Abstract
Body composition analysis (BCA) is useful for evaluating nutritional status, general health, and predicting patient outcomes. Computed tomography (CT)-based BCA has become increasingly popular, with research often focusing on single-transaxial-slice measurements at the L3 vertebral level to estimate volumetric tissue properties. This method assumes the same tissue distribution in all transaxial slices across the body, potentially leading to inaccuracies in clinical assessments.
This retrospective study investigates the reliability of single-slice BCA compared to volumetric measurements across the torso in a cohort of 237 adults. Using the GA-Net architecture for automatic tissue segmentation, we analyzed tissue distribution at 25 craniocaudal levels. Our results reveal significant variability in tissue distribution, with cross-sectional areas at L3 poorly correlating with 3D tissue volumes. For instance, we observed a visceral adipose tissue (VAT) volume difference of over 2,900 cm3 between two subjects with similar L3 VAT areas. These findings indicate that single-slice BCA, particularly at L3, is insufficient for accurate body composition analysis.
This study underscores the need for more precise, volumetric approaches to BCA, which would enhance the accuracy of clinical assessments, disease prognostication, and treatment planning.
Keywords: Body Composition Analysis (BCA), Volumetric Analysis, Subcutaneous Adipose Tissue (SAT), Visceral Adipose Tissue (VAT), Skeletal Muscle Tissue (SMT)
1. Introduction
Body composition analysis (BCA) offers valuable insights into nutritional status and general health, as well as useful information for prognostication of clinical outcomes and treatment planning. Over the past decade, computed tomography (CT)-based body composition research has gained popularity, with extensive investigations into various tissue biomarkers, including area, volume, and attenuation of abdominal and paraspinal skeletal muscles, as well as of abdominal subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT)1.
Current methods of BCA using CT often rely on a single slice (with the L3 mid-vertebral level most commonly used as a reference) or only several slices to estimate the overall properties of the tissues of interest. This approach assumes that the properties obtained from selected slices represent the same body habitus characterization that can be derived from the entirety of the tissues of interest 3-dimensionally, which can lead to potential inaccuracies. The assumption that SAT, VAT, and skeletal muscle tissue (SMT) occur at the same proportion at different slices as at L3 throughout the body is questionable, has never been demonstrated, and may not be valid for most subjects, rendering these methods unreliable for various applications. Furthermore, the optimal selection of slices for assessing specific tissues remains a topic of debate1,2,3.
With the global increase in the aging population and the rising prevalence of sarcopenia and sarcopenic obesity, both of which are strongly linked to decreased overall survival, there is compelling evidence that abnormal body composition (BC) phenotypes predict adverse outcomes. Therefore, it is crucial to transition BCA from the research domain to clinical practice. This study aims to demonstrate the variability in tissue distribution across individuals with similar anthropometric measures, emphasizing that single-slice measurements do not accurately represent comprehensive body composition analysis. It also explores ways of defining body habitus groups based on body composition features. We believe our study is a pivotal first step toward translating BCA into clinical practice3.
2. Methods
This retrospective single-center study was conducted following approval from the Institutional Review Board at the University of Pennsylvania, along with a Health Insurance Portability and Accountability Act waiver.
Cohort Selection and Classification:
Our initial cohort consisted of 245 adult subjects who previously underwent 18F-fluorodeoxyglucose (FDG)-positron emission tomography/computed tomography (PET/CT) imaging for clinical or research purposes, of which 5 were subsequently excluded due to the absence of body mass index (BMI) data. The remaining 240 subjects were then classified into four BMI subgroups: obese, overweight, normal weight, and underweight, based on currently utilized guidelines4. The underweight group, which comprised only 3 individuals, was subsequently excluded due to insufficient sample size. Thus, the final study cohort included 237 subjects, with 141 females and 96 males, as shown in Table 1. The mean age for females is 51.2 years (range: 21–86), for males 48.2 years (range: 21–75), and overall 49.9 years (range: 21–86).
Table 1.
