Literature review

The present section discusses approaches in existing literature on the classification of cervical cells, outlining the different methodologies proposed, their efficiency, and the challenges faced. CNNs are frequently used for cervical cell classification because they are capable of extracting hierarchical features from images of Pap smears [20]. Hemalatha et al. [21] combined fuzzy logic with CNN to classify cervical cell images. The authors claimed that the integration of interpretability of fuzzy systems with the efficiency of CNNs improved the classification accuracy. De Lima et al. [22] proposed the use of Mask R-CNN, which is capable of instant segmentation, for cervical cell classification. Ali et al. [23] proposed a hybrid model by integrating naive Bayes (NB), random forest (RF), SVM, and decision tree (DT) classifiers. The risk factor dataset of the UCI Machine Learning Repository was used for models training and testing. The study [24] explored transfer learning for classifying cervical cells, comparing 13 pre-trained deep CNN models on the relatively small and imbalanced Herlev dataset, containing Pap-smear images. The best classification accuracy of 87.02% was achieved with DenseNet-201.

The federated learning (FL) approach allows hospitals to train models independently and collaboratively improve a global model without sharing sensitive data. Nasir et al. [25] developed a federated ML model with Blockchain for data security and IoMT for data collection. Reference [26] utilised a CNN-based federated learning framework to classify cervical cell images while ensuring patient data confidentiality. The approach demonstrates high classification accuracy in both scenarios. However, limitations such as high computational demands, communication overhead in FL, and reduced accuracy in non-FL scenarios remain challenges. The study [27] explored digital twin along with CervixNet classifier for cervical cancer detection, using the relatively small SIPaKMeD dataset. Although CervixNet demonstrates high accuracy, it has dataset constraints, reducing its generalisability.

An approach was presented by Xie et al. [28] to predict the radiation dose to be administered for cervical cancer radiotherapy. For prediction, beam channel generative adversarial networks (BC-GAN) are proposed, but the dataset used is relatively small. Mathivanan et al. [29] employed deep CNN models with machine learning algorithms to detect cervical cancer using AlexNet, InceptionV3, ResNet-101, and ResNet-152. Due to its reliance on only the SIPaKMeD dataset, the study has generalisation issues. The study [20] combined Vision Transformer (ViT) with SeNet, DenseNet169, and ResNet101 models for the classification of unsegmented images of cervical cells. The authors claim that accuracy can be improved by using approaches such as cervix feature fusion (CFF) and fuzzy feature selection. The datasets used, SIPaKMeD and Mendeley, are relatively small, leading to generalisation and overfitting issues. Study [31] investigated synthetic medical images created using GANs for predicting the reappearance of cervical cancer following radiotherapy. The study aims to expand dataset diversity and availability to help overcome the limitations of real-world medical data.

Madathil et al. [32] trained a deep learning model using multiple modalities like clinical records, medical imaging, and molecular biomarkers. ConvNeXt has been used for feature extraction and CerVital Predict for predicting cervical neoplasia. The study [33] combined CNN with transfer learning for cervical cell classification. They employed pre-trained models for extracting features along with machine learning classifiers, trained using the SIPaKMeD and Herleve datasets. In 2025, Hemalatha et al. [34] proposed a self-supervised learning method in which contrastive learning is used in extracting features from unlabelled cervical cancer cell images. But self-supervised learning approaches are believed to be less accurate in comparison to supervised approaches. Another limitation of these works is the use of relatively small datasets, which may result in lack of generalisation and overfitting.

Wang et al. [35] proposed an AI-based system for detecting cervical cancer. Utilising a comprehensive real-world dataset, ResNet-18 was used for feature extraction and random forest for classification. In study [36] the authors developed the MaxViT architecture by substituting MBConv with ConvNeXtv2 blocks and replacing MLP layers with global response normalisation (GRN)-based MLPs. The study [37] combined InceptionV3 with DenseNet201 to improve cervical cell cancer prediction accuracy. The authors used pre-processing techniques like augmentation, normalisation, dataset splitting, and dimensionality optimisation for feature selection. But these studies use a relatively small database and interpretability remains a challenge.

