Comparison of clinical, radiomics, deep learning, and fusion models for predicting early recurrence in locally advanced rectal cancer based on multiparametric MRI: a multicenter study

Predicting early recurrence (ER) in locally advanced rectal cancer (LARC) is critical for clinical decision-making. This study aimed at comparing clinical, deep learning (DL), radiomics, and two fusion models for ER prediction based on multiparametric MRI. This retrospective study involved 337 LARC patients from four centers between January 2016 and September 2021. Radiomics and DL features were extracted from preoperative multiparametric MRI, including T2WI, DWI, T1WI, and contrast-enhanced T1WI (CET1WI). The extreme gradient boosting (XGBoost) classifier was applied to establish the clinical model, radiomics model, DL model, and two fusion models (the feature-based early fusion model and the decision-based late fusion model). The area under the curve (AUC), DeLong test, calibration curve, and decision curve analysis (DCA) were used to assess models. Kaplan-Meier analysis was conducted to determine the prognostic value of the models by evaluating the differences in recurrence-free survival (RFS) between the high- and low-risk patients of ER. The late fusion model demonstrated the best performance compared with the early fusion model, clinical, radiomics and DL models, with the highest AUC (0.863-0.880) across all cohorts. In addition, the late fusion model exhibited the highest clinical net benefit, and good calibration. Kaplan-Meier survival curves showed that high-risk patients of ER defined by the late fusion model had a worse RFS than low-risk ones of ER (log-rank p < 0.001). The late fusion model can accurately predict ER in LARC and may serve as a clinically useful, non-invasive tool for optimizing treatment strategies and monitoring disease progression.