Multimodal deep learning using preoperative CT and ultrasound for recurrence risk prediction in high-grade serous ovarian carcinoma.

backgroundHigh-grade serous ovarian carcinoma (HGSOC) is associated with a high risk of postoperative recurrence, and the timing of recurrence is closely related to patient survival outcomes and subsequent treatment strategies. However, current postoperative surveillance primarily relies on clinical indicators and routine imaging examinations, which offer limited predictive accuracy. Although deep learning–based image analysis has shown promise in capturing tumor heterogeneity and improving prognostic assessment, studies integrating preoperative contrast-enhanced computed tomography (CE-CT) and ultrasound imaging for recurrence risk prediction in HGSOC remain limited.

methodsThis single-center retrospective study enrolled 293 patients with pathologically confirmed HGSOC, who were randomly assigned to a training cohort and an internal validation cohort. Separate two-dimensional deep learning survival models were developed using preoperative CE-CT and ultrasound images. A Cox partial likelihood–based time-to-event loss function was applied to generate modality-specific deep learning scores (DL-scores). Independent clinical predictors were identified through univariate and multivariate Cox regression analyses and were integrated with CT-DL and US-DL scores to construct a multimodal Cox prognostic model, which was visualized using a nomogram. Model performance was assessed using the concordance index (C-index), time-dependent receiver operating characteristic (ROC) curves, Kaplan–Meier survival analysis, calibration curves, and decision curve analysis. Bootstrap resampling was performed to evaluate model robustness, and Gradient-weighted Class Activation Mapping (Grad-CAM) was used to visualize model attention.

resultsThe multimodal integrated model demonstrated superior prognostic performance, achieving C-index values of 0.840 in the training cohort and 0.722 in the internal validation cohort. Time-dependent ROC analysis showed that the combined model achieved areas under the curve (AUCs) of 0.880, 0.890, and 0.892 for predicting 1-, 2-, and 3-year recurrence-free survival (RFS) in the training cohort, and 0.864, 0.781, and 0.772 in the validation cohort, respectively. Kaplan–Meier survival analysis revealed a significant separation between high- and low-risk groups (log-rank test, all P < 0.001). Calibration and decision curve analyses indicated good agreement between predicted and observed outcomes and a higher net clinical benefit. Bootstrap resampling further confirmed the robustness of the multimodal model. Grad-CAM visualizations suggested that the model primarily focused on tumor and peritumoral regions.

conclusionsThis study proposes a multimodal prognostic framework integrating deep learning features from CE-CT and ultrasound with clinical variables for preoperative prediction of postoperative recurrence risk in HGSOC. Using routinely acquired imaging data, the model shows consistent interpretability and performance, supporting its use as an exploratory decision-support and risk-stratification tool. Prospective, multicenter validation is required before clinical implementation.