Whole-brain neural cell body identification and counting method and device based on machine learning, and storage medium

Abstract

The invention provides a whole brain neural cell body identification and counting method and device based on machine learning and a storage medium. Firstly, whole-brain image data is collected and preprocessed, then a plurality of cubic image blocks are intercepted from different areas of multiple mouse brains, neuron cell bodies and background information are marked, a classifier model is trained, the model is applied to complete whole-brain neuron cell body recognition, whole-brain fine registration is carried out according to a brain map, and the whole-brain image data is obtained. And finally, counting the number of neuron cell bodies in different brain regions. According to the method, a plurality of characteristics of the neuron cell body are automatically learned by adopting the pixel classifier of machine learning, higher accuracy is realized under the condition that the brightness value of the cell body is relatively high, and particularly, the expression is outstanding in a cerebellar region containing a large number of low-brightness granular cells; the efficient cell body recognition and counting model can be established only by a small number of samples, the workload of data preparation is reduced, and a more efficient and accurate solution is provided for recognition and counting of whole brain neural cell bodies.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a whole-brain neuron cell body recognition and counting method based on machine learning. Background Art

[0002] The Fluorescence Microscope Optical Sectioning Tomography System (FMOST) can acquire a complete mouse brain data set at a spatial resolution of 0.35μm×0.35μm×1μm through precise mechanical cutting combined with fluorescence microscopy. This high-resolution data acquisition capability not only enables the morphology of individual neurons to be accurately depicted, but also makes it possible to construct a detailed three-dimensional (3D) mouse brain model.

[0003] However, the amount of whole-brain imaging data generated by the fluorescence microscope optical sectioning tomography system is extremely large, usually reaching the TB level, and the diameter of neurons in the mouse brain is only about 10um, while the number of neurons in the whole brain reaches tens of millions, which poses a severe challenge to the identification and counting of neurons in the whole brain. At present, the statistics of the number of neurons in the whole brain and the analysis of the impact of these neurons on brain function still rely on manual or semi-manual methods. The current methods for identifying and counting the number of neuronal cell bodies in the whole brain are completed with the help of the commercial software IMARIS. The software imports the data after atlas segmentation and uses the brightness difference of image pixels to identify neuronal cell bodies. Although IMARIS provides a certain degree of manual correction function, allowing users to add or delete identified cell bodies to correct errors in the automatic detection process, given that the number of cell bodies in the whole brain can be as high as millions or even tens of millions, this manual correction method is not feasible in actual operation. Therefore, the automatic detection results of IMARIS often cannot meet the needs of accurate statistics of the number of neuronal cell bodies in disease model research. Summary of the invention

[0004] The purpose of the embodiments of the present invention is to provide a whole-brain neuron cell body identification and counting method based on machine learning, so as to realize automatic identification of neuron cell bodies and improve the accuracy of cell body identification and counting.

[0005] The second purpose of the embodiment of the present invention is to provide a whole-brain neuron cell body identification and counting device and storage medium based on machine learning.

[0006] In order to solve the above technical problems, the technical solution adopted by the present invention is a whole-brain neuron cell body recognition and counting method based on machine learning, and the specific steps are as follows:

[0007] Preprocessing of the whole-brain 3D images of multiple rats acquired from the FMOST device;

[0008] A plurality of cubic image blocks are intercepted from a plurality of different regions of the preprocessed three-dimensional whole-brain images of a plurality of mouse brains, and the neuronal cell bodies in the intercepted cubic image blocks are marked, and then a classifier model is trained using the marked cubic image blocks, and then the trained classifier model is used to perform neuronal cell body recognition on the preprocessed three-dimensional whole-brain images;

[0009] Register the whole brain 3D image to distinguish the left and right brain regions;

[0010] The cell bodies of different brain regions were counted based on the results of whole-brain region registration and whole-brain neuron cell body identification.

[0011] Further, the preprocessing includes the following steps performed in sequence:

[0012] Jigsaw puzzle: stitch the collected images of different fields of view into a complete whole-brain image;

[0013] De-striping: using denoising algorithm to remove the stripe noise in the image collected by FMOST equipment;

[0014] Cropping and cutting out: extracting the effective part of the whole brain three-dimensional image through cropping and cutting out;

[0015] Rotation alignment: rotate the whole brain three-dimensional image at a certain angle to obtain an image with an aligned angle;

[0016] Denoising, the image is processed and cleaned through the denoising algorithm.

