Prediction of whole body composition utilizing cross-sectional abdominal imaging in pediatrics

Although body composition is an important determinant of pediatric health outcomes, we lack tools to routinely assess it in clinical practice. We define models to predict whole-body skeletal muscle and fat composition, as measured by dual X-ray absorptiometry (DXA) or whole-body magnetic resonance imaging (MRI), in pediatric oncology and healthy pediatric cohorts, respectively. Pediatric oncology patients (≥5 to ≤18 years) undergoing an abdominal CT were prospectively recruited for a concurrent study DXA scan. Cross-sectional areas of skeletal muscle and total adipose tissue at each lumbar vertebral level (L1-L5) were quantified and optimal linear regression models were defined. Whole body and cross-sectional MRI data from a previously recruited cohort of healthy children (≥5 to ≤18 years) was analyzed separately. Eighty pediatric oncology patients (57% male; age range 5.1–18.4 y) were included. Cross-sectional areas of skeletal muscle and total adipose tissue at lumbar vertebral levels (L1-L5) were correlated with whole-body lean soft tissue mass (LSTM) (R2 = 0.896–0.940) and fat mass (FM) (R2 = 0.874–0.936) (p < 0.001). Linear regression models were improved by the addition of height for prediction of LSTM (adjusted R2 = 0.946–0.971; p < 0.001) and by the addition of height and sex (adjusted R2 = 0.930–0.953) (p < 0.001)) for prediction of whole body FM. High correlation between lumbar cross-sectional tissue areas and whole-body volumes of skeletal muscle and fat, as measured by whole-body MRI, was confirmed in an independent cohort of 73 healthy children. Regression models can predict whole-body skeletal muscle and fat in pediatric patients utilizing cross-sectional abdominal images.


INTRODUCTION
Body composition is of broad interest in human health, however, relevant techniques are not routinely available for use in children [1][2][3][4]. Aberrant body composition (e.g., visceral obesity, sarcopenia, sarcopenic obesity) is an important and well-established predictor of morbidity and mortality across a spectrum of adult cancers [5]. Increasingly, sarcopenia has been associated with treatment-related morbidity and mortality among children with different chronic diseases, including cancer [6][7][8][9][10].
Methods to assess body composition have inherent precision, specificity, strengths, and limitations; their specific application to children entails consideration of the practicability of the method in terms of accessibility of the imaging modality and the minimum time that a child can be expected to lie still. Dual-energy X-ray absorptiometry (DXA) and whole-body magnetic resonance imaging (MRI) can be used to precisely quantify body composition and are considered reference standards [11,12]. Lean soft tissue mass (LSTM) is the sum of body soft tissue and includes skeletal muscle, organs, and body water, but excludes fat and bone [13]. Whole-body LSTM measured by DXA is validated in both adults and children [14,15]. Although DXA is fast and has a low radiation exposure (<4 micro Sieverts), neither DXA, nor whole-body MRI is routinely acquired in children with cancer. Studies in healthy adults, adults with cancer or nonmalignant diseases have shown that data obtained from a single axial slice of an abdominal computed tomography (CT) scan may be used to predict wholebody LSTM and fat mass (FM) [16][17][18][19][20][21]. In adults, tissue crosssectional areas in the lumbar area (e.g., 3rd lumbar vertebrae, L3) are most highly correlated with whole-body values. However, such relationships remain to be characterized in pediatric populations. Abdominal CT images are routinely acquired for almost all children with solid tumors, even those without abdominal tumors, due to the routine use of fluorodeoxyglucose-positron emission tomography (FDG-PET)/CT or metaiodobenzylguanidine (MIBG)/CT scans for disease evaluations, and these images could be utilized to assess body composition with specificity and precision.
We aimed to assess correlations between cross-sectional area of skeletal muscle or total adipose tissue at all lumbar vertebral levels (L1-5) and whole-body LSTM or FM, respectively, as determined by concurrent DXA in prospectively recruited pediatric oncology patients. We determine if clinical covariates (age, sex, height, ethnic group) significantly contribute to optimal multivariable linear regression prediction models for LSTM or FM. We also assess correlations between cross-sectional area of SM (skeletal muscle) or TAT (total adipose tissue) at all lumbar vertebral levels with whole-body skeletal muscle and adipose tissue volumes determined by whole-body MRI in a healthy pediatric population.

