Research On Hippocampus Segmentation In Magnetic Resonance Brain Images And Its Application In Alzheimer’s Disease

In a variety of medical image analysis tasks,research on automatic hippocampus segmentation based on magnetic resonance(MR)brain images and diagnosis of Alzheimer’s Disease(AD)are of great importance.First,improving the accuracy of the automatic hippocampus segmentation algorithm on MR brain images is beneficial to clinicians in disease diagnosis and treatment plan by better characterizing the shape and volume changes of the hippocampus.Second,the integration of deep learning technique and MR brain images in the AD study can facilitate early diagnosis of AD and predict the progression of patients with mild cognitive impairment(MCI).Therefore,two studies were carried out including hippocampal segmentation based on multi-atlas images and AD diagnosis based on deep learning.The multi-atlas image segmentation(MAIS)method has been widely studied and achieved promising results in hippocampus segmentation.The MAIS mainly includes hippocampus registration and the followed label fusion.However,most of the current label fusion algorithms conduct machine learning classification models on each individual voxel,which leads to the neglect of the correlation between voxels when segmenting the target image.In this study,a multi-task learning method and a semi-supervised label propagation algorithm were combined to construct a new hippocampus segmentation method based on multi-atlas images.The multi-task learning and semi-supervised segmentation scheme could effectively make use of the correlation between voxels in the target image to improve the accuracy of hippocampus segmentation.First of all,the multi-task learning method trained a joint classification model for a multi-task-voxel cluster in a target image patch.Then,the probabilistic segmentation result of the central voxel in the voxel cluster were obtained by majority voting method.Finally,the probabilistic segmentation results were used to guide the semi-supervised label propagation algorithm to obtain the final segmentation results.In this study,the experiments were carried out on the publically available dataset of EADC-ADNI(including 135 MR brain images with hippocampus segmentation labels),and compared with five hippocampus segmentation methods based on multi-atlas images,including non-local patch,majority voting,metric learning,local label learning and random local binary pattern.The experimental results showed that this method achieved superior performance on hippocampus segmentation on MR brain images.Deep learning has been widely used in the diagnosis of AD to accomplish the task of disease classification and prediction.However,the whole brain-based AD diagnosis classification strategy is time-consuming,and the repeatability of capturing significant brain regions through the attention mechanism in deep learning networks also needs to be further verified by multi-center datasets.Therefore,a deep learning study based on a 3D attention network(3DAN)was carried out based on multi-center datasets(ADNI,EDSD,OASIS,a total of 2467 subjects).First,the comparison experiments were carried out on gray matter volume images of whole brain,hippocampus,and more brain regions(subcortical nucleus(removal of hippocampal),temporal lobe,insular lobe,frontal lobe,parietal lobe,limbic lobe,and occipital lobe)in AD diagnosis.At the same time,the attention mechanism in the deep learning network was also used to capture significant brain regions from the whole-brain-based AD diagnostic classification strategy and verify its repeatability on multi-center datasets.Second,the AD diagnostic classification models established on the whole brain and hippocampus were further compared by applying them on the classification of stable MCI(s MCI)and progressive MCI(p MCI)in MCI patients.Finally,an improved 3D attentional network(3D Attention Network with Drop Block,3DAN-DB)was proposed to further improve the performance of hippocampal grey matter volume-based AD diagnosis.The experimental results on multi-center datasets showed that 1)the deep learning classification model based on hippocampus gray matter volume achieved more promising results with stable AD diagnosis and fast computation;2)The stability and repeatability of the attention mechanism in capturing significant brain regions in the whole brain-based deep learning model were insufficient,which also had negative impacts on the interpretability and correlation analysis of the captured significant brain regions in AD analysis;3)Regarding the s MCI and p MCI classification,hippocampus also achieved better results than whole brain,which further proved that hippocampus had high potential in AD diagnosis;4)The proposed 3DAN-DB model achieved better performance in the classification and diagnosis of AD based on hippocampus gray matter volume.