[0017] Furthermore, before the cropping and cutting, the whole brain image data after the stripe removal operation is downsampled, specifically: the original 1.3×1.3×0.92 μm resolution image is downsampled to a (1.6~3.6)×(1.6~3.6)×(1.76~3.76) μm resolution image.

[0018] Furthermore, before the cropping and cutting, the whole brain image data after the stripe removal operation is downsampled, specifically: the original 1.3×1.3×0.92 μm resolution image is downsampled to a 2.6×2.6×2.76 μm resolution image.

[0019] Furthermore, the method of intercepting multiple cubic image blocks from multiple different areas of the preprocessed whole-brain three-dimensional images of multiple mouse brains is specifically: intercepting cubic image blocks with a size of 100×100×100 pixels from the cerebral cortex, striatum, cerebellum and pure background areas of the preprocessed whole-brain three-dimensional image data of at least 4 mouse brains, wherein at least 4 cubic image blocks are intercepted from each of the three areas of the cerebral cortex, striatum and cerebellum of each mouse brain, and at least 1 cubic image block is intercepted from the pure background area of ​​each mouse brain.

[0020] Furthermore, the step of marking the neuron cell bodies in the intercepted cubic image block includes:

[0021] The foreground target and background of the cubic image block are distinguished and marked, and the neuron cell bodies in the cubic image block are manually marked. For the marking of the background, the distance transformation algorithm and the threshold filter method are used to extract part of the background information for marking. After each cubic image block is marked, two new cubic image blocks are obtained, one containing the foreground target, i.e., the neuron cell body, and the other containing the background information. The grayscale values ​​of the pixels where the foreground target and the background information are located are both 1, and the grayscale values ​​of the remaining pixels are 0.

[0022] Furthermore, the specific steps of extracting background information using the distance conversion algorithm and the threshold filter are as follows: first, the distance conversion algorithm is used to perform distance conversion processing on the image after the neuron cell bodies are marked, and the shortest distance of each pixel point from all the surrounding neuron cell bodies in three dimensions is obtained, and the shortest distance is used as the brightness value of each pixel point, and then according to the data obtained after the distance conversion, multiple thresholds are set to obtain a threshold filter, and two different threshold filters are applied to the image to obtain filamentous background information.

[0023] Furthermore, the method of training the classifier model using the labeled cube image blocks includes:

[0024] The classifier model is trained using a pixel classification algorithm, and each pixel is independently classified into a predefined class using a random forest model in machine learning;

[0025] The intercepted cube image blocks and the corresponding annotated cube image blocks are used as image data sets to train the random forest model. Each cube image block and its corresponding label are provided as input data to the model. The labels are foreground objects and backgrounds. The model learns how to distinguish foreground objects from backgrounds based on the input features.

[0026] The performance of the model is evaluated using three evaluation indicators: accuracy, recall, and F1 score. The model parameters are optimized based on the evaluation results to finally obtain a classifier model that meets the requirements.

[0027] Furthermore, the use of the trained classifier model to perform neuron cell body recognition on the preprocessed whole-brain three-dimensional image includes: inputting the preprocessed whole-brain three-dimensional image data into the trained classifier model, and the classifier model outputting the whole-brain three-dimensional image in which the foreground target and the background are distinguished and marked with different grayscale values, wherein the classifier model marks all pixel points of the neuron cell body as a grayscale value of 1, and all pixel points of the background as a grayscale value of 0.

[0028] Furthermore, it is characterized in that the brain region registration for distinguishing the left and right brain of the whole brain three-dimensional image includes:

[0029] All brain regions in the brain atlas are merged into sub-regions to form 244 brain regions. Then the new atlas is segmented into left and right brains, and finally a new registration atlas template is generated that distinguishes the left and right brains and contains 244 brain regions in each brain.

[0030] During the registration, the BIRDS plug-in in the Fiji image processing software was used to perform whole-brain brain region registration, and the newly generated registration atlas template was used to perform three-dimensional registration of the whole-brain image to achieve fine segmentation of the whole-brain brain regions. Finally, the IMARIS software was used to realize the visualization and data statistics of the whole-brain brain regions, and the information of 244 brain regions in the left and right brains after registration was obtained through data statistics.

[0031] Furthermore, the counting of cell bodies in different brain regions according to the results of whole-brain region registration and whole-brain neuron cell body identification includes: determining which brain region each cell body belongs to according to the position information of each cell body and the gray value information of each brain region, thereby obtaining the number of cell bodies in each brain region of the whole brain.

[0032] An electronic device adopts the above-mentioned whole-brain neuron cell body recognition and counting method based on machine learning to implement the whole-brain neuron cell body recognition and counting based on machine learning.

