Label Brain Transplant: Utilizing Incomplete Manual Segmentations to Boost Deep Learning Generalization

Training generalizable deep learning models for infant brain MRI segmentation is hindered by limited data and structural inconsistencies in available annotations. Many high-quality manual segmentations lack head labels, rendering them incompatible with synthetic data generation frameworks requiring full anatomical context. We present Label Brain Transplant, a simple yet effective method that preserves high-fidelity brain segmentations by embedding them into structurally complete donor head labels. This enables the generation of anatomically realistic synthetic training data from previously unusable label maps. We validate this technique by comparing models trained on transplanted labels and original headless labels, showing that our method produces superior performance on external infant MRIs.