The paper proposes an improved preoperative model to predict microvascular invasion (MVI) in hepatocellular carcinoma (HCC) using an optimized MR radiomics and machine learning strategy. HCC was manually segmented from abdominal T1-weighted MR images in the arterial phase, creating tumor masks. 125 × 1692 radiomic features were extracted from these images. 125 × N features and 125 × 10 fused features were selected using an optimized 5-fold cross-validation strategy. They built the best model using random forest (RF) with LASSO and SPECTRAL-10. The model achieved an average accuracy of 0.7520 ± 0.0867, sensitivity of 0.7354 ± 0.1863, specificity of 0.6955 ± 0.2203, F1 score of 0.6943 ± 0.1437, and AUC of 0.7962 ± 0.1700. This approach eliminates the need for non-imaging information and subjective determination of the perioperative area.