Deep Learning Algorithm Based on Contrast-Enhanced Ultrasound Potentially Optimizes Treatment Strategies for Solitary...

We aimed to propose a prognostic framework using a dual-branch Vision Transformer (ViT) deep learning (DL) architecture for stratifying recurrence risk in solitary primary hepatocellular carcinoma (HCC) patients undergoing surgical resection (SR) or thermal ablation (TA) and explore its potential value as a tool for aiding therapy-related discussions.

This two-center study enrolled 575 HCC patients with preoperative contrast-enhanced ultrasound (CEUS) images between January 2017 and October 2022. Among them, 365 underwent SR and 210 underwent TA. We combined ViT-extracted CEUS features with clinical features to separately establish models (ViT-SR and ViT-TA) for recurrence-free survival (RFS) stratification across SR and TA cohorts. Finally, Risk reclassification was further performed by applying the baseline-balanced patients to the alternative treatment model to investigate the potential for treatment optimization.

The models exhibited favorable performance, achieving C-Index, AUC, precision, and F1 score of 0.73, 0.76, 0.86, and 0.70 in validation cohorts for ViT-SR, respectively, while corresponding values for ViT-TA were 0.76, 0.83, 0.76, and 0.76. Time-dependent AUC values demonstrated consistent stability across 1-/3-/5-y (ViT-SR: 0.76, 0.77, 0.76; ViT-TA: 0.79, 0.84, 0.82), with both models showing satisfactory calibration in the validation cohort. Treatment strategy conversion analysis indicated that 11.6% (17/146) of SR patients and 4.8% (7/146) of TA patients had a potentially more favorable prognostic profile under the alternative treatment.

ViT-based DL models that integrated multi-phase CEUS images and clinical variables facilitated non-invasive prediction of HCC recurrence probability, and may aid discussions on personalized clinical treatment strategies by simulating treatment conversion.