SUBJECTS AND METHODS Prospective pediatric oncology cohort
Research ethics boards of the two participating centers approved this study (University of British Columbia, Vancouver, Canada: Children's and Women's Research Ethics Board, H12-03232; University of Alberta, Edmonton, Canada: Health Research Board of Alberta-Cancer, HREBA.CC-16-0102). We prospectively recruited children and adolescents (≥5 to ≤18 years) with cancer between April 2013 and December 2017 who were undergoing an abdominal CT scan as a routine part of disease staging or re-evaluation. This age range was included because the methodology for determining whole-body LSTM and FM from DXA scan data has been validated in this group [15]. Patients were eligible for study enrollment at any point during their disease trajectory.
Eligible patients had a confirmed or suspected malignancy and had planned, clinically indicated abdominal CT imaging (CT abdomen, FDG-PET/CT, or MIBG/CT). Patients were a minimum of 7 days since chemotherapy, 2 weeks from surgery, 6 weeks from radiation (6 months from total body irradiation or craniospinal irradiation), and 3 months from stem cell transplantation with no evidence of graft versus host disease. Patients had adequate functional status (Lansky or Karnofsky ≥ 50) and renal function (creatinine clearance/radioisotope GFR > 70 ml/min/1.73 m 2 ) or serum creatinine within normal limits for age/sex). Pregnant or lactating patients and those with uncontrolled infection were ineligible. Patients with BMI >35 kg/m 2 were excluded due to limitations of body composition assessments in patients with severe obesity [22,23]. Weight and height were measured at time of DXA scan. Tanner stage was assessed by each patient's treating pediatric oncologist. Ethnicity was recorded as per patient or parent self-report. Triceps skinfold and mid-upper arm circumference measurements were collected on the date of DXA by oncology dieticians.
Imaging procedure and analysis. Patients provided consent, and assent when applicable, to obtain a DXA scan within two weeks of their scheduled clinical CT scan. As contrast media used in CT imaging can alter DXA readings, DXA scans were conducted either prior to CT, or if after CT then after >10 contrast elimination half-lives had elapsed (i.e., 20 h). DXA images were acquired on Hologic Discovery A system (Site 1 British Columbia: % coefficient of variability (% cv) <2% for both whole body LSTM and FM) or General Electric LUNAR iDXA (Site 2 Alberta: % cv for LSTM 0.39% and FM 0.99%) systems. All CT studies were conducted in the context of routine care and were contrast-enhanced and acquired in porto-venous phase. Overall imaging parameters included acquisition at 0.625-3.75 mm, tube voltages 80-120 kV and tube current 42-224 mA; the latter being scaled to the size of the child [24]. (Supplementary Table S1) CT series were from either a stand-alone contrast-enhanced abdominal CT study, fluorodeoxyglucose-positron emission tomography (FDG-PET)/CT study or MIBG/CT study.
CT images, in DICOM format, were analyzed using Slice-O-matic version 5.0 image analysis software (www.tomovision.com) [25]. Total abdominal skeletal muscle cross-sectional area (anatomically defined rectus abdominis, oblique and lateral abdominal, quadratus lumborum, psoas, and paraspinal muscles) and total visceral plus subcutaneous adipose tissue were manually segmented [12]. Segmentation tools were set on tissue-specific radiodensity thresholds (−29 to +150 Hounsfield Units (HU) for skeletal muscle; −150 to −50 HU for visceral adipose tissue and −190 to −30 HU for subcutaneous adipose tissue [18,26]. The cross-sectional area of skeletal muscle (SM) and total adipose tissue (sum of visceral and subcutaneous adipose tissue), from CT images obtained at the mid-point of each lumbar vertebrae (L1-L5) were determined. A single observer who was blinded to clinical characteristics and trained in soft tissue (musculoskeletal and adipose tissue) radiology conducted all measurements. The intra-observer coefficient of variability for single measures of muscle and adipose tissue cross-sectional area was 1.3% and 2.1%, respectively.