Munshi [38] addressed the issue of class imbalance and missing data by integrating support vector machine (SVM) imputation for handling missing values, adaptive synthetic (ADASYN) sampling for class imbalance, and CNN that performs feature extraction. These are combined within a stacked hybrid ensemble model that integrates the prediction strengths of diverse ML models. The study [39] addressed the challenge of missing values by combining the K-nearest neighbours (KNN) imputer with ensemble voting classifier. Muksimova et al. [40] developed RL-CancerNet, which blends CNNs and reinforcement learning to enhance diagnosis of cervical cancer. The proposed framework was trained and evaluated on the SipaKMeD and Herlev datasets. In these works, the proposed models are evaluated on single, relatively small datasets, limiting the generalisability. Study [41] proposes a model for classifying cervical cytology images based on DenseNet121 improved with Convolutional Block Attention Module (CBAM). The combination is to improve feature representation by highlighting informative spatial and channel areas of the image, thereby attaining high classification accuracy on the SIPaKMeD dataset. But the model is limited by its dependence on a single backbone architecture with no ensemble fusion, which could limit its stability in heterogeneous clinical situations. The research [42] presents a model based on CNN with additional CBAM and parallel branches for enhanced cervical cell image feature extraction. The model was 92.82% accurate on the SIPaKMeD database with enhanced attention to diagnostically critical locations. There was no ensemble learning and external validation. No clinical deployment process was addressed.

Research gaps

According to the literature, the following are identified as the key gaps in research:

Dataset

SIPaKMeD is an open-source dataset, which was used in the study. This dataset contains a total of 4049 single-cell images, which were manually retrieved from 966 cell cluster images of Pap-smear slides. Microscopic images of cells are captured with the help of a CCD camera and an optical microscope. Normal cells, abnormal cells, and benign cells fall into 5 categories: dyskeratotic, metaplastic, parabasal, koilocytotic, and superficial-intermediate (Figure 1).

Figure 1

Sample images from the dataset of each category

Pre-processing

To prepare the dataset images for better feature extraction and classification, various pre-processing techniques were utilised to improve the quality and contrast of images, and a new dataset was created.

Edge enhancement (Laplacian sharpening)

This step was used to sharpens the image by highlighting intensity transitions (edges). The Laplacian operator detects edges and is eliminated from the image to enhance contrast. The Laplacian filter is applied using Eq-7.

(7)

L(x,y)=∇2I(x,y)

where ∇2=∂2∂x2+∂2∂y2.

The enhanced image was obtained by subtracting the fraction α of the Laplacian from the original image, represented in Eq-8.

(8)

Isharp=I-αL

where α controls the sharpening strengths.

Figure 2 shows the flow diagram of creation of new dataset, and Figure 3 shows the sample images from the new dataset achieved after all steps of pre-processing.

Figure 2

Flow diagram for new dataset creation

Figure 3

Sample images from the pre-processed dataset

Random oversampling

Random oversampling (ROS) is a data balancing method that is used in balancing the dataset if it is imbalanced as per defined categories. It expands the dataset by augmenting the minority class samples through random duplication of existing samples, without generating new data. This helps equalise the class distribution, letting the classifier learn patterns from both the classes more effectively.

Algorithm-1 shows the important steps used in the random oversampling. Let X is the feature set, y is the label set, C_maj represents majority class, C_min represents minority class, and N_maj and N_min are samples of majority and minority classes, respectively.

Simple CNN

A convolutional neural network is a deep learning model applied mainly to image processing and computer vision tasks. Inspired by the human visual system, it contains 3 primary kinds of layers: convolutional layers for extracting features from images, pooling layers for down-sampling to reduce spatial dimensions while preserving important features, and classification is done through fully connected layers, also known as dense layers. The parameters used for designing the model are described in Table 1.

InceptionV3

InceptionV3 is a deep CNN used to efficiently extract features and improve classification performance. It has a few inception modules, each of which contains several parallel convolutional branches: factorised convolution, asymmetric convolutions, and auxiliary classifiers.

