Densely connected attention network for diagnosing COVID-19 based on chest CT

Area of Science:

• Artificial Intelligence
• Medical Imaging
• Deep Learning

Background:

• Deep learning models require enhanced feature extraction for accurate COVID-19 diagnosis from chest CT images.
• The self-attention mechanism is crucial for improving deep learning model performance.

Purpose of the Study:

• To develop a densely connected attention network (DenseANet) for enhanced COVID-19 diagnosis using chest CT scans.
• To improve the feature extraction capabilities of deep learning models through novel attention mechanisms.

Main Methods:

• A DenseANet was constructed, densely connecting attention features within and between blocks of the same scale, and across different scales.
• The model utilizes spatial attention features from different layers, densely connecting them to deeper layers.
• Improved U-Net segmented lung fields were used as input for the DenseANet, which outputs COVID-19 probability.

Main Results:

• DenseANet maximizes the utilization of self-attention features across different model depths.
• Experiments on 2993 CT scans demonstrated effective localization of lung lesions in SARS-CoV-2 infected patients.
• The model achieved 96.06% accuracy and 0.989 AUC in distinguishing COVID-19, common pneumonia, and normal controls.

Conclusions:

• The proposed DenseANet generates strong attention features, leading to superior diagnostic performance for COVID-19.
• The method of densely connecting attention features is adaptable to other deep learning models, enhancing their capabilities in related medical imaging tasks.