Enhanced accuracy and stability in automated intra-pancreatic fat deposition monitoring of type 2 diabetes mellitus using Dixon MRI and deep learning

Motivation: Intra-pancreatic fat deposition (IPFD) is linked to the onset and progression of type 2 diabetes mellitus (T2DM). Goal(s): Develop an accurate and automated method to measure IPFD on multi-echo Dixon MRI. Approach: 518 patients who underwent upper abdomen MRI were included.A deep learning radiomics (DLR) model and two radiologist models were constructed. Results: The mean Dice similarity coefficient for pancreas segmentation was 0.959.The DLR model achieved the AUCs of 0.862 (training cohort) and 0.786 (external test cohort),both outperforming the radiologist models.The ICCs between radiologists' pancreatic proton density fat fraction (PDFF) measurements were 0.797 and 0.708 in the two cohorts,suggesting good and moderate reproducibility,respectively. Impact: The DLR model demonstrated superior performance over radiologists, providing a more efficient, accurate and stable method for monitoring IPFD.