Deep dive into deep learning methods for cervical cancer detection and classification

Abstract

Cervical cancer continues to pose a significant global health challenge, highlighting the urgent need for accurate and efficient diagnostic techniques. Recent progress in deep learning has demonstrated considerable potential in improving the detection and classification of cervical cancer. This review presents a thorough analysis of deep learning methods utilized for cervical cancer diagnosis, with an emphasis on critical approaches, evaluation metrics, and the ongoing challenges faced in the field. We explore various deep learning architectures, particularly convolutional neural networks (CNNs), and their applications in the segmentation and classification of cervical cytology images. Key performance indicators, such as accuracy, sensitivity, specificity, and the area under the curve (AUC), are reviewed to assess the effectiveness of these models. Despite advancements, challenges like limited annotated datasets, inconsistencies in medical imaging, and the demand for more resilient models remain. Strategies like data augmentation, transfer learning, and semi-supervised learning are examined as potential solutions. This review synthesizes current research to guide future studies and clinical implementations, ultimately advancing early detection and treatment of cervical cancer through cutting-edge deep learning technologies.

Introduction

According to the Global Cancer Observatory (GCO), 661,000 new cases of cervical cancer were identified worldwide in 2022, with 348,000 lakh associated deaths [1]. Within India, cervical cancer ranks as the second most prevalent cancer among women, with approximately 75,209 new cases with a crude rate of 10.9 and cumulative risk of 1 in 75 reported in 2020 [2]. Annually, India sees between 2000 and 3000 hospitalizations due to cervical cancer, with a majority of cases being diagnosed at advanced stages. India’s cervical cancer mortality rate is more than double that of countries like the United Kingdom, the Netherlands, and Finland, despite efforts such as screening programs and vaccination campaigns. Malignancy treatment continues to be the most expensive illness in India, with the average expense for cancer treatment varying from 100,000 INR to 1,000,000 INR or more (approximately 1,200 USD to 12,000+ USD) depending on several factors, including the stage of the disease, treatment methods, and the healthcare facility [3]. The majority of these costs, amounting to three quarters, are incurred as out-of-pocket expenses [4].

Screening, recommended for women aged 25 to 74, includes methods like colposcopy, where the cervix and vagina are examined under magnification [5]. Early detection is crucial, as delays reduce survival rates. Research indicates that inadequate detection and treatment lead to high mortality rates. Early screening is vital to reducing the global burden of cervical cancer, yet barriers such as lack of awareness and costly procedures limit access in developing nations [6].

Figure 1.

A., C. Cross sectional view of cervix [8]; B. Detection of pre-cervical cancer

Deep learning-based convolutional neural networks, relying on robust feature extraction capabilities, can autonomously identify key features in an image through gradient descent [7]. This advancement significantly enhances object recognition accuracy, establishing deep learning as a prominent tool in computer image processing. Implementing this technology in cervical cancer diagnosis can assist clinicians in making precise auxiliary assessments, decreasing their workload, and enhancing diagnostic accuracy. The growth in computational power, availability of large datasets, and proven efficacy of deep learning in fields like radiology have been key drivers behind the growing favour towards deep learning algorithms for cervical cancer screening. Moreover, the superior accuracy and robustness of deep learning models compared to traditional machine learning methods have made them the preferred choice for researchers working in this domain.

Traditional cervical cell screening depends on pathologist experience, leading to inefficiencies and low accuracy due to the similarity between cancerous and normal cells. Recent advancements include novel human papilloma virus (HPV) detection methods and the use of nanotechnology, biosensors, and deep learning techniques to enhance the analysis of Pap smear images and improve cervical cancer detection.

Despite progress in artificial intelligence (AI) healthcare, its use in cervical cancer remains limited. Our research focuses on developing and validating a portable AI-based cervical cancer screening system. We aim to address two key questions: Which AI algorithms are effective for detecting cervical cancer using pelvic images? Which algorithms show the most promise for diagnosis? To explore these, we conducted a preliminary investigation to synthesize existing research on machine learning algorithms’ accuracy in detecting cervical cancer, paving the way for a potential systematic review.

Materials and methods

Search strategy

Search renowned datasets for studies using computerized or automated methods for cancer diagnosis, following the Arksey and O’Malley approach. The search strategy, developed with an information specialist and adhering to PRISMA criteria, covered publications from database inception to November 2023, with an update in July 2024.

Criteria for approval and rejection

Articles between 2007 to April 2024 have been included as part of the Calman-Hine report, a milestone for cancer management document. Table 2 contains a collection of all Approval and Exclusion Standards.

Table 2.

Approval and rejection standards

Approval standards	Rejection standards
Studies on how cervical cancer affects patients’ life Research that incorporates cervical cancer patient images as a component of a broader study Knowledge of linguistics in English	Professional and scientific papers on the optimal attributes for cervical cancer The study outlines the way pros are observed Viewpoints, judgments, and remarks

Evaluation of studies

Two authors conducted a comprehensive assessment of search results, focusing on screening titles and abstracts to determine relevance. Duplicate entries were removed, and additional manual searches were conducted by examining reference lists. A member of the research team obtained full copies of potentially relevant studies and reviewed them independently. Any discrepancies regarding the inclusion of studies in the final evaluation were resolved through team discussion.

Table 1.

