Automated volumetric MRI quantification of body fat and skeletal muscle in the UK Biobank: technical validation against expert manual segmentations and comparison with single-slice techniques

Abstract

Objectives

The escalating global incidence of obesity, cardiometabolic disease and sarcopenia necessitates reliable body composition measurement tools. MRI-based assessment is the gold standard, with utility in both clinical and drug trial settings. This study aims to validate a new automated volumetric MRI method by comparing with manual ground truth, prior volumetric measurements, and against a new method for semi-automated single-slice area measurements.

Methods

4905 individuals from the UK Biobank with repeat whole-body Dixon MRI scans were selected. MRI data were processed automatically to derive new (1) volumetric and (2) single-slice area measurements at L3 vertebral level for visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and abdominal skeletal muscle (SM). For comparison, prior volumetric measurements of VAT and SAT were included. A separate set of scans from 100 subjects was randomly selected and body composition volumes and areas were manually segmented as ground truth.

Results

The new automated volumetric measurements were found to have excellent agreement with manual ground truth with no substantial bias (ICC ≥ 0.96, CoV ≤ 3.8%). In the cohort of 4905 individuals (49% male, mean age 62 years ± 8, BMI 26 kg/m² ± 4), we confirmed that prior and new volumetric methods of VAT and SAT measurement were very strongly correlated (VAT: ρ = 0.99; SAT: ρ = 1; both p < 0.001). Single-slice L3 area measurements demonstrated very strong correlations with corresponding volumes (for VAT ρ = 0.97, for SAT ρ = 0.94, for SM ρ = 0.95, all p < 0.001), and remained excellent across sex, age, and cardiometabolic characteristics (median ρ = 0.95).

Conclusion

We found robust correlations between manually segmented, automated volumetric, and semi-automated single-slice body composition methods. The interchangeability of these methods suggests that for each application the method should be selected according to practical considerations, operational differences and measurement granularity, rather than technical performance.

Supplementary Information

The online version contains supplementary material available at 10.1007/s00261-025-05170-w.

Introduction

Obesity affects 1 billion people worldwide [1], with increased rates of cardiometabolic disease and escalating healthcare costs in treating them [2–5]. Obesity-associated comorbidities include heart failure, kidney disease, steatotic liver disease, diabetes, arthritis, and increased rates of cancer. Body Mass Index (BMI) is a simple and commonly accepted measure, but fails to account for body fat and muscle distribution [6, 7], which vary by sex, age, ethnicity, genetic profile, and environmental factors, and influence disease progression and outcomes [6–8]. A recent consensus statement from the Lancet Diabetes & Endocrinology Commission on Clinical Obesity recommends that direct measurement of body fat, not BMI, is a diagnostic criterion for clinical obesity and that organ dysfunction is a co-diagnostic criterion [9]. In patients with obesity, dietary, surgical or pharmacologically-assisted weight loss is beneficial, but there are growing concerns about the loss of skeletal muscle, which is associated with poor outcomes [10–13].

Magnetic resonance imaging (MRI) remains the gold standard for in vivo quantification of skeletal muscle and adipose tissue, due to its capability to provide accurate, high-resolution data on tissue distribution and size non-invasively [14–16]. MRI-derived body composition measurements offer quantitative metrics essential for risk stratification, treatment planning and monitoring of therapeutic efficacy, whether through lifestyle, pharmacologic or surgical interventions. The scale of obesity has accelerated the development of automated and semi-automated methods to quantify body fat and muscle to meet the different requirements for granularity and throughput in both clinical practice and clinical trial settings. Both volumetric and single-slice area methods have high technical performance in terms of repeatability and reproducibility [17–19]. While volumetric MRI remains the gold standard for comprehensive body composition assessment and is particularly valuable for tracking longitudinal changes, its extended acquisition and processing times may limit throughput in routine clinical settings. Consequently, semi-automated single-slice approaches require only a single breath-hold acquisition, significantly reducing both scan time and processing demands. However, these methods have not been compared directly to assess their performance, interchangeability, and fit-for-purpose applications in the context of obesity management.

