Multi-sequence MRI deep learning and habitat radiomics for predicting mismatch repair status and prognosis in...

This retrospective study aims to evaluate the habitat radiomics and deep learning models based on multi-sequence MRI in preoperatively predicting mismatch repair (MMR) status, and prognosis in colorectal liver metastasis (CRLM).

The total cohort (including 178 patients) was divided into a training cohort (93 patients), an internal validation cohort (40 patients), and an external validation cohort (45 patients). Axial T2WI, DWI (b = 800), and the BH Axial Dynamic Contrast-Enhanced (portal vein and delay phase) abdominal MRI were performed preoperatively for construction of classical radiomics model, habitat radiomics models, and deep learning model. Kaplan-Meier survival analysis was conducted to investigate prognostic stratification.

Among 178 patients (including 126 males and 52 females), the prevalence of dMMR/MSI-H was 19.1% (34/178). The primary tumor grade and location were the independent clinical predictors of dMMR/MSI-H. The deep learning (DL) model outperformed the classical radiomics and habitat radiomics models in both internal (AUC = 0.817, 95% CI: 0.657 ~ 0.978) and external validation cohorts (AUC = 0.710, 95% CI: 0.519 ~ 0.900). The prognosis of the DL output-high and DL output-low subgroups exhibited significant differences (log-rank test, P = 0.011).

The habitat radiomics and deep learning models based on multi-sequence MRI can effectively predict the MMR status of CRLM. Meanwhile, the DL model demonstrates superior performance which may facilitate prognostic stratification of patients with CRLM, thereby assisting in individualized clinical treatment and prognosis prediction.