Population-Exposure Framework (PEO) investigation plans

PEO element	Title/Focus	Descriptive terms	Improved pursuit terms
Population	Cervical cancer patients	Malignant cervical cancer images, patient conditions	Patients with cervical cancer, women undergoing diagnostic imaging (cervicography, colposcopy), cancer screening
			Combined with AND
Exposure	Deep learning and AI in cervical cancer detection	Deep learning, machine learning, cervical cancer detection methodologies	Convolutional neural networks (CNNs), transfer learning, colposcopy image analysis, AI-enhanced diagnostics
			Combined with AND
Outcome	Performance and diagnostic accuracy	Diagnostic skills, performance evaluation, outcome accuracy	Sensitivity, specificity, precision, F1 score, clinical applicability, AI in clinical workflow
			Related concepts: clinical validation, interpretability of AI models, patient outcomes, diagnostic impact

AI — artificial intelligence

Data extraction

Data was collected through research in various categories, including research objectives (whether focused on cervical cancer specifically or more broadly universal), research methods, sample size, tumor characteristics, objectives, and main results.

Quality appraisal

To evaluate academic quality, the Critical Appraisal Skills Programme (CASP) qualitative checklist was used by two independent authors, with discrepancies resolved through discussion [9]. Pilot studies were included if they reported results and impacts on patient care, with no restrictions on publication date or language. Studies on cervical cancer treatment were excluded due to differences in research settings, as were retrospective studies and those focusing only on patient symptom screening without physician involvement. Relevant articles were selected based on full-text references.

Evaluation, synthesis, assessment, and analytical review

After conducting the de-duplication process outlined in the Bramer method, studies were organized in EndNote and reviewed in Covidence for titles and abstracts, noting reasons for removal. Relevant citations were included in the full-text review. Key factors for data synthesis were extracted from the analyzed articles.

Figure 2 shows the PRISMA flow diagram, which visually represents the stages of study identification, screening, eligibility assessment, and inclusion in systematic reviews and meta-analyses.

Figure 2.

PRISMA flow diagram

Cervical cancer detection and diagnosis approaches

Ongoing screening is essential for cervical cancer prevention and therapy since early detection can improve treatment results significantly. Figure 2 depicts the steps for cervical screening and diagnosis.

Visual inspection with acetic acid (VIA) is a quick screening technique for cervical cancer that employs a 3–5% acetic acid solution to identify pre-cancerous cells. VIA is cost-effective and can be performed by mid-level providers in resource-limited settings, though it has lower sensitivity compared to cytology and colposcopy. Cytology involves pap smears or liquid-based tests to examine cervical cells, while HPV testing uses polymerase chain reaction (PCR) or hybridization techniques to detect HPV DNA. Confirmatory diagnosis of cervical intraepithelial neoplasia (CIN) involves colposcopy and biopsy, where a gynecologist or physician examines tissue for cancer cells. Recent progress in deep learning has notably advanced image analysis tasks, including semantic segmentation, classification, and object detection, leading to better screening and diagnostic outcomes [10].

Multiple deep learning-based approaches for detecting cervical cancer are currently available that have shown promising results in research studies. Some of the notable approaches include:

1. Utilizing convolutional neural networks (CNNs) to analyze cervical cell images obtained from Pap smears or colposcopy exams for the presence of abnormal cells or lesions [11].

2. Developing deep learning models that can accurately classify different stages of cervical cancer based on histopathological images [12].

3. Implementing automated image analysis algorithms to assist pathologists in identifying and diagnosing cervical cancer more efficiently [13].

4. Integrating deep learning techniques with other like magnetic resonance imaging (MRI) or ultrasound, to improve the accuracy of cervical cancer detection [14].

Figure 3.

Key procedures in cervical screening and diagnosis. HPV — human papilloma virus; PCR — polymerase chain reaction

These methods can greatly enhance the accuracy, efficiency, and cost-effectiveness of cervical cancer screening and diagnosis. However, it is essential to continue validating and refining these deep learning-based approaches through rigorous clinical trials and real-world implementation to ensure their effectiveness in clinical practice.

Deep learning in cervical cancer classification

Medical imaging, essential for diagnosis, is often hindered by individual interpretation and resource limitations, especially in low- and developing countries. The request for diagnostic images exceeds practitioner capacity in these regions. Artificial intelligence, particularly deep learning, offers a potential solution by improving computerized diagnosis [15]. Despite initial excitement, thorough studies are needed to address concerns about technology bias, research generality, and real-world applicability. Deep learning has shown exceptional speed and accuracy in medical imaging tasks such as segmentation, classification, and record-keeping. Its multilayer neural networks can uncover complex patterns and enhance diagnostic precision, potentially enabling less experienced healthcare professionals to achieve expert-level evaluations [16]. A major challenge in deep learning models for cervical cancer detection is their ability to generalize across different datasets. Models trained on specific datasets may perform poorly when applied to new populations or imaging conditions due to variations in imaging devices, population demographics, and dataset characteristics. Generalization issues are particularly acute in medical imaging, where high variability exists between datasets.

Various deep learning approaches for cervical cancer classification is shown in Figure 4.

Figure 4.

Various artificial intelligence (AI) algorithms utilized in medical studies

Self-supervision

Self-supervised learning is an advanced method that generates supervisory signals from unlabeled data, reducing the need for human annotation. This approach extracts semantic features from the data itself, which are then used for tasks with limited annotated data. It merges unsupervised learning’s ability to avoid manual labeling with supervised learning’s use of self-generated labels.

Research has shown that applying self-supervised learning to Pap smear images, which are abundant and unlabeled, can outperform standard pre-trained models. The Cervical Cell Copy-Pasting technique exemplifies this by enhancing performance beyond simple transfer methods and benefiting from integration with multiple instance learning techniques.

