Self-Supervised Pretraining Reduces Labeled Data Requirements for Whole-Brain Parcellation of Magnetic Resonance Images

Neurodegenerative diseases such as Alzheimer disease are commonly studied with struc tural magnetic resonance imaging (MRI), in which whole-brain parcellation provides volu metric measures of brain structures. Atlas-based parcellation takes hours per subject, and deep learning models that accelerate it are limited by the cost of volumetric annotation. Self-supervised pretraining addresses this, but is usually demonstrated on large external unlabeled corpora that many laboratories do not have. We ask whether pretraining on the same modest cohort that supplies the labels helps, and at which annotation budgets. Using 148 T1-weighted volumes from OASIS-3 (one per participant, Clinical Dementia Rating 0–2) and 27 FreeSurfer-derived structures, a 3D U-Net was pretrained by masked-volume reconstruction on each cross-validation fold’s training split and fine-tuned with 10, 25, 50 and 104 labeled subjects. The gain from pretraining was largest when annotation was scarce: +0.0610 Dice at 25 labeled subjects against +0.0143 at 104. Under a corrected resampled t-test for dependent folds, it was significant at 25 (p = 0.005) and 10 (p = 0.039), but not at 50 or 104. With 25 labeled subjects, the pretrained model reached 95.7% of the Dice of a randomly initialized model trained on 104. Segmentation accuracy showed no detectable dependence on dementia severity once scanner model was accounted for, whereas scanner model had a large effect (Cohen’s d = 0.99 among cognitively normal participants), and the pretraining benefit did not vary detectably with severity. When annotation is the constraint, pretraining on the unlabeled cohort before annotating a fraction of it is the better use of effort.