Original Research
Pediatric Imaging
August 16, 2023

Automated Deep Learning–Based Segmentation of Abdominal Adipose Tissue on Dixon MRI in Adolescents: A Prospective Population-Based Study

Chinese (audio/PDF) and Spanish (audio/PDF) translations are available for this article's abstract.
Please see the Author Video associated with this article.
Please see the Author Interview associated with this article.
To listen to the podcast associated with this article, please select one of the following:
iTunes, Spotify, Google, or direct download.

Abstract

BACKGROUND. The prevalence of childhood obesity has increased significantly worldwide, highlighting a need for accurate noninvasive quantification of body fat distribution in children.
OBJECTIVE. The purpose of this study was to develop and test an automated deep learning method for subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) segmentation using Dixon MRI acquisitions in adolescents.
METHODS. This study was embedded within the Generation R Study, a prospective population-based cohort study in Rotterdam, The Netherlands. The current study included 2989 children (1432 boys, 1557 girls; mean age, 13.5 years) who underwent investigational whole-body Dixon MRI after reaching the age of 13 years during the follow-up phase of the Generation R Study. A 2D competitive dense fully convolutional neural network model (2D-CDFNet) was trained from scratch to segment abdominal SAT and VAT using Dixon MRI–based images. The model underwent training, validation, and testing in 62, eight, and 15 children, respectively, who were selected by stratified random sampling, with manual segmentations used as reference. Segmentation performance was assessed using the Dice similarity coefficient and volumetric similarity. Two observers independently performed subjective visual assessments of automated segmentations in 504 children, selected by stratified random sampling, with undersegmentation and oversegmentation scored on a scale of 0–3 (with a score of 3 denoting nearly perfect segmentation). For 2820 children for whom complete data were available, Spearman correlation coefficients were computed among MRI measurements and BMI and dual-energy x-ray absorptiometry (DEXA)-based measurements. The model used (gitlab.com/radiology/msk/genr/abdomen/cdfnet) is publicly available.
RESULTS. In the test dataset, the mean Dice similarity coefficient and mean volu-metric similarity, respectively, were 0.94 ± 0.03 [SD] and 0.98 ± 0.01 [SD] for SAT and 0.85 ± 0.05 and 0.92 ± 0.04 for VAT. The two observers assigned a score of 3 for SAT in 94% and 93% for the undersegmentation proportion and in 99% and 99% for the oversegmentation proportion, and they assigned a score of 3 for VAT in 99% and 99% for the undersegmentation proportion and in 95% and 97% for the oversegmentation proportion. Correlations with SAT and VAT were 0.808 and 0.698 for BMI and 0.941 and 0.801 for DEXA-derived fat mass.
CONCLUSION. We trained and evaluated the 2D-CDFNet model on Dixon MRI in adolescents. Quantitative and qualitative measures of automated SAT and VAT segmentations indicated strong model performance.
CLINICAL IMPACT. The automated model may facilitate large-scale studies investigating abdominal fat distribution on MRI among adolescents as well as associations of fat distribution with clinical outcomes.

Highlights

Key Finding
In this population-based prospective study of adolescents, a model for automated SAT and VAT segmentation from Dixon MRI showed strong quantitative performance (Dice coefficients and volumetric similarity relative to manual segmentations: range, 0.85–0.98) and qualitative performance (best possible visual score [3/3] assigned by two independent observers in to 93–99% of assessments).
Importance
The model facilitates future studies of abdominal fat distribution in adolescents and relationships of fat distribution with other obesity measures or clinical outcomes.
Childhood obesity has become a major global health problem that is increasing in prevalence worldwide [1]. Childhood obesity has both physical and psychologic consequences, which may persist into adulthood [2]. Early prevention of childhood obesity is therefore of the utmost importance. Anthropometric measurements have historically been used for diagnosing obesity [1]. Such measurements do not account for weight gain from skeletal muscle, nor do they differentiate subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT). Population-based studies of adults have shown that a higher amount of visceral fat is associated with a higher risk of cardiovascular disease [3], metabolic disease [4], and cancer [5]. A relationship of visceral fat with cardiometabolic risk factors has also been observed in adolescents and younger children [6, 7]. Moreover, a cohort study found significant fluctuations in fat mass during puberty [8]. Hence, accurate noninvasive tools are needed to quantify fat distribution in adolescents.
Previous research has supported the use of CT or MRI for quantification of SAT and VAT [9, 10]. The presence of ionizing radiation makes CT less suitable for use in longitudinal studies and in studies of young individuals [9]. Several MRI techniques have been proposed to visualize body fat compartments, including T1-weighted spin-echo or gradient-echo sequences as well as chemical-shift selective sequences [11]. In addition, Dixon MRI techniques, based on the difference between the resonance frequencies of protons bound in water and fat, have been proposed as a means of obtaining coregistered water and fat images, which may be applied for adipose tissue analysis [12]. Multiple large population studies, including the UK Biobank [13], the German National Cohort [14], the Cooperative Health Research in the Region of Augsburg (KORA)-MRI study [15], The Netherlands Epidemiology of Obesity Study [16], the Generation R Study [17], and the Dallas Heart Study [18], used MRI examinations to enable advanced body composition analysis.
Various automated or semiautomated techniques for image segmentation based on shape and intensity features, such as K-means clustering [19], fuzzy clustering [20], graph cut [21], statistical shape models [22], and active contour models [23], have been proposed. Deep learning methods based on trained fully convolutional neural networks (CNNs) represent state-of-the-art techniques for pixel- or voxel-wise segmentation in general [24] and for abdominal adipose tissue segmentation specifically [2529]. Langner et al. [27] achieved good performance using a standard U-Net for automated segmentations of SAT and VAT on 2D transverse slices of abdominal Dixon MRI acquisitions. However, the network also segmented the arm's adipose tissue and required the manual exclusion of such tissue for any downstream analysis. Estrada et al. [29] proposed the use of 2.5D FatSegNet, a fully automated pipeline to segment adipose tissue inside a consistent, anatomically defined abdominal region on different slice orientations (axial, coronal, and sagittal). Their method was based on a 2D competitive dense fully CNN architecture (2D-CDFNet), which promotes feature selectivity within a network by introducing maximum attention through a maxout activation unit [30]. The method showed high accuracy and reliability on abdominal Dixon MRI acquisitions from the Rhineland Study, a large prospective cohort study of participants 30 years old and older [31, 32]. However, to our knowledge, none of these methods have been tested in adolescents.
The aim of this study was to develop and test an automated deep learning method for SAT and VAT segmentation using Dixon MRI acquisitions from adolescents.

