Introduction
Osteoporosis (OP), a systemic metabolic bone disease defined by diminished bone mass and impaired microarchitecture, is clinically insidious and markedly elevates fracture susceptibility, severely compromising patient survival and quality of life [1], [2]. The intimate relationship between bone and muscle status is well-established in musculoskeletal science. Research has delineated their pathophysiological interplay during aging, driven by close anatomical adjacency and intensive chemical and metabolic crosstalk [3].
The coexistence of osteoporosis and sarcopenia leads to a higher incidence of falls, fractures, and mortality [4], [5]. Despite being the World Health Organization-recommended gold standard for diagnosing OP, dual-energy X-ray absorptiometry (DXA) remains underutilized in practice, leaving a significant portion of the high-risk population unscreened. Against the backdrop of global aging, leveraging routine clinical imaging for early, precise evaluation of musculoskeletal health is of paramount importance.
This method represents a cost-effective and low-radiation strategy, providing a practical means to screen high-risk individuals for OP and associated fractures [6], [7].
However, current research methodologies face significant limitations. In Hounsfield unit (HU) value-based bone mineral density (BMD) evaluation at CT, the reliance on manual region of interest (ROI) placement within a single or limited number of vertebrae leads to incomplete sampling [8]. In comparison to traditional ROI measurements, the HU value derived from a vertebral volume of interest (VOI) accounts for the spatial heterogeneity of BMD within the vertebra and demonstrates stronger correlations with both DXA and quantitative CT (QCT) results [9], [10].
However, its measurement still relies on tedious manual delineation. In MRI, the assessment of paravertebral muscles is also typically restricted to single-slice measurements at specific levels, which provides an incomplete representation of muscle status [11], [12]. The rapid advancement of artificial intelligence has endowed it with remarkable capabilities in medical image segmentation, offering high accuracy and reproducibility that overcome the limitations of manual methods [13], [14].
Yet, a systematic approach to concurrently extract 3D quantitative musculoskeletal information from routinely acquired CT and MRI scans remains lacking.
Here, our study aims to leverage widely available routine clinical images to develop a fully automated pipeline for the precise 3D volumetric quantification of both bone and muscle, enabling an integrated assessment in OP patients. UNet3D can simultaneously perform pixel-level semantic segmentation on lumbar spine CT and MRI images, and it harnesses a gradient-adaptive algorithm to achieve oscillatory separation of cortical and cancellous bone [15]. Additionally, we reconstructed the complete L1-S1 spatial continuum of muscles and intermuscular spaces by parameterizing disc-level data and applying CT-MRI co-registration.
In summary, the model provides fully automated quantification of L1-S1 vertebral VOI HU values, and muscle/intermuscular space volumes, allowing for a holistic analysis of musculoskeletal relationships in OP patients.
Materials and methods
Ethics statement
This study was approved by our institutional review board and compliant with the Declaration of Helsinki. A waiver of consent was granted due to the retrospective nature of the study and implementation of strict data anonymization and de-identification protocols.
Data set preparation
This study was conducted using retrospectively collected data from two independent centers. The cohorts included OP patients, defined as a DXA BMD T-score of ≤ -2.5 at any measured site. The lumbar spine CT, MRI, and DXA scans were acquired from two centers, with data collection periods spanning January 2018 to January 2024 (Center 1) and January to August 2025 (Center 2), respectively.
Inclusion criteria required that patients had: (a) no prior anti-osteoporosis treatment; (b) available lumbar CT (with axial, sagittal, and coronal reconstructions), MRI (including sagittal and axial T2-weighted imaging), and DXA scans; and (c) no evidence of severe/multilevel lumbar spinal stenosis or disc herniation. Exclusion criteria were: (a) history of lumbosacral spine surgery; (b) severe degeneration and scoliosis of the lumbosacral spine; (c) vertebral fractures; (d) poor image quality or artifacts precluding analysis; and (e) other significant lumbar disorders (e.g., infections, tumors). The patient selection flowchart is summarized in Fig.
1.Fig. 1Flowchart diagram for the study. DXA: dual-energy x-ray absorptiometry.
All images conformed to the Digital Imaging and Communications in Medicine standard (DICOM). To uphold data integrity, any incomplete or corrupted datasets were identified and excluded prior to analysis to prevent bias. Dataset from Center 1 were randomly allocated into training (70%), internal validation (20%), and internal test (10%) sets.
The independent dataset from Center 2 was held-out as an external test set to provide an unbiased assessment of model performance.
Lumbar spine CT and MRI image acquisition and annotation
The regions of interest for this study comprised the vertebral bodies, paravertebral muscles, and intermuscular spaces across the L1 to S1 levels. Detailed scanning parameters for CT and MRI are provided in Table E1 (supplemental materials).
All lumbar spine CT and MRI images were annotated slice-by-slice by a research trainee under the supervision of expert radiologists, using Pair software (version 3.0; https://www.aipair.org). All annotations were subsequently reviewed and refined by a spine surgery specialist (Associate Chief Physician) with 20 years of clinical experience. To ensure accuracy and reliability, a final review was performed by a senior spine surgery specialist (Chief Physician) with 30 years of experience.
