Fig. 1

Workflow of this study. (a) data annotation and preprocessing stage; (b) training stage of the artificial intelligence segmentation model; (c) annotation stage using the trained AI model; (d) construction stage of the T2DM prediction model

Fig. 2

Cascade U-Net segmentation network architecture[17]

Fig. 3

Pancreas automatic partitioning demonstration. The lengths in the diagram are all in millimeters. Blue, orange and green represent the tail, body and head respectively

Fig. 4

The PDFF pseudo-color map of the abdomen and pancreas uses axis values to represent signal intensity. Color ranges from blue to red, indicating fat fraction—deeper red areas correspond to higher fat content, while deeper blue areas indicate lower fat content

Table 1

Segmentation results of different models

Fig. 5

Scatter plot of DSC coefficients for segmentation results

Fig. 6

Demonstration of partial segmentation results. The green region represents annotations by radiologists, and the red region represents model annotations

Fig. 7

Correlation analysis map of fat content in different pancreatic regions across the three patient groups, with color ranging from blue to red representing correlation values from 0 to 1. The Healthy, Pre-diabetic, and Type 2 groups represent healthy individuals, pre-diabetic patients, and T2DM patients, respectively

Fig. 8

Correlation between the volume distribution of the pancreatic head, body, and tail and diabetes. H, D, P represents healthy individuals, diabetic, and prediabetic patients, respectively

Fig. 9

Correlation between the fat content distribution of the pancreatic head, body, and tail and diabetes. H, D, P represents healthy individuals, diabetic, and prediabetic patients, respectively

Fig. 10

Classification model results of five classifiers