CMM: A CNN-MLP Model for COVID-19 Lesion Segmentation and Severity Grading

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate segmentation and severity grading of COVID-19 lesions in CT images are crucial for diagnosis and treatment.
• Existing methods may struggle with precise lesion delineation and objective severity assessment.

Purpose of the Study:

• To propose a comprehensive Convolutional Neural Network-Multilayer Perceptron (CNN-MLP) model, termed CMM, for automated COVID-19 lesion segmentation and severity grading in CT images.
• To enhance the accuracy of lesion segmentation by incorporating shape priors and multi-scale feature learning.
• To develop a robust severity grading mechanism using image-derived features.

Main Methods:

• Lung segmentation using UNet, followed by lesion segmentation with a multi-scale deep supervised UNet (MDS-UNet) incorporating shape priors.
• Multi-scale inputs and deep supervision in MDS-UNet to preserve edge information and enhance feature learning.
• Severity grading using a Multi-Layer Perceptron (MLP) with features including weighted mean gray-scale value (WMG), lung area, and lesion area.
• Label refinement using the Frangi vessel filter to improve segmentation precision.

Main Results:

• The proposed CMM demonstrated high accuracy in both COVID-19 lesion segmentation and severity grading on public datasets.
• Integration of shape prior information effectively reduced the search space for segmentation outputs.
• The WMG feature effectively captured lesion appearance related to severity, improving MLP performance.

Conclusions:

• The CMM provides an effective and accurate solution for automated COVID-19 lesion segmentation and severity grading from CT images.
• The developed methods, including MDS-UNet and WMG feature, show significant potential for improving computer-aided diagnosis in respiratory diseases.
• The CMM model offers a promising tool for clinical application in managing COVID-19 patients.