Representative annotated CT and MRI images are shown in Fig. 2.Fig. 2Annotation of Lumbar Spine CT and MRI in a 60-Year-Old Male with Osteoporosis.
The patient had a total lumbar DXA T-score of -2.3 (Segmental scores: L1: -2.3, L2: -2.5, L3: -2.0, L4: -1.8). (A) CT-based annotation of six vertebral bodies (L1 to S1). (a) Axial CT slice.
(b) 3D reconstruction from annotated data. (c) Sagittal and (d) Coronal views used for accurate spinal level identification. (B) Annotation of paraspinal muscles and intermuscular spaces on axial T2-weighted MRI.
Ten bilateral structures (5 muscles and 5 spaces) were demarcated at each of the five intervertebral disc levels (L1/2 to L5/S1), resulting in 20 categories per level and 100 labels in total. Sagittal MRI was referenced for precise level confirmation. (C) Close-up view of the annotated structures on three consecutive axial T2-weighted images at the L1/2 level.
UNet3D model development
We developed UNet3D model based on the U-Net [16] for musculoskeletal segmentation and quantification in OP patients. To address the computational burden of 3D processing while preserving critical anatomical details, two principal architectural modifications were implemented: (a) An innovative dimensional transformation module was incorporated during input processing, which performs two dual-convolution operations prior to downsampling. This hierarchically converts the input 2D data (16×1×512×512) into enriched 3D feature maps (64×512×512).
This strategy achieves significant computational efficiency without compromising essential spatial information. (b) The network's output was structured as a 2D map (num_class×512×512) with a softmax activation layer, producing final categorical masks. This design maintains dimensional alignment with clinical DICOM standards, facilitating seamless integration into clinical workflows.
These enhancements collectively lower computational costs via early-stage dimensionality reduction and improve the delineation of fine anatomical margins, maintaining high segmentation accuracy with superior training efficiency (Fig. 3).Fig. 3UNet3D model architecture.
The model input consists of 16 consecutive 512×512 grayscale images, forming a volume data with 512×512×1×16. Initially, this input undergoes two consecutive convolutions, transforming it into a feature map of size 64×512×512. Each downsampling level in the encoder includes two 3×3 convolutions (stride=1, padding=1) followed by a ReLU activation function for feature extraction, and is then downsampled via a 2×2 max-pooling operation (stride=2).
The channel count along the encoding path starts at 64 and doubles sequentially: 64, 128, 256, 512, and 1024 (the bottleneck layer). The decoding path progressively restores spatial resolution through upsampling. At each level, it incorporates skip-connections by concatenating features from the corresponding encoder layer, followed by two 3×3 convolutions and ReLU activation.
Finally, a 1×1 convolution layer maps the features to the target number of classes.
The UNet3D model was implemented within the PyTorch framework, with all training and evaluation conducted on an NVIDIA Tesla A5000 GPU. We utilized the Adam optimizer for model optimization, setting the learning rate to 0.001 and the batch size to 4. An early stopping strategy, which monitored the performance on the validation set, was adopted to prevent overfitting.
Furthermore, data augmentation techniques such as rotation, translation, and scaling were applied during the training phase to enhance the model's generalization capability and overall robustness.
Quantification at CT and MRI
HU Value Quantification in Vertebral Body VOI.—Subsequent to automated L1-S1 vertebral segmentation by the UNet3D model, trabecular bone HU values were quantified through a dedicated pipeline. The procedure involved: (a) converting pixel values to HU values via inward radial sampling from the segmented vertebral margins; (b) employing a gradient-adaptive algorithm to achieve oscillatory separation of cortical and cancellous bone [15]; and (c) computing the mean HU within the final cancellous VOI. A correlation analysis was then performed between these CT-derived HU values and the corresponding DXA T-scores at the L1-L4 levels.
Paravertebral Muscle and Intermuscular Space Volume Quantification.—A limitation of the original dataset was its restriction of paravertebral muscle and intermuscular space data to the intervertebral disc levels, precluding direct whole-volume measurement. Therefore, UNet3D was employed for 3D reconstruction. The process involved parameterizing the 2D cross-sectional anatomy of the muscles and spaces at each disc level.
These parameterized sections were then spatially integrated using a CT-MRI co-registration framework to establish the continuum of L1-S1. Finally, the UNet3D model automatically calculated the bilateral volumes of the segmented paravertebral muscles and intermuscular spaces for each lumbar segment and for the entire L1-S1 region. Automated optical inspection was used to quantify the fatty infiltration (FI) rate by identifying adipose tissue [17].
A multivariate analysis was performed to evaluate the correlations among key parameters: HU values in vertebral trabecular VOIs, paraspinal muscle volume, intermuscular space volume, and FI rate.
