SAM-NeuroAdapt: an atlas-guided, training-free pipeline for skull-stripping in infant and adult MRI

Skull-stripping, the removal of non-brain tissue from magnetic resonance images, is an early step of most neuroimaging pipelines whose errors propagate to later measurements. However, current methods such as HD-BET, SynthStrip and LifespanStrip depend on task-specific training on large image collections. To test how much of this benefit survives without training, this study presents SAM-NeuroAdapt, a training-free pipeline. First, an age-matched template is selected from a library of 21 atlases and registered to the image. The registered atlas then supplies the prompts of a frozen Segment Anything Model 2, and the mask is refined by classical boundary correction. The pipeline was compared with HD-BET on three whole-head benchmarks, with multiplicity-adjusted p-values. On infants (Baby Open Brains, 71 visits, 51 participants), it achieved higher Dice (0.920 versus 0.905) and a lower average symmetric surface distance (ASSD; 2.07 versus 5.11 mm). On adults, no Dice difference was detected; ASSD was ower on Mindboggle-101 (n = 101) and, under one of two adjustment procedures, on IBSR-20 (n = 20). Furthermore, an ablation showed that the age-matched atlas alone reached the infant Dice of the complete pipeline (0.923 versus 0.920), while the remaining stages reduced infant surface distances by 36–46%; without atlas guidance, 26–100% of cases failed. These findings indicate that atlas matching, rather than the foundation model, drives overlap accuracy and that a well-matched atlas is a competitive training-free baseline. Two eval-uation pitfalls are also documented: a surface-distance definition that rewards over-segmentation and benchmarks distributed already skull-stripped