In this study, we aimed to validate different automated MRI methods for assessing body fat distribution, specifically volumetric measurements and single-slice measurements at L3. Using data from the UK Biobank that represent a relevant population of adults with a range of cardiometabolic risk factors, we first validated a new automated volumetric method against manually segmented data for ground truth. We subsequently performed comparisons with measurements from a prior volumetric method and compared with a new semi-automated method of single-slice measurement.

Methods

Study design and populations

This study utilized data from participants enrolled in the UK Biobank imaging sub-study between January 2016 and February 2020 [20]. MRI scans from the baseline imaging visit of 4905 individuals were included. Body composition measurements were derived using a new fully automated volumetric method and a new semi-automated single-slice area method at the 3rd lumbar vertebra (L3) level (Perspectum Ltd., Oxford, UK) (Fig. 1). Measurements comprised volumes and areas of visceral adipose tissue (VAT), subcutaneous adipose tissue (SAT), and skeletal muscle in the abdominal region (SM).

Fig. 1.

Body composition methods included in this work. (A) Comparison of the new volumetric body composition method against the following: Manual volumetric ground truth, automated prior volumetric method, and new semi-automated single-slice method. (B) Schematic of body composition measurements derived with the new methods (volumetric and single-slice). Volumetric measurements spanned from T9 to the top of the femoral head and single-slice area measurements were centred at the L3 vertebral region

For additional comparison, volumetric measurements from a different vendor, previously processed using an automated method (AMRA Profiler, Medical, Linkoping, Sweden) [13, 21], were also included. The measurements obtained with this processing method, referred to as the “prior volumetric method”, included VAT (Field ID 22407) and SAT (Field ID 22408) data from the same participants and required no additional analysis.

A separate subset of MRI scans from 100 UK Biobank individuals were manually segmented independently of this study to provide the ground truth for body composition analysis. These were randomly selected as a representative cohort, reflecting a diverse range of age, BMI, gender and physical activity levels. UK Biobank has approval from Northwest Multi-Centre Research Ethics Committee (REC reference: 11/NW/0382) and obtained written informed consent from all participants prior to the study [22].

MRI data acquisition

All participants had been scanned at one of the UK Biobank imaging centres, on a Siemens Aera 1.5 T scanner (Siemens Healthineers, Erlangen, Germany), with a 6-min dual-echo Dixon VIBE protocol [23]. This provided an image dataset with separate water and fat blocks from the neck to the knees with an adjacent slab overlap of 2–4 cm. Imaging parameters included a 10° flip angle, an in-plane resolution of 2.23 mm × 2.23 mm, and a slice thickness ranging from 3 to 4.5 mm. These images were used for all segmentations (automated, semi-automated, and manual).

New volumetric and single-slice area body composition methods

The new fully automated volumetric method builds on the approach described in [13, 18] to derive body composition metrics in the abdominal region (from the T9 vertebra to the femoral heads), replacing the original multi-atlas technique with neural networks. These networks were trained on multiple datasets to segment body structures (Supplementary Methods). In addition to VAT and SAT, the method was extended to compute SM and skeletal muscle index (SMI) in the abdominal region, in contrast to previous applications [13, 18] that encompassed the whole body or thigh. SMI was calculated as skeletal muscle CSA (cm²) divided by height squared (m²) for single-slice measurements, and abdominal muscle volume (here in cm³) divided by height squared (m²) for volumetric measurements.

The single-slice area method comprises a new method of single-slice selection and area measurements derived from a separate sequence of neural networks trained and validated on different datasets (Supplementary Methods).

Abdominal volumetric analysis

For volumetric analysis, the region from the centre of the T9 vertebra to the top of the lower of the two femoral heads was selected. To automatically identify the region limits, vertebrae and femurs were segmented in the stitched fat volume using a neural network (configuration described in Supplementary Methods). Each independent vertebra was given a unique label from inferior to superior, starting with L5 upwards. Each femur was given a unique label, and the top of the lower femoral head was identified as the more inferior of the highest labelled slice for the two femurs. The identified slices were used to crop the fat, water and SFF volumes to the analysis region. The cropped fat and water volumes were then used as input to an ensemble of Unet + +) that segmented SAT, muscle tissue, abdominal cavity, and bone (see Supplementary Methods, including Supplementary Table 1, for configuration, hyperparameters, and datasets). The abdominal cavity was defined as the compartment that extended superiorly from below the diaphragm to the rectum within the pelvic bones inferiorly.

