Gynaecological cancers, including ovarian, endometrial, cervical, and vulval/vaginal tumours, represent a major global health burden. Yet compared to cancers like lung or prostate, they have been relatively overlooked in the AI-for-pathology research landscape. This 2025 narrative and scoping review from Christian Medical College (Vellore, India) and the University of Leeds (UK) surveys the current state of artificial intelligence applied to whole slide image (WSI) analysis across the entire female reproductive tract.
The pathologist crisis: A growing global shortage of pathologists is straining diagnostic services, particularly in low- and middle-income countries where limited access to ancillary testing (such as immunohistochemistry and next-generation sequencing) already causes delayed diagnoses and incomplete profiling. The digitisation of histopathology services has created large WSI repositories that form the backbone for developing AI platforms. These tools promise to accelerate diagnoses by decreasing turnaround times, improve patient safety by providing objective second opinions, and streamline workflows by reducing pathologist workload.
Consistency problem: Even among expert pathologists, inter-observer agreement on histopathological evaluation can be poor, with differences in visual assessment and interpretation of clinical data introducing subjectivity. AI tools, when properly deployed, can offer both objectivity and consistency. A recent study found that pathologists responded positively to AI adoption, believing it could increase reporting efficiency and decrease errors, though substantial training would be needed for effective implementation.
Scope and method: The review searched PubMed and Embase for studies combining AI with ovarian cancer, endometrial cancer, cervical cancer, and vulval cancer in the pathology context. Studies focused solely on radiology, clinical metadata, or genomic/transcriptomic profiling were excluded to maintain a focus on digital histopathology. The authors also performed targeted searches for AI work in related histomorphological entities (such as malignant melanoma and squamous cell carcinoma) to identify findings that could be translated to the gynaecological setting.
Over 300,000 new cases of ovarian cancer were diagnosed globally in 2022, with mortality exceeding 200,000. The disease is characterised by late detection and frequent resistance to platinum-based chemotherapy. Epithelial ovarian cancers account for the vast majority of malignant ovarian tumours, with germ cell and sex-cord tumours comprising the remaining 10%. Among all gynaecological cancers, ovarian cancer claims the largest share of AI pathology studies.
Histological classification: Several research groups have explored automated histological classification using computer-aided diagnosis systems on WSIs. Early contextual models showed good concordance with pathologist diagnoses. A 2022 study developed four different deep learning (DL) models using 948 WSIs, finding that a one-stage transfer learning algorithm classifying slides into five morphological carcinoma subtypes was most efficient. Additionally, researchers have combined DL algorithms with multiphoton microscopy to analyse unstained tissue from mouse models of ovarian and upper reproductive tract tissue, successfully distinguishing healthy tissue from serous carcinoma.
BRCA and HRD prediction: BRCA 1/2 mutations occur in approximately 25.7% of high-grade serous carcinomas, and homologous recombination deficiency (HRD) status is critical for treatment decisions, since HRD tumours show heightened sensitivity to platinum-based chemotherapy and PARP inhibitor (PARPi) combination therapy. Bourgade and colleagues developed a Convolutional Neural Network (CNN) approach for identifying BRCA mutations from WSIs using tumour segmentation, greatly reducing manual annotation time. A separate single-blinded study reported an AI model that predicted HRD status from H&E-stained WSIs alone with a remarkable 99.3% accuracy, offering a promising alternative to existing genetic tests that suffer from high failure rates and long turnaround times.
Treatment response and prognosis: Wang and colleagues predicted bevacizumab therapy efficacy by analysing WSIs using a DL-based approach. A CNN-based model predicted the impact of platinum-based chemotherapy on high-grade ovarian cancer with 91% specificity and 73% sensitivity. The Ovarian Cancer Digital Pathology Index (OCDPI), developed using H&E-stained WSIs, stratified patients into high and low risk groups with a significant association to overall survival. Spatial transcriptomics combined with AI identified discrete tumour regions with unique transcriptional signatures, revealing that the proto-oncogene JUN was exclusively upregulated in AI-detected areas linked to rapid recurrence after platinum treatment.