Retrospective healthy pediatric cohort
We undertook a secondary analysis of whole-body contiguous MRI imaging data from a healthy pediatric cohort, earlier described by Shen et al. [27]. Briefly, 73 healthy children were scanned with T1-weighted, spin-echo sequence, TR/TE 210 ms/17 ms, contiguously with 1 cm slice thickness on a 1.5 T General Electric system (6X horizon, Milwaukee, WI). Cross-sectional MR images were segmented using image analysis software (SliceOmatic, Tomovision Inc.) into SM and adipose tissue to determine the influence of between-slice intervals on whole body composition prediction, estimated as volumes of all tissue compartments (SM and adipose tissue compartments). Demographics of this US-based cohort were previously described (61% male, ages 5-17 y: Black, White, and Hispanic races/ethnicities) and all had BMI < 35 kg/m 2 .
Data analysis. Baseline clinical characteristics of study participants (age, sex, weight, height, ethnic group, Tanner stage, and anthropometric measurements: triceps skinfold and mid-upper arm circumference) were summarized as medians, ranges, and proportions. Correlations between cross-sectional SM and TAT areas obtained at the mid-point of each lumbar vertebrae (L1-L5) and whole-body LSTM and FM, respectively, from DXA scan were determined by Pearson's correlation coefficient (r). For MRI studies, landmarking was done at the intervertebral disks. Univariable regression models were constructed to describe relationships between the cross-sectional area of SM (or TAT) at each of the lumbar vertebral levels with whole-body LSTM or FM from concurrent DXA studies and whole-body volumes in MRI studies. For each univariable model, the R value, coefficient of determination (R 2 ) and standard error of the estimate were determined. Before undertaking linear regression, all clinical variables were explored for collinearity using the Durbin-Watson test. Potential independent variables explored for contribution to the multivariable model included age, sex, height, and ethnic group. For variables with high collinearity, the variable with lower correlation to the whole-body value was excluded from the analysis.
Multivariable linear regression analysis was used to explore the relationships between total-body body composition values and lumbar cross-sectional area of SM or TAT measured by CT or MRI at each lumbar level. The strongest prediction equation was developed using general linear regression model analyses. All independent variables with significant correlation to whole body composition were included in the initial multivariable model, and those coefficients with statistically significant contributions to the model were identified. The coefficient of determination (R 2 ) described the goodness of fit of the model. The difference between model-predicted and DXA-determined LSTM and FM was assessed using residual plots and calculation of the root mean squared prediction error. Statistical analyses were completed in SPSS 25.01 with two-tailed tests of significance (P < 0.05).

RESULTS
Ninety-one pediatric cancer patients were enrolled, four patients were excluded for BMI ≥ 35 kg/m 2 and seven were excluded if the routine abdominal CT (n = 2) or study DXA (n = 5) were not completed. Eighty patients were included in the primary analytic cohort with a range of oncologic diagnoses. (Table 1) Median patient age was 13.6 y (range 5.1-18.4 y), 11% of the cohort was underweight (BMI < 5%ile) and 24% were overweight or obese (BMI > 85%ile). Representative segmented CT images and DXA scans across a spectrum of body composition among study participants are shown in Fig. 1. Clinical CT abdominal imaging modality was FDG-PET/CT in 78%, and all had a study DXA within 14 days (median 1 day; interquartile range 1-6 days).
Cross-sectional SM and TAT areas at the midpoint of each lumbar vertebra (L1-L5) correlated strongly with both whole-body LSTM and FM, respectively, as determined on DXA ( Table 2). The cross-sectional area of both skeletal muscle and TAT at all lumbar vertebral levels (L1-L5) were correlated with whole-body LSTM (r 2 = 0.896-0.940) and FM (r 2 = 0.874-0.936) (p < 0.001). Representative plots of the simple linear relationships are illustrated in Fig. 2.
