A convolutional neural network with pixel-wise sparse graph reasoning for COVID-19 lesion segmentation in CT images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• The COVID-19 pandemic necessitates automated tools for analyzing lung CT scans.
• Current deep convolutional neural networks (DCNNs) struggle with segmenting infected regions due to limited receptive fields and global reasoning.
• Accurate segmentation is crucial for assessing disease severity and guiding treatment.

Purpose of the Study:

• To develop an advanced segmentation network for identifying COVID-19 infected areas in lung CT images.
• To introduce a novel pixel-wise sparse graph reasoning (PSGR) module to enhance global contextual understanding.
• To improve the accuracy and efficiency of automated COVID-19 detection in medical imaging.

Main Methods:

• Proposed a novel segmentation network incorporating a pixel-wise sparse graph reasoning (PSGR) module.
• The PSGR module constructs a graph from encoder features, sparsifies connections for uncertain pixels, and performs global reasoning.
• Integrated the PSGR module between the encoder and decoder to capture long-range dependencies.

Main Results:

• The PSGR module effectively captures long-range dependencies within CT images.
• The proposed segmentation network accurately identifies COVID-19 infected regions.
• The model demonstrated superior performance compared to existing widely-used segmentation methods on three public datasets.

Conclusions:

• The novel PSGR module significantly enhances the global reasoning ability of segmentation networks.
• The developed algorithm provides accurate and effective segmentation of COVID-19 lung infections in CT scans.
• This approach offers a promising advancement for automated COVID-19 diagnosis and monitoring.