Adaptive deep learning for quantification and spatial distribution of abdominal adipose tissue from magnetic resonance imaging proton density fat fraction in adults with a body mass index ≥24 kg/m²

Introduction

As the prevalence of overweight and obese individual continues to rise across the world, obesity-related metabolic diseases have emerged as a major public health concern (1,2). Research indicates that both the volume and spatial distribution of adipose tissue are closely linked to individual metabolic risk (3-5). Distinct adipose depots, such as visceral adipose tissue (VAT) and subcutaneous adipose tissue (SAT), exhibit markedly different metabolic functions. Increased VAT accumulation is strongly associated with higher risks of metabolic syndrome, cardiovascular disease, and multiple cancers, whereas SAT expansion is generally considered to correspond to a more favorable metabolic phenotype. Additionally, heterogeneity in VAT deposition across different anatomical regions may carry distinct inflammatory features and metabolic implications (3). Therefore, accurate quantification of abdominal adipose tissue and systematic characterization of its distribution are of considerable importance for obesity phenotype stratification and metabolism-related research.

Several medical imaging techniques facilitate the noninvasive quantification of human adipose tissue (6-13), among which magnetic resonance imaging (MRI) chemical shift-encoded water-fat separation is the most widely used (14-21). Proton density fat fraction (PDFF) images provide voxel-level fat information, overcome the effects of radiofrequency field and main magnetic field inhomogeneities, and have demonstrated good accuracy and repeatability across different field strengths (19-21), thus providing a reliable imaging basis for fat volume quantification. With the establishment of large-scale imaging cohorts (14-16,22,23), MRI-based body composition phenotyping studies have been conducted at the population level. Previous studies have mainly focused on fat volume measurements and their associations with metabolic indices and have investigated age- and sex-related differences in adipose tissue distribution in large populations (14-16), as well as its relationships with metabolic and genetic factors (22,23). However, the spatial distribution characteristics of adipose tissue may vary across populations, imaging coverage, and analytical strategies. Moreover, there is a lack of systematic analyses on abdominal MRI PDFF images that integrate automated whole-slice fat quantification with spatial mapping for examining Chinese individuals with overweight and obese status.

Accurate volumetric quantification and spatial characterization of abdominal adipose tissue depend on reliable image segmentation of adipose depots (24). Abdominal SAT and VAT are elongated, anatomically continuous structures with complex boundaries and substantial interindividual variability. Manual slice-by-slice delineation is time-consuming, labor-intensive, and subject to interobserver variability, making it impractical for large-scale population studies or longitudinal analyses. To address these challenges, various automated adipose segmentation approaches have been proposed, including semiautomated segmentation (25), multi-atlas segmentation (26-28), rule-based fully automated segmentation (29-32), and machine learning algorithms (33-38). Building on this foundation, U-shaped network (U-Net)-based deep learning methods have been widely applied to image segmentation tasks in recent years and have achieved favorable performance. However, differences across models in network architecture design, parameter configuration, and training strategies have increased the complexity of methodological reproducibility and cross-dataset application. The no new U-Net (nnU-Net) framework (39-43) mitigates, to some extent, the dependence of deep learning-based segmentation methods on manual expertise by automatically configuring the network architecture, data-preprocessing pipeline, and training strategy, thereby providing an adaptive solution for adipose tissue segmentation. A previous study demonstrated the high accuracy and consistency of three-dimensional (3D) U-Net and nnU-Net for SAT and VAT segmentation in multicenter longitudinal MRI datasets (40). Beyond volumetric analysis, previous work has systematically investigated the craniocaudal spatial distribution of SAT and VAT based on the nnU-Net framework and revealed sex- and age-related differences (16). However, whether this approach can efficiently and consistently capture spatial adipose phenotypes in Chinese populations who are overweight or obese and characterized by substantial interindividual heterogeneity and high metabolic risk, remains to be validated.

This study primarily aimed to perform methodological validation by evaluating the accuracy and feasibility of an nnU-Net-based adaptive deep learning framework for automated segmentation, volumetric calculation, and PDFF quantification of abdominal SAT and VAT on MRI PDFF images in overweight and obese adults with a body mass index (BMI) ≥24 kg/m². On the basis of automated segmentation, we further characterized spatial abdominal fat distribution patterns and explored depot-specific differences across age and sex groups within this population. This study provides a methodological foundation for the automated imaging-based quantification of adipose phenotypes in Chinese populations and may serve as a reference for future studies in this area. We present this article in accordance with the TRIPOD+AI reporting checklist (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0094/rc).

Methods

Data acquisition

A total of 238 abdominal MRI datasets from overweight or obese adult individuals were included for analysis. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The study protocol was approved by the Ethics Committee of Shanghai East Hospital and included a retrospective study design, which had no impact on the clinical diagnosis or treatment of the participants, and thus the requirement for informed consent was waived. Data were retrospectively collected from individuals who underwent abdominal MRI examinations at Shanghai East Hospital between January 2020 and December 2022. The exclusion criteria were as follows: (I) history of major abdominal surgery resulting in substantial alterations of abdominal anatomy; (II) presence of diagnosed metabolic or systemic diseases, including impaired glucose tolerance, diabetes mellitus, malignancy, hormone-related syndromes, or chronic wasting disorders; and (III) MR images with considerable artifacts or insufficient quality that precluded subsequent analysis.

MRI data were acquired with an Ingenia 3.0-T MRI system (Philips Healthcare, Best, the Netherlands) under a standardized imaging protocol. All participants underwent abdominal MRI examinations in the supine position after fasting for at least 10 hours. The mDixon MRI Quant sequence was applied under the following parameters: repetition time =15 ms, echo times =6 first. echo time/echo spacing =1.15 ms/1.15 ms, flip angle =3° (to minimize T₁ bias in PDFF quantification), field of view =320 mm × 320 mm, slice thickness =3 mm, and acquisition matrix =256×256. 3D axial images covering the region from the diaphragmatic dome to the pelvic floor were acquired during a single breath-hold. After acquisition, multislice two-dimensional (2D) axial images were reconstructed, with water images, fat images, PDFF images, R₂^* maps, and T₂^* maps being automatically generated. Only PDFF images were used for subsequent analysis.