Optimal multivariable regression models for whole-body LSTM and FM using CT cross-sectional areas at all lumbar vertebral levels ( Table 3, Supplementary Table S2) are shown. Due to the collinearity of height and age (Pearson correlation r = 0.940), only height was included in the models. The Durbin-Watson test showed no first-order auto-correlations detected in final model variables, so no transformation was required. Multivariable regression models including cross-sectional areas (SM or TAT) and clinical covariates (height, sex, ethnic group) were tested. Fig. 1 Representative CT and DXA images illustrating variations in fat mass in pediatric oncology patients. For each imaging modality, three female study participants aged 12-13 years are illustrated. In each panel, the patient at left is below the 3rd percentile for fat mass, patient at center is at the 50th percentile for fat mass and patient at right is above the 97th percentile for fat mass. A Segmented CT images. Upper panel: Coronal images were landmarked at the proximal edge of the femoral heads and are shown only for illustrative purposes. Lower panel: Segmented cross-sectional tissue analysis was completed using axial images at the 3rd lumbar vertebra in this example. B Dual energy X-ray images.
Height was significantly, independently predictive in all models at all lumbar levels, and included in final models. Sex was not significant at any vertebral level for the association between lumbar muscle cross-sectional area and whole body LSTM. Female sex associated with higher FM in the upper lumbar region, but there was no sex effect at L4 or L5 (Table 3). Residual plots confirmed normality and the root mean square prediction error between model-predicted and DXA-determined whole body LSTM and FM were determined (Fig. 3).
Among 73 healthy children with contiguous MRI data, crosssectional SM and TAT areas at the intervertebral space below each lumbar vertebra (L1-L5) correlated strongly (p < 0.001, all levels) with both whole body skeletal muscle and fat volumes as determined by MRI (whole-body muscle (R 2 = 0.826-0.897) and fat (R 2 = 0.883-0.934) ( Table 2). Multivariable regression models including cross-sectional area of SM or TAT, height, and sex were tested and reported (Table 3). Similar to the results presented for the pediatric cancer patient cohort, the addition of height significantly improved all models. As seen for LSTM, sex was not significant at any vertebral level for the association between lumbar muscle cross-sectional area and whole body skeletal muscle volume by MRI. For fat volume, female sex was associated with higher FM at upper lumbar levels, but there was no sex effect below L4/L5 and the impact of sex was reversed at L5/S1.

DISCUSSION
Lumbar body composition is highly correlated to whole body composition in our two cohorts of pediatric patients. Crosssectional area of total SM and TAT in single axial images correlated with whole-body values for DXA-derived LSTM and FM, as well as MRI-skeletal muscle and fat volumes. The correlations were already very high at the univariable level for both LSTM (R 2 = 0.896-0.940 for L1-5) and FM (R 2 = 0.874-0.936 for L1-5), and further improved by the addition of height to both models, and sex only for whole body FM at high lumbar levels (LTSM R 2 = 0.946-0.971; FM R 2 = 0.930-0.953). No single lumbar vertebral level was clearly optimal as adjusted R 2 values at all levels were   excellent and as good or better than those previously published for adult prediction models [16,17,19,20]. These prediction models will facilitate future retrospective and prospective analyses of pediatric body composition. Our results are timely, for the reason that great strides are being made in high-precision automated CT image segmentations and these methods are on the cusp of application in clinical practice [28,29]. The observed correlations are a measure of body proportions i.e., how does lumbar body composition relate to that of the whole body? Shen et al. [30] provide a foundational work using a diverse sample of 329 adults, showing that single abdominal skeletal muscle areas were highly correlated with total body muscle (r = 0.924 at L4-L5 intervertebral space) and single slice to whole body correlations for adipose were likewise very strong (r = 0.963 at L4-L5 intervertebral space). Shen et al. included participants of Asian, Black, and White races, male and female, BMI < 18.5 to >30 kg/m 2 , ages 20-70 years. However, the addition of subject sex, age, ethnicity, scanning position, BMI, and waist circumference in the model only barely affected the relationship between lumbar tissue cross-sectional areas and whole body values. Other researchers have repeatedly shown that tissue areas in specific lumbar regions are highly correlated with whole-body values and that this is independent of sex, age, height, and BMI [17,19,20]. Our findings suggest that the proportional body composition is already present at 5-18 y of age, with the slope, intercept, and r 2 values highly similar to those reported in adult patients. In children, sex did not significantly contribute to body composition model prediction of DXA-LSTM or MRI-derived muscle volume. This does not negate the fact that body composition is sexually dimorphic; indeed, sex differences in LSTM and FM are present at birth, with girls having a higher percentage of body fat and FM and lower LSTM than boys. Sex effects on body composition become evident during adolescence, typically at age >11 y when males have higher LSTM and FFM compared to females [31]. However, we observe that the relationship between lumbar crosssectional area and total body composition in children is remarkably similar to adults.
