Issues associated with deploying CNN transfer learning to detect COVID-19 from chest X-rays

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• COVID-19, originating in December 2019, rapidly became a global pandemic.
• Machine learning on radiography images offers potential decision support for radiologists.
• Radiography analysis is crucial for rapid disease diagnosis and management.

Purpose of the Study:

• To critically analyze the applicability of CNNs for COVID-19 detection in chest X-rays.
• To investigate issues associated with using CNNs directly on whole chest X-ray images.
• To evaluate the interpretability of CNN decisions using Class Activation Maps (CAMs).

Main Methods:

• Utilized 12 pre-trained CNN architectures in transfer learning mode.
• Trained a shallow CNN architecture from scratch.
• Employed 3 publicly available chest X-ray datasets.
• Performed qualitative investigation using CAMs to visualize CNN decision-making regions.

Main Results:

• CNNs achieved high classification accuracy in detecting COVID-19 from chest X-rays.
• CAMs visualized the specific image regions influencing CNN predictions.
• Identified potential discrepancies between CNN focus and clinically relevant areas.

Conclusions:

• CNNs can aid in COVID-19 detection from chest X-rays.
• Clinical validation of CNN-identified regions is essential before accepting predictions.
• Interpretability methods like CAMs are crucial for understanding AI diagnostic reasoning.