Differential association of psoas and paraspinal muscle volumes with appendicular lean mass index: AI-derived lumbar computed tomography three-dimensional volumetric indices in older Korean Women: an observational study Authors: Eun Chae Lee, Jang Hyuk Cho, Dong Gyu Lee Published: 2026 ABSTRACT AbstractStudy DesignRetrospective cross-sectional study.PurposeTo investigate the relationship between sarcopenia-related markers and artificial intelligence (AI)–derived three-dimensional (3D) volumetric indices of lumbar muscles.Overview of LiteratureSarcopenia is characterized by loss of skeletal muscle mass and physical function. Although appendicular skeletal muscle mass (ASM) is widely used for sarcopenia assessment, reliable evaluation of lumbar paraspinal muscles remains limited. AI-based 3D volumetric analysis may provide a reproducible approach for muscle assessment.MethodsThis single-center retrospective cross-sectional study included 58 women aged ≥65 years who underwent lumbar computed tomography and sarcopenia assessment. AI-derived 3D volumetric analysis was performed using TotalSegmentator implemented in 3D Slicer. Volumetric indices were calculated as muscle volume (cm3) divided by height3 (m3). Associations between volumetric indices and ASM measured by dual-energy X-ray absorptiometry were evaluated using Pearson correlation and multivariable linear regression analyses.ResultsAmong the 58 participants, 25 were classified as having sarcopenia. Appendicular skeletal muscle mass index (ASMI) was not significantly associated with paraspinal muscle volume indices. In contrast, the iliopsoas muscle volume index showed a significant positive correlation with ASMI (r=0.424, p<0.001). No significant differences in AI-derived 3D lumbar muscle volumetric indices were observed between sarcopenic and non-sarcopenic groups.ConclusionsParaspinal muscle volume indices were not significantly associated with ASMI, whereas iliopsoas muscle volume indices showed a significant association. These findings suggest that the iliopsoas muscle may better reflect systemic skeletal muscle mass than paraspinal muscles in older women.Keywords: Paraspinal muscles, Sarcopenia, Psoas index, Artificial intelligence, Muscle atrophy Graphical Abstract Introduction Sarcopenia poses a big threat to the health of older people as it leads to their deteriorating health and eventual death, while at the same time it drags the society down with a heavy economic burden and other social problems [1]. This disease is characterized by the gradual loss of skeletal muscle tissue accompanied by diminished handgrip strength or reduced physical capabilities. Current consensus guidelines, including those from the European Working Group on Sarcopenia in Older People and the Asia-specific criteria by the Asian Working Group for Sarcopenia (AWGS 2019), standardize the diagnostic criteria and algorithms for diverse populations [2]. In postmenopausal women above 65 years, the rate of occurrence and the effects of sarcopenia are intensified by hormonal changes. Particularly, estrogen depletion contributes to accelerated musculoskeletal decline, heightening the risk of frailty and osteoporotic fractures. For this reason, precise measurement of muscle-related factors is an essential step in the treatment of the highest risk group in terms of functional loss [3,4]. Appendicular skeletal muscle mass (ASM) by dual-energy X-ray absorptiometry (DXA) is the established standard for muscle mass measurement in sarcopenia [5]. In spite of DXA being characterized by low radiation doses and high throughput, no unified imaging approach has been developed for the assessment of lumbar muscle volume. Historically, spinal musculature has been primarily assessed on the basis of a single-slice axial image. Typically, research has relied on cross-sectional area (CSA) measurements from computed tomography (CT) and magnetic resonance imaging to assess muscle status [6]. Regional analysis at the third lumbar vertebra (L3) is especially recognized as a good indicator of overall body composition and, in addition to providing volumetric data, it can highlight the extent of myosteatosis [7]. However, two-dimensional (2D) single-slice methods are restrictive, since the pattern of muscle morphology and fatty infiltration varies widely throughout the spine [8]. Furthermore, CSA manual measurement both demands a high level of expertise and is often susceptible to inter-rater variability. Recently, automated artificial intelligence (AI)-based three-dimensional (3D) volumetric segmentation has been proposed as a novel option to overcome these drawbacks. Automated AI-based 3D segmentation not only avoids operator