[0033] A computer storage medium stores at least one program instruction, and the at least one program instruction is loaded and executed by a processor to implement the above-mentioned whole-brain neuron cell body recognition and counting method based on machine learning.

[0034] Compared with the prior art, the beneficial effects of the present invention include the following points:

[0035] 1. Improved accuracy of cell body identification: Compared with traditional cell body identification methods, such as the automatic detection function used in the commercial software IMARIS, the method of the present invention shows higher accuracy when the cell body brightness value is high. Especially in the cerebellum area, which contains a large number of granule cells with low brightness values, the present invention can effectively identify these weak signal granule cells, which is significantly better than traditional cell body identification methods such as IMARIS, which often produce large errors in such cases.

[0036] 2. Reduced requirements for training data sets: The method of the present invention only requires a small amount of training samples to achieve efficient cell body recognition and counting. Compared with the existing methods that usually require a large amount of training data to build a model, the method of the present invention only requires a dozen to dozens of cubic image blocks as training data sets to quickly build a classifier model. This not only greatly reduces the workload of data preparation, but also makes the method more convenient in practical applications and easy to promote.

[0037] In summary, the present invention not only solves the problem of insufficient accuracy in cell body recognition in the prior art, but also simplifies the preparation process of the training data set, thereby providing a more efficient and accurate solution for the recognition and counting of FMOST whole-brain neuron cell bodies. These advances are of great significance for a deeper understanding of brain structure and function, especially for revealing the mechanism of neuronal changes in disease model research. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0039] Figure 1 is a flow chart of a method according to an embodiment of the present invention;

[0040] Figure 2 Schematic diagrams of intercepting cubic image blocks from different regions of a mouse brain according to an embodiment of the present invention; (a) is a schematic diagram of intercepting cubic image blocks from a cerebral cortex region and a three-dimensional image of four intercepted cubic image blocks, (b) is a schematic diagram of intercepting cubic image blocks from a striatum region and a three-dimensional image of four intercepted cubic image blocks, and (c) is a schematic diagram of intercepting cubic image blocks from a cerebellum region and a three-dimensional image of four intercepted cubic image blocks;

[0041] Figure 3 It is a single-layer image along the Z-axis direction in a cubic image block with neuron cell bodies and background information marked in an embodiment of the present invention; (a) is a single-layer image along the Z-axis direction in a cubic image block with neuron cell bodies marked, and (b) is a single-layer image along the Z-axis direction in a cubic image block with background information marked;

[0042] Figure 4 3D images of the whole brain after preprocessing and the whole brain image after neuron cell body identification according to an embodiment of the present invention; (a) is the 3D image of the whole brain after preprocessing, and (b) is the 3D image of the whole brain after neuron cell body identification;

[0043] Figure 5 is a new registration atlas for distinguishing the left brain from the right brain according to an embodiment of the present invention;

[0044] Figure 6 The three-dimensional visualization of the whole brain region and the half brain region after registration and the distribution of the whole brain cell bodies and the half brain cell bodies are shown in the figure; (a) is the three-dimensional visualization of the whole brain region, (b) is the three-dimensional visualization of the half brain region, (c) is the distribution of the whole brain cell bodies, and (d) is the distribution of the half brain cell bodies;

[0045] Figure 7 : is a comparison effect diagram of cell body recognition and counting between the embodiment of the present invention and the prior art; (a) is the original image of region A, (b) is the manual cell body recognition effect of region A, (c) is the IMARIS cell body recognition effect of region A, (d) is the recognition effect of region A according to the present invention, (e) is an enlarged view of the original image of region A, (f) is an enlarged view of the manual cell body recognition effect of region A, (g) is an enlarged view of the IMARIS cell body recognition effect of region A, (h) is an enlarged view of the recognition effect of region A according to the present invention, (i) is the original image of region B, (j) is the manual cell body recognition effect of region B, (k) is the IMARIS cell body recognition effect of region B, (l) is the recognition effect of region B according to the present invention, (m) is an enlarged view of the original image of region B, (n) is an enlarged view of the manual cell body recognition effect of region B, (o) is an enlarged view of the IMARIS cell body recognition effect of region B, and (p) is an enlarged view of the recognition effect of region B according to the present invention;

[0046] Figure 8 It is the F1 score of the embodiment of the present invention and the existing method for identifying the cell body. DETAILED DESCRIPTION

[0047] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0048] like Figure 1 This embodiment provides a method for identifying and counting whole-brain neuron cell bodies based on machine learning, specifically a method for identifying whole-brain neuron cell bodies and counting brain regions based on a fluorescence microscope optical sectioning tomography imaging system based on machine learning, specifically comprising the following steps:

[0049] S1. Preprocess the whole brain data;

[0050] In some possible implementations, preprocessing of whole-brain data involves integrating image data collected by a fluorescence microscope optical sectioning tomography system into whole-brain three-dimensional data suitable for subsequent steps. The preprocessing steps include sequentially performing jigsaw puzzles, stripe removal, downsampling, cropping, rotation and straightening, and denoising.

