A Transformer-Based Model for Preoperative Early Recurrence Prediction of Hepatocellular Carcinoma with Muti-modality MRI

Abstract

Hepatocellular carcinoma (HCC) is the most common primary liver cancer which accounts for a high mortality rate in clinical, and the most effective treatment for HCC is surgical resection. However, patients with HCC are still at a huge risk of recurrence after tumor resection. In this light, preoperative early recurrence prediction methods are necessary to guide physicians to develop an individualized preoperative treatment plan and postoperative follow-up, thus prolonging the survival time of patients. Nevertheless, existing methods based on clinical data neglect information on the image modality; existing methods based on radiomics are limited by the ability of its predefined features compared with deep learning methods; and existing methods based on CT scans are constrained by the inability to capture the details of images compared with MRI. With these observations, we propose a deep learning transformer-based model on multi-modality MRI to tackle the preoperative early recurrence prediction task of HCC. Enlightened by the vigorous capacity of context modeling of the transformer architecture, our proposed model exploits it to dig out the inter-modality correlations, and the performance significantly improves. Our experimental results reveal that our transformer-based model can achieve better performance than other state-of-the-art existing methods.

G. Zhan and F. Wang—First authors.