Unsupervised learning, which lacks direct supervision, relies on deriving signals without explicit guidance. Table 4 summarizes recent methods in unsupervised learning. Semi-supervised learning (SSL) is a key technique where models learn representations through auxiliary pretext tasks applied later to target tasks. SSL’s effectiveness depends on the design of these tasks, as they introduce biases. It includes four main types: predictive, generative, contrastive, and multi-self-supervision. A conceptual view of generative and contrastive learning in shown in Figure 5.

Table 4.

Advancements in using unsupervised techniques for cervical cancer classification

Reference	Method	Task performed	Description	Dataset used	Result obtained
[37]	3D CNN and ViT	Classification	3D CNN: spatiotemporal features ViT: sequence modeling	Not specified	Accuracy: 98.6%
[14]	Exemplar pyramid deep feature extraction model	Image classification	Uses deep feature extraction based on exemplar pyramids	SIPaKMeD Pap-smear images Training: 70%; validation: 15%; testing: 15%	Accuracy: 92.95%, precision: 93.89%, recall: 92.71%, F1 score: 93.30%
[38]	FSOD-GAN	Object detection and classification	Uses GANs for detecting small cervical spots	Open-source cervical cancer datasets (Herlev database)	Accuracy: 99%, sensitivity: 76.92%, specificity: 90.91%
[39]	3cDe-Net (dilated convolution ResNet and feature pyramid network)	Multiscale feature extraction and classification	Combines dilated convolutions and multiscale feature fusion	Tian-chi initiative and Herlev	Achieved superior performance with a MAP of 50.4%
[40]	Decision Tree with RFE and SMOTETomek	Risk prediction and classification	Uses decision trees with feature selection and class imbalance techniques	Patient medical records	Accuracy of 98.72% and sensitivity of 100%
[41]	Convolutional Neural Networks (CNN) for Histological Images	Image classification	Uses CNNs for classifying histological images of cervical cancer.	Histological images	Improved classification accuracy and AUC: 0.966
[42]	SIPaKMeD and Herlev Dataset Classification	Feature and image classification	Uses a new dataset for classifying normal and pathological cervical cells.	SIPaKMeD and Herlev datasets	Enhanced recognition ability Accuracy: 99.81%, precision: 89.4%
[43]	Hybrid DL Pre-Trained Models with Fuzzy Min–Max NN	Diagnosis and classification	Combines deep learning models with fuzzy neural networks	SIPaKMeD and Herlev dataset	Accuracy of 95.33% is obtained using Resnet-50
[44]	CNN with ELM	Classification	Integrates CNN with ELM for better performance	Herlev database	Accuracy: detection (2-class) — 99.5%, classification (7-class) — 91.2%
[11]	Ensemble Deep Learning Network	Classification	Combines multiple deep learning models for better accuracy	Colposcopy images	Sensitivity: 92.4%, specificity: 962%, Kappa scores: 88%
[45]	RFE and LASSO	Feature selection and classification	Uses RFE and LASSO for feature selection in risk prediction	Cervical cancer risk datasets	Accuracy: RFE — 98.0% and LASSO — 96.0%
[39]	Multiscale Feature Fusion Network and Channel-Wise Cross-Attention	Feature extraction and classification	Uses channel-wise cross-attention and feature fusion for better feature extraction	SIPaKMeD, Herlev and Motic dataset	Higher MAP scores compared to baseline methods
[46]	Deep Learning with Depthwise Separable Convolutions	Image recognition and classification	Uses shuffle net for efficient deep learning	Cervigram dataset	Accuracy: 81.38% AUC: 0.99
[47]	Generative Adversarial Networks (GANs) for Cervical Cell Detection	Detection and classification	Uses GANs for detecting cervical cells in images	Cervical cell images	Accuracy of almost 97% and a loss of less than 1%
[48]	Hybrid Convolutional Neural Network and Transfer Learning	Image classification	Combines CNN with transfer learning for better performance	Pap smear images	Accuracy of 91.46%
[49]	Multiscale Feature Extraction and Classification Network	Feature extraction and classification	Uses multiscale feature extraction techniques for better accuracy	SIPaKMeD dataset	Accuracy of 98.08%
[50]	Deep Learning-for ADSC	Screening and classification	Automates dual stain cytology using deep learning	Cervical screening data	Enhanced accuracy and efficiency in screening

CNN — convolutional neural networks;

Figure 5.

Conceptual view of generative and contrastive learning [17]

Table 3.

Frequently utilized datasets for achieving data-efficient deep learning in cervical cancer analysis

