Hippocampus Image Segmentation Based On Global Spatial Attention And Two-Stage Feature Fusion

Chinese society is gradually aging.Mental diseases represented by Alzheimer’s disease are endangering the life and health of the elderly.As an important part of the brain’s nervous system,the changes in the volume and structure of the hippocampus are one of the clinical features of many early-stage psychiatric diseases.Therefore,the effective segmentation of hippocampal MRI images is of great significance for the diagnosis and treatment of related psychiatric diseases.However,the traditional manual method of segmenting the hippocampus image not only requires high professional level of the operator,but also is susceptible to subjective factors to cause segmentation errors.Therefore,how to use deep learning technology to automatically segment the hippocampus is an important research direction in medical image processing.In this paper,an improved segmentation algorithm based on attention mechanism and boundary loss function is proposed for the automatic segmentation of hippocampus images,which has the features of small size,irregular shape and unclear contour.The main contributions are:(1)Proposed a dual-path Global Spatial Attention mechanism(GSA)for the problem of classimbalance caused by the small size of the hippocampus in the MRI image compared to the whole brain.On one path,based on the idea of non-local mean filtering operation,the interaction between the two pixels of the image is calculated by Gaussian function to directly capture the remote dependence,which breaks the limitations of the adjacent points of the convolution operation,expands the receptive field to the entire feature map,and obtains the global context modeling ability.On the other path,the feature channel is compressed and activated to obtain the spatial weight spectrogram,which is applied to the feature map to complete the spatial information calibration,and finally fused with the global information to achieve precise positioning of the hippocampal target area by the broadcast mechanism.(2)Inspired by discrete optimization techniques for computing gradient flows of curve evolution,a boundary thinning loss function is proposed for the problem of unclear boundary contours of the hippocampus.By integrating the predicted value and the labeled value in the spatial domain carry out the contour measurement and minimize the L2 regular distance between two boundaries to smooth the segmentation process.In addition,the Special Feature deep Conv module(SDC)is proposed for the non-uniform spatial details contained in images of different dimensions,which use bottle-neck operation to extract the unique features while maintaining high resolution,and then supplement the lost spatial information in the upsampling process to improve the segmentation accuracy.(3)A two-stage deep feature fusion network oriented to the hippocampus is proposed,which divides the automatic segmentation of the hippocampus into two stages: In the first stage,based on the Ghost module idea,a series of linear transformations are used at low cost to generate phantom images that can fully reveal the inherent feature information to build a lightweight network model,so that can quickly segment the hippocampus image.In the second stage,the coarse segmentation result is fed to the improved U-Net network to achieve further refinement of the segmentation.Through training and testing on the public data set NITRC,the final results show that compared with U-Net Plus Plus and other networks,the algorithm model proposed in this paper focuses more accurately on the hippocampus region,is more sensitive to the texture of the edge contour,and the DICE similarity is significantly improved.It effectively solves the problems of unbalanced and low precision in the segmentation of hippocampus images,exploring a new intelligent hippocampal segmentation method for the clinic.