Deep learning-based respiratory muscle segmentation as a potential imaging biomarker for respiratory function...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Pulmonology

Background:

• Respiratory diseases are a major cause of global mortality.
• Quantitative analysis of respiratory muscles is vital for assessing respiratory system status and patient quality of life.

Purpose of the Study:

• To develop an automated artificial intelligence (AI) approach for segmenting and classifying three types of respiratory muscles from computed tomography (CT) images.
• To correlate segmented respiratory muscle volumes and densities with pulmonary function test (PFT) parameters.

Main Methods:

• Utilized a large dataset of approximately 600,000 thoracic CT images from 3,200 individuals.
• Employed the Attention U-Net architecture for detailed and focused segmentation of muscle tissue and specific respiratory muscles.
• Calculated muscle volumes and densities from AI-segmented masks and performed correlation analysis with PFT parameters.

Main Results:

• Achieved high dice scores for segmentation models (0.9823 for muscle tissue, 0.9688 for respiratory muscles) and classification (0.9900 generalized dice score).
• Demonstrated high F1 scores for classifying pectoralis (0.9793), erector spinae (0.9975), and intercostal muscles (0.9839).
• Found a strong correlation between respiratory muscle volume and PFT parameters.

Conclusions:

• Respiratory muscle volume, accurately quantified by the AI model, shows potential as a novel biomarker for respiratory function.
• While muscle density showed a weaker correlation, it may hold significance in future medical research.
• The developed AI approach offers an automated and accurate method for respiratory muscle analysis in CT imaging.