MEDLEY: Orchestrating Imperfect Segmentation Through Physics-Aware Multi-Tool Curation for Infant Brain MRI

Accurate infant brain MRI segmentation remains difficult, particularly during early infancy, when rapid myelination drives nonlinear changes in tissue contrast and pushes gray and white matter toward an emerging “isointense” regime on T1-weighted scans. Existing automatic segmentation tools exhibit predictable failure modes, and no single method performs reliably across all structures during early infancy. We present Medley, a physics-aware curation workflow that combines per-tissue T1 intensity enhancement, affine-based structural enlargement, and anatomy-informed rule-based label consolidation. Medley operates on the outputs of standard segmentation tools, correcting their systematic weaknesses rather than replacing them. We validate the workflow on a subset of T1-weighted scans from the Marcus Autism Center (manual QC) and on a subset of the BabyOpenBrains dataset (Dice and HD95), restricting analysis in this paper to infants aged 1-4 months. Within this 1-4-month subset, Medley improves the quality-control (QC) pass rate from 44% (best single tool) to 90%, increases mean Dice, and substantially reduces manual touch-up time. On the manually labeled 1-4-month BabyOpenBrains subset (17 anatomical labels), Medley achieves a macro Dice of 0.74 and a micro Dice of 0.84. This workflow produces curated labels that can facilitate downstream infant segmentation model development and enables more scalable pediatric neuroimaging studies.