Deep learning of pretreatment ascites cytopathology for platinum-resistance risk stratification in advanced epithelial ovarian cancer

Background Platinum resistance is a major determinant of poor outcome in advanced epithelial ovarian cancer, yet reliable predictors available before treatment initiation remain scarce. Ascitic fluid is commonly obtained during diagnostic work-up and directly reflects the peritoneal tumour microenvironment, but its cytomorphological information has not been systematically exploited for treatment-response prediction. Methods We present OVCAP, a multi-scale deep-learning framework that analyses pretreatment ascites cytology whole-slide images to estimate platinum-resistance risk. The study included 438 patients with FIGO stage IIIB–IV epithelial ovarian cancer. Model performance was evaluated in one internal and two independent external validation cohorts. Attention-guided cytopathology review was performed to identify high-risk morphologic patterns, and integrated single-cell RNA sequencing analyses were used to characterise the underlying biological features. Results OVCAP achieved area under the receiver operating characteristic curve (ROC-AUC) values of 0.894, 0.863, and 0.828 in the internal and two independent external validation cohorts, respectively, and outperformed the KELIM score (AUC 0.619). Attention-guided cytopathology review identified recurrent high-risk morphologic patterns in resistant disease: epithelial cytoplasmic vacuolization and interaction-rich malignant aggregates accompanied by immune and mesothelial cells. Integrated single-cell analyses linked these phenotypes to membrane remodelling, lipid reprogramming, hypoxia-associated stress signalling, and reinforced adhesion and immunoregulatory networks. Conclusion These findings support pretreatment ascites cytology as a clinically accessible substrate for early risk stratification before first-line platinum-based therapy.