Dataset	Dataset type	Task performed	Description	Access link
UCI Machine Learning Repository [27]	Tabular (clinical data)	Classification (predicting cancer presence based on clinical features)	Contains clinical and demographic data related to cervical cancer, including attributes like age, number of sexual partners, and number of pregnancies	https://archive.ics.uci.edu/ml/datasets/Cervical
Cervix Cell Dataset (Kaggle) [28]	Image data (cell images)	Image classification (classifying images of cervix cells into different categories)	Includes images of cervix cells that are useful for training deep learning models for cell classification and anomaly detection	https://www.kaggle.com/datasets/prahladmehandiratta/cervical-cancer-largest-dataset-sipakmed
The Cancer Imaging Archive (TCIA) [29]	Medical imaging data (MRI, CT scans)	Image analysis (segmentation and classification of cervical cancer in medical imaging)	Provides imaging data including MRI and CT scans relevant for cervical cancer research	https://www.cancerimagingarchive.net/collection/tcga-hnsc/#citations
SEER Cancer Statistics [30]	Statistical data (cancer incidence and survival data)	Statistical analysis (analyzing cancer incidence, survival rates, and trends)	Provides a comprehensive range of cancer-related statistics, including detailed data on cervical cancer incidence and survival	https://seer.cancer.gov/statistics/
Cervical Cancer Screening Dataset [31]	Tabular (Screening Test Data)	Classification (predicting the risk of cervical cancer based on screening test results)	Contains data related to cervical cancer screening, including results from various tests and demographic information	https://doi.org/10.1016/j.sxmr.2019.09.005
Pap Smear Data [32]	Tabular (Pap smear test results)	Classification (classifying Pap smear test results to detect cervical cancer)	Includes information from Pap smear tests used for diagnosing cervical cancer, with attributes related to cell morphology	https://doi.org/10.24432/C5HS5C
Cervical Cancer Dataset (Kaggle Competitions) [33]	Tabular and/or image data	Various (depending on competition objectives; often classification or prediction tasks)	Data used in Kaggle competitions to develop predictive models for cervical cancer. The exact nature of the dataset can vary between competitions.	https://www.kaggle.com/datasets/ranzeet013/cervical-cancer-dataset
Genomic Data Commons (GDC) Cervical Cancer Dataset [34]	Genomic data	Genomic analysis (studying genetic mutations and expression related to cervical cancer)	Offers genomic and clinical data relevant to cervical cancer, useful for studying genetic factors and molecular mechanisms.	https://portal.gdc.cancer.gov
NCI Genomic Data [35]	Genomic data	Genomic and molecular analysis (analyzing genetic and molecular data across various cancers including cervical cancer)	Provides access to a broad range of genomic data for various cancers, including cervical cancer, for research and analysis	https://gdc.cancer.gov
Papanicolaou Smear Dataset (UCI) [36]	Tabular (Pap smear test data)	Classification (classifying Pap smear test results to detect cervical cancer)	Contains data from Pap smear tests used for detecting cervical cancer, including attributes related to cell characteristics	https://doi.org/10.1016/j.dib.2020.105589

MRI — magnetic resonance imaging; CT — computed tomography

In SSL, the pretext task trains the model on unlabeled data by generating self-created labels, allowing the model to learn representations. These representations are then used as initial weights for the downstream task, which fine-tunes them for specific goals. This workflow is illustrated in Figure 6.

Figure 6.

Overview of the semi-supervised learning (SSL) process. Top: in SSL, an auxiliary task is trained using pseudo-labels derived from a large, unlabeled dataset. Bottom: the features learned from this initial task are then applied to the main task, enabling training with a limited set of data containing true labels. CNN — convolutional neural networks

As the field evolves, the choice and complexity of pretext tasks become increasingly important, as they significantly influence the quality of the learned features. Tasks such as predicting rotations, filling in missing parts of an image, or distinguishing between augmented versions of the same data help the model capture various aspects of the data’s structure. In downstream tasks, these pre-learned features can greatly reduce the need for large annotated datasets, making SSL a powerful tool in areas where labeled data is scarce or expensive to obtain. Researchers continue to develop more sophisticated pretext tasks and refine the transfer process of learned representations to improve the performance of downstream applications.

Predictive self-supervision

In this a pretext task is framed as an identification or analysis problem. Common predictive tasks include relative position calculation, jigsaw completion, and rotational orientation identification. These methods are beneficial in medical image analysis (MIA).

For example, one method for cardiac magnetic resonance imaging (MRI) segmentation uses a pretext task to predict anatomical positions, employing various cardiac perspectives like short-axis and long-axis views. A similar approach for cervical cancer could involve segmenting cervical images to predict anatomical positions, using different views such as colposcopic images and Pap smear slides. This could improve the detection and diagnosis of cervical abnormalities by training networks to predict specific anatomical landmarks.

Contrastive learning

Contrastive learning has shown promising results in improving cervical cancer classification. By leveraging large amounts of unlabeled cervical images, contrastive learning trains a model to distinguish between similar and dissimilar image pairs. This approach enhances the model’s ability to learn robust and discriminative features, leading to better performance in identifying cervical abnormalities.

In contrastive learning as shown in Figure 7, input images are augmented and classified into positive or negative pairs based on labels. A pre-trained CNN encoder extracts features, which are projected into a space to enhance similarity in positive pairs and reduce it in negative ones. This process helps in accurate classification by improving feature representation.

Figure 7.

Steps for the execution of contrastive learning

Generative self-supervision

Generative self-supervision uses generative models, like variational auto-encoder (VAEs) and generative adversarial networks (GANs), to enhance cervical cancer classification by creating meaningful data representations. These models utilize large volumes of unlabeled images to improve accuracy and robustness. VAEs help in learning latent representations, while GANs augment datasets and detect anomalies. Techniques include generating synthetic cervical images to increase data diversity and identifying anomalies by comparing original and reconstructed images. A novel approach as shown in Figure 8 involves swapping patches in an image and using a generative model to restore it, aiding in early detection of cervical abnormalities [18].

Figure 8.

In the convolutional neural networks (CNN) computational structure for self-supervised reconstruction studying, blue, green, and orange represent layered elements, downsampling elements, and upsampling elements, respectively. The CNN’s structure throughout the reconstruction stage can be modified depending on the next assignment [19]

Inexact supervision

Inexact supervision involves using weak or imprecise labels for training models, which is particularly relevant in medical fields like cervical cancer detection, where precise annotations are challenging. This approach reduces the annotation burden by employing partially labeled data or heuristic methods. Techniques like Multiple Instance Learning (MIL), which labels groups of instances rather than individuals, and SSL, which uses both labeled and unlabeled data, are effective in handling the ambiguity in medical images. These methods improve the scalability and performance of cervical cancer diagnostic systems, but ensuring model reliability under inexact supervision remains a challenge [20].