The dataset included several instances of fat–water swapping, which negatively impacted the fat fraction calculation, segmentation performance, and metrics accuracy. Cases with fat swaps were detected using a ResNet [24] (see Supplementary Methods, including Supplementary Table 2, for configuration, hyperparameters, and datasets) and corrected by re-labelling the affected blocks before the stitching step, avoiding the swap artefacts in the whole-body volume.

During pre-processing any fat–water swaps detected were visually confirmed by a manual operator. Post-processing quality control for abdominal volumetric analysis was not directly performed on the images or segmentations, as this had been performed previously in the prior volumetric measurements [13, 18]. Instead, quality control issues were examined and addressed only in cases flagged or excluded by the prior volumetric measurements. This approach ensured that only reliable data, based on both vendors’ predefined quality criteria, was included in the analysis, removing potential bias on the performance assessment. The above steps are illustrated in Fig. 2.

Fig. 2.

Summary of analysis workflow for new body composition methods: (A) abdominal volumetric analysis and (B) single-slice cross-sectional area (CSA) analysis

Single-slice cross-sectional area analysis

For single-slice analysis, axial slices passing through the centre of mass of the L3 vertebra were automatically extracted from the whole-body volumes for measurement. A separate neural network model, distinct from the volumetric segmentation model, was used to segment all vertebrae as a single class (see Supplementary Methods, including Supplementary Table 3 for configuration, hyperparameters, and datasets).

Following segmentation, morphological operations were applied to ensure that the vertebrae were separated from one another. This enabled the subsequent labelling of individual, unconnected elements, which were then counted in a superior direction starting from the sacrum. The third most inferior segmented element was identified as L3.

Within the extracted axial slice at the centre of L3 vertebra, ROI were segmented in a two-stage neural network (see Supplementary Methods, including Supplementary Table 4 for configuration, hyperparameters, and datasets). Firstly, a U-Net neural network classified voxels into categories: SAT, abdominal cavity, bone, and abdominal SM (including coarse segmentations of the right and left back muscles, right and left psoas muscles, and abdominal wall muscles). Secondly, an additional U-Net neural network was applied to a focused sub-region of the image to perform fine-grain segmentations of the psoas muscles.

Two rounds of manual quality control (QC) were performed on the entire population of 4905 participants to ensure model reliability. The first QC step assessed the L3 centre of mass placement, rejecting cases where L3 was poorly visible or repositioning the landmark as needed. The second QC step involved a visual review of single-slice segmentations, rejecting cases with segmentation errors or unclear tissue boundaries (Fig. 2).

Extraction of body composition metrics

Volume or cross-sectional area of fat and muscle tissues was quantified by filtering segmented voxels based on SFF values. Fat tissue was defined as having SFF values ≥ 0.5, and muscle tissue as having values < 0.5, ensuring that any over-segmented voxels from adjacent tissues were excluded.

Manual segmentation of volumetric body composition measurements

For the assessment of the accuracy of the automated volumetric body composition measurements described in this study, we utilized a retrospective subset of 100 UK Biobank cases, each manually and independently segmented by one of three expert operators with  > 20 years’ experience in body composition segmentation and anatomy. Each dataset was annotated by a single expert using the freely available software ITK-SNAP [25]. Preprocessing and stitching were performed as described previously [26, 27]. For manual measurements of VAT, SAT, and SM volumes, the manual segmentations were prospectively cropped by 2 expert operators to match the z-axis range of their corresponding automated datasets. This approach standardises abdominal coverage, focusing on the region between the centre of the T9 vertebra and the top of the most inferior femoral head, thereby minimising measurement bias.

