Background: A accurate preoperative evaluation of microvascular invasion (MVI) in hepatocellular carcinoma can help surgeons choose the appropriate treatment, while it still remains challenge for radiologists. Computer-aided diagnosis with deep learning technology may be one of the solutions to improve the prediction accuracy.
Methods: MRI with six different sequences were included in this experiment. A deep neural network was used to segment hepatocellular carcinoma after a cross-sequence registration preprocess. Radiomics features together with clinical features were used to form the final prediction model. The clinical features used in the final model was chosen by univariate analysis.
Results: In this work, we collected MRI of 420 cases of hepatocellular carcinomas, in which 140 were MVI cases and 280 patients in the non-MVI group. The radiomics features showed good performances in predicting MVI. The radiomics features showed good performance in predicting MVI. The multi-sequences fusion radiomics features get the best AUC (0.794 ± 0.033). Results for sizes in 3-5cm group showed that the AUC was 0.860±0.065.
Conclusions: A fully automatic system is proposed to predict MVI using preoperative MRI and with the fusion between radiomics and the clinical features, the system shows good performance in prediction MVI, especially in 3-5 cm tumor group.