Multiple instance learning for endometrial cancer diagnosis: optimization on weakly annotated whole-slide images and web visualization

Background Endometrial cancer is the 6th most common cancer affecting women worldwide, yet early detection remains challenging due to increasing diagnostic volumes, limited triaging workflows, and a shortage of specialized pathologists. Existing AI approaches are constrained by labor-intensive region-level annotations and single-stage models that do not reflect the hierarchical reasoning used by pathologists in clinical practice. Methodology This research proposes a weakly supervised framework for classifying hematoxylin and eosin-stained endometrial biopsies using only slide-level labels. The dataset was retrospectively curated from archived biopsies and included 490 benign, 425 neoplasia, and 382 hyperplasia cases. The dataset supports multiple diagnostic tasks from broad classification (benign vs. neoplasia vs. hyperplasia and benign vs. non-benign) to fine-grained subtype identification (benign, neoplasia, and hyperplasia subclassifications). Multiple feature extractor–classifier combinations were systematically trained, tuned, and evaluated to determine optimal configurations. Novel data augmentation strategies were also explored to improve robustness. In addition, a web-based interactive platform was developed to deploy the models, enabling clinicians to upload slides, visualize model attention maps, and review case-level reports within a hierarchical diagnostic workflow. Results The models achieved AUCs of 99.39% for three-class classification and 99.89% for binary classification. Subtype-level performance reached AUCs of 98.01% for benign, 95.25% for hyperplasia, and 93.19% for neoplasia. RankMix augmentation specifically improved performance across primary tasks and hyperplasia subclassification. Conclusions This framework demonstrates strong performance for automated endometrial biopsy classification across multiple diagnostic tasks. These findings support its potential as a decision-support tool for future pathology triage workflows, subject to external validation and workflow-level evaluation.