Automated Cervical Cancer Screening Using Single-Cell Segmentation and Deep Learning: Enhanced Performance with Liquid-Based Cytology

Abstract

Cervical cancer (CC) remains a significant health issue, especially in low- and middle-income countries (LMICs). While Pap smears are the standard screening method, they have limitations, like low sensitivity and subjective interpretation. Liquid-based cytology (LBC) offers improvements but still relies on manual analysis. This study explored the potential of deep learning (DL) for automated cervical cell classification using both Pap smears and LBC samples. A novel image segmentation algorithm was employed to extract single-cell patches for training a ResNet-50 model. The model trained on LBC images achieved remarkably high sensitivity (0.981), specificity (0.979), and accuracy (0.980), outperforming previous CNN models. However, the Pap smear dataset model achieved significantly lower performance (0.688 sensitivity, 0.762 specificity, 0.8735 accuracy). This suggests that noisy and poor cell definition in Pap smears pose challenges for automated classification, whereas LBC provides better classifiable cells patches. These findings demonstrate the potential of AI-powered cervical cell classification for improving CC screening, particularly with LBC. The high accuracy and efficiency of DL models combined with effective segmentation can contribute to earlier detection and more timely intervention. Future research should focus on implementing explainable AI models to increase clinician trust and facilitate the adoption of AI-assisted CC screening in LMICs.

1. Introduction

Cervical cancer (CC) markedly affects the mortality rate of females worldwide [1]. This type of cancer is associated with a greater number of cancer-related deaths per year than breast cancer [2]. Various scientific studies have shown the existence of an inequality in the incidence of CC between different parts of the world [3,4], as shown by the fact that this it is the most common type of cancer and the leading cause of mortality from cancer in Latin America [5,6,7]. The incidence rates of CC have been steadily increasing over recent years [7].

CC is 99% linked to the human papillomavirus [8,9,10,11]. CC is an almost asymptomatic disease in its first two stages, which are the only two stages in which it is possible to treat the disease effectively [6,12]. For this reason, the early detection and management of early lesions are essential [10]. However, over recent years, the mortality rate associated with CC has remained high in all Latin American countries, reflecting the inefficiency of programs and screening techniques [3].

The early and accurate detection of CC remains a challenge in developing countries, with the disease continuing to be a major health concern [12,13,14]. The costs of diagnosis, treatment and control are among the highest in medicine; thus, this disease is considered catastrophic both collectively and institutionally [14].

The diagnosis of CC depends on initial screening with Papanicolaou (Pap) smear cytology [3]. The Pap smear consists of collecting cervical cells and examining them under a microscope to identify abnormalities [4]. To date, it has been demonstrated that this type of screening has an efficacy of 70% in reducing the mortality rate of associated with CC and its sensitivity ranges from 50 to 75% [15,16]. The Pap smear has a low sensitivity for pre-neoplastic events, such as grade 1 cervical intraepithelial neoplasia (CINI). For the same reason, numerous clinicians decide to perform a colposcopy with biopsy following a positive HPV test [3]. The Pap smear has several limitations that undermine its performance, such as high human capital and time requirements, since it requires cytotechnologists or trained pathologists to manually review numerous slides. This technique is subjective and inconsistent, as different observers may have different interpretations about the cellular material. It is very prone to human error, such as misclassification, false negatives, false positives, or omissions. It has a low sensitivity and specificity, since subtle or rare abnormalities are usually omitted or it can confuse some benign conditions with malignant ones [4].

In recent years, the improved detection of CC has been observed in developed countries owing to the implementation of liquid-based cytology (LBC). This is a novel method of the preservation and handling of cytological samples that can replace the traditional Pap smear, overcoming its limitations [16,17,18]. In LBC, the sample is transferred to the fixative fluid, which increases the cytological detection of squamous intraepithelial lesions and reduces the number of unsatisfactory smears [13]. In LBC, cells are filtered and transferred to the slide in a thin layer containing only one cell level, considered representative of the entire sample (monolayer). This facilitates the analysis of the sample compared to the conventional Pap test [16], improving the quality and interpretation of the slides, as well as reducing the number of false negatives and inadequate samples, due to fixation defects and masking of cellularity by excess blood, mucus, inflammatory cells or other artifacts [13,17]. Thus, LBC outperforms the Pap smear in terms of representativeness, sample fixation and smear quality.

