Artificial Intelligence-assisted chest X-ray assessment scheme for COVID-19

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging

Background:

• Chest X-rays (CXRs) are crucial for diagnosing Coronavirus disease (COVID)-19.
• Radiologists face challenges in accurately interpreting CXRs, especially for subtle cases.
• Artificial intelligence (AI) offers potential to augment human diagnostic capabilities.

Purpose of the Study:

• To evaluate the effectiveness of a trained convolutional neural network (CNN) in assisting radiologists to differentiate COVID-19 positive from negative patients using CXRs.
• To identify patient subgroups where AI provides the most value.
• To analyze imaging features contributing to AI performance via visualization techniques.

Main Methods:

• An ambispective study involving 487 patients' CXRs classified by radiologists.
• AI analysis of CXRs initially classified as "normal" or "indeterminate" by radiologists.
• Comparison of radiologist performance alone versus AI-assisted performance against RT-PCR as the gold standard.
• Analysis of CNN attention maps to understand AI decision-making.

Main Results:

• Radiologist precision improved from 65.9% to 81.9% with AI assistance.
• Radiologist recall improved from 17.5% to 71.75% with AI assistance.
• AI achieved 92% accuracy in classifying "normal" CXRs for COVID-19 status.
• Attention maps highlighted the cardiac shadow in "normal" radiographs, suggesting subtle cardiac involvement.

Conclusions:

• AI algorithms can effectively complement human experts in diagnosing COVID-19 from CXRs.
• AI demonstrates high accuracy in classifying "normal" CXRs, potentially identifying subtle cardiac changes.
• Human-AI collaboration can lead to improved diagnostic accuracy for COVID-19 detection on CXRs.