Two popular segmentation methods used today are atlas based and graph cut based segmentation techniques. The atlas based method deforms a manually segmented image onto a target image, resulting in an automatic segmentation. The graph cut segmentation method utilizes the graph cut paradigm by treating image segmentation as a max-flow problem. A specialized form of this algorithm was developed by Lecoeur et al [1], called the spectral graph cut algorithm. The goal of this paper is to combine both of these methods, creating a more stable atlas based segmentation algorithm that is less sensitive to the initial manual segmentation. The registration algorithm is used to automate and initialize the spectral graph cut algorithm as well as add needed spatial information, while the spectral graph cut algorithm is used to increase the robustness of the atlas method. To calculate the sensitivity of the algorithms, the initial manual segmentation of the atlas was both dilated and eroded 2 mm and the segmentation results were calculated. Results show that the atlas based segmentation segments the thalamus well with an average Dice Similarity Coefficient (DSC) of 0.87. The spectral graph cut method shows similar results with an average DSC measure of 0.88, with no statistical difference between the two methods. The atlas based method's DSC value, however, was reduced to 0.76 and 0.67 when dilated and eroded respectively, while the combined method retained a DSC value of 0.81 and 0.74, with a statistical difference found between the two methods.