This will reduce computation cost from O(n²k²) to

L_total= L_main+ 0.3L_auxiliary , where L_main is loss of the main classifier, L_auxiliary is loss of the auxiliary classifier, and 0.3 is th eweight of auxiliary loss.

Xception

Xception is a type of deep convolutional neural network (CNN) architecture introduced by François Chollet in the year 2017. It extends the inception architecture by fully substituting conventional convolutional layers with depth-wise distinct convolutions, making the model computationally efficient while retaining high performance. Xception decomposes the standard convolution into 2 independent processes: depth-wise convolution, which uses one filter per input channel, and pointwise convolution (1 × 1 conv), which combines the results of depth wise convolution with a 1 × 1 convolution. This helps to reduce the computational cost and improves learning efficiency but still allows for independent feature extraction. There are 3 key components of the Xception model: entry flow (extracts low-level features using depth-wise distinct convolutions), middle flow (deep feature extraction using repeated depth-wise distinct convolutions), and exit flow (final aggregation of features and classification).

The core innovation of Xception is the depth-wise distinct convolution, which replaces standard convolutions. For a standard convolution, the output feature map at position (i, j, k) is computed as Eq-9

where W_m,n,c,k is the filter weight, x_i+m,j+n,c is the input feature map, F is the kernel size, C is the input channels, and b_k is the bias term for output channel k.

However, in a depth-wise distinct convolution, there are 2 convolutions:

where D_m,n,c is the depth-wise filter

where P_c,k is a point-wise filter.

This reduces the number of multiplications from

Inceptionv3 and Xception with attention

The attention mechanism is used here to enhance the feature extraction from the model. To apply the attention, the feature map is reshaped into a sequence: F′ = Reshape(F)→(49,2048) . This transformation flattens spatial dimensions into a sequence of 49 vectors, each with 2048 channels. After this, the model applies multi-head self-attention to improve the representation of features, as per Eq-12:

where Q = F'W_Q is the query matrix, K = F′W_K is the key matrix, V = F′W_V is the value matrix and d_k is the scaling factor to stabilise gradients. The proposed model uses 8 attention heads – the attention function is applied 8 times in parallel and then concatenated as per Eq-13; this allows the model to attend to different spatial features simultaneously.

After this, Gaussian noise is added as a regularisation technique to improve generalisation, as shown in Eq-14 – this prevents the overfitting.

where noise is sampled from 0 to 0.25.

Ensemble model fuzzy-based distance

The suggested ensemble technique fuzzy-based distance is predicated on the notion that the deviation between recorded and optimal solutions must be minimised. In this context, the recorded solution denotes the confidence scores produced by various models for a specific sample, while the ideal solution signifies a confidence score of 1, reflecting absolute accuracy in categorisation. To quantify this difference, the consensus measure using 3 distance metrics has been calculated – Euclidean distance, Manhattan distance, and Cosine distances. These distance-based measures help in evaluating how close each classifier’s prediction is to the ideal case, ensuring a more robust and reliable decision-making process. Algorithm-2 outlines all the steps of the proposed method. There are 2 steps used by the ensemble process:

Let f = {(f₁,c₁),(f₂ ,c₂ ),− − −,(f_m,c_m)} be the dataset, where each sample f_i ∈Rⁿ belongs to a class c_i ∈{1, 2,−−, n} , where n=5. Let j CSkj(fi) represent the confidence score assigned to sample f_i by the k^th classifier for the j^th class, where k = 1, 2,…..,N .

For each sample f_i and class label j , the ensemble is built by computing the distance between the confidence scores and the ideal solution vector 1={1}i=1N. The difference is defined as Eq-15.

Algorithm-2 Fuzzy distance-based aggregation

Input – original dataset, CATEGORIES=”im_Dyskeratotic”, “im_Koilocytotic”, “im_Metaplastic”, “im_Parabasal”, “im_Superficial-Intermediate”

Output – performance metrics and predictions

Begin

End

The work has considered 3 distance measures; therefore, 3 separate ensemble values are obtained: Euclidean-based ensemble (CjEc(fi)), Manhattan-based ensemble (CjMa(fi)), and Cosine-based ensemble (CjCo(fi)). These values are aggregated as per Eq-16 into a single vector.

