Weakly supervised deep multi-instance learning for classification of endometrial lesions on hematoxylin and eosin-stained whole-slide images.

Endometrial cancer (EC) is the most common gynecological malignancy, yet reliable screening and diagnostic approaches remain limited. We developed a weakly supervised deep multi-instance learning model (DSMIL) to classify hematoxylin and eosin-stained whole-slide images (WSIs) of endometrial tissue. A total of 885 WSIs from 442 patients, including EC, atypical endometrial hyperplasia (AEH), endometrial hyperplasia without atypia (EH), and normal endometrium (NE), were analyzed. DSMIL achieved an average AUROC of 0.9776 for four-class classification, with inter-class AUROCs of 0.9876 for EC, 0.9600 for AEH, 0.9771 for EH, and 0.9855 for NE, and outperformed other algorithms such as TransMSL, CLAM, and ABMIL (average accuracy = 0.8914). Attention heatmaps highlighted regions associated with pathological features, while nnU-Net v2 combined with HoverNet enabled identification of atypical glandular epithelial cells, which showed increased density, size, and perimeter but reduced axis ratios compared with normal cells. These results suggest that DSMIL provides a reliable computational pathology approach for the classification of endometrial lesions and the characterization of atypical cells.