Advancing Head and Neck Cancer Survival Prediction via Multi-Label Learning and Deep Model Interpretation

Area of Science:

• Oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Accurate survival prediction is crucial for personalized Head and Neck Cancer (HNC) management following curative Radiation Therapy (RT).
• Existing models often lack interpretability and the ability to predict multiple survival outcomes simultaneously.
• Developing reliable prognostic tools is essential for optimizing patient treatment strategies.

Approach:

• Proposes IMLSP (Interpretable Multi-Label multi-modal deep Survival Prediction), a novel framework for simultaneous prediction of multiple HNC survival outcomes.
• Utilizes Multi-Task Logistic Regression (MTLR) layers to transform survival prediction into a multi-time point classification task.
• Introduces Grad-Team, a Gradient-weighted Time-event activation mapping technique for visual explanation of deep survival predictions.

Key Points:

• IMLSP outperforms single-modal and single-label models on the RADCURE HNC dataset for all survival outcomes.
• Visual explanations (activation maps) reveal the model's focus on tumor and nodal volumes, varying between high- and low-risk patients.
• Multi-label learning enhances prognostic performance and learning efficiency.

Conclusions:

• The interpretable multi-label survival prediction model shows promise for understanding AI decision-making in oncology.
• This approach facilitates personalized treatment planning for Head and Neck Cancer patients undergoing Radiation Therapy.
• The framework offers a pathway toward more transparent and effective AI-driven cancer care.