Predicting advanced-stage recurrence in resected hepatocellular carcinoma using preoperative gadoxetic acid-enhanced MRI radiomics.

OBJECTIVES

To develop and validate a preoperative gadoxetic acid-enhanced MRI-based radiomics model for predicting the time to advanced-stage recurrence in patients undergoing resection for hepatocellular carcinoma (HCC).

MATERIALS AND

METHODS

This retrospective single-center study included patients with Barcelona Clinic Liver Cancer (BCLC) stage 0/A HCC who underwent resection between 2009 and 2014. Patients were temporally divided into training (n = 272; 2009-2011) and validation (n = 210; 2012-2014) cohorts. Time to advanced-stage recurrence was defined as the interval from resection to macrovascular invasion and/or extrahepatic metastasis. Two radiologists segmented tumors on hepatobiliary phase images, followed by 3-mm peritumoral expansion and propagation to co-registered dynamic T1- and T2-weighted images. After extraction of Image Biomarker Standardization Initiative-compliant features, a radiomics model was developed using least absolute shrinkage and selection operator (LASSO)-Cox regression. Model performance and inter-reader reproducibility of binary risk stratification were evaluated using the concordance index (C-index) and Cohen's kappa, respectively.

RESULTS

Five-year advanced-stage recurrence occurred in 15% of the training (42/272) and validation (32/210) cohorts. In the validation cohort, the radiomics model yielded a C-index of 0.716 (95% CI, 0.631-0.801). This outperformed the preoperative BCLC (0.597; 95% CI 0.551-0.644) and Hong Kong Liver Cancer (0.594; 95% CI 0.516-0.672) staging systems, and was equivalent to the postoperative clinicopathological model (0.716; 95% CI 0.626-0.807). Radiomic-based risk stratification showed good inter-reader reproducibility (kappa = 0.757).

CONCLUSION

The preoperative radiomics model outperformed staging systems and was comparable to the postoperative clinicopathological model in predicting advanced-stage recurrence, relying entirely on noninvasive data.