S2428
Physics - Radiomics, functional and biological imaging, and outcome prediction
ESTRO 2026
Only three models achieved confidence intervals completely above 0.5, i.e. were successfully validated. The meta-analysis revealed a significant association between validation performance and training cohort size (Figure 2).
comprehensive external validation of published radiomics models for rectal cancer to identify potentially generalizable models and to uncover methodological and reporting deficiencies that currently limit reproducibility and clinical translation. Material/Methods: This study included 294 rectal cancer patients who underwent T2-weighted MRI before neoadjuvant radiochemotherapy. Endpoints were treatment response and survival. A systematic literature search identified studies developing radiomics models for outcome prediction. Only studies matching the available data and providing sufficient methodological detail for replication were included. The METRICS score [1] was used to assess documentation quality. Models were either identically replicated when complete model specifications were available, or retrained via bootstrapping when at least feature signatures and model type were reported. Model performance was assessed using the area under the curve (AUC) for classification and the concordance- index (C-Index) for survival analysis, both with 95% confidence intervals. A meta-analysis explored potential associations between validation results and study-level characteristics, including the original validation strategy, publication year, training cohort size, and METRICS score. Results: A total of 58 models from 40 studies were replicated, including 47 for response prediction and 11 for survival analysis. Identical replication was possible for only 28 regression-based models, due to the lack of shared code or model details. Overall, the performance of external validation was poor (Figure 1), with AUC/C-index values ranging from 0.39 to 0.64 for identically replicated models, and from 0.43 to 0.66 for retrained models.
Conclusion: External validation showed limited reproducibility and poor generalizability of published radiomics models for rectal cancer, with none demonstrating performance levels suitable for clinical use. The main barriers for model validation were insufficient methodological reporting and lack of accessible model code. These findings underscore the urgent need for open, standardized, and reproducible research practices to enable robust external validation and accelerate the translation of radiomics models into clinically reliable tools. References: [1] - Kocak B, Akinci D'Antonoli T, Mercaldo N, et al. doi:10.1186/s13244-023-01572-w Keywords: external validation, meta-analysis, rectal cancer Digital Poster 1582 Generalizability of CT Radiomics for NSCLC Survival Prediction: Internal and External Validation Mengcheng Li 1 , Rhodri Smith 2 , Khamael Al-Battat 3 , Prosper Oyibo 1 , Emiliano Spezi 1 1 School of Engineering, Cardiff University, Cardiff, United Kingdom. 2 Faculty of Medicine and Dentistry - Radiology and Diagnostic Imaging Department, Alberta University, Edmonton, Canada. 3 School of Medicine, Cardiff University, Cardiff, United Kingdom
Made with FlippingBook - Share PDF online