Statistical analysis

All statistical analyses were performed in R (version 4.3, R Project for Statistical Computing, Vienna, Austria) [28]. All significance tests were two-tailed, and p < 0.05 was considered statistically significant. Bland–Altman analysis was performed to assess bias and 95% limits of agreement between the automated volumetric method and the manual ground truth. Reliability of the automated methods relative to the manual reference was quantified using intraclass correlation coefficients (ICC), and measurement variability was assessed using coefficients of variation (CoV). Correlations between volumetric and between volumetric and single-slice measurements were assessed using Spearman’s rank correlation coefficient (ρ). Further subgroup analyses were performed by gender, age group, BMI category, and waist circumference (WC) category to assess whether the correlations between methods varied across these subgroups. The World Health Organisation standards for BMI and WC stratification were applied [29].

Results

Data processing

A total of 4905 cases with available body composition images were selected for comparison. In the processing with the new volumetric method (commercially available from Perspectum Ltd., Oxford, UK), 103 cases of 4905 (2.1%) that exhibited fat–water swap issues were detected and corrected. Eight (0.16%) cases for VAT, SAT, and SM were excluded due to problems during the un-swap process or failed segmentations in the intermediate steps. Existing UK Biobank imaging-derived phenotypes (IDPs), using a prior volumetric method (AMRA Medical, Linkoping, Sweden, [13]), had missing values in 4 and 5 cases for VAT and SAT, respectively. Abdominal SM was not calculated by the prior method and thus could not be compared.

With the study new single-slice area method, 37 of 4905 (0.75%) cases were rejected during QC of L3 placement (mostly misplacement of L3 and fat–water swaps), and 235 (4.79%) cases were rejected during QC of the segmentations (mostly due to inaccurate abdominal cavity boundaries or missed psoas muscle).

None of the 100 cases were excluded from the ground truth manual segmentations and the new volumetric method was successful in all cases.

Participant characteristics

4905 individuals were included overall (Supplementary Table S5), with a mean age of 62 ± 8 years, and BMI of 26 ± 4 kg/m². 97% were White, 49% were male, 14% had obesity. The 100 individuals included in ground truth manual segmentations had a mean age of 64 ± 7 years, and BMI of 26 ± 3 kg/m². 98% were White, 47% were male, 4% had obesity. (Supplementary Table S6).

Validation of new automated volumetric measurements against ground truth manual measurements

The new automated volumetric method for VAT, SAT, and SM showed strong agreement with manual ground truth measurements. All measurements showed low variability (CoV 2.4% for SAT and abdominal SM, 3.2% for VAT) and excellent reliability (ICC 0.96 to 0.99) (Table 1). Bland Altman plots showed that the two methods had good agreement with small biases (ranging from 0.21 to − 0.31L and representing 5.3 to 5.7% of the mean values). The differences between methods were even across the measurement ranges, indicating minimal systematic and proportional effects (Fig. 3).

Table 1.

Inter-method agreement and variability: Comparison of manual ground truth measurements vs. new automated volumetric method for VAT, SAT, and abdominal SM, and vs. measurements of VAT and SAT from the prior automated volumetric method

Volume measurement	Mean value—manual method (L)	Mean value—automated method (L)	Bias [95% CI] (L)	Lower LoA (L)	Upper LoA (L)	Within-subject SD of differences (L)	ICC	CoV (%)
New automated volumetric method
VAT	3.6 ± 2.1	3.4 ± 2.1	0.21 [0.19, 0.24]	− 0.09	0.52	0.11	0.99	3.2
SAT	5.9 ± 2.4	6.2 ± 2.4	− 0.31 [− 0.35, − 0.28]	− 0.72	0.09	0.15	0.99	2.4
Abdominal SM	5.1 ± 1.2	5.4 ± 1.2	− 0.29 [− 0.32, − 0.26]	− 0.64	0.05	0.13	0.96	2.4
Prior automated volumetric method
VAT	3.6 ± 2.1	3.4 ± 2.0	0.20 [0.17, 0.24]	− 0.16	0.57	0.19	0.99	3.8
SAT	5.9 ± 2.4	6.4 ± 2.6	− 0.47 [− 0.51, − 0.43]	− 0.93	0.00	0.24	0.98	2.8

CI confidence interval, CoV coefficient of variation (%), ICC intraclass coefficient, L litres, LoA limits of agreement; SD, standard deviation

Fig. 3.