Methods

Study Sample

The present investigation was embedded in the Generation R Study, a population-based prospective cohort study of individuals followed from fetal life onward in Rotterdam, The Netherlands [17]. The Generation R Study was approved by the Medical Ethics Review Committee at Erasmus MC, University Medical Center, in Rotterdam. Written informed consent was obtained from parent(s) while the child was in utero. The Supplemental Methods provide additional detailed information on the Generation R Study.
On reaching age 13 years, children in the Generation R Study were invited to participate in a follow-up phase of the study, which was performed at Sophia Children's Hospital, a research center at Erasmus Medical Center. The children provided written informed consent for participation in the follow-up phase before their first visit to the research center after reaching the age of 13 years. For participating children, anthropometric measurements including BMI were obtained and dual-energy x-ray absorptiometry (DEXA) was performed at the first visit. A range of body composition measurements were derived from the DEXA examinations. Also at the first visit, children were recruited to undergo a whole-body Dixon MRI examination for investigational purposes. The MRI examination required the child to provide additional written informed consent and was performed during a subsequent visit to the research center. The Supplemental Methods provide additional information regarding the anthropometric and DEXA-based measurements.
As part of the Generation R Study, a total of 3041 children successfully completed the whole-body Dixon MRI examination. Of these children, 52 were excluded because the Dixon water-fat images were not satisfactory (12 with low image resolution, 15 with metal artifacts, nine with respiratory motion artifacts, and 16 with water-fat swap). The remaining 2989 children (1432 boys, 1557 girls; mean age, 13.5 years) formed the final sample of the current sample. These children were divided into the manually segmented dataset (n = 85) and the visual evaluation data-set (n = 2904) by use of stratified random sampling. The manually segmented dataset was further subdivided into the training dataset (n = 62), the validation dataset (n = 8), and the test data-set (n = 15), also by use of stratified random sampling. The model's automated segmentations were deemed to show gross errors in five children in the visual evaluation dataset, as described in the Results section. Of the remaining 2899 children in the visual evaluation dataset who did not have gross segmentation errors, 504 were selected for the visual scoring dataset, by use of stratified random sampling. Of these 504 children, 100 were randomly selected for the segmentation evaluation dataset. Of the 2899 children in the visual evaluation dataset who did not have gross segmentation errors, 79 children did not have BMI or DEXA measurements obtained during the first visit to the research center at age 13 years. The remaining 2820 children (with available BMI and DEXA measurements and with satisfactory MRI segmentation) formed the case study dataset. Figure 1 shows the study sample selection process. Table 1 summarizes the characteristics of the children in the various datasets and subsets.
Fig. 1 —Flowchart shows sample selection and study datasets and subsets. DEXA = dual-energy x-ray absorptiometry.
TABLE 1: Characteristics of Participants
Dataset and SubsetSexAgea (y)Weight (kg)Height (cm)BMI
FemaleMale
Entire sample      
All (n = 2989)1557 (52.1)1432 (47.9)13.6 ± 0.353.7 ± 11.5164.3 ± 7.419.9 ± 3.4
Underweight (n = 237)89 (37.6)148 (62.4)13.5 ± 0.740.0 ± 3.9160.9 ± 7.215.4 ± 0.4
Normal weight (n = 2138)1090 (51.0)1048 (49.0)13.6 ± 0.351.0 ± 6.7164.5 ± 7.218.8 ± 1.8
Overweight (n = 614)376 (61.2)238 (38.8)13.7 ± 0.168.9 ± 10.4165.0 ± 6.525.2 ± 3.4
Manually segmented dataset      
Training dataset      
All (n = 62)30 (48.4)32 (51.6)13.5 ± 0.651.7 ± 12.7164.3 ± 7.819.0 ± 3.8
Underweight (n = 27)10 (37.0)17 (63.0)13.5 ± 0.642.7 ± 3.8162.1 ± 5.216.2 ± 0.9
Normal weight (n = 27)16 (59.3)11 (40.7)13.5 ± 0.754.6 ± 8.4166.6 ± 9.419.6 ± 1.4
Overweight (n = 8)4 (50.0)4 (50.0)13.4 ± 0.572.2 ± 16.5163.7 ± 7.826.7 ± 4.3
Validation dataset      
All (n = 8)4 (50.0)4 (50.0)14.1 ± 0.850.8 ± 10.9158.0 ± 7.120.5 ± 5.0
Underweight (n = 3)1 (33.3)2 (66.7)14.3 ± 0.640.4 ± 3.2161.0 ± 8.915.6 ± 1.5
Normal weight (n = 3)2 (66.7)1 (33.3)14.7 ± 0.651.3 ± 3.4156.4 ± 8.021.0 ± 1.3
Overweight (n = 2)1 (50.0)1 (50.0)13.0 ± 0.065.8 ± 2.6155.8 ± 4.827.2 ± 2.8
Test dataset      
All (n = 15)8 (53.3)7 (46.7)13.4 ± 0.353.8 ± 16.2162.7 ± 6.420.1 ± 5.3
Underweight (n = 5)2 (40.0)3 (60.0)13.6 ± 0.436.1 ± 6.8158.4 ± 8.114.3 ± 1.5
Normal weight (n = 5)3 (60.0)2 (40.0)13.6 ± 0.453.8 ± 3.0164.0 ± 2.220.0 ± 0.7
Overweight (n = 5)3 (60.0)2 (40.0)13.3 ± 0.171.4 ± 9.0165.6 ± 5.926.1 ± 3.3
Visual evaluation dataset      
All (n = 2904)1513 (52.1)1391 (47.9)13.6 ± 0.453.8 ± 11.2164.3 ± 7.919.9 ± 3.5
Underweight (n = 202)76 (37.6)126 (62.4)13.5 ± 0.339.7 ± 4.0160.8 ± 7.315.3 ± 0.6
Normal weight (n = 2103)1069 (50.8)1034 (49.2)13.6 ± 0.350.9 ± 7.0164.5 ± 8.018.8 ± 1.6
Overweight (n = 599)368 (61.4)231 (38.6)13.7 ± 0.568.8 ± 10.6165.0 ± 7.525.2 ± 3.1
Visual scoring dataset      
All (n = 504)252 (50.0)252 (50.0)13.6 ± 0.459.7 ± 15.3164.4 ± 7.822.0 ± 5.0
Underweight (n = 168)84 (50.0)84 (50.0)13.5 ± 0.344.6 ± 5.3162.0 ± 7.716.9 ± 1.0
Normal weight (n = 168)84 (50.0)84 (50.0)13.7 ± 0.458.0 ± 7.1166.0 ± 7.721.0 ± 1.6
Overweight (n = 168)84 (50.0)84 (50.0)13.8 ± 0.676.5 ± 10.7165.1 ± 7.328.0 ± 2.8
Segmentation evaluation dataset      
All (n = 100)48 (48.0)52 (52.0)13.7 ± 0.559.8 ± 14.5164.4 ± 7.522.0 ± 4.7
Underweight (n = 6)3 (50.0)3 (50.0)13.8 ± 0.539.7 ± 4.5159.9 ± 7.615.5 ± 0.5
Normal weight (n = 48)26 (54.2)22 (45.8)13.6 ± 0.550.0 ± 6.5163.7 ± 7.418.6 ± 1.8
Overweight (n = 46)19 (41.3)27 (58.7)13.7 ± 0.572.7 ± 9.5166.8 ± 7.326.4 ± 2.6

Note—Data on sex are expressed as number with percentage in parentheses; remaining variables are expressed as mean ± SD. Information for case study dataset is provided in Table 3.

a
At time of MRI examination.

