Elastic Registration Based Neonatal Brain Segmentation

Petronella Anbeek¹, Britt J. M. van Kooij¹, Floris Groenendaal¹, Linda S. de Vries¹, Manon J. N. L. Benders¹

A powerful fully automatic technique for neonatal brain segmentation is proposed, based on elastic registration combined with K-nearest neighbor classification. The method uses two routine diagnostic MR images: 3D-T1- and T2-weighted, and segments 8 brain structures simultaneously: (un)myelinated white matter, cortical gray matter, basal ganglia, cerebro-spinal fluid, ventricles, brainstem and cerebellum. Results are evaluated by comparison with a gold standard, showing high Dice similarity indexes (0.80-0.92) for 7 tissues. This technique is suitable for large and longitudinal neonatal brain analysis studies.