Prediction of microvascular invasion in hepatocellular carcinoma by combining deep attention mechanism with clinical features

Objective: Microvascular invasion (MVI) is an independent factor for postoperative recurrence of hepatocellular carcinoma (HCC). Accurate preoperative prediction of MVI grading is helpful for surgical planning in HCC management. We aimed to investigate the consistency and diagnostic performance of Magnetic resonance imaging(MRI) in assessing the presence of MVI, and the validity of deep learning attention mechanisms and clinical features in MVI grade prediction. Method: A total of 93 patients were selected from the Shunde Hospital affiliated to Southern Medical University in China. Retrospective image data and clinical data (n=93, collected between January,2017 and February,2020) were used to establish single sequence deep learning models and fusion models based on the EfficientNet and attention modules. Among them, the image data is enhanced by conventional MRI sequences (T1WI, T2WI, DWI), enhanced MRI sequences (AP, PP, EP, HBP) and synthesized MRI sequences (T1mapping-pre, T1mapping-20min). Furthermore, high-risk areas of hepatocellular carcinoma microvascular invasion were visualized by deep learning visualization techniques. Result: The fusion model based on T1mapping-20min sequence and clinical features outperforms other fusion models. Accuracy:0.8376; Sensitivity:0.8378; Specificity: 0.8702; AUC:0.8501. And deep fusion models can display MVI high-risk areas. Conclusion: Fusion model based on multiple MRI sequences and were successfully established. The effectiveness of deep learning algorithm was verified combined with attention mechanism and clinical features for MVI grading prediction. Therefore, the combination of deep attention mechanism and clinical features is an effective tool for preoperative prediction of MVI, which has advantages over using only deep features and radiomics.