The Pitfalls of Using Open Data to Develop Deep Learning Solutions for COVID-19 Detection in Chest X-Rays

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Radiology

Background:

• Deep learning models have been developed for COVID-19 identification from chest X-rays.
• The AI community often uses public datasets due to limited hospital data access.
• Prior AI models for pneumonia detection showed promising results.

Purpose of the Study:

• To evaluate deep learning models trained on open-source data against external and hospital datasets.
• To classify chest X-rays into COVID-19, non-COVID pneumonia, and no-pneumonia categories.
• To assess model performance and explore image features using explainability modules.

Main Methods:

• Training impactful models on a widely used open-source dataset (COVIDx).
• Testing models on an external dataset and a hospital dataset.
• Evaluating classification performance using ROC curves, confusion matrices, and standard metrics.
• Implementing explainability modules for feature importance analysis.

Main Results:

• The COVIDx open-source dataset is not representative of the clinical problem, leading to inflated performance results.
• Models trained and tested solely on open-source data may be vulnerable to bias and confounding variables.
• Performance evaluation on real-world data is crucial for clinical viability.

Conclusions:

• Open-source datasets like COVIDx may not accurately reflect real-world clinical scenarios for COVID-19 detection.
• Over-reliance on specific open-source data can lead to biased and unreliable AI models.
• Careful validation on diverse, representative datasets is essential for developing clinically useful AI tools for chest X-ray analysis.