Mind the Shift: Quantifying Morphometry Drift Brain Segmentation Across MRI Scanners
Keywords: Brain Morphometry, MRI, Multi-Scanner Variability, Dice, FreeSurfer, Synth- Seg, Segmentation, ANTs
TL;DR: Even state-of-the-art tools like FreeSurfer 8 with SynthSeg show up to 20% variability in brain morphometry across multi-scanner longitudinal MRI, highlighting the need for standardized preprocessing and robust registration
Abstract: Accurate measurement of brain morphometry is essential not only for detecting abnormalities, but also for tracking subtle structural changes in healthy individuals—such as those linked to stress or neuroplasticity. While segmentation of pathological regions (e.g., BraTS challenge) has benefited from well-established benchmarks and metrics, the reproducibility of morphometric estimates in healthy brains, particularly across MRI scanners and protocols, remains underexplored.
In this study, we assess the consistency of brain volume measurements using FreeSurfer 8 with integrated SynthSeg across 73 longitudinal MRI scans from a single individual (SIMON dataset). We quantify inter-scan variability on segmentation stability using absolute volume difference, Dice, and Surface Dice metrics.
Our results show that even state-of-the-art tools exhibit sensitivity to domain shift, with average variability in regional brain volumes reaching 3.1\% and maximum of 19\% for individual subcortical regions. This highlights a critical limitation for using current segmentation pipelines in personalized brain health monitoring or early detection of conditions such as Alzheimer’s disease.
Submission Number: 112
Loading