Automatic brain extraction methods for T1 magnetic resonance images using region labeling and morphological operations

In this work we propose two brain extraction methods (BEM) that solely depend on the brain anatomy and its intensity characteristics. Our methods are simple, unsupervised and knowledge based. Using an adaptive intensity thresholding method on the magnetic resonance images of head scans, a binary image is obtained. The binary image is labeled using the anatomical facts that the scalp is the boundary between head and background, and the skull is the boundary separating brain and scalp. A run length scheme is applied on the labeled image to get a rough brain mask. Morphological operations are then performed to obtain the fine brain on the assumption that brain is the largest connected component (LCC). But the LCC concept failed to work on some slices where brain is composed of more than one connected component. To solve this problem a 3-D approach is introduced in the BEM. Experimental results on 61 sets of T1 scans taken from MRI scan center and neuroimage web services showed that our methods give better results than the popular methods, FSL's Brain Extraction Tool (BET), BrainSuite's Brain Surface Extractor (BSE) gives results comparable to that of Model-based Level Sets (MLS) and works well even where MLS failed. The average Dice similarity index computed using the “Gold standard” and the specificity values are 0.938 and 0.992, respectively, which are higher than that for BET, BSE and MLS. The average processing time by one of our methods is ≈1 s/slice, which is smaller than for MLS, which is ≈4 s/slice. One of our methods produces the lowest false positive rate of 0.075, which is smaller than that for BSE, BET and MLS. It is independent of imaging orientation and works well for slices with abnormal features like tumor and lesion in which the existing methods fail in certain cases.