| Brain is the most complex and mysterious organ of the human body.The study of brain structure and function is of great significance in understanding how the brain works and revealing the mysteries of the brain.The neural circuit is the basic unit of the brain nervous function.It transmits and processes information in the brain.Neuron is the basic unit of a neuron circuit.Neurons with different functions have different shapes,locations and projection patterns,which morphology distributes across multiple brain regions even the whole brain.Therefore,it is necessary to map the neural circuits in whole-brain at single neuron resolution.With the development of labeling technology and imaging technology,imaging the whole mouse brain at single-neuron resolution has become a reality,which provides important image data for neuroscientists to study neural circuits.However,wholebrain imaging with high-resolution produces massive amounts of data up to TB or even dozens of TB.Traditional brain image processing methods have become difficult to process these large images data.Aiming at large data problem in brain image data processing,this thesis solved two key problems of big data storage and big data access respectively,and established a software platform for massive brain image data processing.(1)Big data storage system and method.A petabyte-scale mass data storage system based on Lustre cluster file system is selected and implemented,which provides the hardware foundation for mass data storage.The raw data archiving method based on tar is designed,which solves the problem that the large number of small files produced by imaging are difficult to store and transmit.TDat data format for 3D data storage is established,which uses multi-level resolution block storage method to replace 2D image sequence storage.An efficient data reformat method is developed to convert the general image format to TDat format.TDat can support 1PB of 3D data and has a performance advantage of up to 32 times compared with similar methods on data reformatting.(2)Big data access method.We put forward two access modes: sparse data accessing and high throughput data accessing.Aim at these two access modes,based on the TDat data format,we designed the caching mechanism to reduce the reading of repeated data during data access,and to accelerate the efficiency of data access using data I/O parallelism.Four access methods of cache-based ROI block parallel access,direction-predicted cache preloaded access,non-overlapping data sequential access and "Tile" data continuously access are established.It satisfies the demand of different data access methods for different brain image processing applications.The access method is provided for processing massive data in TDat data format.(3)Big data processing software platform.Based on TDat data storage method and data access method,a software platform to process massive 3D brain image processing is established.The entire platform adopts modular three-tier architecture,which can be compatible with different computing platforms and common image processing software,and has good scalability,which makes up for the deficiency of existing software tools.Based on this platform,a rigid registration method based on TDat is developed,which uses lowlevel resolution data to acquire the spatial transformation parameters,and high-resolution data blocks are used for registration.This method can quickly register the whole brain image of the order of terabytes with only a few gigabytes of memory consumption.Based on this platform,the method of interactive tracing for long range neural fiber is developed,combined with Amira Fiber tracking module and TDat data storage format for block tracking,and using the traced of nerve fiber to predict the direct of fiber and preload the required data,realizing the efficient tracing of long-range projection neurons in TB-scale dataset.This method takes only one day to process data preparation and tracing,compared to the traditional method,which consumes dozens of days. |
PDF Full Text Request