BMI-adjusted adipose tissue volumes exhibit depot-specific and divergent associations with cardiometabolic diseases

For any given body mass index (BMI), individuals vary substantially in fat distribution, and this variation may have important implications for cardiometabolic risk. Here, we study disease associations with BMI-independent variation in visceral (VAT), abdominal subcutaneous (ASAT), and gluteofemoral (GFAT) fat depots in 40,032 individuals of the UK Biobank with body MRI. We apply deep learning models based on two-dimensional body MRI projections to enable near-perfect estimation of fat depot volumes (R2 in heldout dataset = 0.978-0.991 for VAT, ASAT, and GFAT). Next, we derive BMI-adjusted metrics for each fat depot (e.g. VAT adjusted for BMI, VATadjBMI) to quantify local adiposity burden. VATadjBMI is associated with increased risk of type 2 diabetes and coronary artery disease, ASATadjBMI is largely neutral, and GFATadjBMI is associated with reduced risk. These results – describing three metabolically distinct fat depots at scale – clarify the cardiometabolic impact of BMI-independent differences in body fat distribution.

size if models were sufficiently predictive. Second, gluteofemoral adipose tissue (GFAT)which may serve as an adaptive energy storage depot and a possible modifier of insulin resistancehas not been quantified in most previous imaging studies [18][19][20][21][22] . Third, fat depot volumes tend to be highly correlated with both BMI and one another, making it challenging to isolate depot-specific associations with disease 23 .
In this study, we downloaded raw MRI imaging data from 40,032 participants of the UK Biobank and tested the hypothesis that deep learning models can be used to precisely quantify three fat depot volumes: VAT, abdominal subcutaneous adipose tissue (ASAT), and GFAT. We derived measures of local adiposity burden, each fully independent of BMI, and note significant heterogeneity in risk conferred: VAT adjusted for BMI (VATadjBMI) associated with increased risk of type 2 diabetes and coronary artery disease, ASA-TadjBMI largely risk-neutral, and GFATadjBMI associated with protection.

Results
Among 40,032 participants of the UK Biobank with MRI data available, the median age was 65 years, 51% were female, and 97% were white (Table 1). Median BMI was 26.6 kg/m 2 among males and 25.2 kg/m 2 among females, and the median waist-hip ratio (WHR) was 0.93 among males, and 0.81 among females. 1,901 individuals had been diagnosed with type 2 diabetes (4.7%) and 1956 with coronary artery disease (4.9%) at the time of imaging assessment. VAT, ASAT, and GFAT volumes were previously quantified in 9040, 9041, and 7754 participants, respectively (Supplementary Data 1 and 2) 20,21,24,25 .

Machine learning facilitates near-perfect estimation of fat depot volumes
We set out to test whether convolutional neural network models could be adequately predictive of VAT, ASAT, and GFAT volumes to enable prediction at scale. We noted that three-dimensional MRI data for 40,032 individuals represented a substantial data burden with almost 58 million axial slices across all participants, corresponding to >18 terabytes of imaging dataa level of complexity that limits computational feasibility for training deep learning models.
To simplify the imaging input into the convolutional neural networks, we transformed three-dimensional MRI images for each participant into two-dimensional coronal and sagittal projections, hypothesizing that this input would prove adequate for highly accurate fat depot volume prediction despite an 830-fold reduction in data input size (Fig. 1) 26 . Convolutional neural networkstrained on 80% of the participants with fat depots previously quantifieddemonstrated near-perfect estimation of each fat depot volume in the 20% of held out individuals (R 2 = 0.991, 0.991, and 0.978 for VAT, ASAT, and GFAT, respectively) (Methods, Supplementary Data 3). Similar predictive accuracy was noted across age, sex, BMI, and self-reported ethnicity subgroups, although the sample size was limited in the latter subgroups (Supplementary Data 4). These convolutional neural network models were subsequently applied to the unlabeled remainder of the 40,032 participants to estimate fat depot volumes.
Next, we applied Gradient-weighted Class Activation Mapping (Grad-CAM) to better understand regions of a given MRI projection contributing to predictions of VAT, ASAT, and GFAT volumes 27 . Briefly, Grad-CAM uses gradients entering the final convolutional layer to generate a low-resolution heat map signifying how much a given region contributes to a model's prediction. Separately in males and females, we selected participants from each held out dataset at the 75th, 95th, and 99th percentiles of absolute error and applied Grad-CAM to generate saliency maps. We also selected three participants who were present in all three held out datasets to compare Grad-CAM results for VAT, ASAT, and GFAT. In all cases, Grad-CAM revealed prioritized regions of the MRI projection that were anatomically consistent with the known distribution of VAT, ASAT, and GFAT, even in cases with higher absolute error (Supplementary Figs. 1-4).

