Validating deep learning inference during chest X-ray classification for COVID-19 screening.

Area of Science:

• Medical imaging
• Artificial intelligence
• Infectious disease diagnostics

Background:

• The COVID-19 pandemic highlighted the need for rapid diagnostic tools.
• Chest X-ray (CXR) imaging became crucial for early COVID-19 diagnosis and treatment planning.
• Deep learning methods for lung screening emerged quickly, but quality assurance lagged.

Purpose of the Study:

• To propose protocols for validating deep learning algorithms used in medical imaging.
• To introduce the ROI Hide-and-Seek protocol for assessing classification performance.
• To evaluate the correlation between anomaly detection and radiological signatures in CXR analysis.

Main Methods:

• Development of a novel validation protocol, ROI Hide-and-Seek.
• Systematic testing of deep learning algorithms on public CXR datasets.
• Assessment of classification performance and anomaly detection capabilities.

Main Results:

• Current deep learning techniques show weaknesses in CXR analysis for COVID-19.
• The proposed protocol effectively assesses classification performance and region-of-interest sensitivity.
• Demonstrated limitations of automated radiography with heterogeneous data sources.

Conclusions:

• Validated protocols are essential for reliable deep learning in medical imaging.
• The ROI Hide-and-Seek protocol offers a method to assess algorithm robustness.
• Further research is needed to address limitations in automated radiography analysis.