Background and purpose: Recurrence is the main risk for high-grade serous ovarian cancer (HGSOC) and few prognostic biomarkers were reported. In this study, we proposed a novel deep learning (DL) method to extract prognostic biomarkers from preoperative computed tomography (CT) images, aiming at providing a non-invasive recurrence prediction model in HGSOC.
Materials and methods: We enrolled 245 patients with HGSOC from two hospitals, which included a feature-learning cohort (n = 102), a primary cohort (n = 49) and two independent validation cohorts from two hospitals (n = 49 and n = 45). We trained a novel DL network in 8917 CT images from the feature-learning cohort to extract the prognostic biomarkers (DL feature) of HGSOC. Afterward, a DL-CPH model incorporating the DL feature and Cox proportional hazard (Cox-PH) regression was developed to predict the individual recurrence risk and 3-year recurrence probability of patients.
Results: In the two validation cohorts, the concordance-index of the DL-CPH model was 0.713 and 0.694. Kaplan-Meier's analysis clearly identified two patient groups with high and low recurrence risk (p = 0.0038 and 0.0164). The 3-year recurrence prediction was also effective (AUC = 0.772 and 0.825), which was validated by the good calibration and decision curve analysis. Moreover, the DL feature demonstrated stronger prognostic value than clinical characteristics.
Conclusions: The DL method extracts effective CT-based prognostic biomarkers for HGSOC, and provides a non-invasive and preoperative model for individualized recurrence prediction in HGSOC. In addition, the DL-CPH model provides a new prognostic analysis method that can utilize CT data without follow-up for prognostic biomarker extraction.
Keywords: Artificial intelligence; Auto encoder; Computed tomography; Deep learning; High-grade serous ovarian cancer; Prognosis; Recurrence; Semi-supervised learning; Unsupervised learning.
Copyright © 2018 The Authors. Published by Elsevier B.V. All rights reserved.