MVI-TR: A Transformer-Based Deep Learning Model with Contrast-Enhanced CT for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma

SIMPLE SUMMARY: For early-stage hepatocellular carcinoma (HCC) (size ≤ 5 cm), the prediction of microvascular invasion (MVI) before operation is important for the therapeutic strategy. This study aimed to construct deep learning (DL) models based only on the venous phase (VP) of contrast-enhanced computed tomography (CECT), and to evaluate the performance of these models for preoperative prediction of MVI. A novel transformer-based end-to-end DL model is proposed for the first time, named MVI-TR, to capture features automatically from radiomics and to perform MVI preoperative assessments. For patient cohorts, it achieved superior outcomes in six performance measures of MVI predication status: accuracy, precision, receiver operating characteristic (ROC), area under the curve (AUC), recalling rate, and F1-score. ABSTRACT: In this study, we considered preoperative prediction of microvascular invasion (MVI) status with deep learning (DL) models for patients with early-stage hepatocellular carcinoma (HCC) (tumor size ≤ 5 cm). Two types of DL models based only on venous phase (VP) of contrast-enhanced computed tomography (CECT) were constructed and validated. From our hospital (First Affiliated Hospital of Zhejiang University, Zhejiang, P.R. China), 559 patients, who had histopathological confirmed MVI status, participated in this study. All preoperative CECT were collected, and the patients were randomly divided into training and validation cohorts at a ratio of 4:1. We proposed a novel transformer-based end-to-end DL model, named MVI-TR, which is a supervised learning method. MVI-TR can capture features automatically from radiomics and perform MVI preoperative assessments. In addition, a popular self-supervised learning method, the contrastive learning model, and the widely used residual networks (ResNets family) were constructed for fair comparisons. With an accuracy of 99.1%, a precision of 99.3%, an area under the curve (AUC) of 0.98, a recalling rate of 98.8%, and an F1-score of 99.1% in the training cohort, MVI-TR achieved superior outcomes. Additionally, the validation cohort’s MVI status prediction had the best accuracy (97.2%), precision (97.3%), AUC (0.935), recalling rate (93.1%), and F1-score (95.2%). MVI-TR outperformed other models for predicting MVI status, and showed great preoperative predictive value for early-stage HCC patients.