Contour-enhanced attention CNN for CT-based COVID-19 segmentation

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Accurate COVID-19 detection is crucial for pandemic control.
• Automated analysis of chest CT scans aids in clinical care for COVID-19.
• Existing methods may struggle with precise localization of infected lung tissues.

Purpose of the Study:

• To propose a Contour-aware Attention Decoder Convolutional Neural Network (CNN) for precise segmentation of COVID-19 infected tissues in chest CT scans.
• To introduce novel attention mechanisms for leveraging contour and shape cues to refine segmentation.
• To improve the capture of intricate morphological details and reconstruct high-resolution segmentation maps.

Main Methods:

• Development of a Contour-aware Attention Decoder CNN incorporating a novel attention scheme.
• Utilizing CT contour features to extract boundary and shape cues.
• Employing Cross Context Attention Fusion Upsampling for robust feature reconstruction.
• Evaluation on 3D CT scans from MosMedData and Jun Ma datasets.

Main Results:

• The proposed CNN achieved state-of-the-art performance in segmenting COVID-19 infected tissues.
• Achieved a Dice Similarity Coefficient (DSC) of 85.43%.
• Achieved a recall of 88.10%.

Conclusions:

• The Contour-aware Attention Decoder CNN effectively segments COVID-19 infected lung tissues.
• The novel attention mechanisms enhance the capture of morphological details, improving segmentation accuracy.
• This deep learning model shows significant promise for augmenting clinical diagnosis of COVID-19 from CT scans.