Bland–Altman plots showing inter-method variability between the new volumetric and manual ground truth measurements for VAT (A), SAT (B), abdominal SM (C)

Cross-vendor comparison of automated volumetric body composition measurements

The VAT volumes measured by the new volumetric method showed excellent correlation with VAT volumes derived by the prior volumetric method (ρ = 0.99, p < 0.001) (Fig. 4). For SAT volumes, the values from both methods also had excellent correlation (ρ = 1, p < 0.001). The correlations were maintained regardless of sex, age group, BMI group, waist circumference group, and diabetes status (VAT: median ρ = 0.99, SAT: median ρ = 1, abdominal SM: median ρ = 0.95, all p < 0.0001) (Table 2).

Fig. 4.

Correlations between automated volumetric measurements derived by the new and prior methods for VAT (A) and SAT (B). Each hexagonal bin is colour-coded to reflect the density of data points

Table 2.

Vendor-to-vendor comparisons for volumetric measurements: comparison of VAT and SAT derived by the new and prior methods

	VAT		SAT
	N	Rho (ρ)*	N	Rho (ρ)*
Total	4898	0.99	4897	1
Sex
Male	2382	0.99	2382	0.99
Female	2516	0.99	2515	1
Age (years)
40–49	103	0.99	103	0.99
50–59	1803	0.99	1802	1
60–69	2013	0.99	2013	1
> 70	979	0.99	979	1
BMI (kg/m2)
Normal (18.5 ≤ BMI < 25)	1927	0.98	1927	0.99
Overweight (25 ≤ BMI < 30)	1888	0.99	1888	0.99
Obese (BMI ≥ 30)	626	0.99	626	1
Waist circumference (cm)
Healthy male (< 102)	1769	0.99	1769	0.99
Obese male (≥ 102)	513	0.98	513	0.99
Healthy female (< 90)	1695	0.98	1695	1
Obese female (≥ 90)	710	0.98	709	1
Type 2 diabetes (T2D) status
T2D	266	0.99	266	1
No T2D	4632	0.99	4631	1

*All pairwise correlations showed p < 0.0001

These vendor-to-vendor comparisons showed minimal bias (ranging from − 0.04 to − 0.09L and representing 1.1–1.3% of the mean values) and excellent reliability (ICC of 1), indicating strong agreement (Supplementary Table S7). This was confirmed when the prior volumetric method was compared to ground truth method and, like the new method, showed good agreement with small biases and low variability (CoV 2.8% for SAT and 3.8% for VAT) (Supplementary Fig S1, Table 1).

Comparisons between single-slice and volumetric body composition measurements

Single-slice L3 measurements of SAT, VAT, and abdominal SM demonstrated very strong correlations with volumes of VAT (ρ = 0.97, p < 0.001), SAT (ρ = 0.94, p < 0.001), and abdominal SM (ρ = 0.95, p < 0.001) derived herein (Fig. 5). These correlations were maintained regardless of sex, age group, BMI group, waist circumference group, and diabetes status (VAT: median ρ = 0.97, SAT: median ρ = 0.93, abdominal SM: median ρ = 0.94, all p < 0.001) (Table 3, Supplementary Table S8). In addition, a strong correlation was observed between single-slice and volumetric SMI (SM indexed to subject height squared) across the entire cohort (ρ = 0.86, p < 0.001) and across demographic subgroups (median ρ = 0.81, all p < 0.001).

Fig. 5.

Correlation between the new volumetric and single-slice measurements for VAT (A), SAT (B), and abdominal SM (C). Each hexagonal bin is colour-coded to reflect the density of data points

Table 3.

Correlation analysis comparing this study’s new volumetric and single-slice body composition measurements (VAT, SAT, abdominal SM)