Several researchers have explored methods for optimizing the early detection of cervical cancer (CC) through automated analysis of cytology images. Early work focused on cervical cell segmentation and computational algorithms using traditional machine learning techniques. For example, Gençtav et al. [19] achieved 74.7% accuracy and 93.5% sensitivity using an unsupervised approach with thresholding on liquid-based cytology images. Zhao et al. [20] employed superpixel-based Markov Random Fields (MRFs) for segmentation, while Bhatt et al. [21] investigated multi-class classification on the Herlev dataset, achieving 79.26% accuracy with KNN. While promising, these methods often relied on complex preprocessing and manual feature selection [22].

Deep learning (DL), leveraging deep convolutional neural networks (CNNs), has emerged as a powerful alternative by automating feature extraction. This, combined with advancements in computer vision, has made DL crucial for medical image analysis, including object recognition, segmentation, classification, and diagnostic support [23]. In the context of cervical cytology, various DL approaches have been investigated. Sompawong et al. [24] used Mask R-CNN for nuclear feature detection, achieving 91.7% accuracy. Other studies have explored UNET, Mask-RCNN, and FCN models [25]. Chen et al. [26] developed CytoBrain with CompactVGG, achieving 82.26% accuracy. Yaman and Tuncer [27] evaluated DarkNet and SVM, reporting over 98% accuracy. Rasheed et al. [28] developed C-UNet for nuclei segmentation (92.78% accuracy), while Nazir et al. [29] applied a modified UNET for cytoplasm and nucleus segmentation (94.24% sensitivity). These AI techniques automate cytological image analysis for classification, severity assessment, and lesion typing, offering quantitative, reproducible results with promising specificity, sensitivity, and diagnostic accuracy [30,31]. AI-powered screening offers advantages such as increased precision, reduced time and human resource requirements, and elimination of subjective bias [22]. By processing vast image datasets, AI algorithms can identify subtle patterns and discrepancies imperceptible to humans, potentially reducing invasive interventions, improving patient comfort, and enhancing outcomes [32].

In this context, AI offers a promising avenue for improving CC screening and contributing to enhanced medical diagnoses, with potential for significant societal impact [33,34]. This study aimed to evaluate the performance of a deep learning ResNet-50 model for cervical cell classification using both Pap smear and liquid-based cytology images to enhance early CC detection. We designed and evaluated a preprocessing algorithm for single-cell segmentation to facilitate training of the deep learning classification model. Furthermore, we compared the diagnostic performance of our system with a model trained on liquid-based cytology images from an external repository.

2. Materials and Methods

2.1. Traditional Pap Smear and Liquid Cytology Image Acquisition

For the purposes of the present study, a public dataset of PAP smear images from Center for recognition and inspection of Cells from patients aged between 35 and 65 years, and with diagnoses classified according to the Bethesda system (https://database.cric.com.br/classification) was used for the segmentation procedure, accessed on 5 September 2024. Liquid cytology images were obtained from the public online repository https://data.mendeley.com/datasets/zddtpgzv63/4 used by Sompawong et al., accessed on 15 September 2024 [24].

2.2. Cell Segmentation Algorithm

Subsampled images were processed using ImageJ 1.53 software (https://imagej.net/ij/, accessed on 5 September 2024) to obtain single-cell patches to train the classification models. The theoretical steps of the algorithm attached in Annex 1 are detailed as follows: (i) linear intensity transformation was applied to improve contrast; (ii) median filter with a kernel size of 9 × 9 was applied to eliminate cell noise and debris; (iii) default sharpen filter to highlight cell edges; (iv) RGB rendering and split to HSB (hue, saturation, brightness) color space; (v) 3× copies of saturation channel to apply thresholding methods for binarization (Minimal, Shanbag and Otsu); (vi) watershed on binary results of segmentations, followed by the addition of segments to the ROI (region of interest) manager; (vii) using the segments in the ROI manager, the cropping was made in the original RGB image, and each of them was saved with the sample identifier and a sequential number; (viii) finally, with the tumor cell ROIs generated in Qupath (https://qupath.github.io/, accessed on 5 September 2024), automatic labeling was carried out, where 1 corresponded to cells after the tumor and 0 to the healthy cell.

