Explainable survival analysis with uncertainty using convolution-involved vision transformer.

Area of Science:

• Digital pathology
• Computational oncology
• Precision medicine

Background:

• Whole Slide Images (WSIs) are crucial for cancer diagnosis but pose computational challenges due to their large size.
• Existing models often use subsets of WSIs, potentially losing vital morphological information.
• Enhancing model explainability is critical for clinical adoption and trust in AI-driven predictions.

Purpose of the Study:

• To develop an explainable survival prediction model for cancer diagnosis using complete WSIs.
• To address the computational challenges associated with large-scale WSI processing.
• To improve the accuracy and interpretability of AI-based survival predictions.

Main Methods:

• A novel explainable survival prediction model based on Vision Transformer architecture.
• Utilization of dual-channel convolutional layers to process entire WSIs.
• Incorporation of aleatoric uncertainty to quantify model limitations.
• Development of a post-hoc method for identifying salient image features and patches.

Main Results:

• The proposed model effectively processes complete WSIs, overcoming computational limitations.
• The model demonstrates improved accuracy in cancer survival prediction compared to existing methods.
• The explainability feature successfully identifies key image regions and morphological features supporting predictions.
• Evaluations on two large cancer datasets confirm the model's efficacy and enhanced interpretability.

Conclusions:

• The developed Vision Transformer model offers a computationally feasible and explainable approach for WSI-based cancer survival prediction.
• The model's ability to utilize complete WSIs and provide interpretable results enhances its potential for clinical application in precision medicine.
• Incorporating uncertainty estimation aids in reliable decision-making for cancer diagnosis and treatment planning.