[0051] This embodiment analyzes the mouse brain, and can identify and count other neuronal cell bodies throughout the brain as needed.

[0052] In some specific embodiments, jigsaw puzzle is to stitch the images of various fields of view collected by the fluorescence microscope optical sectioning tomography system into a complete whole-brain size image; stripe removal is to remove the stripe noise in the image collected by the fluorescence microscope optical sectioning tomography system by using a denoising algorithm; cropping and clipping is to extract the image of the effective mouse brain part in the whole-brain three-dimensional image by cropping and clipping, and remove the irrelevant background part, so as to reduce the amount of data; rotation and correction is to rotate the whole-brain three-dimensional image by a certain angle to obtain an image with a correct angle; denoising is to process and clean the image by some image denoising algorithms to improve the signal-to-noise ratio and image quality of the image, so as to further improve the accuracy of the subsequent registration step and cell body recognition step.

[0053] Example 1

[0054] In some specific embodiments, the whole-brain data is downsampled because the original resolution of the image collected by the fluorescence microscope optical sectioning tomography system is 1.3um×1.3um×0.92um, and a whole-brain image of a mouse is about 2T (16-bit image). Generally speaking, the larger the amount of data, the more time-consuming and labor-intensive the subsequent data processing and analysis will be. Considering that the size of a neuron cell body in the mouse brain is 10-50um, the whole-brain data is downsampled by comprehensively considering the data volume and the needs of cell body identification. Specifically, the x-axis and y-axis are downsampled by two times, and the z-axis is downsampled by three times, that is, the original 1.3um×1.3um×0.92um resolution image is downsampled to 2.6um×2.6um×2.76um resolution image. After downsampling, the amount of whole-brain data is controlled at about 150G, and after subsequent cropping and clipping operations, the amount of whole-brain data will be controlled within 100G, which will save time in subsequent cell body identification and counting operations.

[0055] Example 2

[0056] The original 1.3um×1.3um×0.92um resolution image is downsampled to a 3.6um×3.6um×3.76um resolution image.

[0057] Example 3

[0058] The original 1.3um×1.3um×0.92um resolution image is downsampled to a 1.6um×1.6um×1.76um resolution image.

[0059] S2, intercepting multiple 100×100×100 pixel cubic image blocks in the cerebral cortex, striatum, cerebellum and pure background area of ​​multiple mouse brains (whole brain image data), and marking the neuron cell bodies in these cubic image blocks;

[0060] Example 4

[0061] A plurality of 100×100×100 pixel cubic image blocks were intercepted from the cerebral cortex, striatum, cerebellum and pure background area of ​​multiple mouse brains, including only the background but not other information; specifically, a total of 52 100×100×100 pixel cubic image blocks were intercepted from the cerebral cortex, striatum, cerebellum and pure background area of ​​4 mouse brains (since the image resolution is 2.6×2.6×2.76um, the size of each cubic image block is 260×260×276um), as shown in FIG. Figure 2 (a) to (c), four cubic image blocks were captured from each of the three regions of the cerebral cortex, striatum, and cerebellum of each mouse brain, and one cubic image block was captured from the pure background region of each mouse brain, for a total of 52 cubic image blocks.

[0062] In order to ensure the accuracy of recognition, this embodiment captures at least 4 cubic image blocks from each area of ​​each mouse brain, and at least 1 cubic image block from the pure background area of ​​each mouse brain. In addition, at least four mouse brains are selected as samples for capture.

