1077 - Dynamic Delta-Radiomics and Clinical Fusion Model for Predicting Disease-Free Survival In Locally Advanced Cervical Cancer: A Multi-Cohort Study

Purpose/Objective(s): Accurate survival prediction is essential for personalizing radiotherapy in locally advanced cervical cancer (LACC). This study aimed to develop and validate a model integrating clinical variables, pre- and post-radiotherapy CT radiomics, and delta radiomics to predict disease-free survival (DFS).

Materials/Methods: We retrospectively analyzed 1,159 LACC patients from two centers (training n=512, internal validation n=342, external validation n=117). CT scans were acquired pre-radiotherapy and post-radiotherapy. We extracted 107 radiomic features from CTV using deep learning-assisted segmentation. Delta features were calculated. After univariate and correlation analyses, 28 features (7 clinical, 10 pre-RT, 7 post-RT, 4 delta) were selected. Performance was evaluated via C-index, time-dependent AUC, and Kaplan-Meier analysis.

Results: The dynamic model achieved C-indices of 0.713 (internal) and 0.693 (external), outperforming the static model (0.705 and 0.690). Time-dependent AUCs for 12-, 36-, and 60-month DFS were 0.781, 0.755, 0.736 (internal) and 0.859, 0.770, 0.671 (external). Delta features served as interpretable indicators of therapeutic response. Kaplan-Meier analysis confirmed effective risk stratification (p < 0.001).

Conclusion: Integrating dynamic delta radiomics with multi-timepoint CT imaging significantly improves DFS prediction in LACC. This validated framework bridges predictive performance and biological interpretability, offering a non-invasive strategy for precision oncology.