Automated segmentation model

For automated retrospective analysis of abdominal MRI data, a total of 238 abdominal PDFF MRI datasets were included. Manual annotations of SAT and VAT were obtained for all participants. The dataset was split at the participant level into a development set (n=202, approximately 85%) and an independent internal test set (n=36, approximately 15%). The development set was used for model training, with fivefold cross-validation performed within this cohort for model optimization and internal validation. The independent internal test set was reserved exclusively for final model performance evaluation. Imaging data from the same participant were not shared between the development and independent internal test sets to avoid data leakage. During construction of the development cohort, stratified control was applied according to sex and the distribution of manually segmented total adipose tissue (TAT) volume to ensure that the training data were representative of the demographic and adipose distribution characteristics of the overall study population. Table 1 presents the demographic characteristics and adipose tissue volume distributions of the overall study population, development set, and independent internal test set, including the number and percentage of participants by sex, age group, and BMI category, as well as volume information for SAT, VAT, and TAT. These data were used to describe the population composition and distribution of obesity severity across the different datasets. To assess the effect of model dimensionality on abdominal adipose segmentation, both 2D nnU-Net and 3D nnU-Net models were implemented and compared. As illustrated in Figure 1, both models included standard U-Net architectures in which each resolution level consisted of two basic convolutional blocks, each comprising a convolution operation followed by instance normalization and a Leaky rectified linear unit activation. Downsampling and upsampling were achieved via strided convolution and transposed convolution, respectively, to ensure progressive resolution restoration. The 2D nnU-Net processed individual slices and generated slice-wise predictions, whereas the 3D nnU-Net received volumetric patches as input and produced voxel-wise predictions, thereby explicitly leveraging interslice spatial context. The batch size was set to 32 for the 2D model and to 2 for the 3D model due to graphics processing unit (GPU) memory constraints. Both models were trained with an out-of-the-box configuration under fivefold cross-validation. Following the recommendation from the original nnU-Net, ensemble predictions from the five cross-validation folds were used as the final output (39). The 2D and 3D nnU-Net models were trained on the same training data and evaluated on the same independent test set. Performance was quantified via class-specific metrics derived from confusion matrices, including Dice similarity coefficient (DSC), precision, and recall. Additionally, volumetric accuracy was assessed by comparing automated and manual fat volume measurements to compute absolute differences (mL) and relative volume errors (%).

Figure 1 Network architectures generated by nnU-Net for the abdominal PDFF dataset. 2D, two-dimensional; 3D, three-dimensional; IN, instance normalization; nnU-Net, no new net; PDFF, proton density fat fraction.

Manual segmentation

All 238 abdominal MRI datasets underwent segmentation and volumetric quantification of SAT and VAT within the abdominal cavity. Manual segmentation was performed with 3D Slicer software version 5.2.2 (https://www.slicer.org). Manual annotation was performed by 1 doctoral student and 10 undergraduate students under the supervision of 2 radiologists experienced in abdominal imaging. Before annotation, all annotators underwent standardized training and assessment, which covered anatomical identification of abdominal adipose tissue, interpretation of PDFF images, criteria for defining SAT and VAT boundaries, and operation of the segmentation software. After passing the assessment, the annotators were divided into groups to perform adipose tissue annotation and volume quantification on axial MRI. To reduce interobserver variability, the doctoral student uniformly determined the starting and ending slices and segmentation threshold for each participant, ensuring appropriate differentiation between adipose and nonadipose tissues. All annotations were subsequently reviewed and confirmed by the radiologists. The final dataset included 238 participants, with 62 annotated images for each participant. Figure 2A illustrates the annotation range for abdominal subcutaneous and intra-abdominal adipose tissue, and Figure 2B-2D show representative axial examples from three participants.

Figure 2 Manual segmentation of abdominal adipose tissue. (A) The red dashed line indicates the boundary of the manually segmented region. (B-D) Representative axial examples of manually segmented SAT (yellow) and VAT (red). (B) Participant with a BMI of 33.3 kg/m² and a VAT proportion of 16%. (C) Participant with a BMI of 29.2 kg/m² and a VAT proportion of 66.1%. (D) Participant with a BMI of 25.0 kg/m² and a VAT proportion of 35.9%. BMI, body mass index; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

To assess the internal consistency of manual segmentation, five cases were selected from the overall sample by stratified random sampling and were resegmented. For intrareader similarity assessment, the primary annotator repeated the segmentation of these five cases after a 2-month interval. For interreader agreement assessment, another annotator independently segmented the same five cases. Spatial overlap agreement of the segmentation masks was evaluated via the DSC, whereas agreement in volume measurements was assessed according to the intraclass correlation coefficient (ICC). After completion of the initial annotation, radiologists performed quality-control review of a subset of 15 cases. All 15 reviewed cases underwent corrections or boundary refinements to varying degrees, corresponding to a case-level modification rate of 15/238. For cases with unclear SAT and VAT boundaries, substantial interference from intra-abdominal structures, or inconsistent annotations, the annotations were first rechecked by the primary annotator and subsequently revised after radiologist review. The final confirmed segmentation masks were used as the reference standard for model training and performance evaluation.

Anthropometric data

The height and body weight of all participants were measured and recorded with calibrated stadiometers and weighing scales. BMI was calculated as the body weight (kg) divided by the square of the height (m²). According to the Chinese BMI classification criteria, overweight status was considered to be a BMI between 24 and 28 kg/m² and obesity a BMI ≥28 kg/m². Obesity was further categorized as mild obesity (28≤ BMI <32 kg/m²), moderate obesity (32≤ BMI <36 kg/m²), and severe obesity (BMI ≥36 kg/m²)

Fat quantification and spatial distribution analysis

Adipose tissue quantification was performed with PDFF images automatically generated from the mDIXON Quant sequence. PDFF was defined as the proportion of the fat signal relative to the sum of the water and fat signals within each voxel and was calculated as follows: PDFF = F/(W + F) × 100%, where F and W represent the fat and water signals, respectively. According to the Real World Value Mapping information in the original Digital Imaging and Communications in Medicine (DICOM) files of the fat fraction images generated by the mDIXON Quant sequence, the stored pixel values were converted into true PDFF percentage units. Based on the SAT and VAT segmentation masks, PDFF values within the corresponding regions were extracted on a slice-by-slice basis, and the mean and standard deviation (SD) of PDFF were calculated for each adipose tissue compartment. The mean PDFF was used to characterize the average fat content within adipose tissue, whereas the SD of PDFF was used to reflect the dispersion of fat fraction values and intracompartment tissue heterogeneity. The abdominal PDFF images covered the craniocaudal extent from the diaphragmatic level to the iliac crest. To standardize manual annotation and ensure consistency across participants, an identical number of consecutive axial slices within this anatomical range was selected for labeling in all participants (Figure 2A). Based on the same anatomical coverage and slice correspondence, SAT and VAT volumes and PDFF values were quantified on a slice-by-slice basis. For comparison between automated and manual segmentation results, slice-level pairing was established according to the original image slice index and the output slice index from automated segmentation, and agreement metrics for SAT and VAT PDFF measurements derived from the two segmentation methods were calculated separately. To investigate the association between adipose tissue storage patterns and anthropometric variables, abdominal adipose tissue was categorized into SAT and VAT compartments. The absolute volume and proportional contribution of each adipose depot to total abdominal fat volume were quantified and stratified by sex, obesity severity, and age group to characterize depot-specific spatial distribution patterns.