Endometrial cancer is the sixth most common cancer in women globally, with over 400,000 annual cases and approximately 97,000 deaths. It is the most common gynaecological malignancy in the developed world. These cancers are classified by histology and hormone receptor expression, and the recent molecular classification system divides them into four subtypes: POLE mutated (best prognosis), mismatch repair-deficient (MMRd), p53 abnormal (p53abn) (worst prognosis), and no specific molecular type (NSMP). The WHO also classifies endometrial hyperplasia as without atypia (benign) or with atypia (precancerous).
WSI-based diagnosis: Using 467 H&E-stained endometrial specimen WSIs, Zhao and colleagues developed a CNN to diagnose endometrial hyperplasia with over 97% accuracy, externally validated at over 95%. A separate DL model for identifying endometrial cancer from WSI patches achieved 83.7% specificity on prospective specimens, providing a valuable second opinion that prompted pathologists to revisit discordant cases. To address the problem of time-consuming annotation, a weakly supervised clustering-constrained attention-based multiple instance learning (CLAM) approach achieved an AUROC of 95.19%, representing a 4.41% improvement over standard multiple instance learning (MIL).
EndoNet and grading: EndoNet is an AI model that classifies endometrial cancers from hysterectomy specimen WSIs using CNNs for feature extraction and a vision transformer for aggregation. It classifies slides into low-grade (endometrioid grades 1 and 2) and high-grade (endometrioid grade 3, uterine serous carcinoma, or carcinosarcoma) categories, achieving an AUROC of 0.86 on an external test set. Panoptes, a CNN-based multi-resolution tool using 2.5x, 5x, and 10x magnifications, classified histological subtypes with an AUROC of 0.969, analysing each slide in under four minutes. It could also identify molecular patterns not visible to pathologists, such as characteristics associated with driver mutations.
Molecular subtyping from morphology: Im4MEC, a DL model developed using data from the PORTEC trials, combined a self-supervised learning model with an attention-based classification model to establish morphomolecular correlates and elaborate on intra-class heterogeneity from H&E-stained WSIs. AI algorithms have also been applied to predict genomic profiles of individual tumours from cancer genome data, potentially informing treatment decisions such as choice of targeted therapy.
Beyond diagnosis and subtyping, AI is being applied to some of the most clinically consequential questions in endometrial cancer management: predicting lymph node metastasis, guiding immunotherapy selection, and forecasting long-term recurrence risk.
Tertiary lymphoid structures and immunotherapy: Tertiary lymphoid structures (TLSs), together with B cell infiltration, have been shown to correlate with more favourable prognosis in endometrial cancer, likely through their contribution to an intratumoural immunity amplification loop that increases tumour sensitivity to immunotherapy. Suzuki and colleagues developed an AI model that both detected TLSs and determined their spatial locations in endometrial cancer WSIs. Combined with molecular subtyping, TLS identification and positioning was predictive of both progression-free survival and response to immune checkpoint inhibitors. Given the recent incorporation of immunotherapy into endometrial cancer treatment, this platform offers early signs of how AI could facilitate personalised therapy.
Lymph node metastasis prediction: Identifying lymph node metastasis is one of the most significant prognostic factors in endometrial cancer. A deep learning model predicted the probability of lymph node metastasis from perioperative H&E imaging of biopsy specimen WSIs, achieving an AUC of 0.938 in the internal cohort and 0.77 in the external cohort. The generated heat maps visualised which WSI regions contributed most to the metastasis prediction, enabling pathologists to perform targeted slide review and accelerate diagnostic turnaround.
HECTOR for recurrence prediction: The Histopathology-based Endometrial Cancer Tailored Outcome Risk (HECTOR) model is a multimodal DL prognostication tool that derives prognostic information from a combination of WSIs, image-based molecular class, and anatomical stage. Lower HECTOR scores are associated with more favourable markers (such as POLE mutant lesions and grade 1 tumours), while higher scores correlate with poorer prognostic factors (such as oestrogen receptor negative and p53 mutant lesions). Because the model inputs are both accessible and widely used in clinical diagnostics, HECTOR's implementation in routine practice looks particularly promising.
Cervical cancer ranks as the fourth most common malignancy affecting women globally, with 660,000 new cases in 2022. While developed countries have seen declining incidence thanks to screening programmes and HPV vaccination, low- and middle-income nations face starkly different outcomes due to disparities in screening, prevention, and socio-economic factors. Cervical cancer is caused by infection from high-risk HPV subtypes (particularly types 16 and 18) and typically arises from precursor cervical intraepithelial neoplasia (CIN) or cervical glandular intraepithelial neoplasia (CGIN).
