Development and internal validation of a pretreatment magnetic resonance imaging-based radiomics prediction model for 1-year overall survival in treatment-naïve patients with hepatocellular carcinoma receiving transarterial chemoembolization combined with targeted therapy and immune checkpoint inhibitors.

Authors

Abstract

Reliable pretreatment tools to identify treatment-naïve patients with intermediate-to-advanced hepatocellular carcinoma (HCC) who may benefit from this combination are lacking. We developed and internally validated a clinical-radiomics model based on pretreatment T2-weighted fat-suppressed magnetic resonance imaging (MRI) for predicting 1-year overall survival (OS). This retrospective, single-center cohort study included 85 treatment-naïve patients with Barcelona Clinic Liver Cancer (BCLC) stage B/C HCC treated with transarterial chemoembolization (TACE) plus targeted therapy and immune checkpoint inhibitors between December 2021 and April 2024. The last follow-up was April 1, 2026. The primary endpoint was 1-year OS status (death within 12 months vs. survival ≥12 months). Patients were randomly divided into a training set (70%) and a held-out test set (30%) using stratified sampling. Three-dimensional tumor volumes were manually segmented on pretreatment axial T2 spectral attenuated inversion recovery images. Features were first filtered for intra- and interobserver reproducibility, then selected using univariable screening and least absolute shrinkage and selection operator (LASSO) regression before being entered into the machine-learning models. The optimal model was selected via fivefold cross-validation in the training set and evaluated in a held-out test set through use of discrimination, calibration, decision curve analysis, and Shapley additive explanations. A total of 2,525 radiomics features were extracted, and four final predictors were retained: maximum tumor diameter and three filtered radiomics features. Gradient boosting machine (GBM) achieved the highest mean cross-validation area under the receiver operating characteristic curve (AUC) (0.818±0.079). In the held-out test set, the combined GBM model yielded an AUC of 0.91 [95% confidence interval (CI): 0.87-0.93], accuracy of 84.6% (95% CI: 79.3-88.9%), sensitivity of 81.8% (95% CI: 72.2-89.0%), and specificity of 86.7% (95% CI: 79.1-92.1%); the Brier score was 0.04. Decision curve analysis suggested a net clinical benefit across a broad range of threshold probabilities. A pretreatment T2-weighted fat-suppressed MRI-based clinical-radiomics model demonstrated promising performance for predicting 1-year OS in treatment-naïve patients with BCLC stage B/C HCC receiving TACE plus targeted therapy and immune checkpoint inhibitors. However, this single-center finding requires external validation before clinical implementation.

Tags

Topics