Multi-atlas based neonatal brain extraction using atlas library clustering and local label fusion
Résumé
Brain extraction is one of the most important preprocessing steps in cerebral magnetic resonance (MR) image analysis. Brain extraction from neonatal MR images is particularly challenging due to significant differences in head size and shape between neonates and rapid changes in neonatal brain structure in the weeks and months after birth. In this work, a multi-atlas-based neonatal brain extraction method using atlas library clustering and local label fusion (NOBELL) is presented. In NOBELL, an affinity propagation (AP) approach is first applied to cluster images of an atlas library into clusters represented by exemplars, which are used to select best matching clusters for target images. A local weighted voting strategy based on Jacobian determinant ranking is then employed to extract brain from target images using training images in best matching clusters. The performance of NOBELL was evaluated on T2- and T1-weighted scans of 40 neonates aged between 37 and 44 weeks. NOBELL outperformed two popular brain extraction tools, FSL's Brain Extraction Tool (BET) and BrainSuite's Brain Surface Extractor (BSE), and achieved higher accuracy with brain masks very close to manually extracted ones. NOBELL showed an average Jaccard coefficient of 0.974 (0.942) on T2 (T1)-weighted images in comparison with 0.908 (0.602) and 0.845 (0.762) achieved by BSE, and BET, respectively. NOBELL allows for accurate and efficient brain extraction, a crucial step in brain MRI applications such as accurate brain tissue segmentation and volume estimation as well as accurate cortical surface delineation in neonates.