Abstract: Pancreatic health is closely linked to diabetes, making accurate fat quantification crucial for early diagnosis. This study proposes a deep learning-based method for automatic pancreatic segmentation and fat quantification. The nnU-Net model achieves high-precision segmentation on m-Dixon Imaging [Dice similarity coefficient (DSC) is 0.92]. A novel sub-region partitioning and quantification method enables precise delineation of the pancreatic head, body, and tail. Analysis of 256 subjects (healthy, prediabetic, diabetic) reveals a significant association between pancreatic tail fat and type 2 diabetes (p < 0.05). Using random forest classifiers, tail fat content [Area under the curve (AUC) is 0.68] and a composite fat index (AUC = 0.73) effectively predict diabetes risk. This method offers a promising tool for early diagnosis.

Key words: Magnetic Resonance Imaging, Type 2 Diabetes Prediction, Deep Learning, Pancreas, nnU-Net