Detection of COVID-19 Using Deep Learning Algorithms on Chest Radiographs

Area of Science:

• Radiology
• Artificial Intelligence
• Infectious Disease Diagnostics

Background:

• Chest radiographs (CXR) are crucial for diagnosing pneumonia.
• Accurate and rapid detection of COVID-19 is essential for patient management and public health.
• Deep learning (DL) offers potential for automating and improving diagnostic accuracy in medical imaging.

Purpose of the Study:

• To evaluate the diagnostic performance of a deep learning (DL) algorithm for detecting COVID-19 on chest radiographs (CXR).
• To compare the DL algorithm's accuracy against that of experienced radiologists.

Main Methods:

• A retrospective study utilized a DL model trained on over 112,000 CXR images.
• The model was fine-tuned on 509 patients with confirmed COVID-19 status via RT-PCR.
• A test set of 248 individuals (72 positive, 176 negative for COVID-19) was used, with CXR reviewed by both the DL algorithm and three radiologists.

Main Results:

• The DL algorithm achieved an Area Under the Curve (AUC) of 0.81, with 85% sensitivity and 72% specificity for COVID-19 detection.
• The algorithm demonstrated superior performance compared to human readers (P<0.001), particularly in distinguishing COVID-19 from other pneumonias (AUC 0.87).
• Subgroup analysis showed good performance in patients with fever or respiratory symptoms (AUC 0.79).

Conclusions:

• The developed DL algorithm (COV19NET) shows potential as an effective tool for COVID-19 detection using chest radiographs.
• This AI-driven approach could significantly aid in patient triage, especially in resource-limited healthcare settings.
• The algorithm's performance suggests a valuable role in supporting clinical decision-making for suspected COVID-19 cases.