2.3. Malignant Cell Classification AI-Model Based in ResNet 50 Architecture

ResNet50 architecture corresponds to a convolutional neural network (CNN) developed by Microsoft in 2015 and can support very deep neuronal structures with optimal training performance for image classification. ResNet50, named for its 50 layers, is structured in ‘blocks’ that cleverly utilize ‘skip connections’ or ‘residual connections’. Instead of learning only from the previous layer, these connections allow the network to learn from earlier layer outputs as well. This bypasses the vanishing gradient problem often found in very deep networks, enabling efficient training and better feature extraction for improved accuracy in image classification tasks [35]. The advantages of ResNet50 over other CNN models are that it usually resents high performance values in terms of accuracy and recall, avoids the problem of vanishing gradient through residual learning and can robustly represent high- and low-level features without problems in classification [36]. It was programmed in Python 3.10 using TensorFlow an SciKit Learn libraries (links), using four main code blocks: (i) convolution layers to detect the local features of the image (such as more complex edges, textures and patterns); (ii) pooling layers used to reduce dimensionality and, at the same time, preserve the most important characteristics of the image; (iii) rectified linear unit non-linear activation functions, to introduce non-linearities into the network and allow the network to learn more complex representations; and (iv) fully connected layers, responsible for performing the final classification. All input single-cell images were transformed to tensors of 48 × 48 in dimension and divided and randomized in a 80:20 proportion for the total/test set, then a 75:25 proportion for the train/validation set using train-test-split of SciKit. In the model configuration phase, three hyper-parameters were used: (i) optimizer = ‘Adam’ (with a learning rate of 1 × 10⁻⁵); (ii) loss function: binary cross-entropy (loss = ’binary-crossentropy’); and (iii) the AUC metric, which was used for the evaluation of the model in training and validation. Finally, in the training phase of the model, the following hyper-parameters were used: Bach-size = 512 and number of epochs: 100. At least 20 training iterations of each of the algorithms were performed, varying the random state of the training:test data to find the optimal possible performance.

2.4. Statistical Analysis

2.4.1. Descriptive

Through Python libraries for the management of dataframes, arrays and graphs, the main descriptors of the data were obtained depending on whether the variable was numerical or categorical. Means, medians, deviations and quartiles were obtained for continuous and discrete variables, and the tables of frequencies of these were categorized. Frequencies of all variables were obtained, and distribution graphs were prepared to guide the following steps.

2.4.2. Inferential

Primarily using the Python libraries for statistics and graphs, data normality was determined using Shapiro–Wilk and Kolmogorov–Smirnov tests. Based on the results, either parametric or non-parametric statistical analyses were conducted. Non-parametric inference methods employed were the chi-squared test for one sample and the Mann–Whitney or Wilcoxon tests for two samples, depending on their independence or pairing. Their parametric counterparts corresponded to ANOVA, t-test, z-test, and LSD multiple comparison tests.

2.4.3. Diagnostic Performance Metrics

Diagnostic performance was assessed using confusion matrices in the test dataset, generated with Python’s scikit-learn library, to quantify true positives, false positives, true negatives and false negatives. From these matrices, metrics including precision, accuracy, sensitivity, specificity, and positive and negative predictive values were calculated. Subsequently, sensitivity, specificity and receiver operating characteristic (ROC) analyses were performed. The latter involved generating comparative performance graphs, calculating area under the curve (AUC) values, and conducting DeLong tests for statistical comparisons between techniques.

3. Results

The segmentation in the sub-images generated from the original slide-scan revealed that the application of multiple binarization thresholds in the purified cell signal achieves, in the best of cases, the selection of poorly homogeneous groups in most sectors of the slide; the detection of large cell masses in some areas; and, in very few events, single-cell segmentation (Figure 1A). Due to the above, the single-cell patches generated in each of the sub-images are in most cases noisy with a low-definition of the cell edges, which do not allow the cell chromatin patterns to be clearly shown (Figure 1B). By comparing the results generated in the Pap images with the segmentations of the liquid cytology (LCyt), it can be qualitatively appreciated how the generated patches clearly have single, well-defined, noise-free cells with evident nuclear patterns (Figure 1C).

The single-cell patches generated in the Pap and LCyt datasets were used to simultaneously train a ResNet50 deep network-based classification model. Training with 9566 Pap patches, of which 3386 were labeled as malignant cell-positive, revealed medium performance; however, a tendency to over-adjust when observing the change in the AUC and loss metrics through 100 epochs were observed (Figure 2A, validation loss increase). Likewise, when looking at the confusion matrix of the binary classification, the significant presence of false-negative cases (predicted = 0, true = 1) and false positives (predicted = 1, true = 0, Figure 2B) can be observed.

From the confusion matrix, the final performance metrics detailed in

Table 1. Maximum performance values of the ResNet50-Pap classification model.

