Deep learning models for COVID-19 chest x-ray classification: Preventing shortcut learning using feature...

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Deep learning models for COVID-19 detection use chest radiographs (CXRs).
• Existing CXR datasets are small and heterogeneous, leading to shortcut learning and overfitting.
• Shortcut learning causes models to learn spurious features instead of true pulmonary characteristics.

Purpose of the Study:

• To introduce feature disentanglement as a method to improve deep learning model generalization for COVID-19 CXR analysis.
• To mitigate shortcut learning by forcing models to focus on relevant pulmonary features.

Main Methods:

• Implemented feature disentanglement during the training of deep learning models using CXR data.
• Feature disentanglement penalizes models for learning dataset-specific features.

Main Results:

• Models trained with feature disentanglement demonstrated improved generalization on unseen CXR data.
• The best model achieved a 0.13 increase in Area Under the Curve (AUC).
• This approach outperformed non-lung masking and histogram equalization for bias reduction.

Conclusions:

• Feature disentanglement is an effective technique to combat shortcut learning in COVID-19 CXR analysis.
• This method enhances model robustness and clinical utility for pandemic management.
• It offers a superior alternative to existing bias mitigation strategies for medical imaging AI.