78-PUB: Machine Learning–Assisted Segmentation of Pancreas MRI

Introduction & Objective: Pancreas volume (PV) may be a useful biomarker for T1D, but manual segmentation of PV from an MRI is time consuming. The objective of this project was to develop and evaluate the efficiency of a machine-learning (ML) pipeline for pancreas MRI segmentation to calculate pancreas volume. Methods: Manual and ML-assisted PV MRI segmentations were created from a sample of 68 including 15 controls (no diabetes) from an existing dataset (PMID: 30552130). We linked a local 3D Slicer (PMID: 22770690) instance with a hardware-accelerated MONAI Label server instance on the UF HiPerGator cluster for our ML-assisted segmentations. We evaluated the relative efficiency of ML-assisted vs. manual to create high quality segmentation of the pancreas. Results: ML-assisted segmentations are of high-quality. ML-assisted segmentation resulted in a reduction of average time per segmentation by 5 minutes and a reduction of average total mouse-clicks by 28% when completed by a minimally-trained (non-radiologist) staff. Conclusion: ML-assted segmentation can result in significant productivity gains for pancreas segmentation without loss of data quality or accuracy, reducing costs associated with radiological MRI segmentation for PV. Disclosure P.L. Tirado-Velez: None. S. Kang: None. H. Ju: None. M. Campbell-Thompson: None. S. Kim: None. D. Lamb: None. Funding JDRF (3-SRA-2022-1157-S-B)