Body composition—including bones, muscles, and adipose compartments—constitutes a crucial imaging biomarker for cardiometabolic risk stratification, as supported by extensive evidence; however, conventional imaging quantification remains constrained by labor-intensive workflows and limited anatomical coverage, hindering scalable and comprehensive assessment. Meanwhile, existing deep-learning efforts have largely focused on abdominal fat compartments (e.g., VAT and SAT), overlooking critical tissues such as muscle, bone, and IMAT that also predict mortality and adverse cardiovascular events, thereby limiting integrated analyses at the whole-body/multitissue level. “From an imaging standpoint, although MRI offers superior tissue contrast and greater safety than CT, non-standardized acquisition protocols and pronounced anisotropy pose challenges for automated segmentation; in contrast, fat/water dual-echo sequences provide complementary visualization—fat-only images better delineate adipose tissue, whereas water-only images facilitate discrimination of muscle, bone, and vessels—creating a methodological opportunity to enhance tissue-specific segmentation and mitigate 3D anisotropy.” said the author Jianyong Wei, a researcher at Shanghai Jiao Tong University School of Medicine, “Therefore, we developed an automated MRI quantification workflow that covers multiple anatomical regions and targets multiple tissue compartments, in order to support precision management of cardiometabolic disease and population-scale research.”
This article propose BioCompNet, a dual-channel 2D U-Net framework that ingests paired fat-only and water-only MRI: the fat channel enhances adipose-tissue contrast, while the water channel accentuates muscle, bone, and vascular structures. The network adopts a symmetric encoder–decoder with skip connections, taking 512×512 inputs that are progressively downsampled to 8×8 feature maps and then upsampled back to the native resolution. The abdominal model outputs seven segmentation channels, and the thigh model outputs five, with all layers shared except the output heads. Prior to training and inference, 3D fat–water MRI volumes of the abdomen and thighs undergo Z-score normalization, anisotropic resampling to (0.820, 0.820) mm in-plane spacing, and cropping to 512×512 to mitigate voxel-spacing heterogeneity; annotations are physician-led and iteratively refined with semi-automatic tools to ensure accuracy and consistency. Following segmentation, a fully automated post-processing module derives morphometric indices (e.g., volume, perimeter, cross-sectional area) directly from the masks and applies K-means (k=2) within core abdominal and thigh muscle compartments to identify IMAT, thereby yielding an integrated “segmentation-to-quantification” pipeline for body-composition phenotyping.
The dataset used for testing comprises 503 subjects (8,048 MRI slices) for model development and internal evaluation, with an independent external test set of 21 abdominal and 9 thigh cases to assess generalization. On internal tests, competing models perform comparably; however, on the external set BioCompNet attains mean Dice scores of 0.938 (abdomen) and 0.936 (thigh), surpassing 2D/3D nnU-Net and demonstrating stronger out-of-domain robustness. Ablation studies show that dual-sequence input plus data augmentation markedly improves performance: external abdominal Dice increases from 0.907 to 0.938, and external thigh Dice from 0.928 to 0.936; overall, the dual-sequence configuration consistently outperforms single-sequence variants. Continuous quantitative indices exhibit excellent agreement with physician measurements (ICC 0.881–0.999), and automated IMAT quantification shows a significant linear trend with radiologist grades (P_trend < 0.001). In terms of efficiency, the complete pipeline (including abdominal and thigh segmentation and feature analysis) processes each case in 0.12 ± 0.001 minutes, a dramatic reduction compared with manual assessment (128.8 ± 5.6 minutes per case), supporting scalability for clinical deployment and large-cohort studies.
These results highlight the model’s robustness and generalization capability, supporting its suitability for large-scale population studies and real-world clinical applications. However, this study has several limitations. First, the body-composition components currently included and quantified were selected based on existing evidence in cardiometabolic disease (CMD), and clinically meaningful components not covered may still exist, warranting further expansion and validation. Second, although post-processing is fully automated, visible segmentation errors were manually corrected prior to final quantification, indicating the need for a clinician-friendly interactive interface to enable rapid review and fine-tuning. Third, generalizability across centers, scanners, and heterogeneous populations still depends on larger and more balanced multicenter cohorts as well as further architectural optimization. Fourth, the clinical associations and prognostic value of the imaging phenotypes with respect to cardiometabolic risk require confirmation in large-scale, multicenter prospective studies to clarify their practical utility for risk stratification, diagnosis, and therapeutic decision-making. “Therefore, future work will focus on multi-site/multi-device collaborative studies and systematic evaluation to continually enhance model performance and clinical usability.” said Jianyong Wei.
Authors of the paper include Jianyong Wei, Hongli Chen, Lijun Yao, Xuhong Hou, Rong Zhang, Liang Shi, Jianqing Sun, Cheng Hu, Xiaoer Wei, and Weiping Jia.
This work was supported by the National Science and Technology Major Project (2024ZD0537000); the Shanghai Municipal Key Clinical Specialty, Shanghai Key Clinical Center for Metabolic Disease (No. 2017ZZ01013); the Shanghai Research Center for Endocrine and Metabolic Diseases (No. 2022ZZ01002); the Shanghai Key Discipline of Public Health Grants Award (GWVI-11.1-20); the Shanghai Science and Technology Innovation Action Plan Morning Star Cultivation YangFan Project (24YF2732700); the Shanghai Municipal Health Commission Health Industry Clinical Research Project (20244Y0167); and the Clinical Research Project of Shanghai Sixth People’s Hospital (ynhg202421).
The paper, “BioCompNet: A Deep Learning Workflow Enabling Automated Body Composition Analysis toward Precision Management of Cardiometabolic Disorders” was published in the journal Cyborg and Bionic Systems on Aug. 20, 2025, at DOI: 10.34133/cbsystems.0381.
END