Metrics	ResNet50-PAP
Accuracy	0.7352
Sensitivity	0.6887
Precision	0.6293
Specificity	0.7624
F-Score	0.6577

The accuracy, sensitivity, precision, specificity and F1-score were obtained in the test phase of the model from the construction of the confounding matrix.

were obtained. In these metrics, the acceptable level of accuracy and specificity of the model stands out (0.7352 and 0.7624), although with a low level of sensitivity and precision (0.6887 and 0.6293).

Using the same architecture and hyperparameters of CNN-ResNet50, classifier training was evaluated, this time with single-cell patches from liquid cytology segmentation (LCyt set). Initially, the training was carried out with 50 epochs, but as the metrics were close to convergence, the number of iterations was extended to 100 to determine whether the behavior of the model remained stable. As shown in Figure 3A, the AUC and loss metrics throughout time have a more stable behavior than the Pap model and a very low tendency to overfit. This is evidenced by the similarity between the training set and the validation set. In addition, the confusion matrix clearly demonstrated the low incidence of false-negative and false-positive cases in relation to true positives (Figure 3B).

From the confusion matrix, the final performance metrics detailed in

Table 2. Performance values of the ResNet50-LCyt vs. CNN models of Sompawong et al. [24] and Chen et al. [26].

Metrics	ResNet50-LCyt	R-CNN LCyt Sompawong et al. [24]	VGG-LCyt Chen et al. [26]
Accuracy	0.980	NR	NR
Sensitivity	0.981	0.917	0.928
Precision	0.981	0.917	0.822
Specificity	0.979	0.917	0.911
F1-Score	0.981	NR	NR

The accuracy, sensitivity, precision, specificity and F1-score were obtained in the test phase of the model compared to Sompawong and Chen models.

were obtained. In these metrics, the high level of accuracy and specificity of the model stands out (0.980 and 0.979), as well as the high level of precision and sensitivity (0.981 and 0.981). In addition, Table 2 presents the comparison of the two CNN models published with the same dataset of images by Sompawong et al. [24] and Chen et al. [26], where all the ResNet50-LCyt metrics are superior.

4. Discussion

The aim of the present study was to evaluate the performance of a deep learning ResNet 50 model for cervical cell classification using Pap and LBC samples to improve the early detection of CC. The DL model used herein was trained in a binary cell classification, which is essential for the early detection of CC; identifying patients with lesions that will evolve into CC can help determine an adequate treatment strategy and may thus prevent cancer development.

CC remains one of the leading causes of cancer-related mortality in low-income countries, despite being a highly preventable pathology. In Latin America and the Caribbean, CC is the third most common type of cancer, and HPV infection is present in >99% of cases with a worse prognosis [1]. For decades, the standard method for detecting cervical lesions has been the Pap test [37]. However, screening programs based on this technique in low-income countries have rarely been successful in reducing the mortality and incidence rates of CC due to the lack of high coverage and the associated high analytical complexity [38]. Traditional cytology begins with the collection of the sample, reporting of the analytical phase, followed by colposcopy when a biopsy is necessary, and finally, the treatment of the detected lesions [4]. Along with the lengthy and costly Pap smear process [22], in the majority of cases, conventional smears are difficult to interpret due to the uneven distribution and overlap of cells, and the presence of blood or inflammation [39]. In addition, the test has a high rate of false negatives; 50% of preneoplastic lesions of the cervix are missed with a single test [40]. The occurrence of false-negative reports depends on the morphological quality of the cells and on the abnormal cells being present in the sample in a recognizable form. This is found to be difficult in the Pap smear due to its multilayer pattern distribution. As a result, a number of women with CC have a history of one or more negative cervical cytology reports when they are actually carriers of high-grade lesions [41]. In addition, the interobserver reproducibility of cervical cytology is very inaccurate, as previously demonstrated by Stoler and Schiffman [42], in a study analyzing the reproducibility of 4948 monolayer cytologic interpretations. Of the 1473 original interpretations of atypical squamous cells (ASC-US), the second reviewer only concurred in 43.0% of them [42].

