Answers are grounded in this paper and cite their sources.
Background: While histopathological analysis remains the definitive gold standard for diagnosing endometrial pathology, hysteroscopy is the primary clinical tool for direct visualisation of the endometrial cavity and targeted biopsy. Visual interpretation during the procedure remains subjective and highly dependent on experience. Deep learning has been applied to lesion detection, but few studies report performance at the level of individual patients or assess how stable model performance is across different patient groupings. This study aimed to develop and rigorously evaluate a multi-task deep learning system for simultaneous endometrial lesion subtyping and binary clinical triage from hysteroscopic images. Methods: A retrospective, single-centre study was conducted using 1030 hysteroscopic images collected from 130 verified cases between 2021 and 2025. The dataset included images of Endometrial Polyp (n = 269), Hyperplasia without Atypia (n = 303), and a combined High-grade group comprising Atypical Hyperplasia and Endometrial Cancer (n = 458). A Multi-Task ResNet-50 architecture was implemented with dual classification heads to predict both lesion type and a binary triage outcome. To enhance generalisation and reduce overfitting, online data augmentation was implemented during training. Finally, the model was integrated into a Streamlit-based Graphical User Interface (GUI) to demonstrate software integration for research purposes. Results: On the held-out test set (139 images, 26 patients), three-class accuracy was 0.777 (95% CI 0.712–0.842) and macro-F1 was 0.769 (95% CI 0.694–0.838); binary triage accuracy was 0.842 (95% CI 0.784–0.899) with an AUC of 0.845 (95% CI 0.767–0.909). Sensitivity for the high-grade class was 0.65 under the binary head and 0.63 under the three-class head, which is close but not identical, indicating that the two heads, despite sharing a backbone, are reaching their decisions independently rather than one being derived from the other. Specificity was high under both heads. Conclusions: Aggregate accuracy looked similar whether images or patients were treated as the unit of analysis, but the patient-level analysis, the analysis that actually reflects how the model would be judged in practice, revealed both a lower and far less certain sensitivity for high-grade disease than the image-level figures suggested. We report this as a feasibility study and a methodological illustration, not a validated diagnostic aid and argue that patient-level evaluation, at minimum, should be standard practice when reporting hysteroscopy-AI performance on cohorts of this size.
No takes yet. Share an insight, caveat, or question.
Yue Teng (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: