Introduction/Background Tubo-ovarian high-grade serous carcinoma (HGSC) is thought to originate in the fallopian tubes, where precursor lesions such as serous tubal intraepithelial carcinoma (STIC) and serous tubal intraepithelial lesion (STIL) can be identified. Reliable and reproducible diagnosis of STIC and STIL is vital, as STIC is associated with an increased risk for peritoneal carcinoma and has clinical implications in new risk reducing strategies for women with an increased hereditary risk of developing HGSC. However, STIC can be a challenging diagnosis for pathologists.

Methodology A dataset comprising STIC/STIL (n=323) and controls (n=359) was collected and split into three cohorts. Cohort A (n=169) and B (n=327) contained data from overlapping data sources, whilst Cohort C (n=186) contained data from independent sources. A reference standard for cohorts A and B was set by an international panel review amongst 15 gynecologic pathologists. Cohort A then served to train and validate a deep-learning algorithm (U-Net with resnet50 backbone), designed to automatically detect STIC/STIL. Cohort B and Cohort C served as test sets.

Results The performance of the AI model was evaluated at slide level by ROC curve analysis. The model reached an AUC of 0.98 (95% CI: 0.96–0.99) on cohort B, and 0.95 (95% CI: 0.90–0.99) on cohort C, displaying a robust performance. Visual inspection of all cases confirmed accurate detection of STIC/STIL lesions in relation to the morphology, immunohistochemistry, and the reference standard.

Conclusion We present a deep-learning model that can automatically detect STIC and STIl in digitalized whole slide images of fallopian tube specimens. The output of this model may be useful to assist pathologists in detecting STIC and can contribute towards a more reliable and reproducible diagnosis.

Disclosures This project was funded by the Dutch Cancer Society (KWF), project number 12950.