Loading…
Pediatric brain MRI segmentation model trained on synthetic scans, so a single 3D U-Net labels clinical acquisitions from neonates to young adults.
No providers recorded yet. Browse all providers
In the first year of life, gray matter is brighter than white matter on a T1-weighted scan. Myelination then reverses that relationship, passing through a period of isointensity in which the two tissues are barely separable. A network tuned to adult-like contrast does not simply lose accuracy on these scans — it produces masks that fail quality control outright, which is why studies built on hospital MRI archives have routinely discarded every participant under six months of age.
ClinSeg is a whole-brain anatomical segmentation model built to keep those scans. Developed at the Children's Hospital of Philadelphia and the University of Pennsylvania with collaborators at UCL, the Martinos Center and MIT, and posted as a medRxiv preprint in September 2026, it re-aims the domain-randomization idea behind the SynthSeg family at early life. Rather than train on real scans and hope the network generalizes, the authors synthesize a corpus whose anatomical and intensity variation deliberately exceeds what a pediatric radiology department produces, so orientation, voxel size and pulse sequence stop being variables the model must be told about.
Two augmentations carry the infant-specific work. White matter is split into discrete clusters by k-means on T1-weighted intensities and re-synthesized, manufacturing the patchy, tract-by-tract myelination a single white matter label cannot express. Gold standard segmentations are also deformably registered onto real clinical scans of children with ventriculomegaly, Chiari malformation or benign enlargement of the subarachnoid space, so abnormal morphology appears in training rather than first at inference.
ClinSeg is a 3D U-Net trained with the self-configuring nnU-Netv2 framework under 5-fold cross-validation, with real data held out for validation. The corpus totals 3,822 scan and segmentation pairs. Its seed is the Baby Open Brains dataset — 71 manually segmented T1w and T2w scans from infants aged 1–8 months in the Baby Connectome Project — re-registered with SynthMorph into native space, then registered onto 10 normative and 73 abnormal clinical scans to give 710 and 351 real-looking pairs that the brain generator expanded into 1,420 and 702 synthetic pairs. At inference, near-isotropic volumes are kept as acquired and others resampled to 0.8 × 0.8 × 0.9 mm and reoriented to RAS.
On the DIBS infant dataset (47 T1w scans, 0–36 months) ClinSeg reached a mean Dice of 0.909 for cerebral white matter and 0.913 for cortex, against 0.857 and 0.86 for SynthSeg+ and 0.89 and 0.891 for BIBSNet. On CANDI (101 T1w scans, ages 4–17) it scored 0.86 and 0.88, comparable to both baselines, with weaker performance on hippocampus and amygdala. Across 78,508 scans from the CHOP Scans with Limited Imaging Pathology cohort, 195 neonatal scans — 98% of that sample — passed automated quality control under ClinSeg where none passed under SynthSeg+, along with 2,523 additional passing scans aged 1–6 months. SynthSeg+ retains an edge on anisotropic T1w and T2w series in older children, reflecting ClinSeg's infant-weighted training data.
The immediate use is retrospective: turning a children's hospital's accumulated clinical MRI into quantitative neuroimaging data. The authors fit generalized additive models for location, scale and shape to ClinSeg-derived phenotypes from 11,699 individuals aged 0–21, producing reference trajectories against which a single scan can be assigned a centile. Applying those charts to 22q11.2 deletion syndrome, they replicate the known reductions in gray matter volume and surface area and recover an age-by-diagnosis interaction for cortical thickness visible only once infant scans are included.
ClinSeg addresses a narrow but persistent bottleneck: the youngest scans in a clinical archive are the ones a developmental neuroscientist most wants and the ones existing tools most reliably discard. Extending pediatric brain charts across the full 0–21 range from one segmentation model, rather than stitching together epoch-specific pipelines, is the concrete payoff. Manual infant segmentations remain scarce, and both the quality control module and the SynthDist surface predictor it inherits were trained on adult-like synthetic data. The work is a preprint awaiting peer review, and the model, its Docker image, the repository and the growth charts are promised on publication rather than released.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.