Weakly supervised deep learning-based detection of serous tubal intraepithelial carcinoma in fallopian tubes

Serous tubal intraepithelial carcinoma (STIC) is an uncommon, non-invasive carcinoma that occurs more frequently in individuals with germline <i>BRCA</i> mutations and is an established precursor to high-grade serous ovarian carcinoma. STIC can be challenging to detect during pathologist evaluation, as it can manifest as a small focus of atypia in an otherwise benign salpingectomy specimen. There is a clinical need for scalable, weakly supervised computational approaches to aid in the detection of STIC. We developed a deep learning model to identify STIC and serous tubal intraepithelial lesions (STIL) in whole-slide images. We obtained fallopian tube specimens diagnosed as STIC (<i>n</i> = 49), STIL (<i>n</i> = 48), and benign fallopian tube (<i>n</i> = 83) at a single academic medical center. We trained a weakly supervised, attention-based multiple instance learning model and evaluated performance on independent datasets, including an additional unbalanced dataset (<i>n</i> = 40 benign, <i>n</i> = 2 STIL, <i>n</i> = 1 STIC) and cases diagnosed descriptively as benign reactive atypia (<i>n</i> = 53). The model achieved high sensitivity and specificity on the balanced validation cohort, with an area under the receiver operating characteristic curve (AUROC) of 0.96 (95% CI: 0.90-1.00), and demonstrated similarly strong performance on unbalanced validation cohorts (AUROC 0.98). Interpretability analyses indicated that model decisions were based on epithelial atypia. These results support the potential of integrating deep learning screening tools into clinical workflows to augment pathologist efficiency and diagnostic accuracy in fallopian tubes.