Multiple Instance Learning for the Detection of Lymph Node and Omental Metastases in Carcinoma of the Ovaries, Fallopian Tubes and Peritoneum

Background/Objectives: Surgical pathology of tubo-ovarian and peritoneal cancer carries a well-recognised diagnostic workload, partly due to the large amount of non-primary tumour-related tissue requiring assessment for the presence of metastatic disease. The lymph nodes and omentum are almost universally included in such resection cases and contribute considerably to this burden, principally due to volume rather than task complexity. To date, artificial intelligence (AI)-based studies have reported good success rates in identifying nodal spread in other malignancies, but the development of such time-saving assistive digital solutions has been neglected in ovarian cancer. This study aimed to detect the presence or absence of metastatic ovarian carcinoma in the lymph nodes and omentum.

Methods: We used attention-based multiple-instance learning (ABMIL) with a vision-transformer foundation model to classify whole-slide images (WSIs) as either containing ovarian carcinoma metastases or not. Training and validation were conducted with a total of 855 WSIs of surgical resection specimens collected from 404 patients at Leeds Teaching Hospitals NHS Trust.

Results: Ensembled classification from hold-out testing reached an AUROC of 0.998 (0.985–1.0) and a balanced accuracy of 100% (100.0–100.0%) in the lymph node set, and an AUROC of 0.963 (0.911–0.999) and a balanced accuracy of 98.0% (94.8–100.0%) in the omentum set.

Conclusions: This model shows great potential in the identification of ovarian carcinoma nodal and omental metastases, and could provide clinical utility through its ability to pre-screen WSIs prior to histopathologist review. In turn, this could offer significant time-saving benefits and streamline clinical diagnostic workflows, helping to address the chronic staffing shortages in histopathology.