Multiple instance learning (MIL) simplifies the labeling process by assigning labels to groups of instances, making it suitable for heterogeneous medical images where precise annotation is difficult. MIL has improved accuracy in cervical cancer screening by managing variability in cytological images. Strategies include MIL, which uses broad labels for sets of instances, and transfer learning, where models are fine-tuned on specific medical tasks, also helps in developing scalable and cost-effective screening tools, enhancing early detection despite annotation challenges [21].

Incomplete supervision

Incomplete supervision involves limited labeled data and abundant unlabeled data, and is divided into semi-supervised learning, active learning, and domain-adaptive learning. SSL improves model performance by using both labeled and unlabeled data, with techniques like consistency regularization, pseudo-labeling, and generative models to enhance cervical cancer detection.

Active learning, a type of incomplete supervision, improves labeling efficiency by focusing on the most informative samples, especially in resource-intensive tasks like cervical cancer detection (Fig. 9). It selects key samples from large unlabeled datasets for expert annotation, using strategies like uncertainty sampling, where the model queries labels for uncertain samples, and query-by-committee, where models select samples with high disagreement. This approach reduces annotation effort while improving model accuracy and generalizability.

Figure 9.

This map illustrates the abstract approach for self-teaching in semi-supervised learning and active learning. In self-training, a classifier is first trained on a small labeled set and then updated with additional pseudo-labeled data. Active learning involves retraining the segmentation model using original labeled data and new labels obtained through human interaction [22]. A. Self-training; B. Active learning

Domain adaptive learning (DAL) enhances cervical cancer detection by addressing domain shifts between labeled and unlabeled data, which can degrade model performance due to differences in data sources and conditions. DAL techniques and correlation alignment (CORAL), align feature distributions across domains to reduce discrepancies. Self-ensembling methods refine a model’s predictions on unlabeled data to adapt to new domains [23]. These strategies improve the robustness and generalizability of cervical cancer screening models, making them more effective in diverse clinical settings and aiding early detection. A typical DAL framework that involves adapting a pre-trained segmentation model to a new target domain, using a semantic discriminator to maintain consistent image-to-label mappings and ensuring realistic segmentations is shown in Figure 10.

Figure 10.

The designed optimal unsupervised context adaption system is made up of a segmentation framework and an intuitive detector. This system uses adversarial learning to correlate the statistical representations of the source and domains of concern. The segmentation systems for both regions have the same weights, with the initial domain model trained supervised and the intended domain model trained adversarially. The linguistic detector learns disease-specific spatial patterns explicitly by processing primary images, lexical masks, edge maps, and inverted maps [24]

Inaccurate supervision

Inaccurate supervision occurs when discrepancies in label information arise due to human error, inter-observer variability. This can harm deep neural networks’ ability to generalize, making it crucial to develop strategies to mitigate these effects, especially in medical image analysis, where precision is vital. Modern methods to address faulty supervision fall into three categories: robust loss design, data re-weighting, and training procedures.

Robust loss design involves creating loss functions that reduce the impact of mislabeled data during training. Techniques like using noise-tolerant loss functions, such as mean absolute error (MAE) over mean squared error (MSE), or loss correction approaches that adjust for label noise are common. For cervical cancer detection, enhanced loss functions, like an adapted dice loss or adaptive cross entropy (ACE), improve model accuracy despite noisy labels, aiding in the correct identification of cancerous cells [25].

Data re-weighting helps minimize the effects of noisy labels by assigning weights to training samples based on their likelihood of being correct. This technique uses validation sets or iterative updates to adjust sample weights, prioritizing reliable data. In cervical cancer screening, data re-weighting ensures more robust models, improving automated detection and patient outcomes by reducing the impact of incorrect labels [26].

Limited supervision

Limited supervision occurs when only a small amount of labeled data is available, a common challenge in medical imaging due to the high cost and expertise required for annotation. To overcome this, several strategies have been developed to make efficient use of both labeled and unlabeled data. Semi-supervised learning leverages the combined datasets, with techniques like pseudo-labeling, where model predictions on unlabeled data serve as additional labels, and transfer learning, which pre-trains models on large datasets and fine-tunes them on specific tasks like cervical cancer detection. Data augmentation further addresses the challenge of limited annotations by artificially increasing the size and diversity of the training set. This approach is particularly valuable in cervical cancer screening, where the cost of acquiring labeled medical images is high.

Transfer learning plays a crucial role in enhancing cervical cancer detection by using pre-trained models from large datasets. Since obtaining extensive labeled data in medical imaging is often difficult, this method allows models to utilize previously learned features and representations, improving accuracy even with limited annotated samples. Pre-trained models, such as those trained on ImageNet, can be fine-tuned for cervical cancer detection, enhancing the detection of precancerous lesions and other abnormalities [21].

Few-shot learning is another promising technique that addresses the scarcity of labeled data by enabling models to learn from only a few labeled examples. Techniques like meta-learning, which helps models quickly adapt to new tasks, and metric learning, which allows models to classify new examples based on their similarity to reference samples, have shown significant potential in improving cervical cancer detection with minimal data. Together, these strategies enhance early diagnosis and improve patient outcomes.