Statistical analysis
Model Performance Evaluation: The segmentation efficacy of UNet3D was quantified on the test set using pixel accuracy (PA), mean pixel accuracy (mPA), and mean intersection over union (mIoU). Comparative benchmarks against U-Net [16], U2-Net [18], and Mask R-CNN [19] were established.
Data Analysis and Correlation: All statistical computations were conducted in Python 3.10.10. A Pearson correlation analysis evaluated the relationship between vertebral HU values and DXA T-scores, with statistical significance defined as P < .05 (two-sided). The Shapiro-Wilk test was used to verify the normal distribution of the data.
P > .05 indicated normality.
Multivariate Modeling and Group Comparisons: All quantitative results were normalized via Z-score standardization. A multivariate linear regression model was constructed via least squares estimation. Regression coefficients were standardized and assigned weights (ω).
Results
Participant characteristics
Among 2,700 OP patients in Center 1, 2,320 were excluded: anti-osteoporosis treatment (490); incomplete/below-standard CT/MRI scans (1413); incomplete L1-L4 T-scores (92); vertebral fractures/surgery (205); severe lumbar scoliosis /degeneration or other lumbar pathologies (120). Among 500 OP patients in Center 2, 470 were excluded in the same way (Fig. 1).
The final cohort included 410 OP patients, yielding 2460 vertebral bodies (76700 CT slices) and 2050 intervertebral disc sections (6210 axial T2 MRI slices). The 410 OP patients (66 ± 10 years; range 55–75) included 240 women (64 ± 8 years; range 55–75) and 170 men (68 ± 10 years; range 55–75) (Table 1).Table 1Patient characteristics in the training, validation, internal test, and external test sets.ParameterTraining SetValidation SetInternal Test SetExternal Test SetNo. of participants266763830SexFemale156 (58.6)43 (56.6)21 (55.3)20 (66.7)Male110 (41.4)33 (43.4)17 (44.7)10 (33.3)Age (y)65 ± 1064 ± 1065 ± 1166 ± 9Body mass index (kg/m2)23.9 ± 3.3 (18.9, 30.8)23.1 ± 3.5 (18.5, 30.3)23.0 ± 3.2 (19.3, 30.5)23.7 ± 2.8 (18.2, 29.6)DXA total lumbar T-score−2.5 ± −1.3 (−3.7, −0.9)−2.4 ± −1.6 (−3.6, −1.0)−2.5 ± −1.1 (−3.8, −0.9)−2.4 ± −1.7 (−3.7, −1.2)L1−2.6 ± −1.4 (−3.9, −1.7)−2.5 ± −1.7 (−3.7, −1.4)−2.6 ± −1.3 (−3.9, −1.9)−2.5 ± −1.5 (−4.0, −1.7)L2−2.7 ± −1.5 (−3.9, −1.3)−2.6 ± −1.6 (−3.9, −1.2)−2.5 ± −1.7 (−3.7, −1.4)−2.6 ± −1.4 (−3.7, −1.4)L3−2.0 ± −1.6 (−2.9, −0.5)−1.9 ± −1.7 (−2.7, −0.6)−2.0 ± −1.3 (−2.7, −0.2)−1.9 ± −1.8 (−2.9, −0.3)L4−1.7. ± −1.3 (−2.8, −0.2)−1.8 ± −1.5 (−2.9, −0.3)−1.7 ± −1.5 (−2.7, −0.2)−1.8 ± −1.4 (−2.9, −0.6)DXA total hip T-score−2.1 ± −0.7 (−2.5, −1.1)−2.0 ± −1.9 (−3.0, −1.3)−2.1 ± −1.3 (−2.7., −0.9)−2.0 ± −1.5 (−3.0, −1.7)DXA femoral neck T-score−2.5 ± −0.9 (−3.5, −1.2)−2.5 ± −1.1 (−3.3, −0.9)−2.6 ± −1.3 (−3.4, −1.0)−2.6 ± −1.2 (−3.4, −1.2)Note.—Values reported as averages ± SD; data in parentheses were percentages or ranges. Osteoporosis was defined as any dual-energy x-ray absorptiometry (DXA) T-score of -2.5 or less.
UNet3D model performance
On lumbar spine CT images, UNet3D achieved PA, mPA, and mIoU of 0.984, 0.849, and 0.760, respectively, for L1–S1 vertebral segmentation in the internal test set, outperforming U-Net, U2-Net, and Mask R-CNN (Table 2). On the external test set, UNet3D attained PA, mPA, and mIoU values of 0.971 (95% CI [0.968, 0.973]), 0.852 (95% CI [0.849, 0.854]), and 0.741 (95% CI [0.740, 0.744]). As visually depicted in Fig.