Statistical analysis

Continuous variables are presented as the mean ± SD, and categorical variables are expressed as frequency (percentage). Bland-Altman plots were used to visualize the agreement between manual segmentation and automated quantification results. The ICC was further used to quantify the agreement between the two methods for SAT PDFF and VAT PDFF measurements. The relationships between adipose tissue compartments and anthropometric variables, including sex, age, and BMI, were evaluated via linear regression and Pearson correlation analyses. Sex-related differences were compared with the Welch t-test. The effect size for between-group differences was expressed as Cohen’s d, which was used to quantify the standardized magnitude of the difference between two group means. In general, d≈0.2 was considered a small effect, indicating a minor difference between groups; d≈0.5 was considered a moderate effect; and d≈0.8 was considered a large effect. For multiple sex-related comparisons within the same analytical framework, including comparisons of sex differences in BMI, SAT volume, and VAT volume, the Bonferroni method was applied for multiple-comparison correction. The adjusted P value was calculated as the product of the unadjusted P value and the number of comparisons, and an adjusted P<0.05 was considered statistically significant. Differences in SAT and VAT volumes across age groups and BMI categories were compared via the Kruskal-Wallis test. Unless otherwise specified, P<0.05 was considered statistically significant. All statistical analyses were performed with Python software (Python Software Foundation, Wilmington, DE, USA). Regional spatial distribution maps of adipose tissue storage were generated with MATLAB software (MathWorks, Natick, MA, USA).

Implementation details

All experiments were conducted on a workstation running a 64-bit Ubuntu operating system (Canonical Ltd., London, UK). The workstation was equipped with a Core i7-12700 CPU (Intel Corp., Santa Clara, CA, USA), 16 GB of DDR4 memory, a 1-TB solid-state drive, and an RTX A2000 GPU (Nvidia, Santa Clara, CA, USA) with 12 GB of video memory. The programming environment was based on PyCharm (JetBrains, Amsterdam, the Netherlands), with Compute Unified Device Architecture version 12.2 (Nvidia) installed. All implementations were developed with Python, and the design and training of the network models were carried out with the PyTorch framework.

Results

Automated segmentation of adipose tissue on MRI PDFF images

A total of 238 participants were stratified and split at the participant level, among whom 202 were assigned to the development set for model training and fivefold cross-validation and 36 to the independent internal test set for final model performance evaluation. To compare the segmentation performance of different model architectures, 2D nnU-Net and 3D full-resolution nnU-Net were first trained and tested. In addition, to compare the performance of nnU-Net models with conventional U-Net architectures, standard 2D U-Net and standard 3D U-Net baseline models were implemented with Medical Open Network for AI. Both baseline models used the same participant-level data split, preprocessing pipeline, independent internal test set, and evaluation metrics as those of the nnU-Net models. The segmentation performance of all models on the independent internal test set is summarized in Table 2.

Considering that DSC may be influenced by the target volume, we also used absolute volume error, relative volume error, and Bland-Altman analysis to evaluate the performance of the automated segmentation for adipose tissue volume quantification. The best-performing 3D nnU-Net segmentation model achieved mean DSCs of 0.9788±0.0123 (range 0.9354–0.9927) for SAT and 0.9541±0.0307 (range 0.8681–0.9897) for VAT; the corresponding relative volume errors were −0.867%±2.930% and −1.200%±7.322%, respectively; and the absolute volume errors were 81.051±52.027 mL and 128.530±90.547 mL, respectively.

Overall, 2D nnU-Net and 3D nnU-Net showed comparable segmentation performance on the independent internal test set. For VAT segmentation, 3D nnU-Net yielded slightly higher values for some evaluation metrics than did 2D nnU-Net, which may be attributable to its ability to leverage the spatial contextual information of volumetric data. Figure 3 shows the automated segmentation results of the two nnU-Net models in a participant with a BMI of 36.5 and a VAT proportion of 19.9%. Both 2D nnU-Net and 3D nnU-Net clearly distinguished SAT from VAT on axial, coronal, and sagittal views, with generally consistent segmentation extent and spatial distribution. Figure 4 shows the 3D nnU-Net segmentation results from two participants, one with a BMI of 33.3 and a VAT proportion of 16.0% and the other with a BMI of 29.2 and a VAT proportion of 66.1%. The spatial distributions of SAT (yellow) and VAT (red) are presented as 3D views and coronal slices, illustrating interindividual differences in abdominal adipose tissue storage patterns.

Figure 3 Representative segmentation results from 2D nnU-Net and 3D nnU-Net in a participant with a BMI of 36.5 kg/m² and a VAT proportion of 19.9%. SAT is shown in yellow and VAT in red. (A-C) Axial, coronal, and sagittal views of 2D nnU-Net segmentation. (D-F) Axial, coronal, and sagittal views of 3D nnU-Net segmentation. 2D, two-dimensional; 3D, three-dimensional; BMI, body mass index; nnU-Net, no new net; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

Figure 4 Representative 3D nnU-Net segmentation results in two participants. SAT is shown in yellow and VAT in red. (A,B) Participant with a BMI of 33.3 kg/m² and a VAT proportion of 16.0%. (C,D) Participant with a BMI of 29.2 kg/m² and a VAT proportion of 66.1%. 3D, three-dimensional; BMI, body mass index; nnU-Net, no new net; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

Regarding training and inference efficiency, 3D nnU-Net required 10 hours and 56 minutes for training, whereas 2D nnU-Net required 18 hours and 2 minutes. For inference, 3D nnU-Net required approximately 40 seconds per case, whereas 2D nnU-Net required approximately 12.6 seconds per case. Thus, 2D nnU-Net demonstrated faster inference and may thus be more suitable for large-scale data processing or applications with limited computational resources. In contrast, 3D nnU-Net can leverage the spatial contextual information of volumetric data and may have potential advantages in modeling complex anatomical structures, albeit with a longer inference time per case. Therefore, the selection of a model should involve consideration of not only segmentation accuracy but also computational resources, processing efficiency, and deployability within clinical workflows.

