Prediction of microvascular invasion in hepatocellular carcinoma with expert-inspiration and skeleton sharing deep learning.

ABSTRACT

BACKGROUND AND AIMS: Radiological prediction of microvascular invasion (MVI) of hepatocellular carcinoma (HCC) is essential but few models were clinically implemented because of limited interpretability and generalizability. METHODS: Based on 2096 patients in three independent HCC cohorts, we established and validated an MVI predicting model. First, we used data from the primary cohort to train a 3D-ResNet network for MVI prediction and then optimised the model with "expert-inspired training" for model construction. Second, we implemented the model to the other two cohorts using three implementation strategies, the original model implementation, data sharing model implementation and skeleton sharing model implementation, the latter two of which used part of the cohorts' data for fine-tuning. The areas under the receiver operating characteristic curve (AUCs) were calculated to compare the performances of different models. RESULTS: For the MVI predicting model, the AUC of the expert-inspired model was 0.83 (95% CI 0.77-0.88) compared to 0.54 (95% CI 0.46-0.62) of model before expert-inspiring. Taking this model as an original model, AUC on the second cohort was 0.76 (95% CI 0.67-0.84). The AUC was improved to 0.83 (95% CI 0.77-0.90) with the data-sharing model, and further improved to 0.85 (95% CI 0.79-0.92) with the skeleton sharing model. The trend that the skeleton sharing model had an advantage in performance was similar in the third cohort. CONCLUSIONS: We established an expert-inspired model with better predictive performance and interpretability than the traditional constructed model. Skeleton sharing process is superior to data sharing and direct model implementation in model implementation.