Pancreatic disease can be spatially inhomogeneous. For this reason, quantitative imaging studies of the pancreas have often targeted the 3 main anatomical pancreatic parts, head, body, and tail, traditionally using a balanced region of interest (ROI) strategy. Existing automated analysis methods have implemented whole-organ segmentation, which provides an overall quantification, but fails to address spatial heterogeneity in disease. A method to automatically refine a whole-organ segmentation of the pancreas into head, body, and tail subregions is presented for abdominal magnetic resonance imaging (MRI). The subsegmentation method is based on diffeomorphic registration to a group average template image, where the parts are manually annotated. For a new whole-pancreas segmentation, the aligned template's part labels are automatically propagated to the segmentation of interest. The method is validated retrospectively on the UK Biobank imaging substudy (scanned using a 2-point Dixon protocol at 1.5 tesla), using a nominally healthy cohort of 100 subjects for template creation, and 50 independent subjects for validation. Pancreas head, body, and tail were annotated by multiple experts on the validation cohort, which served as the benchmark for the automated method's performance. Good intra-rater (Dice overlap mean, Head: 0.982, Body: 0.940, Tail: 0.961, N=30) as well as inter-rater (Dice overlap mean, Head: 0.968, Body: 0.905, Tail: 0.943, N=150) agreement was observed. No significant difference (Wilcoxon rank sum test, DSC, Head: p=0.4358, Body: p=0.0992, Tail: p=0.1080) was observed between the manual annotations and the automated method's predictions. Results on regional pancreatic fat assessment are also presented, by intersecting the 3-D parts segmentation with one 2-D multi-echo gradient-echo slice, available from the same scanning session, that was used to compute MRI proton density fat fraction (MRI-PDFF). Initial application of the method on a type 2 diabetes cohort showed the utility of the method for assessing pancreatic disease heterogeneity.
Rejection versus error in a multiple experts environment
