Early prediction of hepatocellular carcinoma using a risk-embedded longitudinal attention model

Abstract

Hepatocellular carcinoma (HCC) frequently arises in patients with liver cirrhosis, and biannual ultrasound surveillance is a cost-effective strategy for its early detection. Longitudinal ultrasound images from routine follow-ups offer critical information for clinical HCC prediction, yet existing models struggle to capture their temporal dynamics. We present the risk embedded and longitudinal attention network (ReLANet), a deep learning framework that fuses diagnostic indicators of cirrhosis progression with cumulative risk data through a spatiotemporal architecture. By incorporating an age-dependent cumulative risk embedding and a longitudinal attention mechanism, ReLANet accommodates variable-length image sequences and dynamically evaluates their predictive value. In experiments on 6,170 samples from 619 cirrhosis patients, ReLANet achieved an area under the receiver operating characteristic curve of 80.2% (95% CI: 75.7%–84.4%), with 75.5% accuracy, 71.0% sensitivity, and 75.8% specificity, outperforming contemporary sequence models. These results demonstrate that ReLANet effectively integrates spatiotemporal and cumulative risk information from longitudinal ultrasound data, offering a state-of-the-art tool to enhance early HCC detection in at-risk populations.