Pancreas Volumetry in UK Biobank: Comparison of Models and Inference at Scale

The UK Biobank imaging sub-study enables large-scale measurement of pancreas volume, an important biomarker in metabolic disease, including diabetes. Previous methods utilised a pancreas-specific (PS) 3D MRI UK Biobank acquisition to automatically measure pancreas volume. This may lead to a clinically significant underestimation of volume, due to partial coverage of the pancreas in these acquisitions. To address this, we propose a pipeline for the accurate measurement of pancreas volume using stitched whole-body (WB) 3D MRI UK Biobank acquisitions and deep learning-based segmentation. We implement and compare the performance of six different U-Net-like model architectures, leveraging attention layers, recurrent layers, and residual blocks. Furthermore, we investigate pancreas volumetry in 42,313 subjects, separated by sex, and present novel results concerning the change in pancreas volume throughout the course of a day (diurnal variation). To the best of our knowledge, this is the largest pancreas volumetry study to date and the first to propose a pipeline using the whole-body UK Biobank MRI acquisitions to measure pancreas volume.