CTUSurv: A Cell-Aware Transformer-Based Network With Uncertainty for Survival Prediction Using Whole Slide Images.

Area of Science:

• Computational pathology
• Artificial intelligence in medicine
• Biomedical image analysis

Background:

• Deep learning for survival prediction using whole slide images (WSIs) faces challenges due to image complexity and noise.
• Existing models struggle with the intricate patterns and trustworthiness required for pathological diagnostics.

Purpose of the Study:

• To propose CTUSurv, a novel deep learning model for survival prediction.
• To enhance diagnostic capabilities by capturing cell-cell and cell-microenvironment interactions.
• To improve reliability through region-based uncertainty estimation.

Main Methods:

• CTUSurv employs a region sampling strategy for informative WSI region extraction.
• A cell-aware encoding module models interactions among biological entities.
• An aleatoric uncertainty estimation module provides region-level uncertainty scores.

Main Results:

• The proposed approach demonstrates superior predictive accuracy across four datasets.
• CTUSurv shows enhanced reliability in survival predictions.
• The model effectively captures complex biological patterns and interactions.

Conclusions:

• CTUSurv offers a robust solution for image-based survival prediction in digital pathology.
• The integration of uncertainty estimation improves the trustworthiness of deep learning models for WSIs.
• This work advances the application of AI in augmenting pathological diagnostics.