Brain extraction using isodata clustering algorithm aided by histogram analysis

Magnetic resonance (MR) imaging has a broad application in diagnosis and detection process of different brain related diseases. Manual analysis of MR images is a cumbersome and time consuming task. In order to automatically analyze the brain tissue accurately, non-brain compartments must be removed from magnetic resonance images. This task is known as brain extraction or skull stripping. In this study a brain extraction method is proposed. The proposed method formulates segmentation problem as a clustering problem and its core component is isodata clustering algorithm. Application of isodata algorithm reveals five distinct clusters. Two of these clusters contain voxels belonging to tissues of interest and three of them belongs to non-brain compartments. In order to produce an accurate brain mask, isodata cluster representatives are initialized by histogram analysis of MR volume of the brain. These representatives are mods of histogram of MR volume. The second stage of the proposed method leads to produce more accurate brain mask by somehow removing outliers. In this case, isodata algorithm performs better. Performance of the proposed method is measured by popular performance measures such as Dice similarity coefficient (Dice), Jaccard similarity index (J), sensitivity, and specificity. The proposed method outperforms BET, BSE, and HWA as popular methods by Dice = 0.959 (0.008) and J = 0.921 (0.168). These results are obtained based on BrainWeb dataset.