Radiomics Features at Multiparametric MRI Predict Disease-Free Survival in Patients With Locally Advanced Rectal Cancer

Objective To investigate the potential value of radiomics features based on preoperative multiparameter MRI in predicting disease-free survival (DFS) in patients with local advanced rectal cancer (LARC). Methods We identified 234 patients with LARC who underwent preoperative MRI, including T2-weighted, diffusion kurtosis imaging, and contrast enhanced T1-weighted. All patients were randomly divided into the training (n = 164) and validation (n = 70) cohorts. 414 features were extracted from the tumor from above sequences and the radiomics signature was then generated, mainly based on feature stability and Cox proportional hazards model. Two models, integrating pre- and postoperative variables, were constructed to validate the radiomics signatures for DFS estimation. Results The radiomics signature, composed of six DFS-related features, was significantly associated with DFS in the training and validation cohorts (both p < 0.001). The radiomics signature and MR-defined extramural venous invasion (mrEMVI) were identified as the independent predictor of DFS both in the pre- and postoperative models. In both cohorts, the two radiomics-based models exhibited better prediction performance (C-index ≥0.77, all p < 0.05) than the corresponding clinical models, with positive net reclassification improvement and lower Akaike information criterion (AIC). Decision curve analysis also confirmed their clinical usefulness. The radiomics-based models could categorize LARC patients into high- and low-risk groups with distinct profiles of DFS (all p < 0.05). Conclusion The proposed radiomics models with pre- and postoperative features have the potential to predict DFS, and may provide valuable guidance for the future individualized management in patients with LARC. To investigate the potential value of radiomics features based on preoperative multiparameter MRI in predicting disease-free survival (DFS) in patients with local advanced rectal cancer (LARC). We identified 234 patients with LARC who underwent preoperative MRI, including T2-weighted, diffusion kurtosis imaging, and contrast enhanced T1-weighted. All patients were randomly divided into the training (n = 164) and validation (n = 70) cohorts. 414 features were extracted from the tumor from above sequences and the radiomics signature was then generated, mainly based on feature stability and Cox proportional hazards model. Two models, integrating pre- and postoperative variables, were constructed to validate the radiomics signatures for DFS estimation. The radiomics signature, composed of six DFS-related features, was significantly associated with DFS in the training and validation cohorts (both p < 0.001). The radiomics signature and MR-defined extramural venous invasion (mrEMVI) were identified as the independent predictor of DFS both in the pre- and postoperative models. In both cohorts, the two radiomics-based models exhibited better prediction performance (C-index ≥0.77, all p < 0.05) than the corresponding clinical models, with positive net reclassification improvement and lower Akaike information criterion (AIC). Decision curve analysis also confirmed their clinical usefulness. The radiomics-based models could categorize LARC patients into high- and low-risk groups with distinct profiles of DFS (all p < 0.05). The proposed radiomics models with pre- and postoperative features have the potential to predict DFS, and may provide valuable guidance for the future individualized management in patients with LARC.