4A, UNet3D delivers accurate and continuous segmentation across all target vertebrae.Table 2Segmentation performance of each model on lumbar spine CT and MRI in internal test setsPerformance metricsU-NetU2netUNet3DMask-RCNNCTPA0.982 (0.978–0.985)0.914 (0.909–0.921)0.984 (0.983–0.987)0.975 (0.971–0.977)mPA0.807 (0.798–0.815)0.507 (0.501–0.513)0.849 (0.846–0.852)0.782 (0.780–0.785)mIoU0.712 (0.707–0.720)0.393 (0.385–0.400)0.760 (0.759–0.762)0.686 (0.680–0.692)MRIPA0.973 (0.969–0.985)0.976 (0.968–0.980)0.973 (0.970–0.975)0.981 (0.979–0.982)mPA0.811 (0.802–0.823)0.770 (0.765–0.776)0.849 (0.846–0.852)0.771 (0.768–0.774)mIoU0.712 (0.704–0.727)0.667 (0.664–0.668)0.724 (0.722–0.726)0.702 (0.700–0.705)Note.—PA: pixel accuracy; mPA: mean pixel accuracy; mIoU: mean intersection over union. Performance metrics were reported as the mean with the 95% confidence interval.Fig. 4UNet3D-based Segmentation of Lumbar Spine CT and MRI in a 59-Year-Old Female with Osteoporosis.
The patient's DXA revealed a total lumbar T-score of -2.4 (Segmental scores: L1: -2.5, L2: -2.6, L3: -2.3, and L4: -2.2). (A) Illustration of the CT segmentation process. (a) Original sagittal CT slice.
(b, c) Sequential segmentation outputs for the L4 and L5 vertebrae at an intermediate and the final stage, respectively. (d, e) Corresponding segmentation for the entire L1-S1 vertebral column at intermediate and final stages. (B) Segmentation output demonstrating the separation of (f) cortical and (g) cancellous bone components.
(C) Axial T2-weighted MRI segmentation at the L2/3 (h), L3/4 (i), L4/5 (j), and L5/S1 (k) levels, with the UNet3D prediction (“Predicted”) directly compared to the manual “Ground Truth”.
On lumbar spine MRI images, UNet3D achieved PA, mPA, and mIoU scores of 0.973, 0.849, and 0.724, respectively, for the segmentation of paraspinal muscles and intermuscular spaces across the L1–S1 levels in the internal test set. Although its PA was slightly lower than that of U2-Net and Mask R-CNN, its mPA and mIoU notably exceeded all compared models (Table 2). UNet3D maintained robust performance with scores of 0.969 (95% CI [0.967, 0.972]), 0.834 (95% CI [0.830, 0.836]), and 0.713 (95% CI [0.711, 0.715]) on the external test set.
As depicted in Fig. 4C, the model produces well-delineated segmentation of these structures.
HU values quantification in vertebral VOIs
UNet3D successfully delineated the cortical-trabecular bone interface (Fig. 4B). The mean HU values within trabecular bone VOIs for L1 to S1 vertebrae were 125.75, 80.33, 105.86, 150.35, 182.28, and 195.14 HU, respectively.
Specifically, HU values in L1–L3 were lower than those in L4–S1. Furthermore, positive correlations were observed between HU values in VOIs and corresponding T-scores for L1 to L4 (L1: r = 0.93, P = 0.04; L2: r = 0.93, P = 0.02; L3: r = 0.89, P = 0.02; L4: r = 0.85, P = 0.04).
Volume quantification of paraspinal muscles and intermuscular spaces
The statistical analysis indicated that the data were largely normally distributed at the α = 0.05 significance level (P > 0.05; Table E2 in supplemental materials). The psoas major accounted for the largest total bilateral volume at L1–S1 (173,317.68 mm3), followed by the iliocostalis, longissimus, and multifidus, while the quadratus lumborum constituted the smallest (65,167.78 mm3). Regarding the total bilateral volume of intermuscular spaces at L1–S1, the PV space exhibited the largest volume (124,301.86 mm3), followed by the LM, ML, and IL spaces, while the QI space was the smallest (29,676.63 mm3).
The volumetric distribution of paraspinal muscles and intermuscular spaces varied across different L1–S1 levels, with distinct trends observed along the craniocaudal axis. Bilateral asymmetry was observed in both paraspinal muscles and intermuscular spaces. For total muscle volumes across L1–S1, right-side predominance was observed in the multifidus (L/R: 0.80), longissimus (0.91), iliocostalis (0.95), and psoas major (0.64), whereas the quadratus lumborum was left-dominant (1.16).