MRI Examinations

The MRI examinations were acquired using a 3-T MRI system (Discovery MR750w, GE Healthcare) at the previously noted research center. A whole-body Dixon sequence was acquired in the axial plane from the head to the toes. The acquisition used a 3D gradient-echo 2-point Dixon technique with chemical-shift encoding–based water-fat imaging (Lava Flex, GE Healthcare) with a slice thickness of 2.5 mm and interpolated in-plane resolution of 1.25 × 1.25 mm. Children were instructed to place their arms alongside their body during the examinations. Multiplanar water, fat, in-phase, and out-of-phase images were reconstructed. The Supplemental Methods provide additional information regarding the MRI examinations.
Before model development, one investigator (T.W., a research assistant with 7 years of experience in medical imaging research) manually annotated the lower bounds of the T12 and L5 vertebral levels on all examinations. The intervening slices (mean of 62 slices per participant) were extracted using Python 3.8.10. These extracted slices, representing the abdomen, were used for the remainder of the investigation.

Manually Segmented Dataset

To form the manually segmented dataset, children from three BMI strata were identified, to select 35 underweight, 35 normal-weight, and 15 overweight children based on Dutch reference growth curves (Growth Analyzer 3.0, Dutch Growth Research Foundation) [33, 34]. A larger number of underweight and normal-weight children were selected because initial pilot experiments indicated suboptimal segmentation performance in such children compared with overweight children. Children were also randomly selected from the three BMI strata when the manually segmented dataset was subdivided into the training dataset, validation dataset, and test dataset. The test dataset was evaluated after completion of development and optimization of the automatic segmentation method in the training and validation datasets.
SAT and VAT were manually segmented for all children in the manually segmented dataset. These manual segmentations were performed on the fat images by use of open-source software (ITK-SNAP) [35], in consensus with a radiologist (E.H.G.O., with 15 years of posttraining experience) and the previously noted research assistant (T.W.). SAT and VAT were manually segmented on every fourth slice, yielding for each participant a mean of 16 slices on which manual segmentations were performed. For the purposes of these manual segmentations, SAT and VAT were defined according to the classification criteria described by Shen et al. [36]. The manual segmentations excluded bone marrow.
Automated method for adipose tissue segmentation—A model for automatic SAT and VAT segmentation was developed using the 2D-CDFNet that formed the basis of FatSegNet [29], given that it outperformed three other commonly used encoder-decoder architectures (SD-Net [37], Dense-UNet [38], and U-Net [39]) on the abdominal adipose segmentation task [29]. The model used as input axial fat and water images reconstructed from the Dixon MRI examinations, and it output images showing the SAT and VAT segmentations along with the SAT and VAT volumes expressed in cubic milliliters. The model assumed that the input images had an abdominal range and that the fat and water images were registered. The model also cropped the child's arms, assuming that the arms were alongside the body during acquisition. The model was developed and evaluated in the training, validation, and test datasets with use of only those slices for which manual segmentations were available.
For the present investigation, the original 2D-CDFNet architecture was trained from scratch exclusively on images from the 62 children in the in-house training dataset. This process entailed initial trial-and-error experiments conducted within segments of the training dataset by use of a fourfold cross-validation approach (results not reported). Subsequently, the 2D-CDFNet was trained three times on the complete training dataset. In each of these three training rounds, the model weights were initialized differently (randomly), resulting in a slightly different model after training. The segmentation performance (based on the Dice similarity coefficient and volumetric similarity, as described later) of the three models resulting from these three training rounds (i.e., with different initializations) was evaluated in the validation dataset. The model with the best segmentation performance in the validation dataset was selected as the final trained model. Last, the segmentation performance of the final trained model was compared with that of the Langner et al. [27] model and the original 2D-CDFNet model (although implemented using a 2.5D method in adults) by Estrada et al. [29] in the test dataset.
All 2D-CDFNet versions were trained using the loss function and learning strategy (e.g., data augmentation, learning rate, weight and decay) described by Estrada et al. [29]. The training process was implemented in PyTorch (version 1.5.1, Linux Foundation) [40] using a Docker container (Docker) [41]. Each 2D-CDFNet underwent training for 40 epochs with a batch size of eight, with use of a 12-GB NVIDIA Titan V graphics processing unit (GPU) and an Adam optimizer [42]. After training, model evaluation in the validation and test sets used an 11-GB NVIDIA 2080 Ti GPU. The total training time and the mean inference time per slice in the test data-set were recorded. The final trained model (gitlab.com/radiology/msk/genr/abdomen/cdfnet) is publicly available.
Segmentation performance metrics—To assess segmentation performance in the validation and test datasets, the Dice similarity coefficient and volumetric similarity were calculated as measures of agreement between the manual (M) and automated (A) SAT and VAT segmentations in each child. The Dice similarity coefficient, defined as
measures the relative spatial overlap between the manual and automated segmentations [43]. The volumetric similarity, defined as
measures the similarity between the volumes resulting from the manual and automated segmentations [44]. These metrics were computed using only those slices with manual segmentations available.

