Histopathologic basis of a deep learning pelvic computed tomography model for prognostic prediction among patients with advanced high-grade serous ovarian carcinoma

Background: A deep learning (DL) model based on preoperative computed tomography (CT) has been proposed for estimating recurrence in patients with ovarian cancer. However, the inherent opacity of DL models complicates the interpretation of their output, limiting their clinical applicability. The aim of this study was thus to generate histopathologic evidence supporting such DL prediction models and to construct a clustering-based analytical framework for identifying patients with risk factors for ovarian cancer. Methods: A retrospective study was conducted in which preoperative CT data were collected from patients with high-grade serous ovarian cancer treated with radical tumor resection from January 2013 to December 2019 at three tertiary care centers. Unsupervised clustering was performed with 1,280 DL model-driven features, and the associations between clusters and histopathological features were analyzed. Multivariate regression was used to investigate the added value of DL outputs for histopathologic correlations. Results: A total of 418 patients [median age 55 years, interquartile range (IQR), 30-77 years] were evaluated. Unsupervised clusters 3 and 4 were associated with the positive status of P53, P16, and Ki-67, along with invasion of the omentum, rectum, and pelvic wall (P