The automated segmentation results generated by 3D nnU-Net were compared with those of manual segmentation in all 238 participants and were stratified into categories of overweight/mild obesity, moderate obesity, and severe obesity. As shown in Figure 5, Bland-Altman analysis indicated good overall agreement between automated and manual segmentation for SAT, VAT, TAT, and the VAT:TAT ratio. The mean bias and 95% limits of agreement were, respectively, 30.75 cm³, and −183.77 to 245.27 cm³ for SAT, 21.00 cm³ and −291.40 to 333.41 cm³ for VAT, 51.75 cm³ and −442.00 to 545.50 cm³ for TAT, and 0.04% and −2.36% to 2.43% for the VAT:TAT ratio.

Figure 5 Bland-Altman plots stratified by obesity category. (A) SAT volume. (B) VAT volume. (C) TAT volume. (D) VAT/TAT ratio. Agreement between automated and manual segmentation of SAT and VAT is shown for three groups: overweight and mild obesity (blue triangles), moderate obesity (orange circles), and severe obesity (red squares). SAT, subcutaneous adipose tissue; SD, standard deviation; TAT, total adipose tissue; VAT, visceral adipose tissue.

To assess the reliability of the manual segmentation reference standard, manual segmentation consistency was evaluated in five cases selected through stratified random sampling. The primary reader repeated the segmentation of these cases after a 2-month interval to assess intrareader consistency, whereas another reader independently segmented the same cases to assess interreader consistency. The intrareader DSC values were 0.963 for SAT and 0.951 for VAT, while the interreader DSC values were 0.959 for SAT and 0.911 for VAT. For volume measurements, the intrareader ICC values were 0.996 for SAT and 0.957 for VAT, while the interreader ICC values were 0.995 for SAT and 0.874 for VAT. These findings indicate good overall consistency of manual segmentation, although interreader consistency was lower for VAT than for SAT.

Tissue-level PDFF quantitative characteristics of SAT and VAT

Based on manual and automated segmentation results, PDFF values within SAT and VAT regions were extracted from representative cases across different BMI categories to evaluate the tissue-level fat fraction characteristics of each adipose tissue compartment. PDFF images provide voxel-level fat fraction information and, when combined with segmentation masks, enable slice-by-slice quantitative assessment of SAT and VAT. As shown in Table 3, VAT percentage varied markedly across representative cases of BMI categories, with values of 24.30%, 49.05%, 42.50%, and 22.14% in the overweight, mild obesity, moderate obesity, and severe obesity categories, respectively. In the mild obesity case, which had the highest VAT percentage, the agreement between automated and manual segmentation-derived VAT PDFF was relatively low, with an ICC of 0.188, whereas relatively better agreement was observed in the overweight and severe obesity cases with a lower VAT percentage. These findings suggest that, in individuals with a higher VAT percentage, the boundaries of intra-abdominal adipose tissue and adjacent structures may be more complex, and segmentation errors may have a greater impact on PDFF quantification. Overall, PDFF images can be used not only for adipose tissue segmentation but also to support tissue-level quantitative assessment of SAT and VAT composition when combined with segmentation masks.

Relationship between adipose tissue volume, BMI, and age

Among the 238 participants, BMI was slightly higher in men than in women; however, this difference did not remain statistically significant after Bonferroni correction for multiple comparisons (unadjusted Welch t-test P=0.039; Bonferroni-adjusted P=0.117; Cohen’s d=0.27; a small effect size was indicated). Based on the 3D nnU-Net predictions, SAT volume did not differ significantly between women and men (women: 4,287.30±1,574.86 mL; men: 3,939.67±2,045.56 mL; Figure 6A-6C; unadjusted Welch t-test P=0.14; Bonferroni-adjusted P=0.420; Cohen’s d=0.19). In contrast, VAT volume was significantly higher in men than in women, and this difference remained statistically significant after correction for multiple comparisons (women: 1,757.71±707.32 mL; men: 3,164.72±984.80 mL; Figure 6D-6F; unadjusted Welch t-test P<0.001; Bonferroni-adjusted P<0.003; Cohen’s d=1.62).

Figure 6 Anthropometric associations. Associations and linear regression analyses of adipose tissue compartments with BMI for SAT in males (A) and females (B) and VAT in males (D) and females (E), as well as associations with age for SAT (C) and VAT (F). Men are shown in blue, and women are shown in orange. BMI, body mass index; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

The correlation coefficients between adipose tissue volumes at different anatomical compartments and BMI and age in men and women are summarized in Table 4. Compared with men, women showed a stronger correlation between VAT and BMI. SAT showed a stronger correlation with BMI than did VAT for both sexes.

SAT and VAT volumes had wide distributions across the study population. SAT volume ranged from 1,267.62 to 10,663.36 mL in men and from 1,840.12 to 11,141.79 mL in women. Meanwhile, VAT volume ranged from 796.91 to 5,587.38 mL in men and from 420.78 to 3,987.25 mL in women. To further quantify variability in abdominal adipose tissue volume within the same BMI category, the mean, SD, and coefficient of variation (CV) of SAT and VAT volumes were calculated for each BMI category according to the Chinese BMI classification criteria, as shown in Table 5. The CVs of SAT in the overweight, mild obesity, moderate obesity, and severe obesity groups were 28.85%, 27.58%, 23.52%, and 33.30%, respectively, whereas the corresponding CVs of VAT were 56.67%, 40.63%, 39.15%, and 36.93%, respectively. These results indicate that SAT and VAT volumes varied within the same BMI category, with VAT showing a generally greater degree of relative variability than SAT. Consistent with the data presented in Figure 6D,6E, the scatter distribution of VAT was more dispersed than that of SAT, suggesting that individuals with similar BMI may nonetheless exhibit distinct abdominal adipose tissue distribution patterns, particularly for VAT volume.

With respect to age-related associations, SAT volume was significantly and negatively correlated with age in both sexes, whereas VAT volume was moderately and positively correlated with age. Stratified analysis by decade revealed an increase in VAT volume with advancing age, with higher median VAT volumes observed in middle-aged and older groups compared with younger groups. Notably, substantial interindividual variability was observed within each age group.

Regional spatial distribution patterns of abdominal adipose tissue

Figures 7,8 illustrate the craniocaudal spatial distributions of abdominal SAT and VAT in male and female participants stratified by age group and obesity category, respectively. Overall, SAT and VAT exhibited relatively consistent craniocaudal distribution patterns across age groups, obesity categories, and sex subgroups. Within the anatomical range from the diaphragm to the iliac crest, SAT volume showed relatively gradual changes along the craniocaudal axis, whereas VAT volume was higher near the renal level and progressively decreased toward the hepatic level. The percentage curves showed a relatively increased SAT proportion near the hepatic level, accompanied by an overall decreasing trend in VAT proportion.

