Automated pancreatic segmentation and regional fat quantification suggest tail fat association with type 2 diabetes

Abstract Accurate pancreatic segmentation and quantitative assessment are crucial for investigating diabetes pathogenesis. Despite advancements in deep learning, regional segmentation remains challenging due to the organ’s anatomical complexity. This study established a two-stage analytical framework: (1) nnU-Net-based whole-pancreas segmentation followed by (2) a novel algorithm for semi-automated head-body-tail partitioning with expert verification, enabling regional quantification of volume and fat content. The segmentation network achieved a Dice coefficient of 0.92. Kruskal-Wallis H tests indicated a significant between-group difference in pancreatic tail fat content across glycemic status groups (p<0.05). Using region-specific fat metrics, we constructed five random forest classifiers showing optimal performance with total fat content (Area Under the Curve, AUC = 0.72) and composite fat indices (AUC = 0.73) for distinguishing healthy controls, prediabetic, and diabetic cohorts.