dependency but also significantly decreases processing time and measures the entire muscle volume in the lumbar spine region, which eliminates slice-selection variability and improves the reproducibility of the measurements [9,10]. In addition, no research has been done to investigate the correlation of 3D paraspinal and psoas muscle volume indices with DXA-derived ASM. Consequently, the objective of this study was to determine the association of AI-derived 3D paraspinal and iliopsoas muscle volume indices with sarcopenia-related markers in the elderly female population. Materials and Methods Study cohort and inclusion criteria We analyzed retrospectively and observationally patient data registered at the tertiary hospital spine center within 22 months timeline (February 2023 to December 2024). Our study population consisted of women aged ≥65 years who met the following conditions: undergoing a standardized sarcopenia evaluation (DXA and bilateral handgrip strength) and having a high-resolution lumbar CT available. We excluded participants who presented with confounding pathologies, such as prior spinal intervention, tumors, or infections in the analytic field. The study was conducted in accordance with the Declaration of Helsinki guidelines and approved by the Institutional Review Board of Youngman University Hospital before starting the study (protocol codes: YUMC 2022-10-10-050; December 29, 2022). Informed consent was obtained from all the participants involved in this study. Grip strength The bilateral handgrip strength was measured with a digital hand-held dynamometer (Jamar Plus+Digital Hand Dynamometer; Performance Health, Warrenville, IL, USA). The participants had to sit comfortably. Their shoulder was placed in an adducted and neutral rotation position, and their elbow was flexed at 90°. They performed the test for both hands, and the highest score obtained from the two attempts was taken for the statistical analysis (kg). Diagnostic criteria for sarcopenia Sarcopenia was identified using the AWGS 2019 criteria. In particular, participants, who all were females in this study, were deemed sarcopenic if they had a low appendicular skeletal muscle mass index (ASMI=ASM/height2) below 5.4 kg/m2, a measurement taken by whole-body DXA, combined with a low handgrip strength below 18 kg [2]. Those who failed to meet these criteria were placed in the non-sarcopenia group. Dual-energy X-ray absorptiometry We used a DXA scanner (Hologic Horizon W; Hologic Inc., Marlborough, MA, USA) with software version 13.6.1.2 for body composition analysis. Total lean mass and ASM were the measures supplied by this equipment, and ASMI was obtained by normalizing for height squared (ASMI=ASM/height2). Apart from measuring body weight and height, the system also automatically computed adiposity-related variables such as total body fat percentage, total fat mass, and the height-normalized fat mass index (FMI=fat mass/height2). CT image acquisition SOMATOM Force scanner (Siemens Healthineers, Forchheim, Germany) was used to obtain non-contrast lumbar CT scans. Axial images were reconstructed using a Br59 reconstruction kernel (100 kVp; head-first supine position) with a 5.0-mm slice thickness. The lumbar spine was the anatomical region of interest. 3D CT segmentation and volumetry For digitized morphometric analysis, raw CT DICOM (Digital Imaging and Communications in Medicine) data were uploaded into the 3D Slicer platform ver. 5.9.0 (https://www.slicer.org/). We relied on the TotalSegmentator add-on (revision c2aaf8f, September 10, 2025) to run deep learning-based segmentation of the lumbar muscles [11]. The multi–tissue classification of the framework enabled separate quantification of skeletal muscle and intramuscular adipose tissue (IMAT) volumes. The volume measurements paraspinal skeletal muscle volume (PSMV), paraspinal IMAT volume (IMATV), and iliopsoas muscle volume (IPMV) were then derived by dividing the original volumes by height to reflect inter-individual differences in stature, i.e., volume index=volume (cm3)/height3 (m3). This is similar to the use of allometric scaling methods in modern volumetric research (Fig. 1) [12,13]. (A) Axial view showing automatic segmentation by TotalSegmentator of the paraspinal skeletal muscle (green), iliopsoas muscle (purple), and intramuscular adipose tissue (IMAT) (yellow). (B) Sagittal view. (C) Coronal view. (D) Three-dimensional reconstruction of the paraspinal muscles, paraspinal IMAT, and iliopsoas muscle. To ensure the quality of the data, all automated segmentations were checked visually by a single rater in order to make corrections, if any, for segmentation errors. Statistical analysis Differences