Related Experiment Video
Updated: Aug 5, 2026
09:06
Whole-brain Segmentation and Change-point Analysis of Anatomical Brain MRI—Application in Premanifest Huntington's Disease
Published on: June 9, 2018
AI segmentation requires accounting for brain size to maintain performance on developmental MRI cohorts
Lena Dorfschmidt1,2, Milly Hang Chi Mak1,2, Sophie Adler2,1
1School of Biomedical Engineering & Imaging Sciences, King's College London, UK.
Biorxiv : the Preprint Server for Biology
|August 1, 2026
Summary
Deep learning segmentation tools like SynthSeg struggle with infant brain MRI scans. A rescale and crop method significantly improved accuracy, enabling better computational neuroanatomy analysis in developing brains.
Area of Science:
- Computational neuroanatomy
- Neuroimaging
- Developmental neuroscience
Background:
- The human brain's early development is crucial for function but vulnerable to neurodevelopmental disorders.
- Accurate computational neuroanatomy is challenging due to dynamic changes in brain size, morphology, and MRI contrast during development.
- Deep learning segmentation tools, such as SynthSeg, show promise for MRI analysis but require validation in early development.
Purpose of the Study:
- To evaluate the performance of the SynthSeg deep learning tool for brain MRI segmentation across a wide developmental range, from infancy to adulthood.
- To identify and address challenges in applying automated segmentation to infant neuroimaging data.
- To improve the accuracy and reliability of computational neuroanatomy in pediatric populations.
Main Methods:
- Aggregated a large cohort of 26,000 MRI scans from infants to adults.
- Evaluated SynthSeg performance using automated quality control (QC) scores, visual inspection, and spatial overlap with expert segmentations.
- Developed and tested a 'rescale + crop' preprocessing pipeline to adapt infant scans to adult brain dimensions and field of view.
Main Results:
- Standard SynthSeg application resulted in poor segmentation quality for infant scans, with only 36% passing automated QC.
- The 'SynthSeg rescale + crop' pipeline significantly improved segmentation accuracy and QC pass rates, reaching 91% for infant scans.
- Visible and quantitative improvements in segmentation were observed throughout infancy and childhood using the enhanced pipeline.
Conclusions:
- Preprocessing steps, including rescaling and cropping, are essential for accurate SynthSeg application to infant brain MRI.
- The improved pipeline facilitates robust computational analysis of typical and disrupted neurodevelopment.
- These findings offer critical considerations for developing and training future computational neuroimaging tools for pediatric research.
