Thalamus segmentation using magnetic resonance imaging data 

Presented by Vahideh Mousavi Manarehbazari

Thalamus is a neuroanatomic structure in the brain believed to act as a relay between subcortical areas and the cerebral cortex. The thalamus is composed of a number of anatomically and functionally distinct cell groups or nuclei (sub-regions). The knowledge of thalamic sub-regions is important for example for the treatment of movement disorders or in developing a treatment for drug-resistant epilepsy.

The aim of the thesis was to develop a method for automatic segmentation of the thalamus into its seven functional sub-regions: primary motor, sensory, occipital, pre-frontal, pre-motor, posterior-parietal and temporal. Two segmentation methods were implemented and investigated: k-nearest neighbors and k-means algorithm. The input feature vector consisted of signals  obtained from Magnetic Resonance Imaging, such as T1-, T2-weighted data  and FA-fractional anisotropy. The experiments were performed on both synthetic  data and a real patient data. The resulting voxel segmentation was compared  to the Oxford thalamic connectivity atlas.

The highest segmentation accuracy was close to 90%, estimated by the supervised k-NN method,  and 45% for the segmentation done by standard k-means clustering. To improve performance of the unsupervised segmentation, a weighted k-means using a priori  information needs to be developed.

The project was carried out at MedTech West, Sahlgrenska Hospital.