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

Technical Summary: AI Segmentation Requires Accounting for Brain Size to Maintain Performance on Developmental MRI Cohorts

Problem Statement
Deep learning tools for neuroanatomical segmentation, such as SynthSeg, offer robustness to heterogeneous MRI contrasts but are primarily trained and validated on adult cohorts. The human brain undergoes rapid developmental changes in size, morphology, and tissue contrast (specifically the inversion of grey-white matter contrast in the first year of life). Consequently, standard adult-trained pipelines face significant challenges when applied to infant and pediatric MRI. Specifically, infant brains are substantially smaller (approximately 25% of adult volume at birth) and often acquired with adult-calibrated fields of view (FOV), resulting in large amounts of non-brain anatomy (neck, shoulders) surrounding the brain. The authors hypothesize that these morphological and spatial discrepancies cause SynthSeg to fail in early development, leading to poor segmentation quality and unreliable automated quality control (QC) scores, thereby hindering the creation of continuous, lifespan-wide neurodevelopmental growth charts.

Methodology
The study evaluated and modified the SynthSeg pipeline using two primary datasets:

1. Neurodevelopmental Cohort: A large aggregated dataset of 26,716 MRI scans (T1w and T2w) from 16 open-source studies, spanning ages from 20 weeks post-menstrual age to 35 years.
2. Expert-Labeled Sample (BOBs): A subset of the Baby Open Brains (BOBs) Repository containing 62 matched sessions with expert-curated "ground truth" manual segmentations for infants aged 1–9 months.

The authors assessed the baseline performance of SynthSeg (using the --robust flag) across the lifespan. They then tested four distinct preprocessing pipelines on the infant data to determine which best aligned with adult training distributions:

1. Original: Unaltered SynthSeg input.
2. Cropped: SynthSeg applied to scans cropped to a 25mm box around the brain mask (derived via SynthStrip).
3. Rescaled: Scans rescaled to approximate adult total cerebral volume (TCV) based on normative lifespan trajectories, with voxel size adjusted in the image header.
4. Rescale + Crop: A combination of rescaling to adult volume followed by cropping to match adult FOV.

Performance was evaluated using:

• Automated QC Scores: SynthSeg's internal regression network predicting Dice scores for eight gross anatomical structures.
• Spatial Overlap: Dice coefficients calculated between SynthSeg outputs and the BOBs ground truth labels.
• Visual Inspection: Qualitative assessment of segmentation boundaries.

Key Results

• Baseline Failure in Infancy: The original SynthSeg pipeline demonstrated a severe age-dependent failure rate. Only 36% of scans in the 0–2 year age group passed automated QC (threshold ≥ 0.65), with 0% of scans under 3 months passing. There was a strong positive correlation between age/brain size and QC scores (ρ = 0.88 for ages 0–2).
• Training Distribution Mismatch: Analysis revealed that the TCV distribution of the SynthSeg training data (HCP, ADNI, OASIS), even with standard augmentation (±20% rescaling), did not cover the smaller volumes observed in infants. The QC failure rate closely tracked the distance of a subject's TCV from this training distribution.
• Pipeline Improvements:
  • Rescaling + Cropping: This combined pipeline yielded the most significant improvements. In the BOBs dataset, it increased the mean Dice score by 0.0495 compared to the original pipeline (p=0.004) and, crucially, restored the correlation between segmentation quality and automated QC scores.
  • QC Recovery: While rescaling alone improved Dice scores, it paradoxically lowered automated QC scores due to the inclusion of large non-brain FOV areas in the infant scans, which the QC network interpreted as low-quality regions. Cropping removed these extraneous areas, allowing the QC score to accurately reflect the improved segmentation of the brain tissue.
  • Cohort-Wide Impact: Applying the "Rescale + Crop" pipeline to the full 26k cohort increased the pass rate for the 0–2 year age group from 36% to 91% (2,803/3,069 scans). The median QC score for all age groups remained above the 0.65 threshold, whereas the original pipeline showed a dip below this threshold in early childhood.
  • Contrast Agnosticism: Similar improvements were observed in T2w scans, confirming the pipeline's utility across different MRI contrasts.

Significance and Claims
The paper claims that the primary barrier to applying adult-trained AI segmentation tools to infants is the mismatch in brain size and field of view relative to the training distribution. The authors demonstrate that a simple, non-retraining preprocessing step—rescaling infant scans to approximate adult volumes and cropping to match adult FOV—significantly restores segmentation accuracy and, importantly, the reliability of automated QC metrics.

The significance of this work lies in enabling the use of a single, unified pipeline for neuroanatomical segmentation across the entire human lifespan (from birth to adulthood). This eliminates the need for disjointed, age-specific pipelines (e.g., separate infant and adult FreeSurfer versions) that create methodological discontinuities in growth curves. By facilitating the processing of large, aggregated developmental cohorts, this approach supports the creation of continuous lifespan growth charts and improves the analysis of neurodevelopmental disorders. The authors note that while this approach improves performance, it does not fully capture the unique biological complexities of neonatal brains (e.g., specific sulcal depth or cortical thickness trajectories), suggesting that future models trained on full developmental ranges would be ideal. However, the proposed mitigation offers an immediate, computationally efficient solution for current adult-focused AI tools.