AI-assisted cytology screening: DL/ML tools have demonstrated greater than 90% specificity for high-risk HPV serotypes and detecting high-grade atypia, outperforming their accuracy for low-risk serotypes and low-grade atypia. Several studies have been performed in resource-limited settings, where the shortage of diagnostic pathologists poses a critical barrier to effective cytology screening. The Pap Smear Analysis Tool (PAT) achieved a 0% false negative rate for screening out cytologically normal specimens, enabling cytologists to concentrate on suspicious cases and reducing both workloads and review times. Wang and colleagues developed a fully automated DL system to analyse cervical cytology WSIs, detecting high-grade squamous intraepithelial lesions or squamous cell carcinoma with a precision of 0.93.
HPV subtyping and risk stratification: Few studies have used AI models alongside PCR assays on cytological specimens to differentiate different HPV subtypes, including high-risk types (16, 18, 31, 33, 35, 45, 52, 58) and low-risk types (6, 11, 56, 59, 66). The integration of this approach with genomic profiles and biomarkers has enabled triaging methods and risk stratification. However, existing automated methods face a practical limitation: while they increase the number of slides screened by cytopathologists, accuracy decreases when large numbers of specimens are reviewed in a single day.
Histology vs. cytology challenges: WSIs of histology and cytology specimens present distinct challenges for computer vision. Histology WSIs contain intact tissue architecture showing organised structural layers, while cytology WSIs contain dispersed cells isolated from their tissue context, complicating spatial inference. Cytology slides can also pose scanning challenges due to multiple planes of focus. Despite these hurdles, advances in scanning technologies have helped overcome some of these issues.
Beyond screening, AI has been incorporated into cervical cancer histopathology for lesion classification and prognosis prediction, addressing the need for faster and more consistent diagnostic workflows.
Lesion classification from WSIs: Pre-trained CNNs have been shown to distinguish malignant from non-malignant H&E-stained histological section WSIs of cervical biopsy specimens. Cheng and colleagues developed a sophisticated multi-resolution tool that first screened images with a low-resolution model to locate suspicious regions and generate location heatmaps. Areas with a probability greater than 0.5 were cropped and passed through high-resolution models to identify 10 lesional cells based on probability scores. A recurrent neural network (RNN) then combined the features of these 10 cells to determine the likelihood that the entire slide was positive for malignancy. This cascading architecture balances computational efficiency with diagnostic precision.
Prognostication models: AI has also advanced prognostication in cervical cancer. Prediction models for stratifying disease recurrence risk have been developed based on factors including age, tumour size, stromal invasion, and adjuvant therapy, predicting disease-free survival and overall survival in post-surgery patients with early-stage cervical cancer. A pathological risk score system using information extracted by DL from WSIs has helped personalise the risk of recurrence for individual patients. These prediction models may inform future multimodal approaches that incorporate WSIs as part of integrated diagnostic and prognostic platforms.
Colposcopic imaging: The most significant work using image-based diagnostics for lower genital tract lesions has been a study exploring the use of CNNs to differentiate LSIL and HSIL from vaginal mucosa using colposcopic images, achieving a high specificity of 99.7%. Future AI techniques incorporating both macroscopic and microscopic images may provide a holistic approach combining point-of-care assessment with subsequent histopathological validation.
Vulval and vaginal cancers are rare gynaecological malignancies, with approximately 47,336 and 18,819 new cases reported worldwide in 2022, respectively. Squamous cell carcinoma (SCC) is the most prevalent histological type of vulval malignancy, alongside basal cell carcinomas, malignant melanomas, vulvar Paget's disease, verrucous carcinomas, and adenocarcinomas. HPV is responsible for 30 to 40% of vulval SCCs, while HPV-independent types can evolve on a background of chronic lichen sclerosis.
Limited but translatable research: Given the relative rarity of these cancers, there has been understandably limited research applying AI diagnostic and prognostic solutions. This scarcity reflects both targeted funding toward more common cancers and the limited availability of WSIs on which AI models can be trained and tested. However, the review highlights a promising strategy: adapting platforms developed for similar lesions in other anatomical sites.