Visual Evaluation Dataset and Its Subsets

Model evaluations in the visual evaluation dataset and its subsets were performed using an 11-GB NVIDIA 2080 Ti GPU.
Visual evaluation dataset—The 2904 children in the visual evaluation dataset comprised remaining participants after removal of the 85 patients in the manually segmented dataset. The final trained model was used to generate SAT and VAT segmentations for all abdominal slices for all children in the visual evaluation dataset. The SAT and VAT segmentations of the model were subjectively compared with the MR images in a slice-by-slice fashion with the use of ITK-SNAP software [35]. Visual evaluation was initially performed by the previously noted research assistant (T.W.), who inspected cases for gross segmentation errors and performed additional quality checks. Segmentations without minor errors (e.g., a slight underestimation or overestimation of VAT or SAT) as well as segmentations with minor errors limited to approximately 10 slices were deemed satisfactory. Segmentations with significant segmentation errors (e.g., errors involving the entire spine), regardless of the number of slices involved, as well as segmentations with minor errors on more than approximately 10 slices, were deemed to have gross segmentation errors and were excluded from the visual scoring, segmentation evaluation, and case study datasets. For segmentations with gross errors, likely causes of segmentation failure were recorded.
As part of the quality checks, the investigator confirmed that the child's arms were positioned alongside the body, per the MRI protocol and as assumed by the model. If the arms were instead positioned over the chest and not completely removed from the chest during automatic segmentation, then the investigator manually cropped the arms from the automated segmentation results; such instances were considered to represent satisfactory segmentations, and further analyses used the adjusted segmentations after manual cropping of the arms.
For any uncertain cases, the research assistant consulted the previously noted radiologist (E.H.G.O.), who made the final determination of whether the automated segmentation was satisfactory.
Visual scoring dataset—Of the 2899 segmentations in the visual evaluation dataset without gross segmentation errors, 504 were randomly selected, stratified by sex and BMI category, to form the visual scoring dataset. A radiologist (R.v.G., with 12 years of posttraining experience) and the previously noted research assistant (T.W.) independently evaluated segmentations in the visual scoring dataset by use of a standardized visual scoring procedure. The two observers underwent training in the visual scoring process and criteria (Table 2). The observers assigned two separate scores for both SAT and VAT segmentations, categorizing the undersegmentation proportion (reflecting fat incorrectly excluded from the automated segmentation) and the oversegmentation proportion (reflecting fat incorrectly included in the automated segmentation). Each score was assigned on a scale from 0 to 3, with a score of 3 indicating nearly perfect automated segmentations for the given proportion.
TABLE 2: Scoring Criteria for Visual Scoring Process
ScoreGeneral Description of Automated SegmentationUndersegmentation Proportiona (%)Oversegmentation Proportionb (%)
3Nearly perfect; possible minor discrepancies limited to one or two slices≥ 95< 5
2Minor errors, observed in up to approximately five slices, not significantly impacting data quality90–945–9
1Minor errors, observed in more than approximately five slices and up to approximately 10 slices, not significantly impacting data quality80–8910–19
0Minor errors, observed in more than approximately 10 slices or significant errors on a few slices, significantly impacting data quality< 80≥ 20
a
Proportion of given area correctly included in automated segmentation, as based on subjective visual assessment.
b
Proportion of automated segmentation that was incorrectly included in automated segmentation, as based on subjective visual assessment.
Segmentation evaluation dataset—In the segmentation evaluation dataset, SAT and VAT were manually segmented on a single randomly selected slice by one of the previously noted observers (T.W.). The Dice similarity coefficient, volumetric similarity, precision, and recall were computed between manual and automated segmentations for the single slice. The Supplemental Methods provides definitions for precision and recall.
Case study dataset—The 2820 children in the case study data-set had available BMI and DEXA measurements and satisfactory automated SAT and VAT segmentations (Table 3). The BMI z score was derived for all children. Extracted DEXA measurements included fat mass (expressed as kilograms), total fat percentage (expressed as a percentage), lean body mass (expressed as kilograms), and the ratio of android fat mass divided by gynoid fat mass (the A/G ratio). SAT and VAT volumes were converted to SAT and VAT mass using the specific gravity of adipose tissue (0.9 g/mL). Total adipose tissue (TAT) mass and the ratio of VAT divided by SAT (the VAT/SAT ratio) were also computed. Associations among these various measures were explored, as described in the Statistical Analysis section.
TABLE 3: Characteristics of Children in the Case Study Dataset
CharacteristicUnderweight (n = 181)Normal Weight (n = 2057)Overweight (n = 582)Total Group (n = 2820)
Sex    
Female67 (37.0)1045 (50.8)356 (61.2)1468 (52.1)
Male114 (63.0)1012 (49.2)226 (38.8)1468 (47.9)
Agea (y)13.5 (13.1–14.0)13.5 (13.1–14.5)13.6 (13.1–15.4)13.5 (13.1–14.7)
Height (m)1.6 ± 0.11.6 ± 0.11.7 ± 0.11.6 ± 0.1
Weight (kg)39.7 (31.7–47.6)50.4 (38.7–65.7)67.0 (52.9–91.8)52.2 (37.1–80.9)
BMI15.5 (14.0–16.0)18.6 (16.2–21.8)24.2 (22.1–32.9)19.1 (15.3–28.7)
BMI z score−1.9 ± 0.4−0.0 ± 0.71.8 ± 0.50.3 ± 1.1
DEXA measurement    
Fat mass (kg)7.5 (5.3–12.0)11.4 (6.5–19.5)22.9 (14.6–40.5)12.4 (6.4–31.9)
Total fat percentage (%)19.0 (13.1–28.4)23.4 (13.2–35.5)35.4 (22.8–47.9)24.8 (13.6–42.1)
Lean body mass (kg)30.2 (24.0–37.3)36.0 (27.3–50.2)40.8 (30.7–56.7)36.6 (27.0–51.8)
A/G ratio0.2 (0.1–0.3)0.2 (0.2–0.4)0.4 (0.2–0.6)0.2 (0.1–0.6)
MRI measurement    
TAT mass (kg)0.5 (0.3–1.2)1.0 (0.4–2.5)2.9 (1.2–6.5)1.1 (0.4–4.9)
SAT mass (kg)0.4 (0.2–0.9)0.7 (0.3–2.0)2.5 (0.9–5.8)0.9 (0.2–4.1)
VAT mass (kg)0.1 (0.1–0.3)0.2 (0.1–0.5)0.5 (0.2–1.3)0.2 (0.1–0.9)
VAT/SAT ratio0.4 (0.2–0.7)0.3 (0.1–0.6)0.2 (0.1–0.4)0.3 (0.1–0.6)

Note—Data are count with percentage in parentheses, mean ± SD, or median with IQR in parentheses. DEXA = dual-energy x-ray absorptiometry, A/G ratio = ratio of android fat mass divided by gynoid fat mass, TAT = total adipose tissue, SAT = subcutaneous adipose tissue, VAT = visceral adipose tissue, VAT/SAT ratio = ratio of VAT divided by SAT.

a
At time of MRI examination.

Statistical Analysis

In the test dataset, segmentation performance metrics were compared between models by use of Wilcoxon signed-rank tests. For each child in the test dataset, the intraclass correlation coefficient (ICC) was separately computed for SAT and VAT between manual and automated segmentations for each manually segmented slice. In the visual scoring dataset, agreement between the two observers for visual scores (computed separately for the proportions with undersegmention and oversegmentation for SAT and VAT) was assessed using Cohen weighted kappa coefficients. For children in the case study dataset, Spearman rank-order correlation coefficients were computed among BMI, DEXA measures, and MRI measures. DEXA and MRI measures were compared between children across BMI categories using one-way ANOVA. Multivariable linear regression models were performed, as described in the Supplemental Methods. Comparisons were considered statistically significant at p < .05. Statistical analyses were performed using SPSS software (version 25.0 for Windows, IBM), and R (version 4.1.2).

