Deep learning for automatic segmentation of paraspinal muscle on computed tomography

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Healthcare
• Biomedical Engineering

Background:

• Muscle quantification is crucial for diagnosing sarcopenia.
• Accurate measurement of paraspinous muscles is vital for sarcopenia assessment.
• Current manual segmentation methods can be time-consuming and subjective.

Purpose of the Study:

• To develop and validate an automated machine learning (ML) algorithm for paraspinous muscle segmentation.
• To assess the algorithm's performance on abdominal and lumbar computed tomography (CT) scans.
• To compare the ML algorithm's results against manual segmentation by radiologists.

Main Methods:

• A novel deep neural network (V-net) was developed for automated muscle segmentation.
• CT scans from 504 patients were used, with manual segmentation serving as ground truth.
• Performance was evaluated using Dice Similarity Coefficients (DSCs) and Cross-Sectional Area (CSA) errors.

Main Results:

• The ML algorithm achieved high mean DSCs (0.950-0.970) and low CSA errors (1.3%-4.6%) on test datasets.
• Muscle Fat Infiltration (MFI) and Muscle Area Index (MI) were identified as key factors influencing segmentation accuracy.
• The algorithm demonstrated robust performance across different patient scans.

Conclusions:

• The developed ML algorithm provides accurate and reliable paraspinous muscle segmentation.
• The automated approach favorably compares to manual measurements, offering efficiency gains.
• This technology has the potential to enhance sarcopenia evaluation and patient care.