Noise Removal Of Micro-Optical Sectioning Tomography Image Based On Convolutional Neural Network

The anatomical structure of neurons is very important in the study of brain science.Traditional biomedical imaging systems are unable to achieve large range of high-resolution imaging and meet the research needs of constructing fine neural circuits in the aspect of acquiring neuron structures.The micro-optical sectioning tomography system has the ability to provide three-dimensional high-resolution data sets of the whole brain,thus solving this problem.However,due to the influence of imaging environment and imaging sensor state,the image noise is unavoidable in the mouse brain contour region.In addition,under the influence of sample embedding agent,there was noise in the micro-optical sectioning tomography image data outside the contour region of rat brain.The existence of image noise will interfere with the recognition of weak nerve fiber signal,which will hinder the research of neuroscience.Traditional image denoising methods cannot effectively remove noise and retain weak effective signals for the micro-optical sectioning tomography image data.Compared with only low-level image features of the traditional extraction methods of predefined,convolution neural network as a supervised learning method,can directly from spontaneously learn deep image characteristics in image data,and therefore has the stronger ability of image feature extraction and widely used in image processing and good results have been achieved.A set of image denoising method is designed for the image noise existing in the microoptical sectioning tomography image data in this paper.According to the location of noise,the noise in the micro-optical sectioning tomography image data can be divided into the noise outside the rat brain contour area and the noise inside the rat brain contour area.The U-Net model was used to remove the noise inside the rat brain contour.The U-Net model uses residual unit to transform the learning of image features by neural network into the learning of noise features.In terms of the production of control data set and training data set,the artificial synthesis of images by means of noise model superposing natural images solves the problem that the generation of noiseless control images cannot be obtained in the research of denoising and the generation of training set of convolutional neural network model.For the noise existing outside the rat brain contour,a neural network with local image feature extraction ability is designed to realize the segmentation of the rat brain contour,so as to remove the noise outside the rat brain contour.This paper designs a set of image denoising methods for image data of microoptical slice tomography.The proposed method can effectively remove the image noise and retain the weak effective signal in the image data.Compared with the traditional image denoising method,it has better denoising results and signal retention ability.This method is expected to be applied to other types of biomedical image data.