Cross-site translation opportunities: AI has been used in oral squamous cell carcinomas from histopathological slides with a specificity of 0.92, a method that could be translated to vulval and vaginal SCC. Similarly, DL algorithms for diagnosing malignant melanomas of the eyelid could potentially be adapted for the rarer mucosal melanomas of the vulva. This cross-pollination approach, where AI models trained on more common cancers with similar histomorphology are repurposed for rarer sites, represents one of the most pragmatic paths forward for these underserved malignancies.
Despite promising results, the review identifies a series of persistent technical challenges that limit the clinical readiness of AI in gynaecological pathology.
Single-centre limitations: Most studies use small sample sizes and single-centre-source WSIs, making it difficult to predict how well AI models will generalise across different clinical environments. Technologies developed for resection samples may encounter limitations when applied to biopsies or tissue samples with different histological contexts, such as lymph node metastases, due to distinct architecture and morphology. Less common subtypes of malignancies, such as non-serous ovarian cancers or non-SCC cervical cancers, may be underrepresented.
Scanner and staining variability: Platform agnosia is a significant technical challenge: scanners from different manufacturers produce digitised slides with varying optical and computed properties, which can impact algorithm performance. Differences in H&E staining protocols across histopathology laboratories also introduce variability, though some of these hurdles have been partially overcome by colour balance pre-processing. Studies must also pay attention to the nature of source specimens (frozen versus formalin-fixed), underscoring the value of multidisciplinary investigatory teams that include pathologists, computer scientists, cancer biologists, and biomedical scientists.
Data augmentation and synthetic images: For cancers with few available datasets, such as lower genital tract malignancies, augmentation techniques can help refine data, especially for histopathological images. This has been explored in cervical cancer, where synthetic images with real image similarity have been developed using techniques like HistoGAN. Expanding such approaches to other gynaecological malignancies could improve the functionality and performance of existing AI models.
The research-to-clinic gap: AI tools remain frequently confined to the research environment rather than being properly trialled in real-life clinical settings, which is critical for establishing safety, efficacy, and end-user engagement. Current systems typically complete a single target activity (such as morphological typing or immunohistochemical scoring), whereas real pathological diagnosis is a multi-step process integrating multi-source data. Any clinical AI solution must be resilient to non-standard histology, including tissue folds, crushed cells, and cell debris.
The review concludes by addressing the higher-level challenges that stand between promising research results and actual clinical deployment of AI in gynaecological pathology.
Data drift and explainability: Data drift occurs when AI models exhibit divergent performance in real-world environments compared to their training phase. Even small shifts in data characteristics can invalidate a model for clinical use. Incorporating explainability into AI models is proposed as a partial solution. Rather than dividing WSIs into spatial units based on pixel dimensions, the authors suggest using functional units based on cell structure and type, which could improve interpretability upon pathology review. Manually curated features (quantitative measurements of size, colour, and morphology) combined with ML tools may also improve explainability, though these approaches require specialist time. Additionally, periodic retraining and subsequent locking of models could mitigate data drift effects.
Pathologist training and adoption: An end-user-based study found that high usability, user involvement, and levels of trust play critical roles in pathologists' willingness to adopt AI tools. Active collaboration between data scientists creating the technology and pathologists using it is crucial. Adequate support from leadership, dedicated physical space, staffing, storage infrastructure, and scanners are all important for effective integration. The authors stress that even with full automation, pathologists will remain essential for rare lesion identification, diagnostic probability assessments, quality control, and clinico-legal liability.
Regulatory landscape: Various country-specific organisations, including the US FDA, the UK National Institute for Clinical Excellence, and the European Medicines Agency (EMA), evaluate AI technologies from the standpoints of efficacy, reproducibility, safety, patient benefit, and cost-benefit. These evaluations are critical not only for independent review but also for establishing credibility and reimbursement channels. The FDA-approved Paige Prostate system for automated prostate cancer detection serves as a model, but no equivalent approval exists yet for gynaecological pathology AI tools.
The overarching vision: The authors call for looking beyond the short-term investment required for digitising pathology services and implementing AI, arguing that the long-term benefits in time savings, accuracy improvements, and cost reductions will benefit both pathologists and, most importantly, their patients. Multidisciplinary co-operation among data scientists, clinical pathologists, regulatory bodies, healthcare infrastructure managers, and pharmaceutical companies will be essential to ensure these technologies are appropriately tailored to clinical diagnostic needs.