Construction of musculoskeletal quantitative model based on deep learning and study of musculoskeletal relationship in patients with osteoporosis.
Authors
Affiliations (5)
Affiliations (5)
- Department of Orthopedics, The Second Affiliated Hospital of the Naval Medical University of the Chinese People's Liberation Army, Shanghai 200003, China.
- Department of Orthopedics, The Second Affiliated Hospital of the Naval Medical University of the Chinese People's Liberation Army, Shanghai 200003, China; Department of Radiology, The 905th Hospital of the People's Liberation Army Navy, Shanghai 200052, China.
- UCSI University, Kuala Lumpur 56000, Malaysia.
- First People's Hospital of Lanzhou City, Lanzhou 730050, China.
- Department of Orthopedics, The Second Affiliated Hospital of the Naval Medical University of the Chinese People's Liberation Army, Shanghai 200003, China. Electronic address: [email protected].
Abstract
Lumbar CT and MRI scans are helpful for osteoporosis (OP) screening. Deep learning enhances the efficiency and accuracy of musculoskeletal quantification for OP screening. To develop a deep learning-based model (UNet3D) for musculoskeletal quantification from lumbar CT and MRI in OP patients, and to explore musculoskeletal relationship. Retrospectively analyzing OP patients' lumbar CT, MRI and DXA scans between January 2018 and August 2025 from two center, and UNet3D was developed based on U-Net to simultaneously segment L1-S1 vertebrae on CT and paraspinal muscles/intermuscular spaces on MRI. Performance was compared against U-Net, U<sup>2</sup>-Net, Mask-RCNN using pixel accuracy (PA), mean pixel accuracy (mPA), mean intersection over union (mIoU). Following automated HU values calculation within cancellous bone volume of interests (VOIs), we correlated L1-L4 HU values with T-scores and analyzed their relationship with muscle/space volumes across L1-S1, computing corresponding weighting coefficients. The study analyzed 410 OP patients (65 ± 10 years; 240 women) with 76,700 axial CT and 6030 axial T2 slices, including training (266), validation (76), internal test (38) and external test (30) sets. On the internal test set, UNet3D's accuracy in CT (PA/mPA/mIoU, 0.984/0.849/0.760) and MRI (PA/mPA/mIoU, 0.973/0.849/0.724), outperformed U-Net/U<sup>2</sup>-Net/Mask R-CNN. On the external test set, UNet3D attained PA, mPA, and mIoU values of 0.971, 0.852, 0.741 in CT, and of 0.969, 0.834, 0.713 in MRI. HU values varied across vertebral VOIs, ranging from 80.33 HU (L2) to 195.14 HU (S1), and correlations existed between L1-L4 HU values and T-scores (r ≥ 0.85, P ＜ 0.05). Volumetric measurements across L1-S1 identified the psoas major as the largest muscle (173,317.68 mm<sup>3</sup>) and the quadratus lumborum as the smallest (65,167.78 mm<sup>3</sup>); among spaces, the psoas-vertebral was largest (124,301.86 mm<sup>3</sup>) and the quadratus lumborum-iliocostalis was smallest (29,676.63 mm<sup>3</sup>). Vertebral HU values correlated positively with muscle volumes (ω > 0) and negatively with space volumes and fatty infiltration (ω < 0). UNet3D outperformed U-Net/U<sup>2</sup>-Net/Mask R-CNN on CT/MRI segmentation, enabling accurate musculoskeletal volume quantification in OP patients. Musculoskeletal correlations differed across spinal segments.