between sarcopenic and non-sarcopenic subjects at baseline were analyzed by independent t-tests. Continuous variables are displayed as mean±standard deviation. Performance of Pearson’s product-moment correlation after the Shapiro-Wilk method for normality testing was used to measure associations of ASMI and 3D volumetric indices (i.e., paraspinal skeletal muscle volume index [PSMVI] and iliopsoas muscle volume index [IPMVI]). Multivariable linear regression models were adjusted for the confounders to find out the association between sarcopenia markers (age, grip strength, and ASMI) and volumetric muscle indices. Sarcopenia markers were treated as independent variables, while lumbar muscle volume indices were regarded as dependent variables. A two-tailed p<0.05 indicated statistical significance. Software for the analysis included IBM SPSS Statistics ver. 29.0 (IBM Corp., Armonk, NY, USA) and the Jamovi statistical platform ver. 2.6 (https://www.jamovi.org/) [14]. Results The baseline clinical profile of the study cohort (n=58) is shown in Table 1. Sarcopenic subjects were significantly older (79.9±5.9 years vs. 76.7±5.4 years, p=0.035) and had lower weight (49.6±6.7 kg vs. 59.3±8.0 kg, p<0.001), height (148.6±7.5 cm vs. 152.0±5.4 cm, p=0.046), and body mass index (BMI: 22.6±3.6 kg/m2 vs. 25.7±3.1 kg/m2, p=0.001) than non-sarcopenic subjects. As expected by the diagnostic criteria, handgrip strength (14.2±2.6 kg vs. 20.3±3.9 kg, p<0.001) and ASMI (4.8±0.4 vs. 5.4±0.5, p<0.001) were significantly lower in the sarcopenic group. We also observed significantly lower total body fat percentage (38.6%±4.9% vs. 42.1%±4.3%, p=0.005) and FMI (8.7±2.1 kg/m2 vs. 10.7±2.2 kg/m2, p<0.001) in the sarcopenia group, indicating a generalized reduction in both lean and fat mass components. 3D volumetric indices obtained by AI did not show any significant differences between sarcopenic and non-sarcopenic groups (Table 2). IPMVI showed a significant moderate correlation with ASMI (r=0.424, p<0.001) (Fig. 2). Multivariable linear regression showed that age was significantly associated with both PSMVI and IPMVI. Interestingly, ASMI was found to be an independent predictor of IPMVI, but not PSMVI (Table 3). Grip strength was not statistically significant in its correlation with either lumbar muscle measurement. Table 1Baseline characteristics of the participantsCharacteristicTotal (n=58)Sarcopenia (n=25)Non-sarcopenia (n=33)p-valueAge (yr)78.1±5.879.9±5.976.7±5.40.035*Weight (kg)55.1±8.849.6±6.759.3±8.0<0.001*Height (cm)150.5±6.6148.6±7.5152±5.40.046*Body mass index (kg/m2)24.3±3.622.6±3.625.7±3.10.001*Grip strength (kg)17.6±4.514.2±2.620.3±3.8<0.001*Total body % fat40.6±4.938.6±4.942.1±4.30.005*Fat mass index (kg/m2)9.9±2.38.7±2.110.7±2.2<0.001*Appendicular skeletal muscle mass index (kg/m2)5.1±0.64.8±0.45.4±0.5<0.001*Values are presented as mean±standard deviation.*p<0.05 (statistically significant). Baseline characteristics of the participants Values are presented as mean±standard deviation. p<0.05 (statistically significant). Table 2AI-derived 3D muscle volume indices between sarcopenia and non-sarcopeniaVariable of 3D indicesTotal (n=58)Sarcopenia (n=25)Non-sarcopenia (n=33)p-valuePSMVI (cm3/m3)123.7±37.0117.7±36.4128.3±37.40.283IPMVI (cm3/m3)61.5±15.758.6±17.563.8±14.10.214Paraspinal IMATVI (cm3/m3)32.3±22.632.6±24.332.1±21.60.939Values are presented as mean±standard deviation.AI, artificial intelligence; 3D, three-dimensional; PSMVI, paraspinal muscle volume index; IPMVI, iliopsoas muscle volume index; IMATVI, intramuscular adipose tissue volume index. AI-derived 3D muscle volume indices between sarcopenia and non-sarcopenia Values are presented as mean±standard deviation. AI, artificial intelligence; 3D, three-dimensional; PSMVI, paraspinal muscle volume index; IPMVI, iliopsoas muscle volume index; IMATVI, intramuscular adipose tissue volume index. Scatter plots showing the relationship between appendicular skeletal muscle mass index (ASMI) and artificial intelligence-derived muscle volume indices. (A) Iliopsoas muscle volume index (IPMVI) showed a significant positive correlation with ASMI, (B) whereas paraspinal muscle volume index (PSMVI) was not significantly correlated with ASMI in Pearson correlation analysis. Table 3Multivariable linear regression analysis of factors associated with AI-derived back muscle volume indicesPredictorIPMVIPSMVI β±SE95% CIp-valueβ±SE95% CIp-valueAge−0.810±0.332−1.48 to −0.150.018*1.353±0.5170.32 to 2.390.012* Grip strength−0.336±0.460−1.26 to 0.590.469−0.079±0.717−1.52 to 1.360.913 ASMI12.319±3.4565.39 to 19.25<0.001*5.175±5.384−5.62 to 15.970.341Model statistics. PSMVI model: R2=0.137, adjusted R2=0.089. IPMVI model: R2=0.261, adjusted R2=0.220.AI, artificial intelligence; IPMVI, iliopsoas muscle volume index; PSMVI, paraspinal muscle volume index; SE, standard error; CI, confidence interval; ASMI, appendicular skeletal muscle mass index.