The low sensitivity of the Pap smear requires repeating the test multiple times over the years for it to be effective. This is very costly and not affordable in the majority of Latin American countries. In this scenario, it is essential to develop new methods with which to detect CC earlier with low-cost and high-accuracy automated screening technologies [43]. The alternative to the Pap smear is LBC, which makes immediate fixation easier and leaves the cells better visualized. LBC allows for a monolayer spreading where the majority of the debris, blood and exudate is removed [41]. The benefits of these liquid-based methods include decreased obscuring materials and hemorrhage on the slide, a decrease in cellular misrepresentation, and an even cell distribution on the slide [44]. In general, conventional screening processes with bright-field microscopes use a high human resource, and, in this context, the structure of the cells to be observed is complex, since the nucleus and cytoplasm are difficult to identify due to overlapping cellular areas and undefined boundaries between neighboring cells. However, this issue can be resolved with novel computer strategies that can classify images automatically and rapidly [41].

AI has been progressively applied in recent years for the diagnosis of various pathologies, with successful results when the evidence is image-based [45]. AI models can automatically recognize key features of images and can learn how to classify and process data using efficient algorithms [45,46]. Based on the above, the application of AI in the screening and early diagnosis of CHD is very useful for overcoming the current challenges of the technique [22]. Machine learning in AI is based on several computational models, and one of these is the CNN, which is mainly used for image processing and computer vision tasks. Within this, there are also different architectures, such as Visual Geometry Group 16 (hereinafter VGG16) [47], Residual Network 50 (hereinafter ResNet50) [48], and Mobile Network (hereinafter MobileNet) [26]. Previous research has shown that the CNN ResNet50 is the most effective compared with VGG16 and MobileNet; thus, this is considered the optimal CNN architecture [49]. In addition, recent studies have demonstrated that DL models are robust against changes in the aspect ratio of cervical cells on cytological imaging [22,50].

Early detection of cervical cancer (CC) is critical for improving treatment outcomes and patient survival [1,44]. Artificial intelligence, particularly deep learning, has shown significant promise in this area. This study investigated the performance of a ResNet-50 DL model for classifying cervical cells in both conventional Pap smear and liquid-based cytology images, aiming to enhance early CC detection [19,24,51].

Our findings demonstrate exceptional accuracy with the ResNet-50 model on LBC images, achieving near-perfect sensitivity, specificity, precision, accuracy, and F1-score. This indicates the model’s strong ability to differentiate between benign and malignant cells when trained on LBC images. However, this study also reveals the challenges inherent in using traditional Pap smear images for AI-based CC detection [16,19]. Overlapping cells, staining variations, blood clots, and artifacts hinder accurate segmentation of individual cells in Pap smears, leading to classification errors [19,20]. This is reflected in ResNet-50’s performance with Pap smear images. While achieving reasonable accuracy (73.52%) and specificity (76.24%), sensitivity was significantly lower (68.87%). Specifically, the model correctly identified 276 of 342 healthy cells but only 146 of 232 malignant cells. In contrast, with LBC images, the model achieved a much higher sensitivity of 98.1%, alongside 97.9% specificity and 98.0% accuracy, correctly classifying 714 of 729 healthy cells and 792 of 807 malignant cells. The class imbalance in the Pap smear dataset, with significantly more benign than malignant cell images, likely contributed to the reduced performance. Addressing this imbalance in future research is crucial for improving malignant cell detection in Pap smears.

LBC offers a significant advantage over Pap smears due to its superior cell isolation and distribution, resulting in cleaner images with less overlap and fewer artifacts [16,17]. The segmentation artifacts in Pap smears are difficult to optimize given the complexity of our customized segmentation algorithm, unlike U-Net-type architectures [52]. This is compounded by the inherent limitations of traditional cytology, which is not designed for single-layer cell analysis like LBC [41]. Nevertheless, incorporating U-Net or Attention-U-Net-based segmentation directly within the classification pipeline could improve performance and warrants further investigation with our cohorts [20,24].

LBC’s facilitated precise cell segmentation translates to improved classification accuracy. Despite the higher cost of LBC compared to Pap smears, our findings underscore the importance of LBC for enhanced CC detection [53]. Our ResNet-50 model achieved comparable or superior performance to previous studies using Pap smear images in terms of accuracy and specificity [24,26]. However, other studies have reported higher sensitivity, specificity, and F1-scores with Pap smears, suggesting that our results may reflect the aforementioned segmentation challenges. Conversely, our model consistently outperformed previous studies using LBC images across all performance metrics, including sensitivity, specificity, accuracy, and F1-score [48,54,55]. This highlights the substantial benefits of using LBC for DL-based cervical cell classification and the potential of ResNet-50 as a powerful tool for early CC detection.

