Evaluating deep learning predictions for COVID-19 from X-ray images using leave-one-out predictive densities

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Epidemiology

Background:

• Early COVID-19 detection is crucial for pandemic control.
• Chest X-rays offer accessible imaging but require expert interpretation for severity assessment.
• Existing deep learning models for COVID-19 detection lack sufficient external validation and struggle with data scarcity and bias.

Purpose of the Study:

• To address the limitations of current deep learning models in COVID-19 detection.
• To incorporate model uncertainty into automated COVID-19 detection using chest X-rays.
• To evaluate the effectiveness of stochastic gradient Langevin dynamics (SGLD) in improving model reliability.

Main Methods:

• Four deep learning architectures were trained using stochastic gradient Langevin dynamics (SGLD).
• Models were compared against baselines trained with standard stochastic gradient descent.
• Model uncertainties were assessed using convergence properties and leave-one-out predictive densities.

Main Results:

• The SGLD approach successfully reduced overconfidence in baseline estimators.
• Predictive accuracy was maintained for the best-performing models.
• The study provides a method to better evaluate model predictions in the context of limited data.

Conclusions:

• Stochastic gradient Langevin dynamics (SGLD) offers a promising approach to enhance the reliability of deep learning models for COVID-19 detection.
• Addressing model uncertainty is critical for the external validation and deployment of AI in medical diagnostics.
• This work contributes to more robust and trustworthy AI-powered tools for infectious disease surveillance.