Results

Numerous works highlight significant advancements in the use of DL algorithms for the accurate diagnosis of various cancers, particularly in analyzing cervical cancer images, where these algorithms have achieved high precision rates. Central to these advancements are several frequently utilized datasets that enable data-efficient deep learning for cervical cancer analysis. Other datasets, such as the Cervix Cell Dataset from Kaggle [28], aid in classifying cervix cells into different categories, essential for training DL models for cell classification and anomaly detection. The Pap Smear Data [32] includes results from Pap smear tests critical for detecting cervical cancer, detailing attributes related to cell morphology. Lastly, the Cervical Cancer Dataset used in Kaggle Competitions [33] encompasses both tabular and image data for various tasks, focusing on classification or prediction based on competition objectives. Together, these datasets are invaluable for advancing cervical cancer research through deep learning techniques.

Numerous innovative models have emerged that leverage these datasets. For instance, Jia et al. [66] developed a Strong Feature CNN-SVM network that achieved an accuracy of 99.3% using the Herlev dataset. Similarly, Rahaman et al. [67] introduced the Deep Cervix Framework, which attained high accuracies on the SIPaKMeD and Herlev datasets. Cao et al. [68] utilized an Attention-Guided Convolutional Network, achieving notable sensitivity and specificity in detecting abnormal cells. Furthermore, Pacal et al. [14] proposed the Exemplar Pyramid Deep Feature Extraction Model, which also demonstrated strong performance metrics. Tan et al. [38] advanced the field further with the FSOD-GAN, achieving impressive accuracy and specificity in detecting small cervical spots.

Recent advancements in inexact supervised techniques for cervical cancer classification have yielded significant results. Ali et al [53] employed an ensemble learning approach using SHapley Additive exPlanations (SHAP) and achieved accuracy rates of 98.06% and 95.45% on two datasets. Wang et al. [69] introduced a Multi-Task Learning AI model that effectively predicts various clinical features of cervical cancer.

Chen and his team developed a compact CNN framework designed for use in embedded devices, emphasizing minimal parameters and low processing needs. Their study focuses on how information distillation can enhance the performance of this small-scale CNN model [70]. This framework integrates a CNN module to extract local features, a vision transformer to capture global attributes, and a multi-layer perceptron to combine these elements for precise classification. The goal of this system is to quickly and accurately evaluate pap images. Figure 11 provides a comparison of the results from these new deep learning approaches with those from earlier studies on similar topics.

Figure 11.

A. Illustration of number of articles published in a particular machine learning technique; B. Illustration of number of articles published year wise rages from 2008 to 2024

Evaluation metrics

When assessing how well a deep learning model performs in classifying cervical cancer, several key metrics are used to measure its effectiveness. These metrics include accuracy, precision, recall, and overall classification performance, as detailed in Table 6.

Table 6.

Comparative analysis of the evaluation metrics

Parameter	Mathematical equation	Remarks
Dice [71]	$\frac{2\times TP}{FP+FN+2\times TP}$	It is a metric used to assess the similarity between two sets
Accuracy [72]	$\frac{TP+TN}{TP+FP+TN+FN}\times 100$	Accuracy evaluates the percentage of samples that are classified correctly
Recall [71]	$\frac{TP}{TP+FN}$	Recall assesses the percentage of actual positive cases that were correctly identified by the model
Precision [72]	$\frac{TP}{TP+FP}$	Precision assesses the fraction of true positive predictions compared to the overall number of positive predictions generated by the model
Specificity [71]	$\frac{TN}{TN+FP}$	Specificity assesses the fraction of true negatives out of all negative cases
Score [71]	$\frac{Dice+Recall}{2}$	Score is the average of dice and recall
F1-Score [72]	$\frac{2\times Precision\times \text{Recall}}{Precision+\text{Recall}}$	It is the harmonic mean of precision and recall, which is viewed as a more effective performance indicator compared to the standard accuracy measure
Kappa score [73]	(Observed agreement – Expected agreement)/(1 – Expected agreement)	Cohen’s Kappa, also known as the Kappa Statistic, is a metric used to assess the agreement between raters for categorical data
Area under the receiver operating characteristic curve (AUC-ROC) [74]	$\begin{array}{l}TPR=\frac{TP}{TP+FN}\\ FPR=\frac{FP}{FP+TN}\end{array}$	AUC is a measure of segregation, with values close to 1 indicating an algorithm’s ability to accurately classify patients with and without cancer
Mathew correlation coefficient (MCC) [75]	$\frac{\text{TN}\times TP-FP\times \text{FN}}{\sqrt{(TP+FP)(TP+FN)(TN+FP)(TN+FN)}}$	It varies between +1 to −1, with +1 indicating perfect agreement and 0 showing no better than random predictions
Confusion Matrix [76]	–	It serves as a valuable method for assessing the model’s accuracy across various classes

Discussion

This review examined research articles covering a period of 20 years (2005–2024) that focused on the application of machine learning, including deep learning algorithms, for image classification in cervical cancer screening. The findings suggest a growing favouritism towards deep learning algorithms over time. The majority of these studies appeared in non-clinical publications, emphasizing the importance of sharing these algorithms in healthcare journals to enhance AI’s role as a diagnostic tool in clinical research.