Application of Euclidean, Manhattan, and Cosine distances as fuzzy aggregators draws from their complementary characteristics in handling classifier confidence scores:

The ensemble system’s utilisation of all 3 measures takes advantage of the multiplicity of similarity viewpoints, hence making the system robust to cope with discordant predictions and stabilising classification. The product rule comes into play when considering the final fuzzy score to aggregate such measures so that models having greater agreement (i.e. closer to the solution optimum) are given greater weight.

To integrate the information obtained from distance measures, the product rule in Eq-17 is applied to each class label

The final prediction is made based on the selection of the class with the minimum aggregated distance as per Eq-18.

where ŷ_i represents the predicted class for sample f_i. The product rule acts as a fuzzy measure, ensuring robust de-fuzzification by normalising distance values into a unified scale.

The sample example of the proposed approach by taking a sample f_i (100 images) is presented in Figure 6, the best distance calculated for each base classifier is displayed, and accordingly the fuzzy aggregated score for each category is calculated by the proposed method and then displayed. The final prediction is done from the values of aggregated scores.

Figure 6

Sample example and prediction using proposed method

Evaluation of framework with original dataset

The Adam optimiser dynamically adjusts the learning rate during training, while the sparse categorical cross-entropy measures the difference between the predicted and the actual probabilities in the target dataset. All 5 models underwent training for 20 epochs. Xception with attention outperformed as a standalone feature extractor. After that, the proposed model was applied, and using the fuzzy aggregator method the score found to be 0.94. 80% of the original picture dataset was utilised in training the models, and the remaining 20% was used for validation. Accuracy and loss metrics were then generated for all the models. Simple CNN achieved an 87% average accuracy, Inceptionv3 achieved an average accuracy of 89%, Xception achieved an average accuracy of 90%, Inception with attention achieved an 86% average accuracy, Xception with attention achieved 92% average accuracy, and finally the proposed model achieved 94% accuracy. The model-wise performance metrics are listed in Table 2. For cervical cytology grading, recall is most critical. It directly contributes to patient outcomes by making sure the abnormal and pre-cancerous cells do not remain undetected, preventing delayed diagnosis and treatment. Figure 7 shows the accuracy comparison plot of the models, and Figure 8 is the recall plot of comparison of all the methods with the original dataset.

Figure 7

Comparison plot of accuracy for all the models using the original dataset

Figure 8

Plot of recall comparison using the original dataset

Evaluation of framework with new dataset

This section describes the evaluation of the models. All 5 models were trained on 20 epochs. Xception outperformed a standalone feature extractor. 80% of the new image dataset was utilised to train the proposed models, and the remaining 20% was used for validation. Loss and accuracy metrics were generated for all the models. Simple CNN achieved an 90% average accuracy, Inceptionv3 achieved an average accuracy of 92%, Xception achieved an average accuracy of 95%, Inception with attention achieved an 96% average accuracy, Xception with attention achieved 97% average accuracy, and the proposed model achieved 98.3% average accuracy. Figure 9 shows the plots of validation accuracy for each model, and other metrics including accuracy are listed in Table 3. Figure 10 is the recall plot of comparison of all the methods with the new dataset.

Figure 9

Comparison plot of accuracy for all the models using the new dataset

Figure 10

Plot of recall comparison using the new dataset

Performance comparison

Table 4 gives a comparison with previous work. The suggested fuzzy distance-based aggregation model outperforms all the previous methods, with 98.3% accuracy, compared to the best literature model (~92%). It also has the better recall (97.8%), which is more important in medical diagnosis. The attention-augmented models (Xception + Attention, Inception + Attention) are much better than their base counterparts, indicating the effect of incorporating self-attention in feature extraction. All the suggested models have been validated by SIPaKMeD, Herlev, or small datasets with the possibility of overfitting. The approach presented herein was trained on a large dataset with high generalisation capacity. Combining various models enhances the confidence in decision-making, the highest reported accuracy, recall, and F1 score for cervical cancer classification.