Quality Control Pipeline for MRI Annotations of Polycystic Kidneys

Motivation: Public MRI segmentation models often fail to accurately label polycystic kidneys, reducing total kidney volume measurement accuracy, a key Autosomal Dominant Polycystic Kidney Disease (ADPKD) biomarker. Goal(s): To create a quality control pipeline and open-source high-quality, voxel-level MRI kidney annotation in ADPKD patients. Approach: 100 MRIs from 20 ADPKD patients were segmented using a deep learning model, with subsequent human-expert corrections followed by a quality control pipeline, validating labeling consistency, anatomical constraints, and inter-sequence volume/dimensional agreement. Results: We provide a quality control pipeline for hole-free, stray-voxel-free, largely (93%) inter-sequence-consistent segmentations that complements and improves upon expert annotators of polycystic kidney segmentations on MRI. Impact: This quality control pipeline with open-source polycystic kidney segmentations aims to enhance future public segmentation models, benefiting ADPKD patients, clinicians, and researchers. The segmentation quality assessment also encourages establishing reporting standards for open-source segmentation datasets.