Variation in adipose volumes and association with cardiometabolic diseases
We confirm and extend prior evidence for marked differences in fat depot volume in male versus female participants ( Fig. 2a) 28,29 . Median visceral adipose tissue volume was substantially higher in males as compared to females -4.8 versus 2.3 liters, respectivelywhile subcutaneous and gluteofemoral depots tended to predominate in females (Table 1). A significant correlation between BMI and all three fat depots was noted -Pearson r ranging from 0.77 to 0.91but considerable variation was observed within any clinical BMI category (Fig. 2a, b). Modest variation in the correlation between BMI and each fat depot was noted across self-reported ethnicity groups (Supplementary Data 5).

BMI-adjusted local fat depots and cardiometabolic disease
To disentangle the unique impact of each fat depot from overall BMI, we next generated measurements of VATadjBMI, ASATadjBMI, and GFATadjBMI for each participant by computing sex-specific BMI residuals in 38,680 (97%) of the study population with BMI measurement on the day of MRI imaging available ( Supplementary  Fig. 5). These residuals reflect the difference in an individual's fat depot volume as compared with that expected based on BMI. These metrics were fully independent of BMI and largely independent of anthropometric measures and each other ( Supplementary Fig. 6). Flexibly modeling BMI with a B-spline basis when computing these residuals yielded similar results ( Supplementary Fig. 7, Supplementary Data 9). neural networks to quantify adipose tissue depots from body MRI images. a Two-dimensional projections are created by computing the mean pixel intensity along the coronal and sagittal axes. Two images for each participant were used as inputs into the convolutional neural network: one consisting of the coronal and sagittal two-dimensional projections in the fat phase, and another consisting of the same projections in the water phase. b Convolutional neural networks trained on two-dimensional MRI projections achieved near-perfect prediction of each fat depot volume in the holdout set (Supplementary Table 3).
c Three female participants with similar BMI (ranging from 29.1 to 29.6 kg/m 2 ) but highly discordant fat depot volumes quantified by convolutional neural networks. Fat depot volume percentiles are computed relative to a subgroup of female participants with overweight BMI (25 ≤ BMI < 30). Note that outlines for each fat depot are drawn as a visual aid for each fat depot and do not reflect segmentation. Abbreviations: VAT, visceral adipose tissue; ASAT, abdominal subcutaneous adipose tissue; GFAT, gluteofemoral adipose tissue.
To better understand the gradients in absolute prevalence rates according to BMI-adjusted fat depots, we calculated standardized estimates for the lowest quintile, quintiles 2-4, and the highest quintile within clinical BMI categories of normal, overweight, and obesity.
Using this approach, we note substantial gradients in the prevalence of cardiometabolic diseases according to local adipose tissue burden, even within clinical BMI categories (Fig. 4, Supplementary Data 14-15). As a representative example, males with normal BMI but VATadjBMI in the highest quintile had a predicted probability of type 2 diabetes of 6.6% (95%CI 5.5-7.9), higher than males with overweight BMI with VATadjBMI in the lowest quintile, in whom probability was 2.7% (95% CI: 2.2-3.4). Among females with obesity, estimates of diabetes ranged from 3.5 to 9.2% across quintiles of VATadjBMI but 7.6 to 3.6% for GFATadjBMI. A similar patternwith less pronounced gradientswas observed for coronary artery disease.

