Negar Noorizadeh - Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran
Kamran Kazemi - Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran
Habiballah Danyali - Department of Electrical and Electronics Engineering, Shiraz University of Technology, Shiraz, Iran
Ardalan Arabi - Faculty of Medicine, University of Picardie-Jules Verne, Amiens, France --- Laboratory of Functional Neuroscience and Pathologies (LFNP, EA۴۵۵۹), University Research Center, CHU AMIENS–SITE SUD, Avenue Laënnec, Salouël ۸۰۴۲۰, France

Background and Aim : Magnetic resonance (MR) imaging, a non-invasive method with high spatial resolution, provides useful information about different anatomical structures. The process of removing non-brain tissues including scalp, fat, muscle, neck and eyeballs from cerebral MR images is called brain extraction. Due to rapid changes in neonatal brain size, low spatial resolution and low contrast, brain extraction from head MR images of neonates is really challenging.Methods : We presented a multi-atlas patch-based label fusion method for automatic brain extraction from neonatal head MR images. In this method, a number of atlases were first selected uniformly from training images. Then, a probabilistic gray level-coded brain mask was created by averaging the brain masks of the selected atlases, all aligned nonlinearly to the target image. The certain and uncertain voxels were defined based on their probability in the gray level-coded brain mask. The certain voxels belonged to the brain tissue with a probability of one (needed no more processing), while the uncertain voxels had a probability value of less than one (required further processing). The label assignment for uncertain voxels were done by considering their degree of uncertainty using a modified non-local patch-based label fusion method based on the integration of low-level and in-depth search patch selection strategies. The low-level search strategy was used for intensity and label dictionary construction for uncertain voxels with a lower degree of uncertainty, caused by registration errors. The in-depth search strategy was done by increasing the depth size for dictionary construction for uncertain voxels with a higher degree of uncertainty.Results : Our proposed method was evaluated on T2-weighted MR images of 40 neonates (gestational age 37- 44 weeks) using the leave-one-out cross-validation method. The method achieved an average Dice, Jaccard and Conformity coefficients of 0.993,0.986 and 0.986, respectively. The proposed method achieved higher accuracy and produced less false positives in comparison with two well-known non-learning-based methods (i.e. brain surface extractor (BSE) and FSL’s brain extraction tool (BET)) and two multi atlas-based methods (i.e. conventional Non-local patch-based (NLPB) and Multi-atlas skull stripping MASS).Conclusion : Our multi-atlas patch-based label fusion method can be used to extract brain masks with high accuracy, an important preprocessing step before brain tissue segmentation in neonates.

Neonatal brain MRI; Brain extraction; Multi-atlas; Label fusion