Prediction of prognosis and treatment response in ovarian cancer patients from histopathology images using graph deep learning: a multicenter retrospective study.

Background: Ovarian cancer (OV) is a prevalent and deadly disease with high mortality rates. The development of accurate prognostic tools and personalized therapeutic strategies is crucial for improving patient outcomes.

Methods: A graph-based deep learning model, the Ovarian Cancer Digital Pathology Index (OCDPI), was introduced to predict prognosis and response to adjuvant therapy using hematoxylin and eosin (H&E)-stained whole-slide images (WSIs). The OCDPI was developed using formalin-fixed, paraffin-embedded (FFPE) WSIs from the TCGA-OV cohort, and was externally validated in two independent cohorts from the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO) and Harbin Medical University Cancer Hospital (HMUCH).

Results: The OCDPI showed prognostic ability for overall survival prediction in the PLCO (HR, 1.916; 95% CI, 1.380-2.660; log-rank test, P < 0.001) and HMUCH (HR, 2.796; 95% CI, 1.404-5.568; log-rank test, P = 0.0022) cohorts. Patients with low OCDPI experienced better survival benefits and lower recurrence rates following adjuvant therapy compared to those with high OCDPI. Multivariable analyses, adjusting for clinicopathological factors, consistently identified OCDPI as an independent prognostic factor across all cohorts (all P < 0.05). Furthermore, OCDPI performed well in patients with low-grade tumors or fresh-frozen slides, and could differentiate between HRD-deficient or HRD-intact patients with and without sensitivity to adjuvant therapy.

Conclusion: The results from this multicenter cohort study indicate that the OCDPI may serve as a valuable and labor-saving tool to improve prognostic and predictive clinical decision-making in patients with OV.