	VAT		SAT		Abdominal SM
	N	Rho (ρ)*	N	Rho (ρ)*	N	Rho (ρ)*
Total	4632	0.98	4632	0.93	4632	0.95
Sex
Male	2314	0.96	2314	0.91	2314	0.84
Female	2318	0.97	2318	0.94	2318	0.82
Age (years)
40–49	99	0.98	99	0.94	99	0.95
50–59	1693	0.97	1693	0.94	1693	0.95
60–69	1914	0.98	1914	0.93	1914	0.95
> 70	926	0.97	926	0.90	926	0.93
BMI (kg/m2)
Normal (18.5 ≤ BMI < 25)	1758	0.96	1758	0.88	1758	0.94
Overweight (25 ≤ BMI < 30)	1848	0.96	1848	0.88	1848	0.94
Obese (BMI ≥ 30)	609	0.95	609	0.90	609	0.95
Waist circumference (cm)
Healthy male (< 102)	1719	0.95	1719	0.88	1719	0.82
Obese male (≥ 102)	502	0.90	502	0.90	502	0.86
Healthy female (< 90)	1537	0.95	1537	0.91	1537	0.82
Obese female (≥ 90)	678	0.94	678	0.89	678	0.83
Type 2 diabetes (T2D) status
T2D	257	0.97	257	0.93	257	0.94
Non-T2D	4375	0.98	4375	0.93	4375	0.95

^*All pairwise correlations showed p < 0.0001

Discussion

The increasing prevalence of obesity, sarcopenia, and cardiometabolic diseases underscores the need for reliable, scalable body composition measurement tools for different applications, and MRI is the gold standard. In this study, we tested a new vendor’s method for automated volumetric measurements of fat and muscle against both ground truth manual segmentation and a prior volumetric MRI processing method, and against a new semi-automated single-slice measurement method in a cohort of 4905 individuals. We found excellent agreement between the new fully automated volumetric method and the ground truth manual method, with only small biases observed that were within measurement error for volumetric methods. This performance was similar to the prior volumetric method, that had been validated previously [18] and also in this work. The new method showed near-perfect correlations to volumetric MRI measurements for VAT and SAT available with the prior method, and excellent correlations with the new semi-automated single-slice method for VAT, SAT, and abdominal SM. These metrics have utility in obesity and within subgroups with increased cardiometabolic risk. The high technical performance and interchangeability of all MRI body compositions processing methods, described here for the first time, will enable broader adoption, increasing efficiency in clinical trials and targeted screening in clinical management.

Skeletal muscle index (SMI), calculated by the ratio of muscle area to body height squared (cm²/m²), is frequently used to measure sarcopenia [23]. Lower SMI is a marker of increased risk of metabolic syndrome [30], poor outcomes in cancer [31], Alzheimer’s and Parkinson’s disease [32, 33], mortality [34], and poor quality of life [34]. In clinical guidelines, particular attention has been focused on diagnosis and monitoring of sarcopenic obesity to improve patient management and outcomes [35]. Adding SMI measurement to the standard clinical workflow may be of particular interest to monitor patients on anti-obesity treatment (for example, GLP-1 receptor agonists) and bariatric surgical intervention [36]. In our study, body composition methods correlated well in individuals with obesity, those with overweight, and those with healthy weight; therefore, they may be applied to evaluate muscle loss in these patient populations. This finding is consistent with prior smaller-scale studies comparing volumetric and single-slice methods, by MRI and by CT [37–40].

Consensus on the optimal topographical location of SM measurement for sarcopenic obesity is lacking [41]. SM in the abdomen, whether assessed from a single-slice image or a series of images, serves as a strong marker of whole-body SM [42]. Abdominal imaging of VAT, SAT, and SM/SMI (without incorporating the thighs), as described herein with both volumetric and single-slice new methods, has operational advantages. For example, it eliminates the need for RF coil changes and reduces the likelihood of missing data from partial imaging of one or both thighs [21], streamlining the scanning process for patients and study participants.

Providing interchangeable, modular solutions facilitates both research and clinical settings. When evaluating a workflow, there are several factors to consider beyond that of acquisition time. The cost–benefit ratios depend on the workflow as a whole: the time required for post-processing, segmentation, and reporting, as well as patient comfort. A neck-to-knee body composition workflow with fully automated post-processing and analysis may be best suited to evaluation of nuanced changes in fat and muscle distribution. The present study validates the accuracy of the UK Biobank default prior volumetric body composition measurements and finds that they can be reproduced utilising fully automated algorithms. The new fully automated method achieved excellent and equivalent performance to the prior method, without post-processing manual checks of every case, as described for the prior method. The volumetric methods presented in this study are ideally suited for cardiometabolic clinical trials, such as for anti-obesity therapies, which can achieve rapid weight loss and may impact muscle quality [43, 44].