The integration of deep learning algorithms into cervical cancer diagnosis has marked a transformative phase in oncological imaging, significantly enhancing detection and classification accuracy. Notably, recent architectures, like the 3D CNN and Vision Transformer, achieved an impressive accuracy of 0.986, reflecting a trend towards more sophisticated models that leverage advanced feature extraction techniques. These architectures are well-suited for processing the high-dimensional and complex data typical of medical imaging. The emergence of frameworks such as the CVM-Cervix, which combines CNNs with vision transformers, signifies a shift towards hybrid approaches that capitalize on both local and global feature representations. This adaptability may lead to more nuanced assessments of cervical cytology images, enabling quicker and more accurate evaluations of Pap smear results.

While these advancements are encouraging, challenges persist, including the need for larger, annotated datasets and the management of inconsistencies in medical imaging. For example, models that performed well on specific datasets may not generalize effectively across different populations or imaging conditions. Furthermore, the reliance on high-quality annotations remains a barrier, as many datasets are limited in their breadth and depth.

Table 5.

Advancements in using inexact supervised techniques for cervical cancer classification

Reference	Method	Task performed	Algorithm	Description	Dataset used	Result obtained
[37]	3D CNN and Vision Transformer (ViT)	Classification	3D CNN, ViT	Combines spatiotemporal features from 3D CNN with global context from ViT	SIPaKMeD dataset	Accuracy of 98.6%
(51)	Deep Learning with Coarse Labels	Classification with sparse labels	Two-stage deep learning	Efficient training with coarse labels on gigapixel images	WSIs	High accuracy despite sparse annotations
[52]	Feature Fusion with ShuffleNet	Classification	Deep learning, ShuffleNet	Combines deep learning with ShuffleNet for enhanced feature extraction	Pap smear images, SIPaKMeD datasets	Classification accuracy: 99.1%
[53]	Ensemble Learning-Based Classification using behavioral risk factors	Classification	Ensemble learning, SHapley Additive exPlanations (SHAP)	Utilize a five-fold cross-validation technique	Cervical cytology images	Accuracy: 98.06% for Dataset 1and 95.45% for Dataset 2 AUC: 0.95 for Dataset 1 and 0.97 for Dataset 2
[54]	Deep Learning with Generative Adversarial Networks (GANs)	Small object detection	GANs	Detects small objects in cervical images	Custom dataset	High detection accuracy and AUC value is 0.984
[55]	CNNs with RELMs	Classification	CNNs, RELMs	Combines CNNs for feature extraction with RELMs for classification	Cervical cytology images	High classification accuracy
[56]	EfficientNet-B0 and GRU-	Colposcopy diagnosis	EfficientNet-B0, GRU	Combines EfficientNet-B0 for feature extraction and GRU for sequence modeling	Colposcopy images	Accuracy: 91.18%, sensitivity: 93.6%, specificity: 87.6%
[57]	Ensemble of FDBD Model	Identification	Fuzzy distance metrics, deep learning	Ensemble of deep models using fuzzy distance metrics	PaP smear images	Performance of 96.96%
[58]	Improving Classification with Imbalanced Data Handling Techniques	Classification	Various imbalance handling techniques	Methods to handle data imbalance in cervical cancer classification	LBCPS dataset, SIPAKMeD, and Herlev dataset	Accuracy: LBCPS dataset (4-class) — 98.79%, SIPAKMeD (5-class) — 99.58%, Herlev (7-class) — 99.88%, IS: 3.75, FID: 0.71, Recall: 0.32, Precision: 0.65
[8]	Cervical Transformation Zone Segmentation	Segmentation and classification	Multi-scale feature fusion framework	2-phase method that includes segmenting and augmentation	Colposcopy images	Accuracy: 81.24%, sensitivity: 81.24%, specificity: 90.62%, precision: 87.52%, FPR: 9.38%, F1 score: 81.68%, MCC: 75.27%, Kappa coefficient: 57.79%
[59]	Self-Supervised Learning Approaches	Feature extraction	ConvNeXt	Leverages unlabeled data for feature extraction.	Deep Cervical Cytological Levels (DCCL) dataset	Classification accuracy 8.85% higher than that of previous advanced models
[60]	Weakly Supervised Learning with Partial Annotations	Detection	Weakly supervised learning (LD-WSCCD)	Weakly supervised model for cervical cell detection, based on a local distillation mechanism	Publicly accessible cervical cell dataset	mAP value of 73.6%
[61]	Graph Neural Networks for Cytology Image Analysis	Image analysis	Graph Convolutional Network (GCN)	Examined how cervical cell images might improve classification accuracy	SIPaKMeD and Motic liquid-based cytology image dataset	Better classification performance
[62]	Meta-Learning for Adaptation to New Data	Adaptation	Optimized meta-learning (OML) algorithm	Quickly adapts models to new data	Custom dataset	Accuracy: 99.39%
[63]	Semi-Supervised Learning Using Unlabeled Data	Classification	SSL	Uses labeled and unlabeled data	Tianchi dataset and the Comparison Detector dataset	High classification accuracy
[64]	Federated Learning for Distributed Data	Classification	Federated learning	Uses distributed data without centralizing it	UCI cervical cancer risk factors data	Accuracy: 99.26%, misprediction rate: 0.74%
[65]	Comparison detector for cervical cell/clumps detection	Detector	Faster R-CNN with FPN for proposal-based object detection	Proposed a dataset for comparison to tackle limited data issues	Comparison Detector dataset	mAP: 48.8% AR: 64%

CNN — convolutional neural networks

Strategies such as data augmentation and transfer learning have shown promise in mitigating these issues, facilitating the development of resilient models capable of generalizing better to unseen data. The exploration of inexact supervised techniques and ensemble learning methods further illustrates the ongoing innovation within this domain, pushing the boundaries of what is possible in cervical cancer detection.

