Noninvasive prediction of microvascular invasion in hepatocellular carcinoma using a fusion model of 5-mm optimal peritumoral radiomics and imaging features based on multiphasic gadoxetic acid-enhanced MRI.

Background: Microvascular invasion (MVI) is a remarkable factor used to predict postoperative recurrence and decide the management plan in patients with hepatocellular carcinoma (HCC). In this study, using a multiscale approach (5-, 10-, 15-mm), we investigated the biological optimal scale of radiomics features for gadoxetic acid-enhanced MRI and determined the 5-mm peritumoral region as the most relevant, developing an effective clinical-grade fusion model that can preoperatively predict MVI. Methods: We included 147 patients with histopathologically confirmed HCC, of whom 21 (14.3%) were MVI-positive. Volumes of interest were delineated at the primary lesion and peritumoral regions (5-, 10-, 15-mm) with exclusion of areas containing the tumor. Radiomic features were extracted from seven sequences to generate 28 feature sets for each patient. To identify stable radiomic features, we used LASSO (least absolute shrinkage and selection operator) regression analysis and calculated radiomics scores for each phase. Multivariable analysis was performed to identify significant imaging and clinical factors. Seven different models were developed. By applying receiver operating characteristic (ROC) curves, calibration plots, and decision curve analysis (DCA), we evaluated the model's performance, with internal validation conducted through bootstrapping to compensate for the limitations of the small sample size. Results: The radiomics and imaging feature fusion model showed good discrimination ability and achieved an area under the curve (AUC) of 0.957 (95% CI: 0.906-1) in the training cohort and 0.937 (95% CI: 0.863-1) in the validation cohort. Bootstrapping validated the stability of the fusion model with an AUC value of 0.951 (95% CI: 0.909-0.982). The 5-mm peritumoral radiomics model had excellent predictive power for MVI with higher accuracy compared to the 10-mm peritumoral radiomic model (81.4% vs. 69.8%, Conclusion: The construction of a radiomic signature based on a 5-mm peritumoral region on MRI is optimal to preoperatively predict MVI in HCC. Our outcomes demonstrate that an integrated model incorporating radiomic features associated with the tumor microenvironment at this new biological scale can be a valuable tool for preoperative risk assessment.