Results

Manually Segmented Dataset

In the validation dataset, the best model yielded for SAT segmentation a mean Dice similarity coefficient of 0.95 ± 0.04 [SD] and volumetric similarity of 0.98 ± 0.02 [SD], and for VAT segmentation it yielded a mean Dice similarity coefficient of 0.85 ± 0.04 and volumetric similarity of 0.93 ± 0.05. Table S1 and Figure 2 present the segmentation performance of the final trained model, as well as that of the models of Langner et al. [27] and Estrada et al. [29], in the test dataset. In the test dataset, the mean Dice similarity coefficient and volumetric similarity for SAT segmentation were 0.94 ± 0.03 and 0.98 ± 0.01 for the final trained model, 0.78 ± 0.19 and 0.93 ± 0.05 for the Langner et al. method, and 0.91 ± 0.08 and 0.96 ± 0.04 for the Estrada et al. method; the mean Dice similarity coefficient and volumetric similarity for VAT segmentation were 0.85 ± 0.05 and 0.92 ± 0.04 for the final trained model, 0.66 ± 0.14 and 0.91 ± 0.04 for the Langner et al. method, and 0.63 ± 0.18 and 0.69 ± 0.18 for Estrada et al. method. The final trained model showed, for SAT segmentation, a Dice similarity coefficient and volumetric similarity that were significantly better than those of the Langner et al. method, and for VAT segmentation, it showed a Dice similarity coefficient that was significantly better than those of the Lang-ner et al. method and the Estrada et al. method as well as a volumetric similarity that was significantly better than that of the Estrada et al. method (all p < .05).
Fig. 2A —Segmentation performance of final trained model, Langner et al. model [27], and Estrada et al. model [29] in test dataset.
A, Tukey boxplots show distribution of Dice similarity coefficient and volumetric similarity between manual and automated segmentation performance of three models for subcutaneous adipose tissue (A) and visceral adipose tissue (B) in test dataset. Horizontal lines in boxes indicate medians, ends of boxes denote IQRs, whiskers represent 1.5 times IQRs, and dots beyond whiskers indicate outliers.
Fig. 2B —Segmentation performance of final trained model, Langner et al. model [27], and Estrada et al. model [29] in test dataset.
B, Tukey boxplots show distribution of Dice similarity coefficient and volumetric similarity between manual and automated segmentation performance of three models for subcutaneous adipose tissue (A) and visceral adipose tissue (B) in test dataset. Horizontal lines in boxes indicate medians, ends of boxes denote IQRs, whiskers represent 1.5 times IQRs, and dots beyond whiskers indicate outliers.
The model was trained for a total of approximately 6 hours and achieved a mean inference time of approximately 0.24 second per slice in the test dataset.
Figure 3 presents examples of segmentations by the final trained model on individual slices for children in the test data-set with varying body shapes and varying amounts of abdominal SAT and VAT. Figure S1 shows examples of substantial segmentation errors in the test dataset when the Estrada et al. model [29] was applied. Figure 4 shows the manual and automated SAT and VAT segmentations volumes on a slice-by-slice basis for all 15 children (five underweight, five normal-weight, and five overweight children) in the test dataset. Table S2 shows the ICCs between manual and automated segmentation volumes across slices for each child in the test dataset; across these 15 children, the ICCs between manual and automated segmentations for SAT ranged from 0.954 to 1.000, and for VAT they ranged from 0.827 to 0.985.
Fig. 3A —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
A, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 3B —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
B, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 3C —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
C, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 3D —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
D, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 3E —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
E, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 3F —Examples of results of automated abdominal adipose tissue segmentation into subcutaneous adipose tissue (SAT) and visceral adipose tissue (VAT) conducted using Dixon MRI acquisitions from different children. In fat-only images without (top) and with (bottom) segmentation, blue area denotes SAT, and green area indicates VAT.
F, Dixon MR images show automated abdominal adipose tissue segmentation in six children with varying body sizes and varying amounts of abdominal SAT and VAT for whom Dice similarity coefficient and volumetric similarity, respectively, were reported as follows: 13-year-old underweight girl (A) with values of 0.94 and 0.99 for SAT and 0.85 and 0.92 for VAT; 13-year-old underweight boy (B) with values of 0.91 and 0.96 for SAT and 0.82 and 0.90 for VAT; 13-year-old normal-weight girl (C) with values of 0.97 and 0.98 for SAT and 0.86 and 0.95 for VAT; 13-year-old normal-weight boy (D) with values of 0.97 and 0.99 for SAT and 0.87 and 0.91 for VAT; 13-year-old overweight girl (E) with values of 0.97 and 0.98 for SAT and 0.90 and 0.93 for VAT; and 13-year-old overweight boy (F) with values of 0.98 and 0.99 for SAT and 0.89 and 0.93 for VAT.
Fig. 4A —Comparison of manual and automated segmentations, on slice-by-slice basis, in test dataset.
A, Line graphs show results for subcutaneous adipose tissue (SAT) (A) and visceral adipose tissue (VAT) (B). Slice numbers along x-axis refer to MRI slices in abdominal region, which progress from cranial to caudal. Analysis was performed for every fourth slice in abdominal region. Each line indicates one child. Triangles indicate manual segmentations, and circles denote automated segmentations. Colors indicate BMI category, with purple indicating underweight; green, normal weight; and red, overweight.
Fig. 4B —Comparison of manual and automated segmentations, on slice-by-slice basis, in test dataset.
B, Line graphs show results for subcutaneous adipose tissue (SAT) (A) and visceral adipose tissue (VAT) (B). Slice numbers along x-axis refer to MRI slices in abdominal region, which progress from cranial to caudal. Analysis was performed for every fourth slice in abdominal region. Each line indicates one child. Triangles indicate manual segmentations, and circles denote automated segmentations. Colors indicate BMI category, with purple indicating underweight; green, normal weight; and red, overweight.

Visual Evaluation Dataset

In the 2904 children in the visual evaluation dataset, the automated segmentation was deemed to show gross segmentation errors in five children. In three children, the hip or SAT was misidentified as VAT (likely as a result of anatomic factors or respiratory motion artifact); in two, a significant amount of VAT was not correctly segmented (likely owing to low-contrast MR images). In 61 children, the arms were positioned over the chest instead of alongside the body and were not completely removed by the automated segmentation; in these children, the arms were manually cropped from the automated segmentation.

Visual Scoring Dataset

Tables S3 and S4 summarize the results of the two readers' assessments in the visual scoring dataset. The two observers assigned a score of 3 (i.e., best possible score) for SAT in 94% and 93% for the undersegmentation proportion and in 99% and 99% for the oversegmentation proportion; they also assigned a score of 3 for VAT in 99% and 99% for the undersegmentation proportion and in 95% and 97% for the oversegmentation proportion.
Interobserver agreement, expressed as a kappa coefficient, was as follows: for the SAT undersegmentation proportion, 0.88 (95% CI, 0.78–0.96); for the SAT oversegmentation proportion, 0.92 (95% CI, 0.72–1.00); for the VAT undersegmentation proportion, 0.86 (95% CI, 0.00–1.00); and for the VAT oversegmentation proportion, 0.71 (95% CI, 0.55–0.86).

Segmentation Evaluation Dataset

Table S5 shows results for the 100 patients in the segmentation evaluation dataset. The Dice similarity coefficient and volumetric similarity for this dataset were similar to results for the test data-set. Mean precision for SAT was 0.98 ± 0.02 [SD], whereas for VAT it was 0.94 ± 0.05. Mean recall for SAT was 0.89 ± 0.08 [SD]; for VAT, it was 0.76 ± 0.10.

