Propagating Segmentation Uncertainties in Ultra-Low Field Using Matched Scans and Bootstrapping

Motivation: Portable ultra-low field (ULF) MRI scanning allows for broader and more accessible scanning in a multitude of settings - including rural areas and low- and middle-income countries. However, most normative brain models assessing typical and atypical neurodevelopment are based on high field MRI results using point estimates for segmentation predictions, not accounting for intrinsic uncertainty in ULF pediatric neuroimaging. Goal(s): To provide model and brain-region specific uncertainty estimates for segmentation volumes. Approach: Using bootstrapping to propagate model uncertainties onto segmentation outputs. Results: Bootstrapping estimates revealed significant differences between high field and ULF segmentation estimates which can impact clinical decision-making. Impact: Robust uncertainty estimates for brain anatomy are fundamental for investigating typical and atypical neurodevelopment. Propagating model uncertainties to volumetric estimates by comparing ULF results against gold standard high-field MRI can help guide diagnosis and assessment of brain health.