Contrast-enhanced MRI-based multi-parameter habitats radiomics models to predict early recurrence in early-stage hepatocellular carcinoma following curative resection.

[OBJECTIVES] To establish and validate models including contrast-enhanced MRI habitat features for predicting the early postoperative recurrence of early-stage hepatocellular carcinoma (HCC).

[MATERIALS & METHODS] 263 patients with early-stage HCC who underwent contrast-enhanced MRI before hepatectomy were enrolled in our study. Using k-means clustering method, the arterial phase (AP) and portal venous phase (PVP) images were separately segmented according to the voxel and entropy to create habitats. The volume ratio and radiomics signature of each habitat were calculated and the correlation with early recurrence was assessed. A comprehensive model incorporating the optimal radiomic signature, significant volume ratio and clinicopathological predictors was constructed via logistic regression for early recurrence. Models discrimination was characterized with area under the receiver operating curve (AUC). the utility of the comprehensive model and optimal habitat were further evaluated.

[RESULTS] Three habitats were defined by clustering for each phase. The volume ratios of AP-habitat1 and AP-habitat2 were associated with early recurrence (P < 0.05). Among all radiomic signatures, AP-habitat1 + 2 (AUC, 0.646) and PVP-habitat fusion (AUC, 0.651) showed the optimal predictive performance. On the validation set, the AUC (0.767) of the comprehensive model was significantly higher than clinicopathological predictors (AUC, 0.589-0.613), radiomics-only models, and demonstrated superior performance than both the BCLC (0.647; P = 0.026) and UICC (0.655; P = 0.030) systems. Comprehensive model may stratify patients prognosis.

[CONCLUSION] The comprehensive model showed best performance in the prediction of early recurrence and may be used to identify patients at high risk recurrence and may help guide oncologic care and follow-up.