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

Abstract

Abstract Monitoring the accumulation of renal sinus fat (RSF) by non-invasive magnetic resonance imaging (MRI) holds promise for assessing the risk of nephropathy in individuals with diabetes. Automatic image segmentation using dedicated U-Net models was deployed for accurate quantification of RSF content and renal parenchyma (RP) from different MRI protocols. Therefore, the accuracy of volumetric vs single-slice analysis for quantifying RP and RSF was assessed. Further, the resulting kidney structures obtained from a whole-body MR images acquired with partial kidney coverage were compared to high-resolution MRI protocol with full-kidney coverage, in people with and without diabetes. Quantification of kidney structures showed accurate estimates of both RP and RSF volume across people with different glycaemic status and imaging protocols. A systematic overestimation of the RSF-to-RP ratio was observed when using the conventional single-slice assessment, supporting the need for volumetric kidney analysis, particularly for small structures such as the RSF. Moreover, MR images with interslice-gaps were found to substantially underestimate RSF content, highlighting the need for careful evaluation and correction of estimates from small kidney structures when data are pooled from different MR imaging protocols. In summary, automatic image segmentation enabled us to determine differences in the precision of RSF content obtained using different methodological approaches and MRI sequences with different kidney coverage.