Case Study Dataset

Figure 5 shows correlations among BMI (and BMI z score), DEXA measurements, and MRI measurements in the case study dataset. All correlations were statistically significant (p < .001). Correlation with SAT and VAT, respectively, was as follows: for BMI, 0.808 and 0.698; BMI z score, 0.768 and 0.678; fat mass, 0.941 and 0.801; and VAT/SAT ratio, −0.640 and −0.175. The correlation between SAT and VAT was 0.851.
Fig. 5 —Chart shows correlations between BMI, dual-energy x-ray absorptiometry, and MRI measurements. Values represent Spearman correlation coefficients, with 95% CI in parentheses (all p < .001). Red shading indicates positive correlation, and blue shading indicates negative correlation. Deeper shading indicates stronger positive or negative correlation. A/G ratio = ratio of android fat mass divided by gynoid fat mass, TAT = total adipose tissue, SAT = subcutaneous adipose tissue, VAT = visceral adipose tissue, VAT/SAT ratio = ratio of VAT divided by SAT.
Table S6 and Figure S2 summarize DEXA and MRI measurements in patients stratified by sex and BMI category. In both boys and girls, increasing BMI was associated with higher fat mass, fat percentage, lean body mass, A/G ratio, TAT mass, SAT mass, and VAT mass but with a lower VAT/SAT ratio. For example, in under-weight, normal-weight, and overweight children, fat mass was 7.2, 10.5, and 23.6 kg, respectively, in boys, and 8.6, 13.2, and 24.6 kg, respectively, in girls.
Table S7 and Figure S3 present associations of BMI and DEXA measurements with MRI measurements, when all measurements are expressed in terms of the SD index. Increases in fat mass, fat percentage, and A/G ratio were associated with increases in SAT (with stronger effect estimates seen in boys [range of regression coefficients, 0.80–0.99] than in girls [range, 0.74–0.88]), increases in VAT (with stronger effect estimates observed in boys [range, 0.79–0.93] than in girls [range, 0.69–0.78]), and decreases in the VAT/SAT ratio (range, −0.35 to −0.52). Increases in lean body mass were associated with increases in SAT and VAT (range, 0.22–0.51) and decreases in the VAT/SAT ratio (range, −0.16 to −0.32).

Discussion

In the present study, we trained and evaluated a CNN (2D-CDFNet) on Dixon MR images of adolescents. The trained model showed strong performance in quantitative and qualitative evaluations. For example, agreement was high between automated and manual SAT and VAT segmentations, and two independent observers reported high frequencies of nearly perfect segmentation based on visual assessment of SAT and VAT segmentation. These findings showed good generalizability to different BMI categories. In a case study of 2820 children, measurements derived from automated segmentations showed associations with BMI and DEXA measurements. By reducing manual workload to derive fat distribution measurements, the model may enable future researchers using large population-based datasets to examine questions regarding associations of childhood obesity with a range of factors, such as physical activity, diet, respiratory diseases, and cardiometabolic risk factors.
Although open source tools created by Langner et al. [27] and Estrada et al. [29] have shown some level of generalization, their segmentation quality diminished when applied in adolescents. The Estrada et al. model showed a particular decrease in performance for VAT segmentation. Similar to the present model, that model used a 2D-CDFNet architecture, although it was trained in adults only. This comparison shows the utility of training a deep learning–based segmentation method on representative data (i.e., training on Generation R Study data to perform fat quantification in adolescents). Indeed, body composition and anatomic characteristics undergo significant variations during puberty. A segmentation model trained in adults could lead to various errors when applied in children (e.g., incorrect identification of vertebrae as fat, errors in cropping of the arms, and misclassification of SAT and VAT).
The Estrada et al. model [29], in addition to being trained in adults, used a 2.5D FatSegNet method. That method aggregates three 2D-CDFNets (in axial, coronal, and sagittal planes), possibly yielding more accurate segmentation results by incorporating a broader model understanding of anatomic structures. However, 2.5D FatSegNet requires manual segmentation of continuous slices for training and is thus more labor intensive than the model used in the current investigation, which used every fourth slice. Also, the 2.5D approach requires increased GPU memory compared with the 2D method. The present trained model required 0.24 second of computation time per slice (i.e., approximately 15 seconds of total processing time for 62 abdominal slices). This short time makes the method suitable for processing of large samples and potentially for routine clinical use.
The present CNN was trained from scratch, rather than by fine-tuning, given differences from the original 2D-CDFNet in the Estrada et al. model [29] with respect to the study population and the Dixon MRI sequence used. As previously noted, the current study evaluated adolescents, whereas the original 2D-CDFNet in the Estrada et al. model evaluated adults. In addition, the MR images in the present examination had an axial in-plane resolution of 1.25 × 1.25 mm, which was higher than the 2 × 2 mm resolution of the Estrada et al. model (as well as the 2.07 × 2.07 mm resolution of the Langner et al. model [27]). Thus, the segmentation task was different. We expect that training from scratch allowed us to adapt the network to better leverage the age-specific characteristics and higher-resolution images of the present dataset and thereby address the unique requirements of the dataset. Although the original model's weights for initialization were not used, the trained model achieved performance in adolescents that was similar to that reported for the original 2D-CDFNet in adults.
The case study provided insights into relationships of SAT and VAT with BMI and DEXA measurements in adolescents. Consistent with prior studies in adults, correlations with BMI were higher for SAT than VAT [45, 46]. Correlations were also higher for SAT than VAT for BMI (z score), fat mass, and TAT, indicating stronger overall association with obesity in adolescents for SAT than VAT. Moreover, the VAT/SAT ratio, a potential parameter of cardiometabolic risk described in previous studies, was more strongly associated with SAT than VAT, which is different from previous findings in adults [47, 48]. More studies are warranted to validate the associations and to examine the potential role of the VAT/SAT ratio as a marker of cardiometabolic risk in children.
This study had limitations. First, because we used cross-sectional data when children presented after reaching age 13 years, the model was not validated in younger children. Second, MRI examinations were performed using a single scanner at a single center. Generalizing the findings of the 2D-CDFNet model to other centers, field strengths, and manufacturers requires additional investigation. Third, the number of children in the test dataset was small. Fourth, because of our study aim to develop a model specifically for application in adolescents, we did not create a technique that can be generalized to both adolescents and adults. Fifth, the model's segmentation performance was higher for SAT than for VAT. This difference may reflect the smaller surface-to-volume ratio of SAT compared with VAT, as spatial overlap metrics are typically lower for structures with higher surface-to-volume ratio [43, 49]. Sixth, the arms were manually cropped from the automated segmentations for examinations in which the child's arms were over the chest rather than alongside the body per the protocol. Seventh, in the visual scoring dataset, interreader agreement of the visual scores was lower for VAT than for SAT, possibly reflecting more complex distributions of VAT. Finally, although all examinations in the visual evaluation dataset were reviewed to exclude those with gross segmentation errors, a small number of examinations in the visual scoring dataset received scores of 0, indicating significant segmentation errors impacting data quality.
In conclusion, we developed and internally validated the performance of a deep learning–based model for automated quantification of abdominal fat on MRI examinations in adolescents. The fat segmentations obtained by the model showed strong agreement with manual segmentations as well as associations with anthropometric and DEXA-derived measures. The model may facilitate future large-scale studies investigating abdominal fat distribution on MRI examinations of adolescents as well as associations of fat distribution with other body composition measures or clinical outcomes.