*p<0.05 (statistically significant). Multivariable linear regression analysis of factors associated with AI-derived back muscle volume indices Model statistics. PSMVI model: R2=0.137, adjusted R2=0.089. IPMVI model: R2=0.261, adjusted R2=0.220. AI, artificial intelligence; IPMVI, iliopsoas muscle volume index; PSMVI, paraspinal muscle volume index; SE, standard error; CI, confidence interval; ASMI, appendicular skeletal muscle mass index. p<0.05 (statistically significant). Discussion In this study, we found that ASMI had almost no relationship with PSMVI, while IPMVI was highly related to ASMI. This implies that the iliopsoas muscle might be a better indicator of whole-body skeletal muscle loss over time than the paraspinal muscles. Despite the fact that both sarcopenia and non-sarcopenia groups had lower mean values of PSMVI and IPMVI as measured by AI-derived 3D volumetric indices, these differences between the two groups were not statistically significant. Paraspinal muscles are essential for axial spinal alignment and stability, including the control of kyphosis, scoliosis, and the lumbosacral angle [15]. Aging causes a systemic skeletal muscle involution and also increases fatty infiltration. Degeneration of paraspinal and psoas muscles may be seen as a part of overall aging, but spinal pathologies like spinal stenosis and radiculopathy are more specifically associated with focal muscle atrophy and myosteatosis [16,17]. Besides, patients with chronic low back pain are often less physically active, resulting in selective disuse atrophy of paraspinal muscles [18]. Therefore, a decrease in CSA of multifidus has been found to be closely connected to chronic pain syndromes [19]. This kind of localized, disease-specific remodeling reduces the association of localized paraspinal degeneration with systemic skeletal muscle loss. In line with this, paraspinal muscles generally show a lesser relation with ASMI. Measures of paraspinal muscles generally represent changes due to spinal degeneration only, rather than skeletal muscle in elderly women [20]. Taken together, it can be concluded that paraspinal muscle measures are not reliable substitutes for DXA-derived ASMI in this clinical population. Interestingly, indices of the iliopsoas muscle were more strongly related to ASMI than paraspinal muscle metrics. This difference can be explained in two separate ways. One is that the iliopsoas muscle is relatively preserved in a localized degenerative change of the lumbar spine, and thereby it is a better marker of general musculoskeletal health [21]. Another explanation could be that, as one of the largest hip flexors and core muscles, the psoas major has both metabolic and functional characteristics closely related to the limb muscles. It demonstrates closer physiological coupling with the appendicular skeletal muscle. In contrast, paraspinal muscles are primarily subjected to localized spinal conditions and may show alterations independent of systemic ASM levels [22]. Longitudinal data from the Health ABC Study are in line with our results. Both the psoas and appendicular muscles display parallel age-related decline in size (psoas: −0.4%, quadriceps: −3.3%, hamstrings: −5.9% over 5 years), whereas paraspinal muscles exhibit distinct compensatory hypertrophy (+4.3%) in the eighth decade of life. These results indicate that the psoas muscle is physiologically more associated with the musculature of the extremities than the paraspinal muscles [23]. Steroids are powerful anti-inflammatory drugs, and epidural steroid injection is one of the most popular interventions for spinal pain management [24]. However, axial pain is affected not only by neuroinflammation but also by the decline in function of the skeletal muscle system [25]. One study has shown that handgrip strength is a major factor related to the clinical outcomes of epidural steroid injection for low back pain in the elderly population [26]. The decrease in grip strength might reflect the lack of sufficient musculoskeletal functional reserve to endure axial loading, which could be a factor in the persistence of spinal symptoms. However, in this study, handgrip strength corresponded with no significant differences in lumbar muscle volume indices. Therefore, our results point toward the necessity of functional biomarkers that can more precisely indicate paraspinal muscle performance. Besides normal sarcopenia, sarcopenic obesity (SO) has recently been recognized as