In contrast, most intermuscular spaces were left-dominant, including the LM (1.55), IL (1.61), QI (2.71), and PV (1.81) spaces, whereas the ML space was right-dominant (0.76). The most pronounced asymmetries occurred at the L4–L5 level (Table 3). The longissimus and iliocostalis muscles exhibited substantial FI rates across all L1-S1 segments, and higher infiltration levels in the multifidus and psoas muscles were predominantly observed in the lower segments on the external test set (Table E3).Table 3Paraspinal muscle and intermuscular space volumes at L1-S1 segments and in total.CategoryL1-L2L2-L3L3-L4L4-L5L5-S1L1-S1MultifidiLeft6858.66 ± 2503.129981.30 ± 4157.4016977.83 ± 6039.698598.83 ± 2998.7310890.92 ± 5082.3252063.39 ± 13369.44Right7045.53 ± 3607.979541.41 ± 4012.9716678.93 ± 5336.3423638.71 ± 9016.4011035.25 ± 4550.7366088.75 ± 18888.01L/R1.031.071.010.361.100.80LongissimusLeft29608.47 ± 11224.0524564.81 ± 10136.7217186.97 ± 6054.994505.77 ± 2807.345090.21 ± 1783.4979618.54 ± 26660.07Right29824.07 ± 12472.4524376.61 ± 8875.1117842.01 ± 7109.1412594.34 ± 4309.194506.44 ± 1656.7188688.66 ± 31189.02L/R1.031.001.000.311.210.91IliocostalisLeft28165.20 ± 12175.7031630.17 ± 12754.0330861.09 ± 11512.312852.47 ± 2865.921274.25 ± 1229.7593767.51 ± 35575.01Right24195.31 ± 9386.6927362.49 ± 13769.9727162.48 ± 10234.4920182.43 ± 8028.311016.02 ± 708.1699405.86 ± 38977.80L/R1.151.211.150.121.180.95QuadratusLeft6284.70 ± 2846.7211099.84 ± 4161.9214641.65 ± 5769.065409.16 ± 12438.45495.88 ± 268.7934348.56 ± 11786.13Right5789.35 ± 2292.838758.75 ± 4025.2511536.83 ± 3569.517930.65 ± 11816.88637.98 ± 396.4930819.22 ± 9381.44L/R1.111.331.290.520.781.16PsoasLeft9913.15 ± 6310.2520030.31 ± 8625.0431017.89 ± 11928.215328.31 ± 5971.513259.97 ± 8443.7566965.18 ± 27342.69Right10449.70 ± 6839.2621325.25 ± 8828.3832822.43 ± 11660.6837899.50 ± 23133.153200.93 ± 8407.14106352.50 ± 46188.27L/R0.960.940.940.151.040.64MLLeft2234.85 ± 673.893108.83 ± 1022.365141.88 ± 1993.985312.05 ± 2079.281631.03 ± 2969.0616581.38 ± 4060.28Right2818.39 ± 1043.984468.80 ± 1281.566115.65 ± 2428.228404.84 ± 2008.421386.92 ± 2567.6822218.38 ± 5358.81L/R0.850.740.900.701.250.76LMLeft2445.77 ± 903.572650.90 ± 890.503133.55 ± 1225.0321988.08 ± 10315.4818066.75 ± 9492.5951823.96 ± 13087.69Right2850.91 ± 1197.702831.53 ± 752.613212.84 ± 956.854299.97 ± 1796.2117790.56 ± 9863.9833473.68 ± 7154.79L/R0.920.930.986.321.001.55ILLeft2458.46 ± 901.372529.51 ± 823.222342.20 ± 1162.1511701.65 ± 7283.0010705.16 ± 9964.4028164.98 ± 9920.97Right2393.43 ± 1181.172763.82 ± 1172.842820.40 ± 1687.681941.96 ± 890.3310282.43 ± 10708.2817588.77 ± 5419.63L/R1.130.980.848.891.091.61QILeft2151.10 ± 905.473069.85 ± 832.542823.29 ± 945.7718965.67 ± 9381.442815.54 ± 2968.5517588.77 ± 5419.63Right1955.77 ± 948.273396.59 ± 1255.943019.07 ± 1152.221915.01 ± 1481.492328.73 ± 3076.3412087.86 ± 5010.27L/R1.200.950.9825.691.432.71PVLeft2230.39 ± 1776.653021.52 ± 1139.163133.55 ± 990.3741060.29 ± 18116.6127698.97 ± 11547.9979946.84 ± 20976.96Right1770.24 ± 729.423705.07 ± 1449.463320.86 ± 1092.415873.63 ± 3058.7927660.99 ± 8114.9744355.02 ± 6076.32L/R1.390.921.029.701.011.81Note— Volumetric analysis revealed distinct segmental distribution patterns from L1 to S1.
The multifidus muscle exhibited a biphasic trend—initially increasing then gradually decreasing—with significantly larger volumes in the lower lumbar segments. The longissimus muscle showed a progressive caudal volume reduction, peaking at L1–L2 and minimal at L5–S1. The iliocostalis muscle maintained substantial volume from L1 to L3 but decreased markedly caudally, being minimal at L5–S1.
The quadratus lumborum volume initially increased then decreased, with relatively larger volumes at L2–L4 and the smallest at L5–S1. In contrast, the psoas major volume was significantly greater in the upper lumbar segments. Among intermuscular spaces, the ML was largest at L3–L5, whereas the LM peaked at L4–S1.