Acknowledgments

The Generation R study is managed by the Erasmus Medical Center in close collaboration with the School of Law and the Faculty of Social Sciences at Erasmus University, Rotterdam, The Netherlands; the Municipal Health Service, Rotterdam area; and the Stichting Trombosedienst and Artsenlaboratorium Rijnmond (Star-MDC), Rotterdam. We thank the children and their parents as well as the general practitioners, hospitals, midwives, and pharmacies in Rotterdam.

Footnotes

Provenance and review: Not solicited; externally peer reviewed.
Peer reviewers: Elanchezhian Somasundaram, Cincinnati Children's Hospital Medical Center; Maria Pilar Aparisi Gómez, Auckland City Hospital; additional individual(s) who chose not to disclose their identity.

Supplemental Content

File (23_29570_suppl.pdf)

References

1.
Kumar S, Kelly AS. Review of childhood obesity: from epidemiology, etiology, and comorbidities to clinical assessment and treatment. Mayo Clin Proc 2017:251–265
2.
Engeland A, Bjørge T, Tverdal A, Søgaard AJ. Obesity in adolescence and adulthood and the risk of adult mortality. Epidemiology 2004; 15:79–85
3.
Mahabadi AA, Massaro JM, Rosito GA, et al. Association of pericardial fat, intrathoracic fat, and visceral abdominal fat with cardiovascular disease burden: the Framingham Heart Study. Eur Heart J 2009; 30:850–856
4.
Fox CS, Massaro JM, Hoffmann U, et al. Abdominal visceral and subcutaneous adipose tissue compartments: association with metabolic risk factors in the Framingham Heart Study. Circulation 2007; 116:39–48
5.
Yamaji T, Iwasaki M, Sasazuki S, et al. Visceral fat volume and the prevalence of colorectal adenoma. Am J Epidemiol 2009; 170:1502–1511
6.
Kelly AS, Dengel DR, Hodges J, et al. The relative contributions of the abdominal visceral and subcutaneous fat depots to cardiometabolic risk in youth. Clin Obes 2014; 4:101–107
7.
Pausova Z, Mahboubi A, Abrahamowicz M, et al. Sex differences in the contributions of visceral and total body fat to blood pressure in adolescence. Hypertension 2012; 59:572–579
8.
Ofenheimer A, Breyer-Kohansal R, Hartl S, et al. Reference charts for body composition parameters by dual-energy X-ray absorptiometry in European children and adolescents aged 6 to 18 years: results from the Austrian LEAD (Lung, hEart, sociAl, boDy) cohort. Pediatr Obes 2021; 16:e12695
9.
Ferland M, Després JP, Tremblay A, et al. Assessment of adipose tissue distribution by computed axial tomography in obese women: association with body density and anthropometric measurements. Br J Nutr 1989; 61:139–148
10.
Gray DS, Fujioka K, Colletti PM, et al. Magnetic-resonance imaging used for determining fat distribution in obesity and diabetes. Am J Clin Nutr 1991; 54:623–627
11.
Baum T, Cordes C, Dieckmeyer M, et al. MR-based assessment of body fat distribution and characteristics. Eur J Radiol 2016; 85:1512–1518
12.
Ma J. Dixon techniques for water and fat imaging. J Magn Reson Imaging 2008; 28:543–558
13.
Karthik L, Kumar G, Keswani T, et al. Protease inhibitors from marine actino-bacteria as a potential source for antimalarial compound. PLoS One 2014; 9:e90972
14.
Schlett C, Hendel T, Weckbach S, et al. Population-based imaging and radiomics: rationale and perspective of the German National Cohort MRI Study. Rofo 2016:652–661
15.
Bamberg F, Hetterich H, Rospleszcz S, et al. Subclinical disease burden as assessed by whole-body MRI in subjects with prediabetes, subjects with diabetes, and normal control subjects from the general population: the KORA-MRI study. Diabetes 2017; 66:158–169
16.
Elffers TW, de Mutsert R, Lamb HJ, et al. Body fat distribution, in particular visceral fat, is associated with cardiometabolic risk factors in obese women. PLoS One 2017; 12:e0185403
17.
Kooijman MN, Kruithof CJ, van Duijn CM, et al. The Generation R Study: design and cohort update 2017. Eur J Epidemiol 2016; 31:1243–1264
18.
Neeland IJ, Grundy SM, Li X, Adams-Huet B, Vega GL. Comparison of visceral fat mass measurement by dual-x-ray absorptiometry and magnetic resonance imaging in a multiethnic cohort: the Dallas Heart Study. Nutr Diabetes 2016; 6:e221
19.
Thörmer G, Bertram HH, Garnov N, et al. Software for automated MRI-based quantification of abdominal fat and preliminary evaluation in morbidly obese patients. J Magn Reson Imaging 2013; 37:1144–1150
20.
Zhou A, Murillo H, Peng Q. Novel segmentation method for abdominal fat quantification by MRI. J Magn Reson Imaging 2011; 34:852–860
21.
Christensen AN, Larsen CT, Mandrup CM, et al. Automatic segmentation of abdominal fat in MRI-scans, using graph-cuts and image derived energies. In: Sharma P, Bianchi FP, eds. Image analysis: 20th Scandinavian Conference on Image Analysis. 2017:109–120
22.
Wald D, Teucher B, Dinkel J, et al. Automatic quantification of subcutaneous and visceral adipose tissue from whole-body magnetic resonance images suitable for large cohort studies. J Magn Reson Imaging 2012; 36:1421–1434
23.
Mosbech TH, Pilgaard K, Vaag A, Larsen R. Automatic segmentation of abdominal adipose tissue in MRI. In: Heyden A, Kahl F, eds. Image analysis: 20th Scandinavian Conference on Image Analysis. 2011:501–511
24.
Litjens G, Kooi T, Bejnordi BE, et al. A survey on deep learning in medical image analysis. Med Image Anal 2017; 42:60–88
25.
West J, Dahlqvist Leinhard O, Romu T, et al. Feasibility of MR-based body composition analysis in large scale population studies. PLoS One 2016; 11:e0163332
26.
Newman D, Kelly-Morland C, Leinhard OD, et al. Test-retest reliability of rapid whole body and compartmental fat volume quantification on a widebore 3T MR system in normal-weight, overweight, and obese subjects. J Magn Reson Imaging 2016; 44:1464–1473
27.
Langner T, Hedström A, Mörwald K, et al. Fully convolutional networks for automated segmentation of abdominal adipose tissue depots in multi-center water-fat MRI. Magn Reson Med 2019; 81:2736–2745
28.
Küstner T, Hepp T, Fischer M, et al. Fully automated and standardized segmentation of adipose tissue compartments via deep learning in 3D whole-body MRI of epidemiologic cohort studies. Radiol Artif Intell 2020; 2:e200010
29.
Estrada S, Lu R, Conjeti S, et al. FatSegNet: a fully automated deep learning pipeline for adipose tissue segmentation on abdominal Dixon MRI. Magn Reson Med 2020; 83:1471–1483
30.
Goodfellow I, Warde-Farley D, Mirza M, Courville A, Bengio Y. Maxout networks. Proc Mach Learn Res 2013; 28:1319–1327
31.
Rheinland Study website. Rheinland Study. www.rheinland-studie.de/. Published 2023. Accessed Jul 20, 2023
32.
Breteler MM, Stöcker T, Pracht E, Brenner D, Stirnberg R. MRI in the Rhine-land Study: a novel protocol for population neuroimaging. Alzheimer's Dement 2014; 10:P92
33.
Cole TJ, Bellizzi MC, Flegal KM, Dietz WH. Establishing a standard definition for child overweight and obesity worldwide: international survey. BMJ 2000; 320:1240–1243
34.
Fredriks AM, van Buuren S, Wit JM, Verloove-Vanhorick SP. Body index measurements in 1996–7 compared with 1980. Arch Dis Child 2000; 82:107–112
35.
Yushkevich PA, Gao Y, Gerig G. ITK-SNAP: an interactive tool for semi-automatic segmentation of multi-modality biomedical images. Annu Int Conf IEEE Eng Med Biol Soc 2016; 2016:3342–3345
36.
Shen W, Wang Z, Punyanita M, et al. Adipose tissue quantification by imaging methods: a proposed classification. Obes Res 2003; 11:5–16
37.
Roy AG, Conjeti S, Sheet D, Katouzian A, Navab N, Wachinger C. Error corrective boosting for learning fully convolutional networks with limited data. In: Descoteaux M, Maier-Hein L, Wells WWM, et al., eds Medical Image Computing and Computer-Assisted Intervention: MICCAI 2017. Springer, 2017:231–239
38.
Roy AG, Conjeti S, Navab N, Wachinger C; Alzheimer's Disease Neuroimaging Initiative. QuickNAT: a fully convolutional network for quick and accurate segmentation of neuroanatomy. Neuroimage 2019; 186:713–727
39.
Ronneberger O, Fischer P, Brox T. U-net: convolutional networks for biomedical image segmentation. In: Navab N, Hornegger J, Wells WM, Frangi AE, eds: Medical Image Computing and Computer-Assisted Intervention: MICCAI 2015. Springer, 2015:234–241
40.
Paszke A, Gross S, Chintala S, et al. Automatic differentiation in PyTorch. OpenReview website. openreview.net/forum?id=BJJsrmfCZ. Modified Oct 28, 2017. Accessed Jul 20, 2023
41.
Merkel D. Docker: lightweight Linux containers for consistent development and deployment. Linux Journal website. www.linuxjournal.com/content/docker-lightweight-linux-containers-consistent-development-and-deployment. Published May 19, 2014. Accessed Jul 20, 2023
42.
Kingma DP, Ba J. Adam: a method for stochastic optimization. arXiv web-site. arxiv.org/abs/1412.6980. Submitted Dec 22, 2014. Last revised Jan 30, 2017. Accessed Jul 20, 2023
43.
Dice LR. Measures of the amount of ecologic association between species. Ecology 1945; 26:297–302
44.
Cárdenes R, de Luis-García R, Bach-Cuadra M. A multidimensional segmentation evaluation for medical image data. Comput Methods Programs Biomed 2009; 96:108–124
45.
Camhi SM, Bray GA, Bouchard C, et al. The relationship of waist circumference and BMI to visceral, subcutaneous, and total body fat: sex and race differences. Obesity (Silver Spring) 2011; 19:402–408
46.
Nattenmueller J, Hoegenauer H, Boehm J, et al. CT-based compartmental quantification of adipose tissue versus body metrics in colorectal cancer patients. Eur Radiol 2016; 26:4131–4140
47.
Miller KD, Jones E, Yanovski JA, Shankar R, Feuerstein I, Falloon J. Visceral abdominal-fat accumulation associated with use of indinavir. Lancet 1998; 351:871–875
48.
Kaess BM, Pedley A, Massaro JM, Murabito J, Hoffmann U, Fox CS. The ratio of visceral to subcutaneous fat, a metric of body fat distribution, is a unique correlate of cardiometabolic risk. Diabetologia 2012; 55:2622–2630
49.
Taha AA, Hanbury A. Metrics for evaluating 3D medical image segmentation: analysis, selection, and tool. BMC Med Imaging 2015; 15:29

