Proposing a novel deep network for detecting COVID-19 based on chest images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Epidemiology

Background:

• Coronavirus outbreaks necessitate rapid and accurate diagnostic tools.
• Existing deep neural network models for chest imaging lack sufficient reliability and sensitivity.
• Insufficient diagnostic infrastructure hinders effective global response.

Purpose of the Study:

• To develop a highly sensitive deep neural network model for coronavirus detection using CT-scan images.
• To create a large-scale, publicly available dataset of chest CT scans for research.
• To extend the model's applicability to chest X-ray (CXR) images via transfer learning.

Main Methods:

• Creation of a dataset with over 13,000 chest CT scans from patients with suspected coronavirus.
• Development of a deep neural network incorporating a pixel-wise attention layer for enhanced feature extraction.
• Application of transfer learning to adapt the model for chest X-ray analysis.

Main Results:

• The proposed model achieved an Area Under the Curve (AUC) of 0.886 on CT scans, outperforming competitors (AUC 0.843).
• On a separate benchmark, the model demonstrated an AUC of 0.899.
• Achieved a sensitivity of 0.858, significantly higher than the competitor's 0.81, highlighting the effectiveness of the attention mechanism.

Conclusions:

• The developed deep learning model shows superior performance in detecting coronavirus from chest imaging.
• The pixel-wise attention strategy is crucial for enhancing model sensitivity and accuracy.
• The model and its extension offer a reliable, efficient tool to aid clinicians in coronavirus diagnosis.