a separate clinical phenotype characterized by increased fat mass and reduced muscle mass and function at the same time [27]. In our present single-center study, sarcopenic subjects had significantly lower BMI and fat mass percentage (FM%) compared to non-sarcopenic counterparts (Table 1). These results suggest that this particular cohort was made up of non-obese sarcopenics. Generally, the frequency of SO is much lower than that of pure sarcopenia without associated obesity. For example, the Dutch Lifelines cohort discovered a 1.4% prevalence of SO in women, while Korean data indicate a much wider variation (0.8% to 22.3%) depending on the diagnostic cut-offs used [28,29]. Therefore, the absence of SO individuals in our sample is not a matter of surprise and corresponds to well-known epidemiological patterns. Among other limitations, this research is constrained by a small number of participants (n=58) and single-center recruitment, which hampered the ability to generalize the findings. Post hoc power analyses indicate that the research was well suited to detecting large effects (Cohen’s d=0.74), in which case smaller differences between the sarcopenic and non-sarcopenic phenotypes could have been missed. Furthermore, the findings pertain to elderly females only and may not be indicative of males or wider community-based cohorts. As for the imaging parameters, the use of a 5-mm slice thickness for CT acquisition likely caused partial volume artifacts, which in turn could have affected the accuracy of the 3D volumetric reconstructions and automated delineation of muscle boundaries. Although thinner sections might enhance the accuracy of segmentation, that would require much higher radiation doses to keep the same level of image quality [30]. Therefore, the 5-mm slice thickness is a clinically justified optimization. Moreover, 3D automated segmentation by TotalSegmentator treats the iliopsoas as one unit, even though traditional 2D methods usually split the psoas major. The resulting volumes do not line up exactly when compared across techniques, and muscles like the iliacus get misaligned in the process. Consequently, this mismatch might create consistent errors in measurements. Models that separate the psoas from iliacus or include manual adjustments should be explored to tighten up results. Conclusions Iliopsoas muscle parameters corresponded considerably more with ASMI as compared to paraspinal muscle metrics. While AI-driven 3D muscle volumetry has the potential to, at least, help in reducing inter-rater variability, it is important to remember that CT-based assessments of the lumbar musculature do not directly replace well-established sarcopenia markers such as ASMI and grip strength. As a result, more research is needed to identify particular biomarkers of back muscle function that contribute to frailty. Key Points Artificial intelligence-based three-dimensional volumetric analysis enables detailed quantification of lumbar muscle volume. Artificial intelligence-based three-dimensional volumetric analysis enables detailed quantification of lumbar muscle volume. The iliopsoas muscle volume index is strongly related to the appendicular skeletal muscle mass index. The iliopsoas muscle volume index is strongly related to the appendicular skeletal muscle mass index. The volume of paraspinal muscle is not a depend-able indicator of systemic sarcopenia. The volume of paraspinal muscle is not a depend-able indicator of systemic sarcopenia. There were no significant changes in lumbar muscle volumes between sarcopenic and non-sarcopenic subjects. There were no significant changes in lumbar muscle volumes between sarcopenic and non-sarcopenic subjects. Muscle variables derived from lumbar computed tomography scans have proven to be poorly per-forming alternatives to dual-energy X-ray absorptiometry for the clinical evaluation of sarcopenia. Muscle variables derived from lumbar computed tomography scans have proven to be poorly per-forming alternatives to dual-energy X-ray absorptiometry for the clinical evaluation of sarcopenia. Footnotes Conflict of Interest No potential conflict of interest relevant to this article was reported. Funding This work was supported by the 2024 Yeungnam University Research Grant. Data Availability The datasets generated during and/or analyzed during the current study are available from the corresponding author on reasonable request. Author Contributions Conceptualization: DGL. Data curation: ECL, DGL. Formal analysis: JHC, DGL. Investigation: JHC, DGL. Visualizations: ECL, DGL. Supervision: DGL. Writing–original draft: ECL. Writing–review & editing: DGL. Final approval of the manuscript: all authors. References