Deep learning detects premalignant lesions in the Fallopian tube
Joep Bogaerts, John‐Melle Bokhorst, Michiel Simons, Majke H.D. van Bommel, Miranda P. Steenbeek, Joanne A. de Hullu, Jasper Linmans, Joost Bart, Jessica L. Bentz, Tjalling Bosse, Johan Bulten, Yen‐Wei Chien, Mohamed Mokhtar Desouki, Ricardo R. Lastra, Tricia Numan, J. Kenneth Schoolmeester, Lauren E. Schwartz, Ie‐Ming Shih, T. Rinda Soong, Gulisa Turashvili, Russell Vang, Mila Volchek, Jeroen van der Laak
Abstract
Abstract Tubo-ovarian high-grade serous carcinoma is believed to originate in the fallopian tubes, arising from precursor lesions like serous tubal intraepithelial carcinoma (STIC) and serous tubal intraepithelial lesion (STIL). Adequate diagnosis of these precursors is important, but can be challenging for pathologists. Here we present a deep-learning algorithm that could assist pathologists in detecting STIC/STIL. A dataset of STIC/STIL ( n = 323) and controls ( n = 359) was collected and split into three groups; training ( n = 169), internal test set ( n = 327), and external test set ( n = 186). A reference standard was set for the training and internal test sets, by a panel review amongst 15 gynecologic pathologists. The training set was used to train and validate a deep-learning algorithm (U-Net with resnet50 backbone) to differentiate STIC/STIL from benign tubal epithelium. The model’s performance was evaluated on the internal and external test sets by ROC curve analysis, achieving an AUROC of 0.98 (95% CI: 0.96–0.99) on the internal test set, and 0.95 (95% CI: 0.90–0.99) on the external test set. Visual inspection of all cases confirmed the accurate detection of STIC/STIL in relation to the morphology, immunohistochemistry, and the reference standard. This model’s output can aid pathologists in screening for STIC, and can contribute towards a more reliable and reproducible diagnosis.