Abstract
Background: Accurate preoperative evaluation of microvascular invasion (MVI) in hepatocellular carcinoma (HCC) is crucial for surgeons to make informed decisions regarding appropriate treatment strategies. However, it continues to pose a significant challenge for radiologists. The integration of computer-aided diagnosis utilizing deep learning technology emerges as a promising approach to enhance the prediction accuracy. Methods: This experiment incorporated magnetic resonance imaging (MRI) scans with six different sequences. After a cross-sequence registration preprocess, a deep neural network was employed for the segmentation of hepatocellular carcinoma. The final prediction model was constructed by combining radiomics features with clinical features. The selection of clinical features for the final model was determined through univariate analysis. Results: In this study, we analyzed MRI scans obtained from a cohort of 420 patients diagnosed with HCC. Among them, 140 cases exhibited MVI, while the remaining 280 cases comprised the non-MVI group. The radiomics features demonstrated strong predictive capability for MVI. By extracting radiomic features from each MRI sequence and subsequently integrating them, we achieved the highest area under the curve (AUC) value of 0.794±0.033. Specifically, for tumor sizes ranging from 3 to 5 cm, the AUC reached 0.860±0.065. Conclusions: In this study, we present a fully automatic system for predicting MVI in HCC based on preoperative MRI. Our approach leverages the fusion of radiomics and clinical features to achieve accurate MVI prediction. The system demonstrates robust performance in predicting MVI, particularly in the 3-5 cm tumor group.