Automated MRI segmentation may make abdominal fat measurement faster and less dependent on manual annotation. A recent publication in the Journal of Imaging Informatics in Medicine evaluated a deep learning vision transformer for quantifying subcutaneous and visceral adipose tissue from T1-weighted abdominal MRI. The dataset included 107 cognitively normal adults aged 40–61 years, with expert-edited masks used as the reference standard. The model was compared with established deep learning architectures, including U-Net, V-Net and UNETR. SwinUNETR48 achieved strong agreement with manual labels, particularly for subcutaneous fat, while visceral fat remained more complex because of its irregular distribution. External testing on an independent dataset also assessed whether the model could generalise beyond the original scans.
Why Abdominal Fat Segmentation Matters
Abdominal adipose tissue measurement has clinical relevance because fat accumulation is linked with imaging biomarkers associated with chronic disease. Visceral adipose tissue is particularly important because it has been associated with insulin resistance and increased systemic and cerebral inflammation. Higher visceral fat levels have also corresponded to increased cerebral amyloid burden in midlife individuals in PET imaging work. These links make abdominal fat distribution a relevant imaging target, especially when the aim is to move beyond a simple body mass index measurement.
MRI offers a non-invasive and relatively affordable method for assessing adipose tissue distribution. Its value lies partly in showing where fat is located, rather than only estimating overall body size. Subcutaneous fat sits under the skin, while visceral fat surrounds organs in the abdominal cavity. The distinction matters because the two compartments differ in shape, distribution and segmentation difficulty. A reliable imaging workflow therefore needs to separate these compartments consistently across abdominal slices.
Manual annotation remains the reference standard for measuring these tissues on abdominal imaging. The process requires trained readers to outline structures slice by slice. This is tedious and time-consuming, especially for visceral fat, and results may vary between readers or between separate readings by the same person. These limitations create a clear need for automated methods that can reduce manual burden while preserving agreement with expert-edited labels.
How the Model Was Built and Tested
The abdominal MRI volumes came from adults enrolled in an NIH-funded prospective project on links between abdominal adiposity and dementia biomarkers. Participants were selected from a health volunteer registry and the surrounding community. Exclusion criteria included previous participation in an obesity project, prior bariatric surgery, enrolment in a bariatric project and contraindications to MRI. The scans used T1-weighted abdominal images acquired from a 3 T scanner.
Ground truth masks were produced with an in-house MATLAB programme and then manually inspected and corrected. The process labelled subcutaneous fat, visceral fat and background tissue. Fat around the spine and outside the visceral compartment was removed where needed. Reliability checks used repeated manual segmentations by two analysts across a subset of participants, allowing the reference labels to be assessed for consistency.
SwinUNETR was used as the main segmentation approach. The model combines transformer and convolutional components, allowing it to capture local image detail and broader spatial context. Two versions were tested, with SwinUNETR48 using a larger feature size than SwinUNETR24. Performance was compared with U-Net, V-Net and UNETR using standard segmentation and agreement measures. Training used fivefold cross-validation, image preparation steps and data augmentation to support model generalisation. Evaluation included overlap, boundary and volume agreement measures, alongside statistical comparison with competing models.
Must Read: MRI Fat Distribution Patterns Linked to Brain Health
Performance Across Internal and External Data
SwinUNETR48 delivered the strongest overall results across the tested models. Average Dice scores reached about 97% for subcutaneous fat and 88% for visceral fat during internal validation. Jaccard and precision results also supported strong overlap with expert-edited labels. Subcutaneous fat produced more consistent results than visceral fat, reflecting its more stable shape and clearer boundaries across MRI slices. Across all models, visceral fat showed greater variability, reinforcing the difficulty of segmenting tissue with a subtle and irregular abdominal distribution.
Agreement testing showed small underestimation of both subcutaneous and visceral fat volumes. Most differences between automated and reference measurements remained within the expected limits of agreement. Visual examples also showed that the automated segmentations were close to the ground truth across different body sizes and abdominal shapes. SwinUNETR48 performed slightly better than the smaller SwinUNETR version, indicating that increased model capacity improved segmentation in this task.
External validation used an independent dataset with different MRI acquisition parameters. The model maintained high volume agreement for both fat compartments, although visceral fat segmentation was less accurate than in the internal dataset. Errors included inconsistent removal of spine fat, confusion at the boundary between visceral and subcutaneous fat and inclusion of arm fat as subcutaneous tissue. These issues indicate that differences in MRI parameters can affect generalisation. Manual review and editing can take up to an hour in difficult cases, while the automated model processed each MRI far more quickly on both graphics processing and standard computer hardware. The FatViT graphical interface was also developed to support automated segmentation and quantification.
SwinUNETR48 provides a strong automated option for segmenting abdominal fat on MRI while reducing dependence on time-intensive manual annotation. The approach performs especially well for subcutaneous fat and remains useful for visceral fat despite greater structural complexity. External validation shows that MRI acquisition differences can reduce performance, particularly for visceral fat. The overall results support continued development of automated tools for measuring abdominal fat distribution and for helping evaluate links between midlife adiposity, insulin resistance and inflammatory markers.
Source: Journal of Imaging Informatics in Medicine
Image Credit: iStock
References:
Kassani SH, Patel K, Commean PK et al. (2026) Vision Transformer–Based Segmentation of Abdominal Subcutaneous and Visceral Fat on MRI. J Digit Imaging Inform med: Online first.
