Gd-BOPTA MRI-based nomogram incorporating habitat features for preoperative prediction of high-risk hepatocellular carcinoma: a multicenter retrospective study

High-risk hepatocellular carcinoma (HCC), defined by the MSE score (micro/macro-vascular invasion, satellite nodules, and Edmondson–Steiner grade), is associated with poor postoperative prognosis. This study aimed to evaluate a Gd-BOPTA MRI habitat fusion model for noninvasive preoperative prediction of MSE-defined high-risk HCC. 583 patients from four centers were retrospectively enrolled. 527 patients from Center A were split into 7:3 training and internal validation sets, and 56 patients from three other centers formed the external validation cohort. Tumor habitat features were extracted by partitioning tumor voxels into distinct subregions using k-means clustering on arterial phase and hepatobiliary phase ratio maps. Independent predictors were identified by multivariable logistic regression, and a nomogram integrating clinical, radiological, and habitat features was constructed. Performance was assessed by the area under the receiver operating characteristic curve (AUC), DeLong test, and decision curve analysis. Kaplan–Meier analysis compared postoperative recurrence-free survival (RFS) between predicted high- and low-risk groups. Multivariable analysis identified elevated AFP (OR 4.40, 95% CI 2.44–7.94), non-rim arterial phase hyperenhancement (OR 0.38, 95%CI 0.23–0.65), corona enhancement (OR 4.54, 95% CI 2.47–8.35), peritumoral hypointensity (OR 10.96, 95% CI 2.02–59.57), and habitat score (OR 44.46, 95%CI 8.04–245.76) as independent predictors of high-risk HCC (all P < 0.05). The BOPTA habitat fusion model achieved optimal predictive performance with AUCs of 0.844 (95% CI: 0.804–0.884), 0.797 (95% CI: 0.718–0.876), and 0.820 (95% CI: 0.669–0.972) in the training, internal, and external validation cohorts, respectively. It significantly outperformed all individual models in the training and internal validation sets, while retaining the highest AUC in the external validation cohort. Predicted high-risk patients had significantly shorter median RFS than low-risk counterparts in both training (18 vs. 56 months, P < 0.001) and internal validation cohorts (19 vs. 69 months, P = 0.0047), supporting prognostic stratification by the nomogram. The Gd-BOPTA MRI habitat fusion model demonstrated promising performance for noninvasive preoperative prediction of MSE-defined high-risk HCC and may assist postoperative prognostic stratification. However, given the retrospective design, large‑scale prospective external validation is required before clinical application.