Conclusion and future scope

This review has highlighted the critical role of diverse datasets in facilitating these developments and underscored the importance of selecting appropriate models that integrate both local and global features for optimal performance. As illustrated by the comparative performance metrics, various innovative architectures, including CNNs and hybrid models, have achieved impressive results, indicating a promising direction for future research.

Despite these advancements, challenges such as limited annotated datasets and the need for more robust models remain. Addressing these issues through strategies like data augmentation, transfer learning, and the incorporation of explainable AI techniques will be essential for further improving diagnostic accuracy and generalizability across diverse populations.

Ultimately, ongoing research in this field must continue to prioritize the translation of these technological advancements into clinical practice, ensuring that deep learning approaches not only enhance diagnostic capabilities but also lead to better patient outcomes in cervical cancer care. By synthesizing current findings and exploring new methodologies, we can pave the way for more effective and accessible diagnostic solutions in the fight against cervical cancer.

Table 7.

Comparative analysis of the performance metrics

Model	No. of images	Precision	Accuracy	Recall	Dice	Specificity	F1-Score	AUC
Convolutional Neural Network [49]	Training — 7,498 images, validation — 1,884, and testing — 1,884 images	0.924	–	–	–	0.962	–	–
RCNN [77]	906	0.7387	0.9291	0.7912	0.7431	0.9589	–	–
Gradient boosting classifier [78]	199 patients images	0.80	0.938	–	–	0.964		0.85
IDCNN-CDC [79]	–	0.9784	0.9796	0.9820	–	–	0.9781	–
GBT [79]	–	0.9618	0.9565	0.9699	–	–	0.9636	–
XGBoost [79]	–	0.9421	0.9515	0.9663	–	–	0.9577	–
ELM [79]	–	0.9367	0.9407	0.9599	–	–	0.9520	–
DCNN [80]	488 images: 117 cancer patients. 509 images: 181 non-cancer patients	0.883	0.908	–	–	0.933	–	0.932
DCNN [81]	13775 images	0.994	–	–	–	0.348	–	0.67
ResNet and combining ResNet with clinical features [82]	15,276	ResNet + clinical features: 0.932 Only ResNet: 0.901	ResNet + clinical features: 0.886 Only ResNet: 0.882	–	–	ResNet + clinical features: 0.846 Only ResNet: 0.867	–	ResNet + clinical features: 0.953 Only ResNet: 0.945
Compare KNN, SVM, and DT [83]	2000	0.989	0.993	–	–	0.994	–	–
Assessment of VIA, VILI, and GIVI contrast combinations [84]	1426	Resnet-v2: 0.833 Resnet-152: 0.852	Resnet-v2: 0.877 Resnet−152: 0.878	–	–	Resnet-v2: 0.886 Resnet−152: 0.882	–	Resnet-v2: 0.932 Resnet-152: 0.947
CNN: AlexNet and VGG-16 with modified filters [48]	4753	0.9822	0.9613	–	–	0.9509	–	–
Convolutional Neural Network CAIADS [85]	240	KNN: 0.75 SVM: 0.727 DT: 0.636	KNN: 0.783 SVM: 0.742 DT: 0.758	–	–	KNN: 0.803 SVM: 0.75 DT: 0.829	–	KNN: 0.807 SVM: 0.744 DT: 0.767
SVM based on radiomics and deep learning [73]	4753	0.9568	0.9687	–	–	0.9872	–	–
CNN-SVM [86]	5679	VGG-19: 0.33 CYENET: 0.924	VGG-19: 0.733 CYENET: 0.923	–	–	VGG-19: 0.0.79 CYENET: 0.962	–	–
Convolutional Neural Network [87]	253	0.9560	0.9410	–	–	0.8330	–	0.963
Convolutional Neural Network [88]	Training — 111 (44 + LVSI and 67 — LVSI) Validation — 56 (23 + LVSI and 33 – LVSI)	Training cohort: 0.773 Validation cohort: 0.739	–	–	–	Training cohort: 0.776 Validation cohort: 0.667	–	Training cohort: 0.842 Validation cohort: 0.775
CCNet [89]	–	0.7264	0.9191	0.7179	0.6849	0.9560	–	–
DarkNet53 [90]	–	0.8541	0.8477	0.8401	–	–	0.8470	–
SVM using VIA and VILI [91]	200	0.813	0.80	–	–	0.786	–	–
SVM and DT [92]	102	For SVM: 0.981 For DT: 0.95	For SVM: 0.9833 For DT: 0.97	–	–	For SVM: 0.985 For DT: 0.9867	–	–
Xception [93]	–	0.8678	0.8511	0.8695	–	–	0.8686	–
ResNeXt50 [94]	–	0.8908	0.8633	0.8658	–	–	0.8781	–
SegNet [95]	–	0.6867	0.9097	0.7057	0.6600	0.9517	–	–
DeepLabV3+ [96]	–	0.6889	0.9083	0.6828	0.6416	0.9545	–	–
ResNet18 [97]	–	0.8637	0.8254	0.8287	–	–	0.8458	–
FCN8x [71]	–	0.7102	0.9094	0.6434	0.6311	0.9522	–	–
UNet [98]	–	0.6941	0.9073	0.6593	0.6307	0.9575	–	–
Multilayer perceptron Networks [99]	170	0.6978	–	–	–	0.68	–	0.73

AUC — ???; KNN — ???; SVM — ???; DT — ???; VGG-19 — ???; CYENET — ???