Related Experiment Video
Updated: Sep 26, 2026
Automated Joint Space Detection Improves Bone Segmentation Accuracy
Published on: November 28, 2025
Automated framework for volumetric musculoskeletal segmentation and assessment of torso CT images in large-scale
Sanaa Amina Gourine1, Mazen Soufi2,3, Yoshito Otake4
1Division of Information Science, Nara Institute of Science and Technology, Ikoma, Nara, 630-0192, Japan. gourine.sanaa_amina.gt9@is.naist.jp.
Purpose:
Automated assessment of musculoskeletal (MSK) structures in large-scale torso CT databases remains challenging due to high anatomical complexity and imaging variability. This study proposes an automated framework for the volumetric segmentation and quantitative analysis of 19 torso muscles and two fat structures, validating the integration of predictive uncertainty as a quality control tool for large-scale analysis.
Methods:
We validated two model architectures incorporating structure-wise predictive uncertainty: a 2D Bayesian UNet (2D BUNet) using Monte-Carlo dropout sampling and a 3D patch-based nnUNet utilizing entropy of prediction probabilities, named 3D MSKSegmenter (3DMS). We additionally compared the accuracy against TotalSegmentator (TS), a state-of-the-art public whole body segmentation model. The predictive uncertainty was employed as a quality control tool to filter reliable cases for downstream feature analysis, and the investigation of structure volume and density [mean Hounsfield Units (HU)] associations with age and sex across a large-scale database.
Results:
The segmentation accuracy was validated via fivefold cross-validation on 20 manually labeled CT images. 3DMS achieved a superior average Dice coefficient (DC) of 0.89 ± 0.03 compared with 2D BUNet and TS (p < 0.001). Strong negative correlations between the predictive uncertainty and DC (- 0.870 for 3DMS and - 0.955 for 2D BUNet) were observed. Furthermore, an external test on a single-institutional database of approximately 1200 patients confirmed the robustness of our framework for large-scale analysis.
Conclusions:
We developed and validated an automated framework for volumetric segmentation and analysis of 21 torso structures. By integrating the predictive uncertainty, our method ensures reliable biomarker extraction in large-scale studies. Our findings revealed significant age- and sex-related trends in muscle density and volume, demonstrating the framework's potential as a scalable solution for automated MSK assessment in CT images.
