AN ALGORITHM TO SEGMENT THE MID-BRAIN STRUCTURES USING MULTIRESOLUTION NON-RIGID REGISTRATION

Abstract

In present study, a reference based technique has been proposed to segment midbrain structures in 3D Magnetic Resonance (MR) images using multi-resolution non-rigid registration. The proposed scheme segmented the mid-brain structures in two stages. During first stage, a target image was registered to an already segmented reference image while in second stage, segmentation information was mapped from reference to target medical image. Using this scheme, the aim was to segment nine mid-brain structures including left ventricle, right ventricle, third ventricle, anterior commissure, posterior commissure, left putamen, right putamen, left caudate nucleus and right caudate nucleus. Results of automatic segmentation achieved through reference based non-rigid registration were validated by comparing with the manual segmentation, considered as ground truth. Three metrics namely Dice Coefficient, Sensitivity and Positive Predictive Values were used for assessing the segmentation results. The proposed algorithm showed the satisfactory segmentation of mid-brain structures by achieving a mean sensitivity of 89. 85 % for five target MR images. The results indicated that the proposed scheme performed accurate segmentation of mid-brain structures.