[0063] In some possible implementations, manual marking of neuron cell bodies in the intercepted cubic image blocks is to distinguish and mark the foreground targets (neuron cell bodies) and background (non-neuron cell body parts) of the cubic image blocks. Neuron cell bodies in cubic image blocks can be directly marked manually. As for the marking of the background part, since the amount of information of the background part in the cubic image block is large, if all background information is completely marked, it will lead to too much marked data volume, affecting the algorithm performance. Therefore, in this implementation, the distance conversion algorithm and the threshold filter method are used to extract part of the background information for marking. Specifically, firstly, the distance conversion algorithm is used to perform distance conversion processing on the image after the neuron cell bodies are marked, so that the shortest distance of each pixel point from all the surrounding neuron cell bodies in three dimensions can be obtained. The shortest distance is used as the brightness value of each pixel point, and then according to the data obtained after the distance conversion, multiple thresholds are set to obtain a threshold filter. Two different threshold filters are applied to the image to obtain filamentous background information. Through this process, each cubic image block can finally obtain two new cubic image blocks, such as Figure 3 (a) to (b) are cubic image blocks containing foreground targets, i.e., neuron cell bodies, and cubic image blocks containing background information. The grayscale values ​​of the pixels where the foreground targets and background information are located are both 1, and the grayscale values ​​of the remaining pixels are 0.

[0064] S3, inputting the above-mentioned cube image blocks and corresponding labels into the algorithm as training data sets to train the classifier model;

[0065] In some possible implementations, the algorithm is a pixel classification algorithm. The pixel classification algorithm uses a random forest model in machine learning to independently classify each pixel into a predefined class. Random forest is an integrated learning method for solving classification and regression problems. It is based on a decision tree algorithm and improves the performance and robustness of the model by building multiple decision trees and combining them.

[0066] In some specific implementations, S2 has manually marked the neuron cell bodies in all intercepted cubic image blocks, and obtained cubic image blocks with marked cell bodies and cubic image blocks with marked background information, respectively. Next, the cubic image blocks and the corresponding labeled cubic image blocks are used as image data sets to train the random forest model, which means that each cubic image block and its corresponding label (foreground and background) are provided as input data to the model, and the model will learn how to distinguish foreground targets and backgrounds based on input features. The performance of the model is evaluated using the three evaluation indicators of accuracy, recall rate, and F1 score, and the model is optimized based on the evaluation results, including adjusting the parameters of the random forest, such as the number of decision trees, the number of features considered when splitting nodes, etc., and finally a satisfactory classifier model is trained.

[0067] S4. Use the trained classifier model to identify cell bodies in the whole brain;

[0068] After the training is completed, the trained machine learning classifier model is used to identify neuron cell bodies in the whole brain. The input data is the preprocessed three-dimensional image of the whole brain, and the output data is the three-dimensional image of the whole brain in which the foreground target and the background are distinguished and marked with different gray values. For example, all pixels of the foreground target (neuron cell body) are marked with a gray value of 1, and all pixels of the background are marked with a gray value of 0. Figure 4 (a), (b).

[0069] S5. Align the left and right brain regions of the whole brain;

[0070] In order to obtain accurate statistics of the number of brain cell bodies in the whole brain, it is necessary to finely segment the spatial location information of the brain region, that is, to complete the whole brain region registration operation. In addition, in order to obtain the statistics of the number of brain cell bodies in the left and right brain regions respectively to analyze the differences between the left and right brains, a new registration atlas that distinguishes the left and right brains is generated to perform statistics on the identification of brain cell bodies in the left and right brain regions respectively.

[0071] In some specific implementations, all of the more than one thousand brain regions included in the brain atlas are merged into sub-brain regions to form 244 brain regions, and then the new atlas is segmented into left and right brains, and finally a new registration atlas template is generated that distinguishes the left brain from the right brain and each of which contains 244 brain regions. The brain atlas selected in this implementation covers more than one thousand brain regions, and the specific number depends on the type of atlas selected; Figure 5 shown.

[0072] In the registration process, the BIRDS plug-in in the Fiji image processing software is used to perform the whole-brain brain region registration. The newly generated registration atlas template is applied to perform three-dimensional registration of the whole-brain image to achieve fine segmentation of the whole-brain brain region. The visualization and data statistics of the whole-brain brain region are realized by connecting to the IMARIS software. Through data statistics, the information of the 244 brain regions of the left and right brain after registration can be obtained, including the volume size, brightness, etc. of each brain region. As shown in the attached figure Figure 6 (a) and (b) in the figure show the three-dimensional visualization effects of the whole brain area and half brain area after registration, respectively, where each brain area is marked with a unique color.

[0073] S6. Count the cell bodies in different brain regions based on the results of whole-brain region registration.

[0074] After obtaining the cell body identification results in the whole brain and the accurate registration results of each brain region, the number of cell bodies in each brain region can be statistically analyzed based on the information of these cell bodies and the brain region where they are located. In some specific embodiments, the algorithm is written in matlab or python programming mode, and the brain region to which each cell body belongs is determined based on the position information of each cell body and the gray value information of each brain region, so as to obtain the number of cell bodies in each brain region of the whole brain, wherein the left and right brain regions are counted separately.