Information & Authors

Information

Published In

American Journal of Roentgenology
PubMed: 37584508

History

Submitted: April 28, 2023
Revision requested: May 15, 2023
Revision received: July 20, 2023
Accepted: August 6, 2023
Version of record online: August 16, 2023

Keywords

  1. adolescents
  2. deep learning
  3. MRI
  4. subcutaneous adipose tissue
  5. visceral adipose tissue

Authors

Affiliations

Tong Wu, MD
The Generation R Study Group, Erasmus MC, University Medical Center, Rotterdam, The Netherlands.
Department of Radiology & Nuclear Medicine, Erasmus MC, University Medical Center, PO Box 2040, 3000 CA, Rotterdam, The Netherlands.
Santiago Estrada, MSc
Image Analysis, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany.
Population Health Sciences, German Center for Neurodegenerative Diseases (DZNE), Bonn, Germany.
Renza van Gils, MD
Department of Radiology & Nuclear Medicine, Erasmus MC, University Medical Center, PO Box 2040, 3000 CA, Rotterdam, The Netherlands.
Ruisheng Su, MSc
Department of Radiology & Nuclear Medicine, Erasmus MC, University Medical Center, PO Box 2040, 3000 CA, Rotterdam, The Netherlands.
Vincent W. V. Jaddoe, MD, PhD
The Generation R Study Group, Erasmus MC, University Medical Center, Rotterdam, The Netherlands.
Department of Pediatrics, Erasmus MC, University Medical Center, Rotterdam, The Netherlands.
Edwin H. G. Oei, MD, PhD
Department of Radiology & Nuclear Medicine, Erasmus MC, University Medical Center, PO Box 2040, 3000 CA, Rotterdam, The Netherlands.
Stefan Klein, PhD [email protected]
Department of Radiology & Nuclear Medicine, Erasmus MC, University Medical Center, PO Box 2040, 3000 CA, Rotterdam, The Netherlands.

Notes

Address correspondence to S. Klein ([email protected]).
First published online: Aug 16, 2023
Version of record: Nov 15, 2023
The study sponsors had no role in the study design, data analysis, interpretation of data, or writing of this report.
T. Wu and S. Estrada contributed equally to this work.
The authors declare that there are no disclosures relevant to the subject matter of this article.
Supported by the Erasmus Medical Center; the Erasmus University Rotterdam; and The Netherlands Organization for Health Research and Development (to The Generation R Study). T. Wu receives support from the China Scholarship Council PhD Fellowship (scholarship 201906260304) for Doctor of Philosophy study at Erasmus Medical Center.

Metrics & Citations

Metrics

Citations

Export Citations

To download the citation to this article, select your reference manager software.

Articles citing this article

View Options

View options

PDF

View PDF

PDF Download

Download PDF

Media

Figures

Other

Tables

Share

Share

Copy the content Link

Share on social media