High-throughput Identification System For The Studies Of Brain-wide Neuronal Distribution

The brain is the most important organ of the human body,and neurons of different types and huge numbers are the basis for the Brain complex functions.Therefore,accurately drawing the distribution atlas of different types of neurons in the brain-wide,which is essential for the correct understanding of the corresponding relationship between brain structure and function.Scientists can obtain clear neuron distribution images by means of specific neuronal markers and advanced micro-optical imaging techniques.However,Because of the neuron use different marker methods,the grayscale and texture characteristics displayed by different neuron images are different.In addition,contemporary neuroscience applications are producing TB-level data sets in an industrialized manner,and still use single neuron identification methods will face problems such as limitations,low accuracy,low efficiency and so on.In order to solve the above problems,this paper studies the images of neuron nucleus markers and soma markers,and proposed a systematic neuron brain-wide recognition scheme.Aiming at the problem of non-uniform signal to dense distributed neuron image of nucleus marker,the idea of minimum cut and decision tree is used to eliminate oversegmentation obstacles caused by non-uniform signal in the image.This method introduces the gradient weighted distance transformation method and the three-dimensional watershed algorithm to perform nucleus segmentation,which solves the problem of adhesion nucleus segmentation in dense distributed neuron images.Based on the idea of graph theory,the nuclear segmentation results are further optimized,avoiding over-segmentation caused by non-uniform nucleus signal.The proposed method can achieve 94% accuracy in dense neuron nuclear recognition.Aiming at the various problems of soma morphology,labeling characteristics,and interference signals in soma marker images,the semantic segmentation network and the weighted mechanism of multi-algorithm fusion are combined to enhance the robustness of different neuronal soma images recognition.This method extracts the soma signal in the image as a weakly supervised manner,avoiding the interference with extra signals such as fibers,blood vessels,and so on.Further introduction to ensemble learning ideas,with a combination of multiple methods to get a more comprehensive model,which significantly improves neurons recognition accuracy.The ensemble method can be applied to a wider range of whole brain neuron data,with a recognition accuracy 96%,which is far better than the identification effect of a single algorithm.In response to accurate recognition is required in the studies of brain-wide neuron distribution,the ideas of TDat divided and conquer as well as merge verification are combined,which have advantages in high efficiency and high accuracy respectively,and a high-throughput neuron recognition system is further designed.Based on this system,studied the location of Ca MK? neurons,and the location and distribution of CRH neurons in the whole brain of mice,as well as the soma morphology of giant pyramidal neurons in the whole brain of ferrets.In summary,this paper constructed the neuron identification system provides a highly accurate and high-throughput recognition scheme for the study of neuron distribution in the whole brain,which is an indispensable technical means in the study of neuron distribution maps.