BMI-based classification of study subjects.
| Females | Males | ||||
|---|---|---|---|---|---|
| Normal | Overweight | Obese | Normal | Overweight | Obese |
| 70 | 30 | 41 | 26 | 47 | 23 |
Definitions of Body Regions and Tissues:
It is important to first precisely define the body region and tissue regions of interest for meaningful quantification and standardized usage to study diseases and compare methods in a common frame of reference. In our work, we first produced computationally feasible and anatomically meaningful definitions of the body regions of interest. Similarly, since SAT and VAT components of adipose tissue and SMT are interconnected 3-dimensionally with many open areas (no anatomic boundaries), these tissues are defined precisely so that when there are open boundaries between tissue regions, we establish computationally feasible definitions5,7.
Torso: The region of the body comprised of the thorax, abdomen, and pelvis, which we defined as extending from the axial slice along the superior aspect of the T1 vertebra to the axial slice through the inferior aspect of the ischial tuberosities.
Subcutaneous adipose tissue (SAT): Adipose tissue superficial to skeletal muscle (and sometimes bone or cartilage) and deep to skin tissue, currently excluding adipose tissue in the axillary and ischiorectal fossa regions. SAT extends from the superior aspect of the thorax body region to the inferior aspect of the pelvis body region. In inferior pelvis, when midline anatomical structures connect left and right halves of the pelvis, then the adipose tissue in between the left and right pelvic/proximal thigh muscles is not included. More inferiorly, when no midline anatomical structures connect left and right halves of the pelvis, then the adipose tissue in between the left and right pelvic/proximal thigh muscles is included.
Visceral adipose tissue (VAT): Adipose tissue deep to body wall skeletal muscle, bone, and cartilage. VAT extends from the superior aspect of the thorax body region to the inferior aspect of the pelvis body region. In the thorax, this includes adipose tissue posterior to anterior chest wall musculature that surrounds the internal mammary vessels. In the abdomen, this includes adipose tissue internal to fascial tissue that extends anteriorly to the skin in the umbilical region as well as when any external hernia (i.e., a hernia that extends through the body wall skeletal musculature) is visualized. In the pelvis, this includes adipose tissue in between the psoas and iliacus muscles when widely separated and excludes such adipose tissue when closely opposed. Also, ischiorectal adipose tissue external to levator ani musculature is currently excluded.
Skeletal muscle tissue (SMT): Skeletal musculature excluding macroscopically visible adipose tissue. SMT extends from the superior aspect of the thorax body region to the inferior aspect of the pelvis body region.
Segmentation:
A pre-trained deep learning network, GA-Net (geographical attention neural network) in our recent work6, was used for segmenting tissues of interest (SAT, VAT, SMT) automatically in the 3D image. GA-Net is a tool designed for BCA with a dual-decoder schema of tissue decoder and area decoder. The tissue decoder segments the body composition tissues, while the area decoder segments the body areas as an auxiliary task. The features of body areas and body composition tissues are fused through a soft attention mechanism to gain geographical attention relevant to the body tissues. The GA-Net has been evaluated on 50 low-dose unenhanced CT images from an in-house dataset with outstanding DICE coefficients (%) for SAT (95.44±2.07), VAT (95.91±1.73), SMT (93.70±3.26), and skeleton (Sk) (97.25±1.50). More details can be found in6. In our current study, GA-net was first used to create auto-segmentations of SAT, VAT, and SMT. Subsequently, mending of the auto-segmentation as needed was performed by a trained operator using CAVASS8 software under the supervision of a board-certified radiologist (DAT) with expertise in thoracic and abdominopelvic radiology (Figure 1).
Fig. 1.
shows representative slices at the L3 vertebral level with autosegmentation results for VAT (A), SAT (B), and SMT (C).
