Deep Learning Predicts Survival and Guides Therapy in Proliferative Hepatocellular Carcinoma

Overview

A multitask deep learning model was developed to non-invasively identify proliferative hepatocellular carcinoma (HCC) from contrast-enhanced CT scans and predict survival outcomes after intra-arterial therapies. The model demonstrated strong diagnostic accuracy and enabled personalized treatment selection between hepatic arterial infusion chemotherapy (HAIC) and transarterial chemoembolization (TACE), improving prognostic stratification in unresectable HCC patients.

Background

Hepatocellular carcinoma (HCC) is a heterogeneous malignancy with a proliferative subtype characterized by aggressive behavior and poor prognosis. This subtype is difficult to identify without invasive tissue sampling, limiting tailored treatment approaches. Intra-arterial therapies such as TACE and HAIC are commonly used for unresectable HCC, but their efficacy varies by tumor biology. Recent advances in deep learning offer potential for non-invasive phenotyping and survival prediction to guide therapy selection.

Data Highlights

Key Findings

• The deep learning model using nnUNet segmentation and Prototype Mamba Net architecture achieved high accuracy in detecting proliferative HCC from CT imaging.
• Prognostic nomograms combining radiomic and clinical data outperformed traditional staging systems in survival prediction (AUC 0.83–0.87 vs lower for traditional).
• Among low-risk patients, survival did not significantly differ between HAIC and TACE treatments.
• High-risk patients with proliferative HCC showed significantly improved survival with HAIC compared to TACE (P < 0.001).
• Proliferative HCC was associated with female sex, multiple lesions, and elevated AFP levels.
• Chronic hepatitis B virus infection was the predominant underlying etiology in the unresectable HCC cohort (90.2%).

Clinical Implications

This deep learning approach enables non-invasive identification of proliferative HCC, facilitating risk stratification and personalized intra-arterial therapy selection. Specifically, it supports preferential use of HAIC over TACE in high-risk proliferative HCC patients to improve survival outcomes. Incorporating such AI-driven tools into clinical workflows may optimize treatment decisions and prognostication in unresectable HCC.

Conclusion

The study demonstrates that a multitask deep learning model can accurately detect proliferative HCC and predict survival after intra-arterial therapies, enabling personalized treatment strategies. This approach holds promise for improving outcomes in patients with aggressive unresectable HCC.

References

1. Bao et al. 2021 -- CT-based predictors of proliferative hepatocellular carcinoma
2. Kang et al. 2020 -- MRI features linked to proliferative HCC phenotype
3. Vaswani et al. 2017 -- Transformer architecture for deep learning