RECIST^Surv: Hybrid Multi-Task Transformer for Hepatocellular Carcinoma Response and Survival Evaluation

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Hepatocellular Carcinoma Research

Background:

• Transarterial Chemoembolization (TACE) is a key treatment for hepatocellular carcinoma (HCC) in non-resectable patients.
• Clinical outcomes following TACE are highly variable, necessitating better prediction tools.
• Accurate assessment of tumor response and long-term survival prediction remains a challenge.

Purpose of the Study:

• To introduce RECIST^Surv, a novel Transformer model for HCC treatment response and survival prediction.
• To develop an efficient strategy for analyzing longitudinal, multi-center imaging data.
• To enhance personalized treatment planning for HCC patients undergoing TACE.

Main Methods:

• Developed RECIST^Surv, a response-driven Transformer model integrating multi-task learning and co-attention.
• Utilized longitudinal Computed Tomography (CT) imaging data from multi-center HCC cohorts.
• Incorporated a response-driven co-attention mechanism to model pre- and post-TACE feature interactions.

Main Results:

• RECIST^Surv demonstrated superior prognostic precision compared to state-of-the-art methods (C-indexes 0.595-0.780).
• RECIST^Surv proved to be an independent prognostic factor when integrated with multi-modal data (HR 1.693-20.7, P=0.001-0.042).
• The model simultaneously performs image segmentation, predicts treatment response, and estimates overall survival.

Conclusions:

• RECIST^Surv offers a powerful, interpretable tool for personalized HCC treatment and outcome prediction.
• The model effectively captures complex relationships between imaging, treatment response, and survival.
• Publicly available code facilitates further research and clinical application.