Selecting Vertebral Levels:
The process began with sagittal reformation of the axial CT scan images using CAVASS software. Manual selection of the central point of each vertebral body, as well as osseous landmarks including the symphysis pubis and ischial tuberosities, was performed from the 1st cervical vertebra (C1) superiorly to the ischial tuberosities inferiorly. The following anatomic levels were identified: thoracic vertebrae (T1-T12), lumbar vertebrae (L1-L5), sacral vertebrae (S1-S5), 1st coccygeal vertebra (Co1), mid-level through the symphysis pubis (SP), and level through the inferior aspect of the ischial tuberosities (IT). The corresponding slice number on the original axial plane CT scan was identified for each point selected on the sagittal reformation.
Measurement and analyses:
We defined the following area and volume measures for the 3 tissues of interest SAT, VAT, and SMT, and generated summary statistics over our subject population:
Vr(s): Volume of tissue r in body region B for subject s. In this paper, B = torso body region.
Ar(L, s): Area of tissue r (one of SAT, VAT, SMT) at slice level L in B for subject s.
Ar(L3, s): Area of tissue r (one of SAT, VAT, SMT) at slice level L = L3 in B for subject s.
Our rationale for the new measures is as follows: (i) Volumetric measure Vr(s) would capture information about SAT-VATSMT quantity and body-wide distribution better than the area measure Ar(L3, s). (ii) In the same vein, the full profile of Ar(L, s) as a function of slice location L craniocaudally would be more useful than Ar(L3, s).
3. Results
To illustrate the fact that Ar(L3, s) is not adequate for BCA, we show in Table 2 measures for representative subjects with very similar cross-sectional areas at the L3 vertebral level for the three different body tissues (SAT, VAT, SMT) but with considerable variation in their tissue volumes. For instance, there is a difference of over 2,900 cm3 in the torso volume of VAT between Subject 3 and Subject 4 even though they are both within the same normal weight BMI group.
Table 2:
Comparisons among representative subjects with similar cross-sectional tissue areas (cm2) at the L3 vertebral level for SAT, VAT, and SMT but with very different volumes in the torso (cm3).
| ASAT(L3, s) (cm2) | VSAT(s) (cm3) | AVAT(L3, s) (cm2) | VVAT(s) (cm3) | ASMT(L3, s) (cm2) | VSMT(s) (cm3) | |||
|---|---|---|---|---|---|---|---|---|
| Subject 1 | 57.07 | 3971.15 | Subject 3 | 19.07 | 3607.18 | Subject 5 | 94.26 | 7773.26 |
| Subject 2 | 57.95 | 5515.60 | Subject 4 | 19.08 | 694.34 | Subject 6 | 94.61 | 6194.41 |
As another illustration, we generated “chess-piece” plots of anatomic level in the torso (y-axis) vs. tissue cross-sectional area (x-axis) in 3 representative study participants to visually illustrate the variability of craniocaudal distribution of SAT throughout the torso within the overweight BMI category. The x-axis data were intentionally plotted symmetrically with 0 at the center to portray the general shape of the body. Figure 2 illustrates the distribution of SAT. These plots reveal significant variability in tissue distribution within the same BMI group. Specifically, the L3 cross-sectional area showed variability among subjects with similar anthropometric measurements, indicating that it cannot reliably estimate the 3D volume of any tissue.
Figure 2:
ASAT(L, s) for three different female subjects from the overweight BMI group are depicted. SAT distribution across the torso for the three subjects shows considerable variation craniocaudally.
4. Discussion
CT-based body composition analysis (BCA) has become a valuable tool for phenotyping patients and prognosticating outcomes in various clinical scenarios. Multitudes of studies have utilized the L3 vertebral level as a standard reference for quantifying skeletal muscle mass, subcutaneous adipose tissue (SAT), and visceral adipose tissue (VAT) as proxies for whole-body composition. Several alternative vertebral levels have less frequently been proposed in the literature for such assessments.
