A comparison study to assess U-Net driven volumetric versus single-slice analysis and MRI sequences with different volume coverage to detect renal sinus fat in people with and without diabetes

Figures

• 

  Fig. 1. Flowchart for training and testing of U-Net models and downstream analysis. Schematic representation of the partitioned datasets used in the study, including abdominal dual-echo gradient echo (GRE) and whole-body turbo-spin echo (TSE) images that were used for the training of 3D- and 2D-U-Net models, respectively. For the analysis, 178 individual datasets were selected to compare the quantification of kidney structures using either whole-kidney or single-slice approach. A total of 99 individuals that underwent both imaging protocols were used for the comparison of two imaging protocols.

• 

  Fig. 2. Accuracy of U-Net for the segmentation of renal sinus fat from abdominal GRE images. (A) Representative coronal projection view of water- and fat-reconstructed images derived from the dual-echo GRE images are displayed with the respective masks overlaid for renal parenchyma (RP) and renal sinus fat (RSF), which were derived manual or using the trained 3D-U-Net model. Bland-Altman plots and intraclass correlation coefficients (ICC) are shown to determine the agreement (B) RP and of (C) RSF volume (cm3) from the test set. The mean differences (red dashed line) and limits of agreement (blue dashed line) are shown for both tests. (D) Segmentation of RP and RSF volume from water- and fat-reconstructed axial images and comparison with manual segmentation for RP and RSF volume with each voxel labelled as true negative (TN, green), false positive (FP, purple) or true positive (TP, grey).

• 

  Fig. 3. Accuracy of U-Net for the segmentation of renal sinus fat from whole-body TSE images. (A) A representative coronal projection view of TSE images with the respective overlaid masks produced either from the manual analysis or by the trained 2D-U-Net model. Bland-Altman plots and intraclass correlation coefficients (ICC) are shown for the quantification of (B) RP and (C) RSF volume (cm3) on the test set. Mean differences (red dashed line) and limits of agreement (blue dashed line) are displayed. (D) Masks of RP and RSF volume derived from the original axial images along with the comparison with manual segmentation showing voxels labelled as true negative (TN, green), false positive (FP, purple) or true positive (TP, grey).

• 

  Fig. 4. Comparison between whole-kidney and single-slice analysis. The scatter plots display the estimates for (A) renal parenchyma (RP), (B) renal sinus fat volume (cm3) and (C) RSF-RP ratio calculated using either the central slice (single-slice) or all the slices (whole-kidney). Linear regression (red line) and Pearson’s r coefficients with 95% confidence intervals (CI) are displayed. (D) Bland-Altman plots and intraclass correlation coefficient (ICC) for RSF-RP ratio calculated between the two analytical methods, with mean difference (red dashed line) and limits of agreement (blue dashed line). Data points are colour labelled using the RSF volume (cm3) calculated using all the slices of the kidneys.

• 

  Fig. 5. Comparison of GRE and TSE images with different kidney coverage: estimates of kidney RSF and RP across persons with different glycaemic states. (A) The scatter plot displays the estimates of renal sinus fat-to renal parenchyma ratio (RSF-RP ratio) calculated from the whole-body turbo spin-echo (TSE) acquired and abdominal dual-echo gradient echo (GRE) images. Linear regression (red line) and identity lines (black line) are displayed. The half-violin plots and the corresponding scatter and boxplots display the percentage difference between the original values and the adjusted values for (B) RP, RSF volume and (C) the RSF-RP ratio (%) obtained from the whole-body TSE images.