To enhance clinical applicability, body composition assessments must balance accuracy with practicality. While volumetric MRI is highly accurate and comprehensive, its extended acquisition and processing times may limit throughput in clinical settings. Herein, a single-slice area-based body composition approach is semi-automated and provides measurements that mirror those of the volumetric methods but requires just a single breath-hold to acquire data. The L3 region has previously been identified as an optimal site for single-slice measurement of visceral adiposity [37]. In this study, for the first time, it was validated at scale using MRI, alongside SAT and abdominal SM. The shorter acquisition and processing time with included manual quality checking of single-slice body composition assessment should enable SMI to be measured as part of routine abdominal MRI if screening, for example, patients with obesity and Crohn’s disease.

The consequences of weight loss with glucagon-like peptide-1 (GLP-1) agonists are pleiotropic and improve the health of other organs, including the heart and liver [45]. Body composition assessment combined with multi-organ imaging could enable efficient and precise patient phenotyping for clinical trial screening. Such approaches, already adopted in the UK Biobank [23, 46, 47] and in cohorts with cardiometabolic disease [48, 49] can enhance efficiency of research trials and advance holistic medicine. This is particularly important as obesity impacts a wide range of organ systems, necessitating comprehensive evaluation strategies [10, 50, 51]. A clinical workflow for obesity that includes direct measurement of body fat, and signs, symptoms or tests for organ dysfunction or substantial limitations for daily activities has been recommended by recent expert consensus [9]. Future multi-organ outcomes studies incorporating the methods described herein are warranted.

Limitations

We did not compare to DEXA nor bioimpedance, as this has been done previously [52], with additional literature showing the variance and shortcomings of these techniques compared to MRI [14]. The UK Biobank cohort comprises predominantly White participants aged > 40 years, and in our study only approximately 1% (N = 41) had severe obesity (BMI > 40 kg/m²), so caution must be exercised in applying these results to other populations and age groups. However, this study supports that substitution of either volumetric or single-slice protocols could enable comparisons to data previously collected, facilitating both interpretability and translatability of new research. These findings should enable broader adoption of body composition MRI with less reliance on specific vendor algorithms.

Conclusion

This research validates robust, interchangeable tools that enable consistent data analysis, promoting the broader adoption of MRI-based body composition assessment as obesity increases in prevalence and cardiometabolic treatments continue to evolve. For measurement of full volume in clinical trials, methods are interchangeable to assess fat volume, and the new volumetric method also measures abdominal skeletal muscle. For clinical adoption, further longitudinal data with the single-slice area method in the real-world are needed.

Supplementary Information

Below is the link to the electronic supplementary material.

Author contributions

Methodology was developed by: L.N., C.E.H., J.M. and M.D.R. Data curation was performed by M. Niglas, M. Nowak, C.B.B., L.N., C.E.H. and. Formal analysis of the data was performed by M. Nowak and M. Niglas. Supervision was by M.D.R, H.T.B., E.L.T. and J.D.B. The first draft of the manuscript was written by M. Nowak and H.T.B. The final manuscript was conceived by H.T.B. All authors contributed to manuscript review and editing.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Data availability

Summary data is included in the manuscript or uploaded as online supplemental information. Anonymized individual patient data can be shared upon request or as required by law and/or regulation and/or governance by and within the rules of UK Biobank access with qualified external researchers. Approval of such requests is at the discretion of the study sponsors and is dependent on the nature of the request, the merit of the research proposed, the availability of the data, and the intended use of the data. Source code for methods herein is not disclosed as it is commercially sensitive.

Declarations

Competing interests

Magdalena Nowak, Luis Núñez, Charles E. Hill, Tim Pagliaro, John McGonigle, Matthew D. Robson, Helena Thomaides Brears are employees at Perspectum Ltd. John McGonigle, Matthew D. Robson, Helena Thomaides Brears are also shareholders of Perspectum Ltd. Louise E. Thomas and Jimmy D. Bell are consultants for Perspectum Ltd. The remaining authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.