AI segmentation requires accounting for brain size to maintain performance on developmental MRI cohorts

Abstract

The human brain undergoes rapid developmental changes through early life, underpinning the emergence of function but also marking a period of vulnerability to a range of neurodevelopmental disorders. With dynamic changes to brain size, morphology, and imaging contrast, consistent and accurate computational neuroanatomy remains a challenge. Deep learning tools for segmentation, like SynthSeg, offer robustness to heterogeneously acquired MRI contrast but remain unproven in early development. Here, we aggregated a large cohort (26k) of MRI scans spanning infant to adult development, and evaluated SynthSeg performance. Automated quality control scores, visual inspection, and spatial overlap with expert-segmented MRI scans revealed poor quality output segmentations during development. In the infant period only 36% of scans (1094/3069) passed automated QC. Rescaling infant scans to adult brain sizes significantly improved spatial overlap, and cropping scans to match adult fields of view retrieved automated quality control. Evaluation of the SynthSeg rescale + crop pipeline demonstrated visible and quantitative improvements in segmentation throughout infancy and childhood. There were marked increases in successful segmentations in infant scans, with 91% of scans now passing QC (2803/3069). These findings facilitate computational analysis of typical and disrupted neurodevelopment and should be considered when training the next generation of computational tools.