Ovarian cancer is one of the deadliest cancers of the female reproductive system, with poor prognosis, especially when diagnosed at an advanced stage. Accurate prognostic prediction and timely treatment are critical for improving patient outcomes. The aim of this study was to develop a prognostic prediction model for ovarian cancer based on pathological images. 158 In-house and 105 TCGA-OV pathological slides were processed with Macenko's algorithm for stain normalization and patch extraction (256 × 256 pixels). The CLAM framework was applied to construct a prognostic model validated via time-dependent ROC and survival analysis. The model achieved AUCs of 0.93 (internal) and 0.70 (external), demonstrating its potential for clinical translation. In addition, the prediction model was analysed in combination with patients' clinical characteristics and transcriptomic data. The results showed a significant prognostic difference between high and low risk groups. Our model can accurately predict the prognosis of ovarian cancer patients, which provides certain reference value for clinical diagnosis and treatment, especially when integrated with biomarkers like CA-125 for personalized risk stratification.
Keywords: Deep learning; Histopathology; Ovarian cancer; Prognostic scoring model.
© 2025. The Author(s).