Study On The Nonlinear Registration Method For Whole-Brain Mesoscopic Neural Connectivity Atlas

From the basic human body movements to the complex emotional memory,they are all controlled by brain networks formed by the cooperation of different types of neurons or neural circuits.Hence,deconstructing fine neural structures and spatial connections at the mesoscopic level is critical in understanding brain networks' mechanisms accurately.Neuroscientists are able to delineate the brain regions and nuclei in which neurons and neural circuits are located with the help of a stereotactic brain atlas.However,mainly due to the differences in individual animals,the correspondence of brain structures with an atlas is highly dependent on personal experience and proficiency.Moreover,the combination of different whole-brain optical labeling methods and imaging technologies makes the image characteristics of specific regions are different at mesoscopic level.The linear and uniform registration methods are challenging to match these heterogeneous datasets into a standard brain atlas accurately.Additionally,various neuroscience applications are also producing massive brain datasets in large-scale and industrialized,still using traditional manual positioning method will face problems of heavy workload,inefficiency.To solve the above problems,two key properties of accuracy and automation of threedimensional nonlinear registration algorithm are studied and a mesoscopic brain image registration framework is constructed.The specific contents of this dissertation are as follows.(1)A three-dimensional registration method based on anatomical regional features is proposed,which aims to solve the problem of accuracy of mesoscopic brain image registration.This method uses anatomical regional features as the positioning landmark for registration,which can obtain a large number of accurate and reasonably distributed reference points for describing complex deformations in the whole-brain.Then,combined with the regional feature based symmetric diffeomorphic registration algorithm,the positioning landmark is matched and deformed;hence,the deformation of each position in the brain can be flexibly recorded with a non-parameter model.The registration results are highly accurate in brain region level(Dice score > 0.9).There is no significant difference between registration results and anatomist expert results in the identifiable nuclei level at a 10 ?m resolution(P > 0.05),which is significantly better than other gray level and point based registration methods such as BrainAligner,aMAP and ClearMap.(2)Another three-dimensional registration method based on a convolutional neural network is proposed,which aims to meet the automation of mesoscopic brain image registration.First,a patch-based neural network registration model for mesoscopic brain images is constructed to predict the deformation field corresponding to the two inputted images directly.Second,a self-feedback training strategy that adjusts the weights of the training samples is designed to make it possible to accurately select regions with larger and more complex deformations for additional “special training”.Third,a dual-hierarchical registration network is proposed to generate training labels to learn the large and small deformation in the brain respectively.Finally,the proposed method can automatically obtain the registration accuracy of Dice scores above 0.9 in the brain region level.The average Dice scores are 0.82 in 140 nuclei level,which is significantly better than other deep learning registration methods such as Voxel Morph,AVSM and RDMM(P < 0.05).(3)A post-registration revision scheme is proposed,which can flexibly combine the accuracy and automation advantages of the above two registration methods.Only a small amount of manual revision is needed,and the accuracy of the revisions scheme is consistent with that of manual positioning(P > 0.05).Futhermore,a mesoscopic brain registration framework is constructed to meet different dataset registration requirements in the actual neuroscience applications.Through the cooperations of the manual registration module,the intelligent registration module,and the feature extraction module,the registration tasks of the distributions of type-specific neuron of the mouse brain,the neuron projections in specific brain regions of the mouse brain,rat brain dataset,and macaque brain dataset are realized.In a word,the registration methods and framework proposed in this study can provide highquality and efficient spatial positioning scheme for multi species brain information dataset,which is an indispensable part in the construction of mesoscopic neural connectivity atlas.