Interpretable multimodal deep learning for time-resolved survival prediction after hepatocellular carcinoma resection

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Hepatocellular carcinoma (HCC) shows significant patient variability, leading to inconsistent outcomes despite similar staging.
• Current staging systems and predictive models offer limited accuracy and fail to capture dynamic survival risks.
• Genomic assays and pathology-based factors are not practical for timely clinical decisions.

Purpose of the Study:

• To develop and validate TEMPO-HCC, a multimodal deep survival model for predicting individualized overall survival risk trajectories in HCC patients post-resection.
• To move beyond static, binary predictions towards dynamic, time-varying risk assessment for improved clinical management.
• To create a model with hierarchical interpretability, linking imaging and histopathology to survival outcomes.

Main Methods:

• Curated a six-center cohort of 1475 HCC patients.
• Integrated multiphasic MRI, postoperative H&E whole-slide images, and perioperative predictors into a deep survival model.
• Employed a discrete-time survival head to generate 1, 2, 3, and 5-year survival probabilities.

Main Results:

• TEMPO-HCC demonstrated superior performance compared to unimodal models and existing staging systems.
• Achieved a C-index of 0.751 and strong time-dependent AUCs (0.836 at 12 months, 0.680 at 60 months) in external validation.
• Augmenting guideline staging with TEMPO-HCC enhanced predictive discrimination for personalized management.

Conclusions:

• TEMPO-HCC offers a paradigm shift towards actionable, temporal risk prediction in HCC.
• The model's interpretability provides a clear link between radiologic and histopathologic features and patient outcomes.
• This tool enables personalized postoperative surveillance and risk-adapted management strategies for HCC patients.