[0075] In some possible implementations, the IMARIS commercial software is used to count the number of cell bodies in each brain region based on the location information of each cell body and the gray value information of each brain region, and to present the results visually. Figure 6 (c) and (d) show the distribution of cell bodies in the whole brain and half of the brain, where the cell bodies in each brain region are identified and distinguished using different colors.

[0076] In some possible implementations, the above method is implemented by a whole-brain neuron cell body recognition and counting system based on machine learning. Specifically, the system includes a preprocessing module, a recognition module, a partition registration module, and a partition statistics module.

[0077] In some specific embodiments, the main task of the preprocessing module is to perform a series of preprocessing operations on the whole brain three-dimensional images of multiple mouse brains obtained from the FMOST device. These operations include but are not limited to splicing images to form a complete whole brain image, removing streak noise, cropping and cutting out processing to extract effective image parts, rotating and straightening images to obtain the correct angle, and applying a denoising algorithm to further clean up the image. In addition, the preprocessing module may also include downsampling the image after stripe removal to achieve a resolution suitable for subsequent processing.

[0078] The recognition module receives the preprocessed 3D whole-brain image as input and uses a pre-trained classifier model to identify the neuron soma in the image. This model is trained by intercepting multiple cubic image blocks from the preprocessed image and marking the neuron soma in these image blocks. During the recognition process, the classifier marks the neuron soma as the foreground target and assigns a specific grayscale value, while the background is assigned a different grayscale value for subsequent distinction and statistics.

[0079] The partition registration module performs partition registration on the whole brain 3D image based on the brain atlas. Specifically, it will segment the whole brain image into left and right brain, merge all brain regions in the brain atlas into sub-brain regions, and finally generate a new registration atlas template that distinguishes the left and right brain. Use specific image processing software plug-ins, such as the BIRDS plug-in in Fiji, to perform 3D registration on the whole brain image to achieve fine segmentation of brain regions.

[0080] The regional statistics module counts the cell bodies in different brain regions based on the results of whole-brain region registration and whole-brain cell body identification. By analyzing the location information of each cell body and the gray value information of each brain region, the specific brain region to which each cell body belongs is determined, and the number of cell bodies in each brain region is counted.

[0081] Through the collaborative work of the above four modules, the present invention realizes the effective identification and counting of neuronal cell bodies in the whole brain, providing important technical support for a deep understanding of the structure and function of the brain.

[0082] In order to test the accuracy of the whole-brain neuron cell body recognition method using machine learning, multiple cubic image blocks were intercepted from the cerebral cortex, striatum and cerebellum regions of the preprocessed mouse brain data, and then the cell body recognition method of machine learning in the present invention and the traditional cell body recognition method of IMARIS software were used to recognize and count the cell bodies of these cubic image blocks. The recognition results were then compared with the manual counting results to calculate the accuracy, recall rate and F1 score of cell body recognition, thereby evaluating the effect of cell body recognition. Figure 7The figure shows the comparison effects of cubic image blocks and cell body identification and counting using the manual cell body counting method, the machine learning cell body recognition method, and the traditional cell body recognition method of IMARIS software, wherein (a) to (h) are the comparison effects of cubic image blocks randomly selected from the striatum area of ​​the brain after being subjected to the manual cell body counting method, the machine learning cell body recognition method, and the traditional cell body recognition method of IMARIS software for cell body identification and counting, (i) to (p) are the comparison effects of cubic image blocks randomly selected from the cerebellum area after being subjected to the manual cell body counting method, the machine learning cell body recognition method, and the traditional cell body recognition method of IMARIS software for cell body identification and counting. The various colored balls in the figure represent the identified cell bodies, and the arrows indicate the recognition effects on certain cell bodies in each case. It can be seen from Figures (a) to (p) that the cell bodies identified by the cell body recognition method based on machine learning in the present invention are closer to the cell bodies manually identified, while the cell bodies identified by the traditional cell body recognition method using IMARIS software will miss some cell bodies. These missed cell bodies are actually cell bodies with weaker signals. This situation is more significant in the cerebellum region because the cell bodies in the cerebellum region are very dense and many of them are cell bodies with weaker signals.

[0083] This embodiment also provides an electronic device, which uses the above-mentioned whole-brain neuron cell body identification and counting method based on machine learning to implement the above-mentioned whole-brain neuron cell body identification and counting method based on machine learning. A computer storage medium, wherein the storage medium stores at least one program instruction, and the at least one program instruction is loaded and executed by a processor to implement the above-mentioned whole-brain neuron cell body identification and counting method based on machine learning.

