Novel coronavirus pneumonia detection and segmentation based on the deep-learning method.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate segmentation of COVID-19 lesions in CT scans is challenging due to lesion variability and image quality.
• Manual analysis of numerous CT scans is time-consuming and prone to diagnostic errors.
• Existing methods struggle with low contrast and complex background tissues in CT images.

Purpose of the Study:

• To develop an advanced deep learning model for precise segmentation of COVID-19 lesions in CT scans.
• To improve the efficiency and accuracy of pneumonia diagnosis and quantitative analysis.
• To address the limitations of current segmentation techniques in medical imaging.

Main Methods:

• A novel deep learning approach using a U-net architecture was employed for lung segmentation.
• The residual attention U-shaped network (ResAU-Net), incorporating attention, residual, and sub-pixel convolution modules, was developed for lesion segmentation.
• Minimum circumscribed rectangle clipping was used to define regions of interest for analysis.

Main Results:

• The ResAU-Net model demonstrated state-of-the-art performance in segmenting pneumonia lesions on a dataset of 100 chest CT scans.
• Achieved high segmentation accuracy with mIoU of 73.40%±2.24% and Dice coefficient of 84.5%±2.46%.
• The model enables fast, real-time processing, significantly aiding in rapid diagnosis.

Conclusions:

• The developed ResAU-Net model effectively overcomes challenges in segmenting COVID-19 lesions, improving accuracy.
• Segmentation results provide valuable assistance for medical staff in diagnosing and quantifying infection severity.
• The model enhances the overall efficiency of pneumonia screening and diagnosis.