Figure 7 Adipose tissue profiles along the craniocaudal axis stratified by sex and age for all test participants. Subcutaneous and visceral adipose tissue are shown as absolute volumes in milliliters (solid lines) and as slice-wise percentages (dashed lines). The mean profiles (lines) and one standard deviation (shaded colored areas) around the mean are illustrated. SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

Figure 8 Adipose tissue profiles along the craniocaudal axis stratified by BMI and sex for all test participants. Subcutaneous and visceral adipose tissue are shown as absolute volumes in milliliters (solid lines) and as slice-wise percentages (dashed lines). The mean profiles (lines) and one standard deviation (shaded colored areas) around the mean are illustrated. BMI, body mass index; SAT, subcutaneous adipose tissue; VAT, visceral adipose tissue.

In the age-stratified analysis, SAT volume differed significantly across age groups (Kruskal-Wallis test P<0.001), with a general decrease in SAT volume with increasing age, as shown in Figure 7. In contrast, VAT volume did not differ significantly across age groups (Kruskal-Wallis test P=0.096), indicating no clear age-related difference in VAT volume in this cohort. In the sex-stratified analysis, VAT curves in men were consistently higher than those in women across the 18 to 30-, >30 to 50-, and >50 to 70-year age groups, whereas SAT curves in women were relatively higher, particularly in the 18 to 30- and >30 to 50-year age groups. In the overall sex comparison, VAT volume was significantly higher in men than in women (Welch t-test P<0.001), whereas SAT volume showed no significant sex-related difference (Welch t-test P=0.140).

In the obesity category-stratified analysis, both SAT and VAT volumes differed significantly across BMI categories (Kruskal-Wallis test P<0.001). As shown in Figure 8, SAT and VAT absolute volumes generally increased with greater severity of obesity, with the severe obesity group showing overall higher SAT and VAT curves than the overweight and mild-to-moderate obesity groups. Sex-stratified analysis further demonstrated consistently higher VAT curves in men than in women across all BMI categories. In contrast, SAT curves were relatively higher in women than in men in the overweight and mild-to-moderate obesity groups, whereas SAT distributions were more similar between sexes in the severe obesity group.

Discussion

With the continuing rise in obesity prevalence worldwide (1,2), efficient, accurate, and reliable quantification of abdominal adipose tissue and assessment of its spatial distribution have garnered heightened interest in both medical imaging and metabolic research. This study systematically evaluated the feasibility and stability of an nnU-Net-based deep learning framework for automated segmentation and volumetric quantification of abdominal SAT and VAT on MRI PDFF. Numerous studies have applied deep learning approaches for adipose tissue segmentation (13,14,16,17,20,21,33,36,37,40), and nnU-Net, as a self-configuring segmentation framework, has demonstrated robust and generalizable performance across a variety of medical image segmentation tasks. Previous work has verified its feasibility in adipose segmentation across different cohort scales (16,21), and our results further confirm the stable performance of nnU-Net in abdominal PDFF for overweight and obese individuals with a BMI ≥24 kg/m².

Previous studies that have examined MRI-based body fat segmentation performance are summarized in Table 6. Although some studies have reported high Dice coefficients for SAT and VAT, differences in study cohorts, BMI distributions, age composition, and MRI sequences might have influenced the segmentation difficulty and DSC performance for SAT and VAT. This is particularly relevant in individuals who are overweight or obese, in whom adipose tissue volume and boundary morphology may differ from those in normal-weight or mixed-BMI cohorts. Therefore, these results should be interpreted as contextual references rather than strict head-to-head performance benchmarks.

In terms of the segmentation performance, the Dice coefficient values in our study were 0.9788±0.0123 for SAT and 0.9541±0.0307 for VAT, with absolute volume errors of 81.051±52.027 mL and 128.530±90.547 mL, respectively. These findings suggest that the proposed method provides favorable internal performance under the cohort characteristics and imaging conditions of this study. Notably, VAT showed lower segmentation performance and greater variability than did SAT. This may be related to the more complex anatomical boundaries of VAT, its discontinuous spatial distribution, substantial interindividual variability, and potential interference from bowel loops, mesentery, and adjacent intra-abdominal structures. In contrast, SAT is mainly located along the abdominal wall periphery and has relatively continuous boundaries, rendering it more amenable to consistent segmentation. In addition, DSC is dependent on target volume, and larger structures with simpler geometry, such as extensive and continuous SAT in individuals with obesity, may inherently yield higher DSC values. Therefore, absolute volume error should be interpreted together with DSC.

From the perspective of volumetric quantification, the mean bias between automated and manual segmentation was small, indicating that the model did not show an obvious systematic tendency toward overestimation or underestimation. The relatively wide limits of agreement for VAT and TAT may be attributable to the complex anatomical boundaries and discontinuous spatial distribution of VAT, as well as substantial interindividual variability in adipose tissue volume. Given the considerable biological variability in SAT, VAT, and TAT volumes within the cohort examined in this study, the observed level of agreement supports the use of automated adipose tissue volume quantification and spatial distribution analysis in population-level studies. However, for individual-level clinical risk assessment or detection of subtle volumetric changes during longitudinal follow-up, further evaluation with repeat scans, external validation, and longitudinal data is needed to determine measurement stability and the minimum detectable change.

Regarding segmentation efficiency, the model required only approximately 40 seconds to segment data from a single participant, shorter than the 1–2 hours of expert manual work typically required for manual segmentation. The feasibility of this approach was further supported by the low complete failure rate in the internal test set and its ability to accommodate certain variations in image quality and anatomy. However, some 3D segmentation results showed slice-level misalignment between the original images and predicted masks, requiring manual global shift correction. This finding suggests that in clinical or large-scale research applications, automated segmentation workflows should still incorporate basic quality-control procedures to identify abnormalities related to image scaling, spatial registration, or slice correspondence. Therefore, basic quality-control review by trained radiologists, imaging technicians, or researchers familiar with abdominal MRI anatomy and adipose tissue segmentation would still be required before clinical or large-scale research application. Although these issues could be corrected with simple adjustments to the procedures of this study and do not preclude quantitative analysis, future work should further optimize data preprocessing and output-space consistency checks to reduce the need for manual intervention.

