Atlas based segmentation of brain structures in 3D MR images

Abstract

In this paper, we propose a robust atlas-based automatic scheme for segmentation of subcortical region and cerebellum. Two separate schemes are presented, each for registration of subcortical structures and cerebellum respectively. Registration between atlas and subject MR image is performed using thin plate spline as non-rigid transformation, normalized mutual information as image similarity measure and powell's method as optimization technique. The final transformation of subcortical structures and cerebellum is used to map the segmentation data-set from atlas to subject image. Results obtained through automatic segmentation are validated for seven structures in ten subjects using sensitivity, positive predictive value and dice coefficient. These structures include left and right ventricles, caudate nuclei, putamena and cerebellum. The values of sensitivity, positive predictive value and dice coefficient range from 0.84 to 0.98, 0.82 to 0.99 and 0.87 to 0.96 respectively.