Dose-response effect of aggressive MRI features and imaging subtypes in HCC early recurrence.

PURPOSE: Hepatocellular carcinoma (HCC) exhibits substantial early recurrence heterogeneity that current staging systems inadequately capture. We investigated whether gadoxetate disodium (Gd-EOB-DTPA)-enhanced MRI aggressive features exhibit a dose-response relationship with early recurrence and whether unsupervised clustering identifies distinct imaging subtypes with differential outcomes. METHODS: This retrospective study included 307 HCC patients from two centers (development: n = 245; validation: n = 62) who underwent preoperative Gd-EOB-DTPA MRI and curative hepatectomy. Nine aggressive MRI features were assessed. Dose-response analysis used Cochran-Armitage trend tests and Spearman correlation. K-means clustering was applied to derive imaging subtypes, with consistency validated via hierarchical clustering and Gaussian mixture models (adjusted Rand index). Predictive performance was compared with BCLC and CNLC staging using bootstrap-derived 95% confidence intervals for AUC differences. RESULTS: Early recurrence occurred in 37.1% (114/307). Recurrence rates increased monotonically from 18.7% (0 features) to 79.6% (≥ 4 features; Spearman P = 0.413, P < 0.001; Cochran-Armitage Z = 7.547, P < 0.001). The MRI feature count achieved AUC 0.741 (95% CI 0.678-0.798), significantly outperforming BCLC staging (0.672; ΔAUC + 0.067, 95% CI + 0.011 to + 0.126). K-means clustering identified three subtypes: indolent (58.4%, recurrence 19.6%), aggressive (38.4%, 52.1%), and infiltrative (3.3%, 87.5%; 95% CI 62.5-100%; P < 0.001). Clustering consistency was confirmed (K-means vs. hierarchical: ARI = 0.505; NMI = 0.481). In multivariate analysis, the MRI count was independently predictive (OR 1.59, 95% CI 1.29-1.97, P < 0.001) after adjustment for age, tumor size, AFP, and MVI. The dose-response trend was replicated in external validation (0 features: 33.3% → ≥4: 83.3%; AUC 0.642). In time-to-event analysis, 24-month recurrence-free survival decreased across risk tiers (78.4% to 20.4%; log-rank P < 0.001), and the feature count remained an independent predictor in Cox regression (HR 1.32 per feature, 95% CI 1.17-1.49, P < 0.001; C-index 0.738). CONCLUSION: Aggressive MRI features demonstrate a robust dose-response relationship with early HCC recurrence, fulfilling a key criterion for causal inference. Data-driven imaging subtypes provide prognostic stratification that complements existing staging systems, offering a simple and clinically interpretable framework for risk-adapted surveillance.