In addition, the assessment of artifacts and image quality degradation in this study was mainly based on mild-to-moderate image quality variations naturally present in the internal test set. Cases with severe motion artifacts, marked field inhomogeneity, or substantial cross-device imaging differences were not systematically included. Therefore, the robustness of the model in more complex clinical scenarios remains to be further validated in multicenter, cross-scanner, and cross-protocol datasets. By transforming abdominal adipose tissue quantification from a highly expertise-dependent task into a scalable computational workflow, this approach provides practical feasibility for retrospective and prospective imaging studies in large populations and is consistent with the technical direction of large-scale MRI-based fat quantification research.

With respect to adipose tissue distribution, previous studies have reported significant associations of intra-abdominal fat with sex and age, with visceral fat being particularly closely related to metabolic risk (3). Several studies have shown that men generally have a greater VAT volume than women (14,16). Based on automated segmentation of whole-abdominal slices, our study confirmed that VAT volume was significantly higher in men than in women among Chinese participants who are overweight or obese. Moreover, substantial interindividual differences in VAT were observed at similar BMI levels, and SAT volume tended to decrease with increasing age. Our findings do not support the replacement of cost-effective conventional abdominal obesity measurements with expensive MRI techniques. Rather, they constitute the technical validation of the feasibility of automated whole-slice adipose tissue segmentation, volume quantification, and spatial distribution characterization based on MRI PDFF. This framework provides a reproducible approach for quantitative research on abdominal adipose tissue spatial distribution and offers an imaging-based methodological platform for future studies integrating metabolic, biochemical, and clinical outcome measures for investigating obesity phenotypes and metabolic risk.

Despite the favorable results, there were certain limitations to this study that should be acknowledged. First, we employed a single-center retrospective design, and all data were acquired with the same MRI system and a uniform imaging protocol. Although the model showed high performance in the internal test set, independent external validation was still lacking. Therefore, its generalizability across different scanner vendors, magnetic field strengths, imaging protocols, image resolutions, and population compositions was not assessed. Different imaging conditions may alter the intensity distribution, signal-to-noise ratio, adipose tissue boundary sharpness, voxel-level fat fraction measurements, and artifact characteristics of PDFF images, thereby affecting the stability of SAT and VAT recognition and the accuracy of volumetric quantification. Future studies should include multicenter, cross-device, cross-protocol, and larger external validation cohorts to systematically evaluate model robustness and clinical generalizability. Second, manual segmentation, which served as the reference standard for model training and evaluation, may be influenced by reader experience and differences in anatomical boundary interpretation. Although standardized training, a unified annotation workflow, and radiologist quality-control review were implemented in this study, VAT annotation remained less consistent than SAT annotation because of its complex anatomical boundaries, discontinuous spatial distribution, and susceptibility to interference from bowel loops, mesentery, and adjacent soft-tissue structures. Therefore, variability in VAT annotation might have introduced label noise and affected both model training and performance evaluation. Future studies may further increase the proportion of expert-reviewed cases and adopt independent annotation by multiple experts or consensus-based annotation strategies to improve the reliability of the reference standard. Third, DSC, as an overlap-based metric, may be influenced by target volume and geometric morphology. Because our study included individuals who were overweight or obese, SAT volume was relatively large and spatially continuous, and the high SAT DSC may therefore be partly related to the large target volume. Consequently, the DSC results of this study should not be directly extrapolated to normal-weight or lean individuals. For abdominal adipose tissue volumetric studies, absolute volume error, relative volume error, and Bland-Altman agreement analysis should be interpreted together with DSC. Fourth, metabolic markers, biochemical parameters, and clinical outcome data were not included in this study. Therefore, the associations of automated adipose tissue volume, PDFF values, and spatial distribution indices with metabolic risk or clinical outcomes could not be directly evaluated. Future studies incorporating multimodal data are needed to further clarify the relationship between imaging-based adipose phenotypes and metabolic risk.

Conclusions

The nnU-Net-based automated segmentation approach enables the efficient and rapid assessment of abdominal adipose tissue volumes and fat topographic maps from magnetic resonance PDFF, with an accuracy comparable to that of expert manual segmentation. Using an automated analysis framework applied to a moderately sized cohort of overweight and obese individuals, this study not only confirmed sex-, age-, and obesity-related differences in abdominal fat compartments but also revealed heterogeneity in fat distribution that cannot be captured by conventional anthropometric measures. The proposed method provides a methodological foundation for the automated quantification and spatial phenotypic characterization of abdominal adipose tissue in individuals who are overweight or obese. It also offers a technical basis for future studies on metabolic risk assessment, longitudinal follow-up, and individualized management. Further multicenter external validation is needed to confirm the generalizability of this approach and the clinical applicability across different clinical settings.

Acknowledgments

The authors sincerely acknowledge the support from Shanghai East Hospital for this study, and thank the hospital for providing the abdominal MRI image dataset used in the research. We also extend our sincere appreciation to all individuals who contributed to this work through their technical assistance and valuable suggestions.

Footnote

Reporting Checklist: The authors have completed the TRIPOD+AI reporting checklist. Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0094/rc

Data Sharing Statement: Available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0094/dss

Funding: The study was supported by the Scientific Research Special Fund of Shanghai University of Medicine & Health Sciences (No. E4610115002) and the Joint Clinical Research Project of the Pudong New District Health Committee (No. PW2024D-01).

Conflicts of Interest: All authors have completed the ICMJE uniform disclosure form (available at https://qims.amegroups.com/article/view/10.21037/qims-2026-1-0094/coif). All authors report that this study was supported by the Scientific Research Special Fund of Shanghai University of Medicine & Health Sciences (No. E4610115002) and the Joint Clinical Research Project of the Pudong New District Health Committee (No. PW2024D-01). The authors have no other conflicts of interest to declare.

