Reliable and Interpretable Visual Field Progression Prediction with Diffusion Models and Conformal Risk Control.

Area of Science:

• Ophthalmology
• Artificial Intelligence
• Medical Image Analysis

Background:

• Accurate visual field progression prediction is crucial for glaucoma management.
• Current methods lack predictive accuracy and reliable uncertainty quantification.
• Interpretable forecasting is needed for personalized treatment strategies.

Purpose of the Study:

• To develop a robust and interpretable framework for forecasting visual field deterioration in glaucoma.
• To enhance predictive accuracy and uncertainty quantification in visual field progression.
• To provide clinicians with interpretable insights into potential disease progression patterns.

Main Methods:

• Utilized diffusion models to predict future visual fields based on historical patient data.
• Implemented a novel archetypal-based conformal risk control method for trustworthy predictions.
• Ensured finite-sample coverage guarantees on intervals of archetypal contributions.

Main Results:

• The framework achieved target archetypal contribution coverage.
• Provided tighter prediction intervals compared to existing baseline methods.
• Visualizations demonstrated interpretable insights into disease progression patterns and uncertainty.

Conclusions:

• The combined approach of diffusion models and conformal methods enhances AI-assisted visual field forecasting reliability.
• The framework offers robust and interpretable forecasts, supporting improved clinical decision-making in glaucoma.
• This method aids in understanding and managing visual field deterioration by quantifying uncertainty effectively.