Discussion
In this study, we demonstrated that a deep learning approach based on two-dimensional MRI projections is adequately predictive to quantify VAT, ASAT, and GFAT volumes at scale. By then moving away from raw fat depot volumeswhich are driven largely by BMI and overall adiposityto BMI-adjusted measurements, we demonstrated a consistent trend of VATadjBMI associated with increased risk of type 2 diabetes and coronary artery disease, ASATadjBMI largely risk-neutral, and GFATadjBMI conferring protection. These results have at least four implications. First, machine learning can enable insights from large-scale data repositories of difficult-to-measure phenotypes. In this study, convolutional neural network models were used to precisely measure fat depot measurements from MRI images, considered the gold standard modality for the volumetric measurement of adipose tissue 16,31 . Hypothesis-informed simplification of the input datain this study moving from three-dimensional MRI images to twodimensional MRI projectionswas necessary to ensure computational feasibility. This work adds to several recent studies of machine learning-derived phenotypes, including aortic size, liver fat, and cardiac trabecular structure [32][33][34] . Although population-based assessment of fat distribution using MRI is unlikely to be practical, these results lay the scientific foundation for efforts to quantify such measures using other datasuch as DEXA images or abdominal CT scans already embedded in the electronic medical record for some patientsor even static images of body silhouette, as might conceivably be obtained with a smartphone camera 35,36 . Abdominal imaging may also be useful for learning hidden variables of biological significance, such as age 37 .
Second, these results support a growing appreciation that various fat depotsrather than serving as an agnostic sink for energy storagehave distinct metabolic profiles. Previous work has noted significant functional differences in adipocytes according to specific fat depot, ascribed in part to site-specific expression of developmental genes associated with adipogenesis 38,39 . While VAT tends to be the primary site for immediate storage of dietary-derived fat via adipocyte hypertrophy and has a higher rate of lipid turnover, GFAT is a more stable fat depot that primarily expands via adipocyte hyperplasia and may spare expansion of harmful visceral or ectopic fat depots. These and other studies support a natural order of fat deposition, whereby a primary driver of high VAT in specific individuals may reflect an inability to adequately expand ASAT or GFAT depots 13,40 . In rare Mendelian lipodystrophiesas occurs in individuals who harbor pathogenic LMNA mutationsan extreme example of this paradigm leads to marked reduction of ASAT and GFAT but increased VAT and increased rates of severe insulin resistance 41 . Whether individuals in the extreme tails of low GFATadjBMI and ASATadjBMI or high VATadjBMI might be enriched for genetic perturbations in lipodystrophy genes or the inherited component to these metrics is largely 'polygenic'due to the aggregate effects of many common DNA variants, each of modest effect size warrants further study [42][43][44] . Sex differences will also be important to Odds ratios per standard deviation with 95% confidence intervals are shown for prevalent type 2 diabetes and coronary artery disease. Sample sizes for each model are shown as the denominator in the second column. P-values correspond to two-sided tests for the indicated independent variable in an adjusted logistic regression. Logistic regression models were adjusted for age, sex (except in sex subgroup analyses), BMI, the other two fat depots (e.g. ASATadjBMI and GFATadjBMI for VATadjBMI), and MRI imaging center. Source data are provided as a Source Data file. Abbreviations: VATadjBMI, visceral adipose tissue adjusted for body mass index (BMI); ASATadjBMI, abdominal subcutaneous adipose tissue adjusted for BMI; GFATadjBMI, gluteofemoral adipose tissue adjusted for BMI. consider in future studies on local adiposityfor example, here we demonstrate that ASATadjBMI and GFATadjBMI are more correlated in male participants than in female participants, which may point to sexdependent fat depot specificity.
Third, changes in measures of local adiposityindependent of weight and body-mass indexmay serve as reliable proxies of cardiometabolic benefits of a given intervention, and warrant consideration as additional endpoints for future clinical trials. Most studies to date of obesity interventions have focused on reduction in overall weight or BMI as the primary outcome, consistent with FDA regulatory guidance 45 . However, at least two classes of drugs appear to have a selective VAT reduction effect in clinical trials: thiazolidinediones and a synthetic form of growth hormone-releasing hormone 46,47 . Whether these therapies might be repurposed from their original indicationstype 2 diabetes and HIV-associated lipodystrophyor new agents might prove useful in a subset of individuals with VAT-driven increases in cardiometabolic risk warrants further study. In such studies, "adjBMI" or similar measures of local adiposity may prove useful for quantifying BMI-independent changes in fat distribution. Considering measures of local adiposity may be particularly important for individuals with normal or low BMIin this study, we observed a trend of amplified associations with type 2 diabetes in participants with BMI less than 25 kg/m 2 , consistent with a prior study examining the association of waist circumference and waist-hip ratio with mortality 48 .  Fourth, although our data suggests similar performance of our deep learning models across self-reported ethnicity subgroups, we were underpowered to study disease associations in non-White subgroups. Additional validation across ancestrally and geographically diverse populations would be of considerable value, especially given prior evidence of significant variability in fat distribution indices across ethnicity groups 49,50 . An important example relates to the South Asian population, where abnormal fat distribution has been postulated as a key driver of the markedly increased rates of cardiovascular disease and diabetes observed, often in the context of a relatively normal BMI 51,52 .
Our study has several limitations. First, this study was a crosssectional analysis of individuals with a median age of 65 years at time of imaging. Future studies of individuals across the lifespanespecially those that include repeat imaging assessmentsare warranted. Second, although we note striking associations of BMI-adjusted fat depots with cardiometabolic disease, these observational data do not definitely prove causation or that modification of fat distribution will lead to therapeutic gain. Third, while two-dimensional MRI projections are a useful simplification of three-dimensional MRI images for the task of predicting adipose tissue compartment volumes, they are unlikely to be appropriate for predicting "density-like quantities" such as liver fat percentage, where a single axial cross-section performs well 33 . This highlights the importance of choosing an appropriate simplification for the desired task. Fourth, while we were able to achieve good performance using CNN-based regression models and saliency mapping results were anatomically reasonable, we were unable to directly compare our approach to segmentation-based models.
In conclusion, we used a machine learning approach based on two-dimensional projections of body MRI data to compute VAT, ASAT, and GFAT volumes at scale in 40,032 individuals of the UK Biobank. BMI-adjusted fat depot measurements displayed divergent associations with cardiometabolic diseases and were shown to alter risk within BMI subgroups. These BMI-adjusted metrics may serve as useful additional endpoints for obesity interventions to more completely capture metabolic health associated with body composition.