The IL showed a progressive caudal increase, with larger volumes at L4–S1 than at L1–L3. The QI was maximal at left L4–L5, and the PV was greatest at L4–S1. Volumetric differences between left and right sides were also observed for both muscles and spaces.ML: space between the multifidus muscle and the vertebral lamina; LM: intermuscular space between the longissimus and multifidus muscles; IL: intermuscular space between the iliocostalis and longissimus muscles; QI: intermuscular space between the quadratus lumborum and iliocostalis muscles; PV: space between the psoas major muscle and the vertebral body; L/R: left-to-right volume ratio for each paraspinal muscle or intermuscular space.
All volumes are expressed in mm3.
Analysis of weighting coefficients between HU values in vertebral VOIs and volumes of paraspinal muscles/intermuscular spaces
Global analysis revealed positive correlations (ω > 0) between total paraspinal muscles volumes and vertebral HU values, whereas total intermuscular spaces volumes were inversely correlated (ω < 0). This correlation was stronger on the left side. Among individual muscles, the multifidus, longissimus, and iliocostalis demonstrated the strongest associations with HU values in vertebral VOIs.
For the intermuscular spaces, the highest correlations were observed in the LM, IL, and ML regions. These musculoskeletal relationships were more pronounced in the upper (L1-L3) than in the lower lumbar segments (Table 4).Table 4Weight coefficients between HU values in vertebral VOIs and volumes of paraspinal muscles/muscle spacesCategoryL1L2L3L4L5S1L_IL−0.016−0.025−0.009−0.005−0.005−0.005L_Iliocostalis0.0080.0180.0060.0070.0010.006L_LM−0.014−0.000−0.011−0.010−0.006−0.007L_Longissimus0.0170.0030.0150.0080.0050.011L_ML−0.020−0.004−0.000−0.003−0.007−0.005L_Multifidi0.0160.0090.0150.0080.0130.004L_Psoas0.0030.0090.0070.0020.0010.008L_PV−0.003−0.005−0.010−0.002−0.000−0.007L_QI−0.009−0.018−0.000−0.006−0.003−0.004L_Quadratus0.0020.0090.0050.0010.0010.002R_IL−0.000−0.003−0.002−0.003−0.003−0.001R_Iliocostalis0.0030.0040.0020.0000.0000.004R_LM−0.005−0.003−0.013−0.009−0.002−0.014R_Longissimus0.0060.0100.0100.0040.0030.012R_ML−0.002−0.001−0.006−0.001−0.003−0.001R_Multifidi0.0070.0120.0050.0020.0020.001R_Psoas0.0030.0010.0080.0010.0000.003R_PV−0.008−0.013−0.002−0.001−0.000−0.004R_QI−0.009−0.017−0.0090.000−0.0030.000R_Quadratus0.0080.0150.0040.002−0.005−0.006Note—Abbreviations for Category follow a consistent nomenclature: each code begins with 'L_' or 'R_' to denote left or right side, followed by the anatomy or acronym. Coefficient weighting was applied to enable equitable comparison among variables with differing units and scales by transforming them into comparable, dimensionless quantities.
The weighting coefficient (ω) reflects the relative importance of each variable, where the absolute value |ω| indicates the magnitude of the effect (with larger values denoting stronger effects). The sign of ω indicates the direction of association: ω > 0 signifies a positive/promoting association, ω < 0 indicates a negative/risk association, and ω ≈ 0 suggests a negligible relationship. These weights imply statistical importance and do not establish causality.
The segmental volumes of the paraspinal muscles and intermuscular spaces exhibited stronger correlations with vertebral HU values than the total volume across L1-S1. In general, the volume of each paraspinal muscle at each segment was positively correlated with the HU value (ω > 0), whereas the intermuscular space volume was negatively correlated (ω < 0). Furthermore, the strength of these correlations differed between the left and right sides (Table E4, in supplemental materials).
Analysis of the external test set revealed a significant inverse correlation between paraspinal muscle FI rates and HU values in vertebral VOIs. Of these, the multifidus and psoas muscles demonstrated the strongest negative correlations (Table E5).
Discussion
Early screening and timely intervention for OP can significantly reduce the risk of OP-related complications such as fractures [20]. However, over half of patients with osteoporotic fractures never undergo OP screening, which underscores the paramount need for early and precise diagnostic strategies [21]. In recent years, end-to-end trainable convolutional neural networks have substantially advanced semantic segmentation.
For medical image analysis, U-Net has emerged as one of the most successful architectures [22], [23], [24], [25]. In this study, while U-Net demonstrated competent segmentation performance, its design primarily focuses on extracting local features from 2D images, resulting in a limited capacity for modeling 3D spatial relationships. To this end, our UNet3D model employs two dual-convolution specifically designed to integrate sequential 2D slices into a coherent 3D volumetric representation at the input stage.
This architecture enables the early capture of anatomical semantics and critical spatial context, thereby establishing foundational semantic consistency and spatial coherence for the subsequent analysis. Additionally, we utilize the softmax activation function in place of sigmoid to align the model output with multi-class segmentation requirements, thereby achieving better compatibility with the loss function and accelerating model convergence. Diverging from the symmetrical encoder-decoder structure of the standard U-Net, the U2-Net architecture is constructed using Residual U-Block (RSU) as its fundamental building modules [18].
