for all the 5 classes and the value is over 0.97. Trustworthy: The model has not only lower parameters but also got highest accuracy. So, to get a more reliable result we visualized the results in GradCam and guided GradCam. Trustworthy: The model has not only lower parameters but also got highest accuracy. So, to get a more reliable result we visualized the results in GradCam and guided GradCam. Although the proposed model achieved the highest performance in terms of accuracy and F1 score having very low amount of parameters. However, this research has some limitations also. Limited dataset: One of the challenges in this domain is the lack of publicly available large datasets. Though Sipakmed and Herlev are publicly available but their size is very small. Limited dataset: One of the challenges in this domain is the lack of publicly available large datasets. Though Sipakmed and Herlev are publicly available but their size is very small. Model architecture: The proposed model is lightweight in terms of parameters but it is complex in architecture. Sometimes complex architecture takes more training time. Model architecture: The proposed model is lightweight in terms of parameters but it is complex in architecture. Sometimes complex architecture takes more training time. Conclusion and future works Cervical cancer is a major concern for women in underdeveloped nations, as it is one of the leading causes of death for them. overall health. The key to reducing its catastrophic effects is prompt, automated diagnosis and treatment in the early stages. Cervical cancer detection at an early stage is more efficiently treatable. Not only does it ensure better recovery chances, but also the treatment choices are less rigorous. PAP smears and HPV testing intervene in the progression toward the advanced stages of cervical cancer, preventing further invasion. Timely treatment can reduce the common aftereffects of cancer treatment, such as future morbidity, complications, and side effects. A patient is less burdened with the expenses of treatment with a higher life expectancy if proper detection is accomplished initially. To smoothen the automatic detection process with better efficiency and cost-effectiveness, we proposed CCanNet, a deep learning model. This model analyzes medical images with high accuracy, enabling it to detect subtle abnormalities. This customized network undergoes training on a wide range of datasets, achieving a high level of specificity in distinguishing between benign and malignant lesions. An automated screening reduces the risk of false positives and false negatives. The model surpassed other state-of-the-art models such as VGG16, VGG19, MobileNetV2, AlexNet, ConvNeXT, DeiT_tiny, MobileViT, and Swin Transformer with a healthy accuracy value. This model not only achieved the highest accuracy but also had a smaller number of parameters than the others. In the future, we will apply our custom CCanNet model to other medical fields, such as breast and prostate cancer detection. It will provide an enhanced mammogram analysis platform in radiology for breast cancer detection. Also, in high resolution MRI and ultrasound image analysis for prostate tumor identification, this model can leave an impact significance. For faster performance on portable devices, we plan to simplify and optimize the model. Acknowledgements Acknowledgments should be brief and should not include thanks to anonymous referees and editors, or effusive comments. Grant or contribution numbers may be acknowledged. Author contributions M.H.K.M. contributed to model development, coding and generating results. M.K., S.A. and M.A.A prepared draft of the paper. M.F.M. and Z.A. reviewed the draft and made correction. M.H.K.M. and M.K. drawn the figures and formated the paper in latex. Z.A was responsible for funding this project. Finally, all the authors revised the final manuscript, read and agreed to the published version of the manuscript. Data availability The Cervical Cancer largest dataset named SipakMed is publicly available on kaggle in the following link: https://www.kaggle.com/datasets/prahladmehandiratta/cervical-cancer-largest-dataset-sipakmed. Competing interests The authors declare that they have no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Contributor Information M. F. Mridha, Email: firoz.mridha@aiub.edu Zeyar Aung, Email: zeyar.aung@ku.ac.ae. References Associated Data Data Availability Statement The Cervical Cancer largest dataset named SipakMed is publicly available on kaggle in the following link: https://www.kaggle.com/datasets/prahladmehandiratta/cervical-cancer-largest-dataset-sipakmed.