Study population
The UK Biobank is an observational study that enrolled over 500,000 individuals between the ages of 40 and 69 years between 2006 and 2010, of whom 43,531 underwent body MRI imaging between 2014 and 2020 as part of an imaging substudy 53,54 . Images were acquired using the Dixon method, an MRI sequence that can be used to isolate fat signals from water signal 55 . Each participant's MRI data consisted of 244 axial slices acquired from the neck to the knees in four sequences: in-phase, out-of-phase, fat-only, and wateronly. After the exclusion of 3489 (8.0%) imaging scans based on technical problems or artifacts, 40,032 participants remained for analysis, 19,435 males and 20,597 females (Supplementary Methods). This analysis of data from the UK Biobank was approved by the Mass General Brigham institutional review board and was performed under UK Biobank application #7089.

Machine learning to measure fat depot volumes
Among the 40,032 individuals with MRI imaging data available, a subset had visceral adipose tissue (VAT) volume, abdominal subcutaneous adipose tissue (ASAT) volume, and total adipose tissue (TAT) volume between the top of vertebrae T9 and the bottom of the thigh muscles, quantified and made available as previously described (N = 9040, 9041, 7754 participants, respectively) 20,21,24,25 . Gluteofemoral adipose tissue (GFAT) volume was derived by computing the difference between TAT and the sum of VAT and ASAT (Supplementary Methods). For each participant, we transformed threedimensional MRI images into two-dimensional coronal and sagittal projections by computing the mean intensity projection in each orientation. For example, a given pixel on a coronal two-dimensional projection represents the mean intensity across all pixels making up a line oriented in the anterior-posterior direction perpendicular to the coronal plane (Supplementary Methods). This procedure was done for the fat-only and water-only MRI sequences, and the resulting images were jointly used as the imaging input for a given participant.
Individuals with previously quantified fat depot volumes were randomly split into 80% for training and a 20% holdout sample for testing. For each of VAT, ASAT, and GFAT, a CNN was trained on a pair of fat phase and water phase MRI images to predict each fat depot volume, where each image was composed of (a) a coronal twodimensional projection and (b) a sagittal two-dimensional projection of the body MRI. Each CNN was developed with the DenseNet-121 architecture pre-trained on ImageNet as the base model 56,57 . The last dense block output was flattened using a global average pooling layer and then fed into three fully connected layers of size 64, 256, and 1, with the last layer having no activation function (linear mapping). All other activation functions use the ReLU non-linearity. All models were trained using the Adam optimizer with a learning rate set to a cosine decay policy decaying from 0.001 to 0 over 100 epochs, a shrinkage loss function using the hyperparameters a = 10.0 and c = 0.2, and a batch size of 32 58,59 . For all training data, the following augmentations (random permutations of the training images) were applied: random shifts in the XY-plane by up to ±16 pixels, rotations by up to ±5 degrees around its center axis, and the coronal view horizontally flipped with a probability of 50%. Each view (coronal and sagittal) were separately pre-normalized by its z-score (0 mean, standard deviation of 1), followed by joint normalization following concatenation side-by-side.
A five-fold cross-validation scheme was used within the 80% training data set. Performance was determined in a 20% holdout sample that was unseen to the model prior to evaluation. The five folds were used to determine the mean and standard deviation of performance metrics, then a single fold was randomly selected to take forward for predicting fat depot volumes in the remaining participants with raw MRI imaging data but without labels. Additional information can be found in the Supplementary Methods. Hazard ratios with 95% CI in parentheses are shown for VATadjBMI, ASATadjBMI, and GFATadjBMI in Cox proportional-hazard models adjusted for age, sex, BMI, the other two fat depots, and MRI imaging center. P-values correspond to two-sided tests for the indicated independent variable in the adjusted models. Median follow-up time for both incident type 2 diabetes and coronary artery disease was 2.8 years from the date of imaging. Note that two participants in the prevalent disease analyses are not included in incident disease analyses because they withdrew consent in the interim period.