RSU was explicitly engineered for Salient Object Detection, excelling at capturing multi-scale features within images, such as object edges. Nevertheless, the nested U-structure tends to over-segment local features, thereby disrupting the global spatial relationships of vertebrae and paraspinal muscles. Mask R-CNN has been employed for vertebrae and muscle segmentation [26], [27].
Its two-stage architecture, which generates proposals via a Region Proposal Network and RoIAlign before performing mask prediction, provides good robustness for small objects [19]. However, its practical application is limited by high computational complexity, prohibitive hardware demands, and suboptimal training and inference speed. On the internal test set, UNet3D achieved a mIoU gain of 6.7% over U-Net, 93.4% over U2-Net, and 10.8% over Mask R-CNN for lumbar spine CT segmentation.
For MRI, it outperformed U-Net, U2-Net, and Mask R-CNN by 1.7%, 8.5%, and 3.1% in mIoU, respectively. The narrow 95% confidence intervals observed for all performance metrics further attest to the robustness and reproducibility of the UNet3D model across diverse CT and MRI segmentation tasks. Furthermore, the model demonstrated robust performance on an independent external test set, indicating its strong potential for early automated screening and quantitative analysis of OP patients.
Osteoporotic changes are most pronounced in trabecular bone, which exhibits a metabolic activity approximately eight times greater than cortical bone and more accurately reflects true changes in BMD [28]. Although CT imaging provides high resolution of bone architecture, it remains challenging to visually delineate the precise boundary between trabecular and cortical bone within vertebrae. In OP patients, differential metabolic changes further obscure this distinction, rendering traditional threshold-based methods ineffective, which is highly susceptible to partial volume effects.
To address this, we innovatively applied a gradient-adaptive algorithm for dynamic boundary delineation. Substantial evidence supports the correlation between HU values of vertebral trabecular bone and DXA measurements. Moreover, CT offers advantages in screening due to its relative insensitivity to spinal degenerative changes [6], [29], [30].
However, the conventional ROI method for measuring HU values fails to account for the spatial heterogeneity of BMD within the vertebra. In this study, UNet3D enabled automated quantification of HU values in VOIs. The strong positive correlation with DXA T-scores (r ≥ 0.85) aligns with prior findings [10], supporting its feasibility for BMD assessment.
While QCT is the gold standard for volumetric BMD, DXA’s clinical primacy for diagnosis and fracture risk assessment is based on its accessibility, low radiation, and cost-effectiveness. The high correlation between DXA and QCT at L1-L4 (0.662 ≤ r ≤ 0.728) [31], [32] justifies its use as our reference standard in this retrospective context, a position further supported by studies linking volumetric HU values to QCT-BMD, such as Zhou et al. [9]. We also observed varying HU values across the L1-S1 VOIs in OP patients.
Future biomechanical studies are warranted to elucidate the stress distribution patterns across vertebral levels.
Bone and muscle are the two primary components of the musculoskeletal system. They interact through a combination of genetic factors, mechanical loading from gravitational forces, and endocrine crosstalk affecting the metabolism of both tissues [3], [33]. While increasing evidence links bone and muscle cross-sectional area [34], muscle volume better reflects global muscle function.
It avoids confounders like posture and degeneration common in single-level assessment [35], [36], making it more clinically relevant [37]. Although Wesselink et al. [38] and Wang et al. [39] successfully automated the segmentation of key muscles across L1-L5 using deep learning, they did not perform volumetric analysis. The UNet3D model automatically quantifies L1-S1 segmental and total volumes of paraspinal muscles and intermuscular spaces.
We observed that the relative muscle volumes on MRI followed the expected anatomical hierarchy, providing further evidence for the UNet3D's efficacy. Notably, the enlarged LM, ML, and IL intermuscular spaces likely indicate significant atrophy of the multifidus, longissimus, and iliocostalis muscles. The volumes of the paraspinal muscles and intermuscular spaces exhibited level-specific variations from L1 to S1, with different changing trends.
Future longitudinal clinical studies are warranted to investigate the causes and progression of these muscular changes.
In this study, paraspinal muscle volume was positively correlated with vertebral HU values in OP patients, whereas intermuscular space volume showed a negative correlation, which supports the theory of muscle-bone interactions [40], [41]. Chiapparelli et al. [34] also reported a significant positive correlation between the functional cross-sectional area of paraspinal muscles and vertebral bone mineral density (BMD). The correlation between HU values and volume of individual paraspinal muscles/intermuscular spaces varied distinctly and was asymmetric between sides.
These correlations also demonstrated segmental specificity, which may align with the hypothesis that different spinal levels experience unique mechanical loading and localized stress stimulation from adjacent muscles. Consequently, regional variations should be considered when interpreting lumbosacral bone-muscle relationships. Mechanistically, muscle FI is a key indicator of impaired muscle function in OP.
