An interpretable model based on weakly supervised learning for uterine smooth muscle tumor diagnosis: A multi-center study
Uterine smooth muscle tumors (USMTs) are the most common tumors of the female reproductive system, but remain diagnostically challenging due to morphological overlap among leiomyosarcoma (LMS), various leiomyoma (LM) subtypes, and smooth muscle tumors of uncertain malignant potential (STUMP). This study aimed to develop a weakly supervised artificial intelligence (AI) model for classifying USMTs as benign or malignant using only slide-level labels. A multi-center dataset comprising 94 LMS cases (408 whole-slide images, WSIs) and 634 benign cases (1389 WSIs) was used for model training and internal testing, with an independent external test set including 27 LMS cases (54 WSIs) and 90 benign cases (248 WSIs). The CAMEL2-based model achieved excellent diagnostic performance, with AUCs of 0.9976 and 0.9889 on the internal and external test sets, respectively, and accuracies exceeding 0.97. Heatmaps frequently highlighted regions enriched for key pathological features of LMS, while benign tissues were consistently assigned low-suspicion regions. In STUMP cases, heatmaps emphasized morphologically suspicious regions that often overlapped with areas identified by pathologists, supporting their potential as decision-support visualizations. In the AI-human collaboration study, model assistance was associated with improved diagnostic accuracy and reduced diagnostic time. This represents the first weakly supervised learning model for USMT diagnosis, achieving high accuracy and interpretability with minimal annotation requirements.