In our cohort, 25% of patients were overweight or obese based on BMI percentile, similar to other recent pediatric oncology cohorts in high-income countries and prevalence observed in LSTM, lean soft tissue mass; SM, skeletal muscle cross-sectional area (cm 2 ); TAT, total adipose tissue cross-sectional area (cm 2 ); L1, first lumbar vertebral level; L5, fifth lumbar vertebral level; S1, first sacral vertebral level; MSE, mean square error. Fig. 3 Representative multivariable models in pediatric oncology cohort. All models are for 3rd lumbar vertebral tissue areas (n = 80).
A Prediction of whole body lean soft tissue mass (LSTM) (utilizing L3 skeletal muscle area). B Prediction of whole body fat mass (utilizing L3 total adipose tissue area). C Residual plots for multivariable prediction of whole body lean soft tissue. D. Residual plots for multivariable prediction of whole body fat mass.
healthy Canadian children [32]. In adult cancers, sarcopenia and sarcopenic obesity have been strongly associated with toxicity and inferior survival outcomes across many histologic types and stages [5,33]. Assessment of sarcopenia in pediatrics has been limited to small studies utilizing a range of surrogate markers including psoas muscle area or skeletal muscle index at lumbar vertebral levels, paraspinal muscle areas or intramuscular adipose tissue [8,9,34]. Among pediatric cancer patients, differences in body composition may be important prognostic factors for both morbidity, mortality, and drug-related toxicity [35,36]. Patients undergoing cancer therapy and childhood cancer survivors have been shown to have higher FM, FM index, and percent body fat than controls [37,38]. Pharmacokinetic data suggest differential exposures to active drugs in obese compared to non-obese children [39], which may impact toxicity outcomes. Among children with acute lymphoblastic leukemia, obesity is associated with an increased risk of hepatotoxicity and pancreatitis and a higher rate of minimal residual disease at the end of induction [40,41]. In children with high-risk neuroblastoma or hepatoblastoma, loss of skeletal muscle mass, as assessed by cross-sectional psoas muscle area, was associated with inferior survival outcomes [6,42]. An ability to routinely assess whole body composition in pediatric cancer patients is an essential step in understanding its impact on toxicity and survival. In addition to understanding toxicity, routine access to body composition metrics may improve drug dosing strategies in pediatrics. Chemotherapy doses in childhood are based on body surface area (BSA), which is easily estimated using height and weight data. However, BSA is poorly correlated with body composition and may not be a good surrogate for understanding drug metabolism and toxicity. Drug-related toxicity may be driven by chemotherapy dose per unit of either lean or fat mass, which could vary drastically between two children with the same BSA.
Limitations of our study include sample size and heterogeneity of our cohort with respect to cancer primary site and time since diagnosis. Accrual was limited to cancer types routinely requiring cross-sectional imaging as part of routine care and to children ≥5 years without morbid obesity (BMI < 35). Sample size and level of ethnic diversity in western Canada also limited our ability to fully assess the impact of ethnic background on body composition. Although significant differences in lean tissue, muscularity, and fat mass have been characterized among White, Black, and Asian populations, it is unclear if the proportional relationship between lumbar tissue cross-sectional area and whole body composition varies by ethnic group.
Use of cross-sectional imaging to predict whole body composition in pediatric oncology will facilitate future research to assess changes in body composition over time and to further determine the impacts of body composition on morbidity and mortality in childhood cancer, as well as in other pediatric cohorts. There is a relative paucity of sarcopenia-related research in children, and future prospective studies should optimally include age-appropriate measures of muscle function in addition to body composition assessment [10]. Validation of our models in other pediatric cohorts may facilitate an improved understanding of the impact of obesity and sarcopenia on long-term health and toxicity outcomes.

DATA AVAILABILITY
De-identified data generated by this project are available to investigators upon request.