Several studies have emphasized the importance of moving beyond cell classification to improve patient-level diagnostic accuracy. Yu et al. [56] demonstrated the effectiveness of combining HPV typing with computer-interpreted cytology, achieving performance comparable to conventional LBC with HPV typing. Xue et al. [57] showed that AI-assisted LBC had higher specificity than cytologists with similar sensitivity, reducing unnecessary colposcopies. Yang et al. [58] found AI-assisted cytology to have comparable or superior diagnostic efficacy to other screening strategies, suggesting its potential as a primary screening method. Finally, Fu et al. [59] demonstrated the benefits of integrating colposcopy, cytology, and HPV data through deep learning, highlighting the potential of multimodal AI systems. These studies underscore the potential of AI to enhance clinical decision-making and improve cervical cancer screening programs.

Integrating AI-assisted LBC into clinical diagnostic workflows offers a promising pathway to significantly enhance cervical cancer screening [45,46]. This approach could be implemented as a supplementary tool for cytotechnologists, providing a second opinion to improve diagnostic accuracy and reduce human error [42]. Alternatively, AI-assisted LBC could be used as a primary screening tool, particularly in resource-limited settings [3] where access to trained cytotechnologists is limited [37]. This would expedite the screening process, allowing for faster triage and treatment for patients with suspicious findings [38]. The improved accuracy and efficiency offered by AI-assisted LBC have the potential to transform cervical cancer screening, leading to earlier diagnosis and improved patient outcomes while optimizing resource allocation [43]. This transition, however, requires careful validation through large-scale clinical trials and integration with existing healthcare infrastructure [1,12].

While our study demonstrates promising results with LBC, certain limitations must be acknowledged. The class imbalance in our Pap smear dataset negatively impacted model performance. Future research should explore techniques like data augmentation, focal loss, and resampling to address this. Our model is currently limited to CC detection in cytology, though cytopathology is also used for other cancers [60,61], where AI applications are emerging. Achieving model generalization is another challenge. Many studies utilize public databases like Herlev, SIPaKMeD, Cx22, and Mendeley, demonstrating good performance [21,25,26,47,48,62,63,64,65]. However, relying on single-region data limits generalizability to other areas like Latin America due to varying demographic, genetic, and environmental factors [4]. Furthermore, patient age, BMI, histological cancer type, and disease stage can influence model accuracy. Wang et al. [66] demonstrated an MTL-based AI model’s ability to predict these factors, aiding personalized treatment planning.

Despite significant advancements in medical image classification using AI, understanding how these algorithms function remains a challenge for both researchers and clinicians. This “black box” problem poses a significant hurdle in AI research, particularly within the medical field, and has fueled the emergence of Explainable Artificial Intelligence (XAI) [67]. Numerous publications over the recent decade have explored this issue. Grad-CAM has been employed as an explainability mechanism for ResNet-34, ResNet-101, EfficientNet-B3, and other proposed CNN architectures [21,68]. These studies demonstrated that Grad-CAM generated heatmaps allow for the visualization of the image regions (nucleus/cytoplasm), contributing to the classification of cells as normal or abnormal in both conventional Pap smears and liquid-based cytology. Alternatively, SHAP values have been used for interpretability of Random Forest (RF) models and hybrid models combining SVM, Logistic Regression (LR), and RF [69,70,71]. These studies utilized the Cervical Cancer Risk Factors dataset from the UCI Machine Learning Repository (https://archive.ics.uci.edu/dataset/383/cervical+cancer+risk+factors, accessed on 10 September 2024), which contains 36 features associated with cervical cancer diagnosis. Through SHAP analysis, these studies identified “Hormonal Contraceptives (years)”, “Schiller test”, “Colposcopy”, “Number of pregnancies”, “Age”, and “Diagnosis of genital herpes” as the most important features influencing the models’ predictive ability. Future research should prioritize model explainability. XAI provides crucial insights into the decision-making process of AI models, offering an additional layer of trust and transparency, which is particularly valuable in critical domains like medical diagnosis [67].

Despite these limitations, our findings strongly suggest the transformative potential of DL, particularly with LBC, for CC screening. The ability of DL to rapidly analyze large datasets and detect subtle abnormalities offers a significant advantage, especially in resource-constrained settings. In conclusion, our study demonstrates the ResNet-50 model’s effectiveness for accurate cervical cell classification with LBC images, highlighting its potential to improve early CC detection. Continued development of DL models and strategies to address data imbalance hold promise for transforming CC detection and improving patient outcomes.