Nattenmüller et al. studied muscle and adipose tissue distribution at L2/3 on CT before and after first-line chemotherapy in 200 lung cancer patients. Their results showed that a high pre-treatment VAT/SAT area ratio was a predictor of poor survival (HR 1.27, p = 0.0008), and high pre-treatment muscle attenuation was predictive of improved survival (HR 0.93, p < 0.05). A decreased skeletal muscle index after chemotherapy was associated with poor survival (HR 1.06, p < 0.001). Very sarcopenic patients received fewer chemotherapy cycles and had worse survival9. Kang et al. retrospectively analyzed 85 breast cancer patients who underwent surgery and chemotherapy. Body composition was assessed using CT scans at the L3 vertebral level. Their findings revealed that a lower skeletal muscle index (SMI) at diagnosis was significantly associated with increased mortality (P = 0.019) and a higher incidence of grade 4 neutropenia (P = 0.008). Additionally, patients with lower SAT at diagnosis showed a trend toward higher mortality, though this was not statistically significant. Overall, lower skeletal muscle mass at diagnosis was strongly associated with increased mortality and chemotherapy-induced complications in breast cancer patients10.
These studies, among others, underscore the importance of leveraging the extensive data obtained from CT scans to screen patients, monitor treatment responses, and predict the prognosis in various clinical scenarios. However, inherent limitations exist in the current CT-based BCA methods. Notably, the reliance on a single slice as representative of the threedimensional distribution of adipose tissue and muscle in the body fails to account for inter-patient variability and conditionspecific differences in adipose tissue and muscle distribution.10
Some studies have assessed the correlations of CT-based BCA performed at various vertebral levels and tissue volumes. For instance, Tong et al. investigated the potential of slices at different anatomical locations in the thorax, abdomen, and thigh to serve as proxies for volumetric measurements. Their findings demonstrated that the Pearson correlation coefficient (PCC) between thoracic adipose tissue volume and thoracic adipose tissue slice area was highest at the T8 level for SAT (PCC = 0.97) and at the T7 level for VAT (PCC = 0.86). In contrast, the correlation between thoracic adipose tissue volume and abdominal slice adipose tissue area was more modest for both SAT (PCC = 0.73) and VAT (PCC = 0.75)11. Similarly, Cheng et al. reported that VAT area measurements at the L2/L3 level showed the strongest correlation with total abdominal VAT volume compared to other slices (r = 0.98, P < 0.001)12. Another study identified the T12-L1 level as the site of maximum correlation for SAT volume of the abdominal region and the L3-L4 level for VAT volume of the abdominal region2. Importantly, these findings highlight that the optimal anatomic slice locations for estimating SAT and VAT volumes vary between studies.
Our study evaluates the reliability of single-slice BCA compared to volumetric measurements across the torso in a cohort of 237 adults. The findings reveal significant variability in cross-sectional area measurements of body composition, which did not consistently correlate with tissue volume. Furthermore, confusion matrix analysis highlighted discrepancies between BMI-based classifications and those derived from volumetric measurements. Notably, neither BMI-based grouping nor single-slice measurements at the L3 level aligned with classifications determined using volumetric data. We also demonstrated a representative example of the variability in tissue distribution among individuals of the same gender and BMI group.
As such, these findings indicate that reliance upon single-slice measurements as representative of tissue volumes for the purpose of body composition analysis (BCA) can be misleading. More precise and reproducible methods are needed to capture and utilize valuable yet underexplored data to advance this critical area of CT-based BCA toward clinical applicability.
5. Conclusion
CT-based body composition analysis (BCA) shows great potential for utility in clinical practice. However, performance of BCA upon a limited number of slices or from a single CT slice, most commonly at the L3 vertebral level, may not accurately represent the properties of the tissues of interest throughout the body, and as such may lead to misleading results. In our study, we analyzed SAT, VAT, and SMT in 237 patients by measuring the cross-sectional area at 25 different craniocaudal levels based on skeletal anatomic landmarks, and compared these with the 3D volumes of these tissues in the torso. We observed significant variability in cross-sectional areas that did not consistently correlate with tissue volumes. This suggests that using a single slice as a reference for body composition analysis and correlation with disease outcomes may be unreliable. Our results underscore the need for more precise and comprehensive methods for CT-based BCA to ensure its effective implementation in clinical practice.