Saliency maps
We used Gradient-weighted Class Activation Mapping (Grad-CAM) 27 to generate saliency maps for selected participants to obtain "visual explanations" for decisions from our CNN-based regression models used for estimating fat depot volumes. Briefly, Grad-CAM uses the gradients flowing into the final convolutional layer to produce a lowresolution localization heat map highlighting important regions (red) and less important regions (blue). In other words, the importance signifies how much a specific area contributes to the overall prediction.

Cardiometabolic disease definitions
Type 2 diabetes was defined on the basis of ICD-10 codes, self-report during a verbal interview with a trained nurse, use of diabetes medication, or a glycated hemoglobin greater than or equal to 6.5% before the date of imaging. Coronary artery disease was defined as myocardial infarction, angina, coronary revascularization, or death from coronary causes as determined on the basis of ICD-10 codes, ICD-9 codes, OPCS-4 surgical codes, nurse interview, and national death registries.

Statistical analysis
We generated BMI-adjusted fat depot measurements by computing residuals from sex-specific linear regression models using BMI to predict each fat depot volume, analogous to prior studies of waist-hip ratio adjusted for BMI 60,61 . Logistic regression models were used to test the association of BMI-adjusted fat depot measurements with prevalent disease in models adjusted for age, sex (except in sex subgroup analyses), BMI, the other two fat depots (e.g. ASATadjBMI and GFA-TadjBMI for VATadjBMI), and MRI imaging center. Cox proportionalhazard models with the same covariates were used to test associations of BMI-adjusted fat depots with incident type 2 diabetes and coronary artery disease. To predict the gradient of prevalent disease across clinical categories, we used logistic regression models separately in males and females including age, BMI, sex-specific quintiles of VATadjBMI, ASATadjBMI, GFATadjBMI, MRI imaging center, and interaction terms between the local adiposity quintiles and BMI. Models were standardized to the median of all predictor variables (except for the MRI imaging center variable, where the mean was used). Effect sizes are reported per sex-specific standard deviation. All analyses were performed with the use of R software, version 3.6.0 (R Project for Statistical Computing).

Reporting summary
Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Data availability
This research has been conducted using the UK Biobank Resource under Application Number #7089. The raw UK Biobank data -including the anthropometric data reported here -are made available to researchers from universities and other research institutions with research inquiries following IRB and UK Biobank approval (https:// www.ukbiobank.ac.uk/enable-your-research/apply-for-access). Visceral, abdominal subcutaneous, and gluteofemoral adipose tissue volume predictions have been submitted to the UK Biobank and will be available for download by researchers (https://www.ukbiobank.ac.uk/ enable-your-research/research-analysis-platform). All other data generated in the study are available in the Supplementary Data. Source data for Figs. 3 and 4, Supplementary Fig. 8 and Supplementary Fig. 9 are provided as a Source Data files. Source data are provided with this paper.