Related Experiment Video
Updated: Sep 28, 2025
Segmentation and Linear Measurement for Body Composition Analysis using Slice-O-Matic and Horos
Published on: March 21, 2021
Deep learning for automatic segmentation of paraspinal muscle on computed tomography
Ning Yao1, Xintong Li1, Ling Wang1
1Department of Radiology, 66526Beijing Jishuitan Hospital, Beijing, PR China.
Background:
Muscle quantification is an essential step in sarcopenia evaluation.
Purpose:
To develop and evaluate an automated machine learning (ML) algorithm for segmenting the paraspinous muscles on either abdominal or lumbar (L) computed tomography (CT) scans.
Material And Methods:
A novel deep neural network algorithm for automated segmentation of paraspinous muscle was developed, CT scans of 504 consecutive patients conducted between January 2019 and February 2020 were assembled. The muscle was manually segmented at L3 vertebra level by three radiologists as ground truth, divided into training and testing subgroups. Muscle cross-sectional area (CSA) was recorded. Dice similarity coefficients (DSCs) and CSA errors were calculated to evaluate system performance. The degree of muscle fat infiltration (MFI) recording by percentage value was the fat area within the region of interest divided by the muscle area. An analysis of the factors influencing the performance of the V-net-based segmentation system was also implemented.
Results:
The mean DSCs for paraspinous muscles were high for both the training (0.963, 0.970, 0.941, and 0.968, respectively) and testing (0.950, 0.960, 0.929, and 0.961, respectively) datasets, while the CSA errors were low for both training (1.9%, 1.6%, 3.1%, and 1.3%, respectively) and testing (3.4%, 3.0%, 4.6%, and 1.9%, respectively) datasets. MFI and muscle area index (MI) were major factors affecting DSCs of the posterior paraspinous and paraspinous muscle groups.
Conclusion:
The ML algorithm for the measurement of paraspinous muscles was compared favorably to manual ground truth measurements.
