Automated analysis of paraspinal muscles: segmentation and multi-parameter quantification in lumbar CT using...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Manual segmentation of lumbar paraspinal muscles is laborious and inconsistent.
• Current methods limit comprehensive analysis and large-scale studies.

Purpose of the Study:

• Develop a deep learning algorithm for automatic segmentation of eight lumbar paraspinal muscles.
• Enable multi-parameter quantification of these muscles using CT images.

Main Methods:

• Collected CT scans (L1 to S1 vertebrae) and partitioned into training, validation, and test sets.
• Employed six convolutional neural networks for automatic segmentation of psoas major, quadratus lumborum, erector spinae, and multifidus.
• Evaluated models using Dice Similarity Coefficient (DSC), Hausdorff Distance (HD), and mean Intersection over Union (mIoU). Calculated muscle cross-sectional area, volume, fat infiltration, CT density, and paraspinal muscle index.

Main Results:

• TransUNet achieved the highest overall DSC (0.903) and mIoU (0.841).
• Individual muscle segmentation showed high accuracy, with the right multifidus reaching a DSC of 0.936.
• Mean Intraclass Correlation Coefficient (ICC) for all quantified parameters was 0.931, indicating strong agreement with manual segmentation.

Conclusions:

• The developed deep learning tool provides accurate and automatic segmentation of lumbar paraspinal muscles.
• Quantification of muscle parameters using this tool shows high agreement with manual measurements.
• This facilitates large-scale epidemiological studies on paraspinal muscles and spine-related diseases.