Prediction of colorectal cancer liver metastasis through an MRI radiomic model

The aim of this study was to investigate the efficacy of a magnetic resonance imaging (MRI) radiomic model in predicting colorectal cancer liver metastasis (CRLM). Two independent cohorts consisting of 194 patients with pathologically confirmed colorectal cancer (CRC) who underwent baseline MRI examinations were recruited from the Affiliated Hospital of North Sichuan Medical College (Unit 1, n = 159) and Nanchong Central Hospital (Unit 2, n = 35) and divided into a training cohort (Unit 1) and an independent external validation cohort (Unit 2). The clinical risk factors for all patients were examined via univariate and multivariate analyses to identify independent clinical risk factors for CRLM. Radiomic features from oblique axial or axial fat-free T₂-weighted imaging (T₂WI) and diffusion-weighted imaging (DWI) sequences were extracted. Least absolute shrinkage and selection operator (LASSO) regression was subsequently used to screen the optimal radiomic features of each sequence. Finally, logistic regression was used to construct a prediction model according to the features from each sequence (e.g., T₂WI and DWI models), a combined radiomic model (M) integrating the features of both the T₂WI and DWI sequences, and a combined imaging-clinical model (U) integrating the radiomic features of both sequences with the independent clinical risk factors. The area under the receiver operating characteristic curve (AUC) was calculated to evaluate the predictive performance of each model. Among the 194 CRC patients enrolled, 86 had liver metastasis, and 108 did not. The tumor marker carcinoembryonic antigen was identified as a clinically independent risk factor for CRLM. Eleven optimal radiomic features were screened from each of the T₂WI and DWI sequences through LASSO regression analysis. The AUC values of the clinical, T₂WI, DWI, M, and U models were 0.755, 0.834, 0.844, 0.853, and 0.890, respectively, in the training cohort and 0.697, 0.786, 0.750, 0.808, and 0.842, respectively, in the validation cohort. The predictive performance of the combined models was better than that of the single-sequence models, and the U model performed best in terms of predicting CRLM. The results of this study suggest that the MRI radiomic model based on the imaging features of primary CRC lesions integrated with clinical risk factors can accurately predict CRLM, which may assist clinicians in the development of individualized treatment plans for patients with CRC.