Accurately predicting relapse after surgery remains a major challenge in stage II and III localized colon cancer, where conventional clinicopathological staging does not fully capture individual recurrence risk. This uncertainty can complicate decisions around adjuvant chemotherapy.
This multicenter real-world study evaluated whether AI-powered radiomics from routinely acquired preoperative CT scans could improve relapse prediction. The analysis included 278 patients and compared clinical-pathological, imaging-only, and combined models.
The best-performing model integrated radiomic, fractal and clinicopathological features, achieving a C-index of 0.68 and a mean time-dependent AUC of 0.69 in the independent test set, outperforming models based on clinical-pathological variables or imaging alone.
Importantly, the model was translated into a new Risk Classification score that clearly separated patients into low- and high-risk groups. This score emerged as the strongest predictor of relapse, with an adjusted HR of 11.74, exceeding the prognostic power of individual conventional clinical variables.
By extracting additional prognostic information from CT scans already acquired as part of standard care, this approach offers a non-invasive opportunity to refine preoperative risk stratification and potentially support more personalized treatment planning in localized colon cancer.
Prieto-de-la-Lastra, C., Carbonell-Asins, J. A., Bueno, A., Gómez-Alderete, A., Busto, M., Alcolado-Jaramillo, A. B., Jimenez-Pastor, A., Monzonís, X., Cuñat, A., Montagut, C., Moreno-Ruiz, P., Huerta, M., Roda, D., Gimeno-Valiente, F., Estepa-Fernández, A., Bellvís-Bataller, F., Fuster-Matanzo, A., Gibert, J., Roselló, S., Martinez-Ciarpaglini, C., Vidal, J., Alberich-Bayarri, Á., Cervantes, A., & Tarazona, N. (2025). A radiomics-based artificial intelligence model to assess the risk of relapse in localized colon cancer. ESMO Open, 10(8), 105495.