Research On System For The Region Localization And Labeled Nerve Cells Counting Of Brain Slice Images

In brain science research,in order to determine the complex network connection relationship of nerve cells inside and outside the brain regions,it is necessary to quantitatively analyze and compare the labeled neural cells and neural network connection modes of each brain region in a brain slice,that is,first determining the location coordinate distribution information of each brain region in the brain slice image based on the Atlas digital brain map(brain slice image region positioning),and then counting the number of labeled nerve cells in each brain region,in this way a quantitative analysis of various connection modes is performed.However,due to the inevitable introduction of various distortion and deformation and anamorphose during the brain slice image surgery and imaging stages,the current regional positioning and nerve cell counting work still requires a lot of human intervention.Therefore,the development of an automatic brain slice region positioning and labeling nerve cell counting system that can adapt to various distortion and deformation and anamorphose will greatly improve the efficiency of brain science related research.Brain slice-Average Template brain map registration and detection of labeled nerve cells in brain slices are the keys to realizing brain slice area localization and nerve cell counting automatically.First of all,facing the problems such as distortion and deformation and anamorphose in brain slice images,and large modal differences from Average Template brain atlas images,we propose an unsupervised registration algorithm of convolutional neural network combined with PCANet to implement a more accurate brain slice-Average Template brain map registration,and use the registration deformation field to deform the Atlas digital brain atlas image containing the distribution information of each brain region to obtain the location information of each brain region in the brain slice;Then,in view of the problems of missed detection and misdetection in the conventional method in the detection of labeled nerve cells,the improved Faster R-CNN detection algorithm is used to detect labeled nerve cells in brain slice images,which has realized accurate automatic detection of labeled nerve cells and obtained the location information of each labeled nerve cell;Finally,combining the location information of each labeled nerve cell and the location information of the brain slice area,the labeled nerve cell count of each brain region is realized.The specific work is as follows:(1)Brain slice image preprocessingIncluding cropping and scaling,background removal,graying processing,affine transformation and so on.Preprocessing can remove the interference factors in the image,reduce the gap between the brain slice and the brain atlas,and improve the accuracy of brain location of the brain slice image.(2)Pre-processed brain slice-Average Template brain map registration and brain region locationAverage Template Brain map and Atlas Digital Brain map are standard brain map of different modalities of the same object.The Average Template brain map is closest to the brain slice modality,but does not contain brain area labeling information and spatial coordinate information.The Atlas digital brain map contains brain area positioning information but is much different from the brain slice modality,therefore,through the brain slice-Average Template brain map registration,a spatial deformation field can be obtained,and according to this spatial deformation field,Atlas digital brain map brain region positioning information can be mapped onto the brain slice.Since the number of labeled neurons on the brain slice needs to be counted,it is necessary to maintain the invariance of the brain slice,and use the brain slice as a reference image,and register the brain map as a floating image on the brain slice.The registration and brain region positioning steps are first based on mouse fluorescently labeled brain slices,with the assistance of brain science experts,the Atlas digital brain map and Average Template brain map in the ARA(Allen Reference Atlases)database corresponding to the brain slices are obtained.After preprocessing,we use the PCANet-based structure representation network to convert the Average Template brain map and mouse brain slices into the same modality,then use Reg-net and STN spatial transformation network to achieve unsupervised registration,and finally use the obtained spatial deformation field deforms the Atlas digital brain map to obtain the regional positioning information of the brain slice.(3)Detection of labeled nerve cells in brain slicesAiming at the missed detection and misdetection of the traditional threshold detection algorithm for labeling nerve cell detection,an improved Faster R-CNN target detection algorithm,namely the IRN Faster R-CNN(Swish)detection algorithm,is used to detect labeled nerve cells in various brain regions to obtain coordinate information of each labeled nerve cell.The algorithm combines Res Net and Inception Net as the backbone network,improves the network's ability to extract multiscale features,and replaces the Re LU activation function with the Swish activation function,improves the model's nonlinear fitting ability,and combines RPN(Region Proposal Network,RPN)network finally improves the detection effect of the algorithm.(4)Labeled nerve cells counting in various brain regions of brain slices and GUI system implementationThe obtained location information of the brain slice area is compared with the coordinate location information of the labeled nerve cells,thereby obtaining the number of labeled nerve cells in each brain area.A Py Qt5-based brain slice image area localization and labeled nerve cells counting system is constructed to facilitate brain science researchers to use the algorithm model proposed in this thesis.By comparing with the experimental results,it is proved that the method proposed in this thesis can perform more accurate and rapid brain region location on brain slice images.Compared with the Voxelmorph registration model results,the proposed unsupervised registration model of convolutional neural network combined with PCANet reduces the Root Mean Square Error(RMSE)by 1.63%,and the Correlation Coefficient(CC)and Mutual Information(MI)are improved 2.31%,0.63% respectively,and the registration time is shortened to less than 1 second;Labeled nerve cells in each brain region are accurately detected.The m AP value of the proposed IRN Faster R-CNN(Swish)detection model reaches 98.9%,and the detection accuracy is better than other target detection algorithms.