Adipose tissue directly replaces contractile fibers, diminishing muscle quality and strength [42], which weakens essential mechanical stimulation for bone maintenance. Furthermore, fat secretes inflammatory cytokines that exacerbate chronic inflammation in the bone microenvironment, directly activating osteoclasts and promoting bone resorption [43]. Elevated fat mass is thus an established risk factor for OP [33].
Our findings confirm that paraspinal muscle FI rates inversely correlate with vertebral volumetric HU values, aligning with existing evidence [44], [45], [46]. We further observed that infiltration exhibits distinct muscle- and segment-specific patterns. Although multiple studies support the use of T2-weighted MRI for body composition analysis, it should be noted that dedicated water-fat separation techniques, such as Dixon sequences, provide higher sensitivity and specificity for the quantitative assessment of lipid infiltration [47].
The synergy between declining muscle function and increased bone fragility significantly elevates the risk and severity of vertebral fractures [44]. Patients with osteosarcopenia face higher risks of frailty, disability, and falls [2], [3], [48], [49] underscoring the necessity for a comprehensive “musculoskeletal integrated care” approach, especially in the frail elderly.
In summary, our UNet3D model enables the synergistic extraction and quantification of bimodal imaging features within a unified 3D anatomical space. This integration is crucial for investigating the spatially coupled interplay between muscle and bone. The approach represents a shift from 2D to 3D assessment, improving precision and efficiency.
The model can not only output quantitative values but also display visual segmentation maps, thereby enhancing the interpretability of its decision-making process. Critically, by leveraging the widespread availability of routine CT and MRI, this automated tool can transform existing scans into objective musculoskeletal metrics to aid clinical decision-making. It holds promise for facilitating primary screening and prevention of OP in broader populations, and offers a tool for future research on fracture risk and long-term dynamic monitoring, promoting more precise OP management.
Nevertheless, challenges remain for widespread clinical adoption, including handling data heterogeneity from diverse sources and validating effectiveness in prospective trials.
This study has several limitations. First, its retrospective nature introduces an inherent risk of selection bias. Second, despite the exclusion of severe degenerative and scoliotic cases, the employment of DXA as the BMD reference standard is nevertheless constrained by the fundamental limitations of 2D projection imaging.
QCT is less affected by spinal degeneration, but its high radiation dose, cost, technical demands, and limited clinical availability restrict data accessibility. Following the acquisition of specialized calibration phantoms and analysis software for QCT measurements at our institution, we will employ synchronous or asynchronous calibration methods to obtain volumetric BMD results for further analysis. Furthermore, our cohort comprised exclusively OP patients and focused specifically on the L1-S1 region, which represents the primary site of osteoporotic involvement.
Future work should involve multi-center prospective studies that incorporate QCT data and more diverse datasets with Dixon sequences, including non-osteoporotic individuals, various ethnicities, and different age groups, and confounding factors such as overall health, physical activity, or nutritional status, to validate and enhance the robustness and generalizability of UNet3D. The UNet3D model itself is based on a universal segmentation architecture and has good scalability. It can employ transfer learning strategies and adaptive training using annotated data of target musculoskeletal regions in the thoracic or cervical spine to learn their specific features, which enabling a comprehensive investigation of musculoskeletal relationships of the entire spine.
Conclusion
In conclusion, the proposed UNet3D demonstrates superior segmentation performance for lumbar spine CT and MRI compared to established models (U-Net, U2-Net, Mask R-CNN). It enables simultaneous and accurate 3D quantification of musculoskeletal parameters in OP patients. Our analysis reveals that in OP, vertebral BMD exhibits a positive correlation with paraspinal muscle volume and a negative correlation with muscle FI, with these relationships showing segment-specific variation.
By facilitating such detailed analysis, UNet3D emerges as a promising tool for early opportunistic screening, risk stratification, and individualized OP management. Ultimately, it is expected to provide novel insights into OP pathophysiology and fracture risk prediction, with the potential to guide personalized clinical decisions based on a holistic assessment of musculoskeletal synergy
Compliance with ethics requirements
This study was approved by The Second Affiliated Hospital of the Naval Medical University of the Chinese People's Liberation Army institutional review board (number: 2024SL097) and compliant with the Declaration of Helsinki. A waiver of consent was granted due to the retrospective nature of the study and implementation of strict data anonymization and de-identification protocols.
CRediT authorship contribution statement
Zhenzhen Guan: Conceptualization, Data curation, Methodology, Validation, Formal analysis, Investigation, Writing – original draft, Writing – review & editing, Visualization. Yijin Wang: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Data curation, Writing – review & editing, Supervision, Project administration. Shiyu Zhang: Conceptualization, Visualization, Investigation, Data curation, Writing – review & editing.
Yanyang Zhang: Investigation, Visualization, Writing – review & editing. Liang Wang: Investigation, Writing – review & editing, Visualization, Supervision, Project administration. Yixin Chen: Investigation, Visualization, Writing – review & editing.
Xuhua Lu: Conceptualization, Writing – review & editing, Resources, Supervision, Project administration, Funding acquisition.
Declaration of competing interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.