Abstract 4364895: Automated External White Adipose Tissue Segmentation Using Routine Magnetic Resonance Imaging and Artificial Intelligence

Introduction: Quantifying adiposity, a key biomarker of metabolic health, typically requires imaging that involves radiation, high costs, and manual effort. We developed an AI framework to segment external white adipose tissue (EWAT) from routine non-contrast MRI, offering a radiation-free, low-effort alternative. Hypothesis: We hypothesized that combining classical image processing with deep learning would enable accurate, robust EWAT segmentation from routine T1/T2-weighted MRI, without specialized sequences or manual labeling. Methods: In 105 Type 1 diabetes patients, T1/T2-weighted axial abdominal MRI scans at the aortic bifurcation were used to develop three segmentation approaches: Region Growing with automatic seed selection, iterative pixel aggregation and adaptive thresholds; UNet CNN trained on 52 masks from region-growing results, with Dice&Binary Cross-Entropy loss; and, Fine-Tuned UNet, optimized on 48 complex cases using extensive augmentations (flips, crops, brightness shifts, Gaussian noise) to enhance robustness and generalizability. Key challenges like artifacts, low fat volume, and anatomical overlap were addressed via local adjustments and hyperparameter tuning. Three independent clinicians scored segmentation quality (0–3) for anatomical alignment (Accuracy), circumference capture (Completeness), target area segmentation (Coverage), and boundary continuity (Smoothness). Results: Table 1 summarizes the mean clinical evaluation scores across all patients and metrics. The UNet and Fine-Tuned UNet consistently outperformed Region Growing in all four metrics, with mean accuracy of 2.81 and 2.80, respectively, versus 2.16 for Region Growing. Figure 1 shows example segmentations for each method. Figures 2 and 3 visualize model performance, highlighting mean scores in complex cases and the percentage of perfect (3/3) segmentations. The Fine-Tuned UNet had the highest mean accuracy (2.80) in challenging images, while the base UNet had the most perfect scores overall (73.3%). Conclusion: This unsupervised AI framework enables accurate, radiation-free EWAT segmentation from routine MRI. All methods, including deep learning, were trained without manual labeling, using region-growing outputs as pseudo ground truth. Clinical evaluations confirmed that the UNets achieved superior accuracy, completeness, coverage, and smoothness, particularly in complex cases. This scalable, cost-effective approach supports broader validation in cardiometabolic populations.