COVID-19 Chest CT Image Segmentation Network by Multi-Scale Fusion and Enhancement Operations

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• COVID-19 spread necessitates rapid diagnostic tools, with CT scans serving as a vital alternative to RT-PCR.
• Manual segmentation of COVID-19 infections in CT images is time-consuming and challenging due to diverse imaging characteristics.
• Existing medical image segmentation methods struggle with the accuracy required for COVID-19 detection.

Purpose of the Study:

• To develop a novel deep convolutional neural network (CNN) for accurate and automatic segmentation of COVID-19 infections in chest CT images.
• To address the limitations of current segmentation techniques in handling diverse and subtle COVID-19 imaging features.
• To create a robust automated system for identifying COVID-19 on CT scans, aiding in clinical decision-making.

Main Methods:

• A large dataset of 165,667 annotated chest CT images from 861 COVID-19 patients was curated.
• A novel deep CNN incorporating a feature variation (FV) block was designed to adaptively adjust global feature properties.
• Progressive Atrous Spatial Pyramid Pooling was introduced to fuse multi-scale features for improved segmentation of complex infection areas.

Main Results:

• The proposed deep CNN achieved state-of-the-art performance in segmenting chest CT images.
• High Dice similarity coefficients were obtained: 0.987 for lung segmentation and 0.726 for COVID-19 infection segmentation.
• The method demonstrated effective performance across datasets collected in China and Germany, indicating generalizability.

Conclusions:

• The developed deep CNN significantly enhances the ability to segment COVID-19 infections from chest CT scans.
• The proposed FV block and multi-scale feature fusion effectively address challenges posed by diverse infection appearances.
• This automated segmentation contributes to improved COVID-19 diagnosis and management strategies.