Hippocampus Segmentation Via Deep Learning

With the aggravation of the aging of the social population,the neurodegenerative diseases represented by Alzheimer’s disease have seriously affected the life,health and spirit of the elderly.Alzheimer’s disease,mild cognitive impairment and temporal lobe epilepsy are often accompanied by shrinkage of hippocampus volume of patients.Therefore,the analysis of hippocampus volume is the key to the diagnosis and research of these diseases,whereas the primary condition for the analysis is to segment hippocampus from magnetic resonance images.Because the hippocampus is irregular in shape,small in size and blurred in edge,it is often difficult for general segmentation methods to obtain satisfactory segmentation results.Based on the theory and technology of deep learning,this paper proposes a hippocampus segmentation method combining sequence learning and u-like-net and a two-stage hippocampus segmentation method.The method combining sequence learning and U-like-net for hippocampus segmentation introduces the dependent information between image sequences into the two-dimensional semantic segmentation network,so as to make full use of the three-dimensional context of image.In addition,for the possible problem of sample balance in the task,this method improves this by applying class weight parameters in the training phase.The experiment on ADNI dataset verifies the performance and stability of the method.Compared with ordinary U-like-net,the experiment verifies the validity of introducing sequence learning structure,and the network visualization technology verifies the rationality of network learning.The two-stage hippocampus segmentation method based on deep learning includes localization stage and segmentation stage.Each stage consists of two steps.In the localization stage,a trained classification network is used to fetch slices containing hippocampus.The fetched slices are then used to generate multi-size candidate regions.In the segmentation stage,candidate regions are input into the segmentation network to generate prediction probability maps containing multi-context information.Then the probability maps are fused to generate a segmentation mask in a view.The above process is repeated in three orthogonal views.Finally,segmentation masks of three views is aggregated through multi-view decision to obtain the final segmentation.Quantitative evaluation on ADNI dataset proves the validity and rationality of the method.In comparison with the related work,the method also achieves good results.