Ethical Statement: The authors are accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved. The study protocol was approved by the Ethics Committee of Shanghai East Hospital. This study was conducted in accordance with the Declaration of Helsinki and its subsequent amendments. The requirement for informed consent was waived by the ethics committee because of the retrospective nature of the study.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. World Health Organization. Obesity and overweight. 2025. [accessed 15 August 2026]. Available online: https://www.who.int/news-room/fact-sheets/detail/obesity-and-overweight
2. Chen K, Shen Z, Gu W, Lyu Z, Qi X, Mu Y, Ning Y. Prevalence of obesity and associated complications in China: A cross-sectional, real-world study in 15.8 million adults. Diabetes Obes Metab 2023;25:3390-9. [Crossref] [PubMed]
3. Sakers A, De Siqueira MK, Seale P, Villanueva CJ. Adipose-tissue plasticity in health and disease. Cell 2022;185:419-46. [Crossref] [PubMed]
4. Wang L, Li Y, Li R, Luan J, Cao K, Liu T, Hu H, Chen S, Bu L, Liu L, Wang H, Lu Q. Diverse associations between pancreatic intra-, inter-lobular fat and the development of type 2 diabetes in overweight or obese patients. Front Nutr 2024;11:1421032. [Crossref] [PubMed]
5. Chait A, den Hartigh LJ. Adipose Tissue Distribution, Inflammation and Its Metabolic Consequences, Including Diabetes and Cardiovascular Disease. Front Cardiovasc Med 2020;7:22. [Crossref] [PubMed]
6. Wan Y, Wang D, Li H, Xu Y. The imaging techniques and diagnostic performance of ultrasound, CT, and MRI in detecting liver steatosis and fat quantification: A systematic review. Journal of Radiation Research and Applied Sciences 2023;16:100658.
7. Park HJ, Shin Y, Park J, Kim H, Lee IS, Seo DW, Huh J, Lee TY, Park T, Lee J, Kim KW. Development and Validation of a Deep Learning System for Segmentation of Abdominal Muscle and Fat on Computed Tomography. Korean J Radiol 2020;21:88-100. [Crossref] [PubMed]
8. Pan YL, Wu YB, Zhang LH, Yu TH, He D, Wang YJ, Cai YK, Yang PF. Comparison of association between intrathoracic and abdominal visceral adipose tissue areas with metabolic syndrome and cardiometabolic risk factors: a cross-sectional study. Quant Imaging Med Surg 2025;15:11499-511. [Crossref] [PubMed]
9. Mecherques-Carini M, Albaladejo-Saura M, Vaquero-Cristóbal R, Baglietto N, Esparza-Ros F. Validity and agreement between dual-energy X-ray absorptiometry, anthropometry and bioelectrical impedance in the estimation of fat mass in young adults. Front Nutr 2024;11:1421950. [Crossref] [PubMed]
10. Jin J, Gong P, Yang Q, Hui D, Zhang H, Qiu C, Wang N, Yi S, Zheng R, Yang Y, Ren J, Chen S. Noninvasive, quantitative evaluation of hepatic steatosis of donor livers by reference frequency method: A preliminary study. Eur J Radiol 2021;143:109909. [Crossref] [PubMed]
11. Fischer JGW, Maushart CI, Becker AS, Müller J, Madoerin P, Chirindel A, Wild D, Ter Voert EEGW, Bieri O, Burger I, Betz MJ. Comparison of (18)FFDG PET/CT with magnetic resonance imaging for the assessment of human brown adipose tissue activity. EJNMMI Res 2020;10:85. [Crossref] [PubMed]
12. Nikiforaki K, Marias K. MRI Methods to Visualize and Quantify Adipose Tissue in Health and Disease. Biomedicines 2023;11:3179. [Crossref] [PubMed]
13. Schneider D, Eggebrecht T, Linder A, Linder N, Schaudinn A, Blüher M, Denecke T, Busse H. Abdominal fat quantification using convolutional networks. Eur Radiol 2023;33:8957-64. [Crossref] [PubMed]
14. Küstner T, Hepp T, Fischer M, Schwartz M, Fritsche A, Häring HU, Nikolaou K, Bamberg F, Yang B, Schick F, Gatidis S, Machann J. 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. [Crossref] [PubMed]
15. Wald D, Teucher B, Dinkel J, Kaaks R, Delorme S, Boeing H, Seidensaal K, Meinzer HP, Heimann T. Automatic quantification of subcutaneous and visceral adipose tissue from whole-body magnetic resonance images suitable for large cohort studies. J Magn Reson Imaging 2012;36:1421-34. [Crossref] [PubMed]
16. Haueise T, Schick F, Stefan N, Schlett CL, Weiss JB, Nattenmüller J, et al. Analysis of volume and topography of adipose tissue in the trunk: Results of MRI of 11,141 participants in the German National Cohort. Sci Adv 2023;9:eadd0433. [Crossref] [PubMed]
17. Kway YM, Thirumurugan K, Tint MT, Michael N, Shek LP, Yap FKP, Tan KH, Godfrey KM, Chong YS, Fortier MV, Marx UC, Eriksson JG, Lee YS, Velan SS, Feng M, Sadananthan SA. Automated Segmentation of Visceral, Deep Subcutaneous, and Superficial Subcutaneous Adipose Tissue Volumes in MRI of Neonates and Young Children. Radiol Artif Intell 2021;3:e200304. [Crossref] [PubMed]
18. Marunowski K, Świętoń D, Bzyl W, Grzywińska M, Kaszubowski M, Bandosz P, Khrichenko D, Piskunowicz M. MRI-Derived Subcutaneous and Visceral Adipose Tissue Reference Values for Children Aged 6 to Under 18 Years. Front Nutr 2021;8:757274. [Crossref] [PubMed]
19. Cheng C, Zhang N, Gong J, Wan L, Peng H, Wan Q, Huang B, Zeng H, Liu X, Zheng H, Zou C. Reproducibility of automatic adipose tissue segmentation using proton density fat fraction images between 1.5 and 3.0 T magnetic resonance. Quant Imaging Med Surg 2025;15:537-52. [Crossref] [PubMed]
20. Wang Z, Cheng C, Peng H, Qi Y, Wan Q, Zhou H, Qu S, Liang D, Liu X, Zheng H, Zou C. Automatic segmentation of whole-body adipose tissue from magnetic resonance fat fraction images based on machine learning. MAGMA 2022;35:193-203. [Crossref] [PubMed]
21. Somasundaram A, Wu M, Reik A, Rupp S, Han J, Naebauer S, Junker D, Patzelt L, Wiechert M, Zhao Y, Rueckert D, Hauner H, Holzapfel C, Karampinos DC. Evaluating Sex-specific Differences in Abdominal Fat Volume and Proton Density Fat Fraction at MRI Using Automated nnU-Net-based Segmentation. Radiol Artif Intell 2024;6:e230471. [Crossref] [PubMed]