The fundamental issue in BCA is: How to define body habitus and derive measures for its quantitative description so that individuals can be classified into a desired number of body habitus groups? We believe that the shape of the profiles A′(L, s) and Ar(L, s) contain information needed for such a definition.
Acknowledgments
The research reported in this paper is supported by an NCI grant R01CA255748.
References
- 1.Tong Angela, and Magudia Kirti. “One Step Forward in Opportunistic Screening for Body Composition.” Radiology vol. 307,5 (2023): e231003. doi: 10.1148/radiol.231003 [DOI] [PubMed] [Google Scholar]
- 2.Tong Y, Udupa JK, Torigian DA. Optimization of abdominal fat quantification on CT imaging through use of standardized anatomic space: a novel approach. Med Phys. 2014;41(6):063501. doi: 10.1118/1.4876275 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 3.Zopfs D, Theurich S, Große Hokamp N, et al. Single-slice CT measurements allow for accurate assessment of sarcopenia and body composition. Eur Radiol. 2020;30(3):1701–1708. doi: 10.1007/s00330-019-06526-9 [DOI] [PubMed] [Google Scholar]
- 4.Weir CB, Jan A. BMI Classification Percentile And Cut Off Points. In: StatPearls. Treasure Island (FL): StatPearls Publishing; June 26, 2023. [Google Scholar]
- 5.Liu T, Pan J, Torigian DA, et al. ABCNet: A new efficient 3D dense-structure network for segmentation and analysis of body tissue composition on body-torso-wide CT images. Med Phys. 2020;47(7):2986–2999. doi: 10.1002/mp.14141 [DOI] [PubMed] [Google Scholar]
- 6.Dai J, Liu T, Torigian DA, et al. GA-Net: A geographical attention neural network for the segmentation of body torso tissue composition. Med Image Anal. 2024;91:102987. doi: 10.1016/j.media.2023.102987 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 7.Liu T, Udupa JK, Miao Q, Tong Y, Torigian DA. Quantification of body-torso-wide tissue composition on low-dose CT images via automatic anatomy recognition. Med Phys. 2019;46(3):1272–1285. doi: 10.1002/mp.13373 [DOI] [PubMed] [Google Scholar]
- 8.Grevera G, Udupa J, Odhner D, et al. CAVASS: a computer-assisted visualization and analysis software system. J Digit Imaging. 2007;20 Suppl 1(Suppl 1):101–118. doi: 10.1007/s10278-007-9060-5 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Nattenmüller J, Wochner R, Muley T, et al. Prognostic Impact of CT-Quantified Muscle and Fat Distribution before and after First-Line-Chemotherapy in Lung Cancer Patients. PLoS One. 2017;12(1):e0169136. Published 2017 Jan 20. doi: 10.1371/journal.pone.0169136 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 10.Kang H, Kim I, Park H, Ahn W, Kim SK, Lee S. Prognostic value of body composition measures in breast cancer patients treated with chemotherapy. Sci Rep. 2024;14(1):23309. Published 2024 Oct 7. doi: 10.1038/s41598-024-74060-w [DOI] [PMC free article] [PubMed] [Google Scholar]
- 11.Tong Y, Udupa JK, Torigian DA, et al. Chest Fat Quantification via CT Based on Standardized Anatomy Space in Adult Lung Transplant Candidates. PLoS One. 2017;12(1):e0168932. Published 2017 Jan 3. doi: 10.1371/journal.pone.0168932 [DOI] [PMC free article] [PubMed] [Google Scholar]
- 12.Cheng X, Zhang Y, Wang C, et al. The optimal anatomic site for a single slice to estimate the total volume of visceral adipose tissue by using the quantitative computed tomography (QCT) in Chinese population. Eur J Clin Nutr. 2018;72(11):1567–1575. doi: 10.1038/s41430-018-0122-1 [DOI] [PMC free article] [PubMed] [Google Scholar]