Interpreting Uncertainty in Model Predictions for Covid-19 Diagnosis

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• COVID-19 diagnosis requires rapid tools beyond lab testing.
• Chest X-rays reveal lung changes indicative of COVID-19, detectable by deep neural networks.
• Traditional networks lack uncertainty quantification, limiting diagnostic reliability.

Purpose of the Study:

• To develop a visualization framework for interpreting uncertainty in AI-driven COVID-19 detection from chest X-rays.
• To quantify and decompose predictive uncertainty into model and data components.
• To enhance radiologist trust and understanding of AI diagnostic assistance.

Main Methods:

• Utilized Bayesian Convolutional Neural Networks for uncertainty computation in COVID-19 predictions.
• Developed a novel visualization framework to interpret uncertainty and its sources.
• Focused on understanding feature contributions to predictive uncertainty in chest X-ray analysis.

Main Results:

• The framework visualizes uncertainty components, aiding interpretability.
• Demonstrated the tool's utility in assisting radiologists with chest X-ray interpretation.
• Showcased how feature importance relates to predictive uncertainty.

Conclusions:

• The developed framework enhances the reliability and interpretability of AI diagnostic tools for COVID-19.
• This approach aids radiologists by clarifying AI predictions and highlighting significant diagnostic regions.
• Bayesian deep learning combined with visualization offers a promising path for assistive diagnostic technologies.