Intratumoral Spatial Heterogeneity at Dynamic Contrast‐Enhanced MRI for Assessing Tertiary Lymphoid Structures in Hepatocellular Carcinoma

BACKGROUND: Tertiary lymphoid structures (TLS) are associated with enhanced antitumor immune activity in hepatocellular carcinoma (HCC). Dynamic contrast-enhanced (DCE)-MRI-based habitat imaging may enable noninvasive TLS identification. PURPOSE: To develop and validate a DCE-MRI-based habitat model to noninvasively identify TLS status in HCC. STUDY TYPE: Retrospective. SUBJECTS: Three hundred and thirty-four patients (mean age, 53.59 years ±11.38; male = 296; training set:test set = 200:134) with pathologically confirmed HCC. FIELD STRENGTH/SEQUENCE: 1.5-T/3.0-T, contrast-enhanced three-dimensional gradient-recalled-echo T1-weighted sequence. ASSESSMENT: Tumor habitats were identified from DCE-MRI via k-means clustering of voxel-wise enhancement patterns. Habitat-derived features and radiomics features were extracted. Five identification models were developed: clinical model, radiomics model, habitat model, clinical-radiomics model, and clinical-radiomics-habitat (hybrid) model. Model performance was evaluated using the area under the receiver operating characteristic curve (AUC). STATISTICAL TESTS: Receiver operating characteristic (ROC) curve analysis, area under the precision-recall curve (AUPRC), DeLong test, decision curve analysis, Spiegelhalter's Z test. Significance level: p < 0.05 (two-sided). RESULTS: Four habitats with distinct dynamic enhancement patterns were identified. The hybrid model demonstrated the highest discrimination in the training and test sets (AUC = 0.83 [95% CI: 0.78, 0.89] and 0.84 [95% CI: 0.78, 0.91]). It significantly outperformed the clinical-radiomics model (AUC = 0.78 [95% CI: 0.72, 0.84] and 0.77 [95% CI: 0.69, 0.85]), radiomics model (AUC = 0.76 [95% CI: 0.70, 0.83] and 0.75 [95% CI: 0.67, 0.83]), and clinical model (AUC = 0.70 [95% CI: 0.63, 0.78] and 0.68 [95% CI: 0.58, 0.77]) in both sets. Compared to the habitat model (AUC = 0.78 and 0.80), the hybrid model showed significantly better training performance but comparable test performance (p = 0.09). CONCLUSION: A hybrid model combining clinical, radiomics, and habitat-derived features was developed for identifying TLS status. TECHNICAL EFFICACY: Stage 2.