22. Agrawal S, Wang M, Klarqvist MDR, Smith K, Shin J, Dashti H, Diamant N, Choi SH, Jurgens SJ, Ellinor PT, Philippakis A, Claussnitzer M, Ng K, Udler MS, Batra P, Khera AV. Inherited basis of visceral, abdominal subcutaneous and gluteofemoral fat depots. Nat Commun 2022;13:3771. [Crossref] [PubMed]
23. Sorokin EP, Cule M, Thanaj M, Basty N, Whitcher B, Sattar N, Thomas EL, Bell JD, Yaghootkar H. Genetic subtypes of type 2 diabetes are distinguished through the lens of abdominal MRI. Front Genet 2025;16:1605721. [Crossref] [PubMed]
24. Borga M. MRI adipose tissue and muscle composition analysis-a review of automation techniques. Br J Radiol 2018;91:20180252. [Crossref] [PubMed]
25. Middleton MS, Haufe W, Hooker J, Borga M, Dahlqvist Leinhard O, Romu T, Tunón P, Hamilton G, Wolfson T, Gamst A, Loomba R, Sirlin CB. Quantifying Abdominal Adipose Tissue and Thigh Muscle Volume and Hepatic Proton Density Fat Fraction: Repeatability and Accuracy of an MR Imaging-based, Semiautomated Analysis Method. Radiology 2017;283:438-49. [Crossref] [PubMed]
26. Shen J, Baum T, Cordes C, Ott B, Skurk T, Kooijman H, Rummeny EJ, Hauner H, Menze BH, Karampinos DC. Automatic segmentation of abdominal organs and adipose tissue compartments in water-fat MRI: Application to weight-loss in obesity. Eur J Radiol 2016;85:1613-21. [Crossref] [PubMed]
27. Borga M, Thomas EL, Romu T, Rosander J, Fitzpatrick J, Dahlqvist Leinhard O, Bell JD. Validation of a fast method for quantification of intra-abdominal and subcutaneous adipose tissue for large-scale human studies. NMR Biomed 2015;28:1747-53. [Crossref] [PubMed]
28. Iglesias JE, Sabuncu MR. Multi-atlas segmentation of biomedical images: A survey. Med Image Anal 2015;24:205-19. [Crossref] [PubMed]
29. Addeman BT, Kutty S, Perkins TG, Soliman AS, Wiens CN, McCurdy CM, Beaton MD, Hegele RA, McKenzie CA. Validation of volumetric and single-slice MRI adipose analysis using a novel fully automated segmentation method. J Magn Reson Imaging 2015;41:233-41. [Crossref] [PubMed]
30. Würslin C, Machann J, Rempp H, Claussen C, Yang B, Schick F. Topography mapping of whole body adipose tissue using A fully automated and standardized procedure. J Magn Reson Imaging 2010;31:430-9. [Crossref] [PubMed]
31. Sadananthan SA, Prakash B, Leow MK, Khoo CM, Chou H, Venkataraman K, Khoo EY, Lee YS, Gluckman PD, Tai ES, Velan SS. Automated segmentation of visceral and subcutaneous (deep and superficial) adipose tissues in normal and overweight men. J Magn Reson Imaging 2015;41:924-34. [Crossref] [PubMed]
32. Hui SCN, Zhang T, Shi L, Wang D, Ip CB, Chu WCW. Automated segmentation of abdominal subcutaneous adipose tissue and visceral adipose tissue in obese adolescent in MRI. Magn Reson Imaging 2018;45:97-104. [Crossref] [PubMed]
33. Greco F, Mallio CA. Artificial intelligence and abdominal adipose tissue analysis: a literature review. Quant Imaging Med Surg 2021;11:4461-74. [Crossref] [PubMed]
34. Maddalo M, Zorza I, Zubani S, Nocivelli G, Calandra G, Soldini P, Mascaro L, Maroldi R. Validation of a free software for unsupervised assessment of abdominal fat in MRI. Phys Med 2017;37:24-31. [Crossref] [PubMed]
35. Shen N, Li X, Zheng S, Zhang L, Fu Y, Liu X, Li M, Li J, Guo S, Zhang H. Automated and accurate quantification of subcutaneous and visceral adipose tissue from magnetic resonance imaging based on machine learning. Magn Reson Imaging 2019;64:28-36. [Crossref] [PubMed]
36. Langner T, Hedström A, Mörwald K, Weghuber D, Forslund A, Bergsten P, Ahlström H, Kullberg J. Fully convolutional networks for automated segmentation of abdominal adipose tissue depots in multicenter water-fat MRI. Magn Reson Med 2019;81:2736-45. [Crossref] [PubMed]
37. Estrada S, Lu R, Conjeti S, Orozco-Ruiz X, Panos-Willuhn J, Breteler MMB, Reuter M. FatSegNet: A fully automated deep learning pipeline for adipose tissue segmentation on abdominal dixon MRI. Magn Reson Med 2020;83:1471-83. [Crossref] [PubMed]
38. Masoudi S, Anwar SM, Harmon SA, Choyke PL, Turkbey B, Bagci U. Adipose Tissue Segmentation in Unlabeled Abdomen MRI using Cross Modality Domain Adaptation. Annu Int Conf IEEE Eng Med Biol Soc 2020;2020:1624-8. [Crossref] [PubMed]
39. Isensee F, Jaeger PF, Kohl SAA, Petersen J, Maier-Hein KH. nnU-Net: a self-configuring method for deep learning-based biomedical image segmentation. Nat Methods 2021;18:203-11. [Crossref] [PubMed]
40. Kafali SG, Shih SF, Li X, Kim GHJ, Kelly T, Chowdhury S, Loong S, Moretz J, Barnes SR, Li Z, Wu HH. Automated abdominal adipose tissue segmentation and volume quantification on longitudinal MRI using 3D convolutional neural networks with multi-contrast inputs. MAGMA 2024;37:491-506. [Crossref] [PubMed]
41. Reinders FCJ, Savenije MHF, de Ridder M, Maspero M, Doornaert PAH, Terhaard CHJ, Raaijmakers CPJ, Zakeri K, Lee NY, Aliotta E, Rangnekar A, Veeraraghavan H, Philippens MEP. Automatic segmentation for magnetic resonance imaging guided individual elective lymph node irradiation in head and neck cancer patients. Phys Imaging Radiat Oncol 2024;32:100655. [Crossref] [PubMed]
42. Baudin PY, Balsiger F, Beck L, Boisserie JM, Jouan S, Marty B, Reyngoudt H, Scheidegger O. A Minimal Annotation Pipeline for Deep Learning Segmentation of Skeletal Muscles. NMR Biomed 2025;38:e70066. [Crossref] [PubMed]
43. Liu J, Balamuralikrishna PTS, Tan S, Mukherjee P, Mathai TS, Pickhardt PJ, Summers RM. Improved muscle and fat segmentation for body composition measures on quantitative CT. Int J Comput Assist Radiol Surg 2025;20:1889-98. [Crossref] [PubMed]