From Tumor to Tumor-Spleen: MRI Habitat Heterogeneity for Predicting Immunotherapy Outcome in Advanced Hepatocellular Carcinoma.

OBJECTIVE: To develop and validate a pretreatment MRI-based radiomics model that integrates tumor and spleen habitat heterogeneity features for predicting objective response to immunotherapy and stratifying progression-free survival (PFS) in patients with advanced hepatocellular carcinoma (HCC). METHODS: In this retrospective study, 107 patients with advanced HCC receiving first-line immunotherapy were included. Tumor and spleen habitats were independently identified on portal venous phase MRI using K-means clustering. High-throughput radiomic features were extracted from each habitat, and their spatial heterogeneity was quantified to generate predictive signatures. These features were integrated into a weighted fusion model (RadCVTS) to predict objective response. Model performance was evaluated in a temporal validation cohort, with comparison against a tumor-only heterogeneity model (RadCVT) and a conventional tumor radiomics model (RadT). Association with PFS was assessed using Kaplan-Meier analysis and Cox regression. RESULTS: The RadCVTS demonstrated superior predictive performance compared to both the RadCVT and the conventional RadT,with AUCs of 0.958 vs 0.828 vs 0.789, respectively, in the temporal validation cohort. Patients stratified as high-risk by the model had significantly shorter PFS than the low-risk group (hazard ratio: 11.44, p < 0.05). Exploratory analysis of the top two features from the tumor and spleen, respectively, revealed that patients exhibiting a "Double-High" imaging phenotype had a 22.5 times greater odds of response than the "Double-Low" group. A simplified model using only these two features also provided significant incremental value for PFS prediction (AUC = 0.802). CONCLUSION: The tumor-spleen combined habitat heterogeneity model demonstrated superior predictive performance over both the tumor-only heterogeneity model and the conventional tumor radiomics model, and can noninvasively predict response to immunotherapy and stratify PFS in advanced HCC, highlighting the value of the liver-spleen axis and providing a promising indicator to support personalized treatment decision making.