Segmentation of hippocampus based on ROI atlas registration

Since the change of hippocampal volume is an early symptom for Alzheimer´s disease, hippocampus segmentation of brain image can be used to assist the diagnosis of the Alzheimer´s disease. This paper presents a multi-atlas based segmentation method which produces higher segmentation accuracy than single atlas-based segmentation. After the extraction of the ROI, a two-step registration from atlas images to target image is applied to obtain initial segmentations which are combined to generate the final segmentation by several atlas fusion strategies. In the experiments, four different strategies including single maximum, majority voting (MV) and simultaneous truth performance level estimation (STAPLE) and consensus level, labeler accuracy and truth estimation (COLLATE) are tested. Dice overlap between automatic and manual segmentation is employed to compare the results of these atlas selection strategies. The experimental results show that the COLLATE can achieve the better segmentation result in our study.