[0084] By statistically evaluating the cell body recognition of twelve cubic image blocks in the cerebrum, brainstem and cerebellum regions of four mouse brain samples (see Table 1), we compared the performance of the cell body recognition method based on machine learning proposed in the present invention with the traditional cell body recognition method of IMARIS software in terms of accuracy, recall rate and F1 score. The results showed that in all the tested image blocks, the accuracy and recall rate of the method proposed in the present invention were both maintained at a high level of more than 0.85, and the F1 score, which is a comprehensive indicator for measuring the accuracy of the model, was always maintained at a high level of more than 0.88, and was always better than the performance of IMARIS software (see Table 1). Figure 8 ), which shows that the machine learning cell body recognition method of the present invention can significantly improve the accuracy of cell body recognition, which is of great significance for the statistics of the number of cell bodies in the whole brain and their distribution analysis in different brain regions. It is worth noting that in the cerebellum region (i.e., blocks 9 to 12), the method of the present invention shows significant advantages over the IMARIS software in terms of accuracy, recall rate and F1 score, and its maximum value can be higher by about 0.5 or more.

[0085] Table 1 Precision, recall and F1 score of two cell body recognition methods

[0086]

[0087]

[0088]

[0089] In summary, the present invention provides a complete set of fMOST (fluorescence microscopy optical sectioning tomography technology) whole-brain data processing and analysis procedures, covering key steps such as whole-brain data preprocessing, whole-brain neuronal cell body identification and counting, whole-brain brain region three-dimensional registration, and whole-brain cell body distribution statistics in various brain regions. In particular, the present invention proposes a fMOST whole-brain neuronal cell body identification and sub-brain region counting method based on machine learning, which uses a machine learning pixel classifier algorithm to automatically learn multiple features of neuronal cell bodies, thereby achieving accurate identification of neuronal cell bodies. Compared with the traditional IMARIS whole-brain cell body identification method, the method of the present invention can significantly improve the accuracy of cell body identification and provide more accurate and reliable data support for subsequent research work.

[0090] Each embodiment in this specification is described in a related manner, and the same or similar parts between the embodiments can be referred to each other, and each embodiment focuses on the differences from other embodiments. In particular, for the system embodiment, since it is basically similar to the method embodiment, the description is relatively simple, and the relevant parts can be referred to the partial description of the method embodiment.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the protection scope of the present invention. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention are included in the protection scope of the present invention.

Claims

1. A method for identifying and counting whole-brain neuron cell bodies based on machine learning, characterized in that: The specific steps are as follows: Preprocessing of the whole-brain 3D images of multiple rats acquired from the FMOST device; A plurality of cubic image blocks are intercepted from a plurality of different regions of the preprocessed three-dimensional whole-brain images of a plurality of mouse brains, and the neuronal cell bodies in the intercepted cubic image blocks are marked, and then a classifier model is trained using the marked cubic image blocks, and then the trained classifier model is used to perform neuronal cell body recognition on the preprocessed three-dimensional whole-brain images; Register the whole brain 3D image to distinguish the left and right brain regions; The cell bodies of different brain regions were counted based on the results of whole-brain region registration and whole-brain neuron cell body identification.

2. A whole-brain neuron cell body recognition and counting method based on machine learning according to claim 1, characterized in that: The pre-processing includes the following steps: Jigsaw puzzle: stitch the collected images of different fields of view into a complete whole-brain image; De-striping: using denoising algorithm to remove the stripe noise in the image collected by FMOST equipment; Cropping and cutting out: extracting the effective part of the whole brain three-dimensional image through cropping and cutting out; Rotation alignment: rotate the whole brain three-dimensional image at a certain angle to obtain an image with an aligned angle; Denoising, the image is processed and cleaned through the denoising algorithm.

3. A whole-brain neuron cell body identification and counting method based on machine learning according to claim 2, characterized in that: Before the cropping and cutting, the whole brain image data after the stripe removal operation is downsampled, specifically: the original 1.3×1.3×0.92 μm resolution image is downsampled to a (1.6-3.6)×(1.6-3.6)×(1.76-3.76) μm resolution image.

4. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 2 is characterized in that: Before the cropping and cutting, the whole brain image data after the stripe removal operation is downsampled, specifically: the original 1.3×1.3×0.92 μm resolution image is downsampled to a 2.6×2.6×2.76 μm resolution image.

5. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 1, characterized in that: The method of intercepting multiple cubic image blocks from multiple different areas of the preprocessed three-dimensional images of the whole brains of multiple rat brains is specifically as follows: cubic image blocks with a size of 100×100×100 pixels are intercepted from the cerebral cortex, striatum, cerebellum and pure background areas of the preprocessed three-dimensional image data of the whole brains of at least 4 rat brains, wherein at least 4 cubic image blocks are intercepted from each of the three areas of the cerebral cortex, striatum and cerebellum of each rat brain, and at least 1 cubic image block is intercepted from the pure background area of ​​each rat brain.

6. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 1, characterized in that: The step of marking the neuron cell bodies in the intercepted cubic image block comprises: The foreground target and background of the cubic image block are distinguished and marked, and the neuron cell bodies in the cubic image block are manually marked. For the marking of the background, the distance transformation algorithm and the threshold filter method are used to extract part of the background information for marking. After each cubic image block is marked, two new cubic image blocks are obtained, one containing the foreground target, i.e., the neuron cell body, and the other containing the background information. The grayscale values ​​of the pixels where the foreground target and the background information are located are both 1, and the grayscale values ​​of the remaining pixels are 0.

7. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 6 is characterized in that: The specific steps of extracting background information using the distance conversion algorithm and the threshold filter are as follows: first, the distance conversion algorithm is used to perform distance conversion processing on the image after the neuron cell bodies are marked, and the shortest distance between each pixel point and all the surrounding neuron cell bodies in three dimensions is obtained, and the shortest distance is used as the brightness value of each pixel point. Then, according to the data obtained after the distance conversion, multiple thresholds are set to obtain a threshold filter, and two different threshold filters are applied to the image to obtain filamentous background information.

8. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 1, characterized in that: The method of training the classifier model using the labeled cube image blocks comprises: The classifier model is trained using a pixel classification algorithm, and each pixel is independently classified into a predefined class using a random forest model in machine learning; The intercepted cube image blocks and the corresponding annotated cube image blocks are used as image data sets to train the random forest model. Each cube image block and its corresponding label are provided as input data to the model. The labels are foreground objects and backgrounds. The model learns how to distinguish foreground objects from backgrounds based on the input features. The performance of the model is evaluated using three evaluation indicators: accuracy, recall, and F1 score. The model parameters are optimized based on the evaluation results to finally obtain a classifier model that meets the requirements.

9. The whole-brain neuron cell body recognition and counting method based on machine learning according to claim 1, characterized in that: The method of using the trained classifier model to perform neuron cell body recognition on the preprocessed whole-brain three-dimensional image includes: inputting the preprocessed whole-brain three-dimensional image data into the trained classifier model, and the classifier model outputting the whole-brain three-dimensional image in which the foreground target and the background are distinguished and marked with different grayscale values, wherein the classifier model marks all pixel points of the neuron cell body as a grayscale value of 1, and marks all pixel points of the background as a grayscale value of 0.

10. The whole-brain neuron cell body identification and counting method based on machine learning according to claim 1, characterized in that: The brain region registration for distinguishing the left and right brains of the whole brain three-dimensional image includes: All brain regions in the brain atlas are merged into sub-regions to form 244 brain regions. Then the new atlas is segmented into left and right brains, and finally a new registration atlas template is generated that distinguishes the left and right brains and contains 244 brain regions in each brain. During the registration, the BIRDS plug-in in the Fiji image processing software was used to perform whole-brain brain region registration, and the newly generated registration atlas template was used to perform three-dimensional registration of the whole-brain image to achieve fine segmentation of the whole-brain brain regions. Finally, the IMARIS software was used to realize the visualization and data statistics of the whole-brain brain regions, and the information of 244 brain regions in the left and right brains after registration was obtained through data statistics.

11. The whole-brain neuron cell body recognition and counting method based on machine learning according to claim 1, characterized in that: The cell body counting of each brain region according to the results of whole-brain brain region registration and whole-brain neuron cell body identification includes: determining which brain region each cell body belongs to according to the position information of each cell body and the gray value information of each brain region, thereby obtaining the number of cell bodies in each brain region of the whole brain.

12. An electronic device, characterized in that: Whole-brain neuron cell body recognition and counting is achieved by using a whole-brain neuron cell body recognition and counting method based on machine learning as described in any one of claims 1 to 11.

13. A computer storage medium, characterized in that: The storage medium stores at least one program instruction, and the at least one program instruction is loaded and executed by the processor to implement a whole-brain neuron cell body identification and counting method based on machine learning as described in any one of claims 1-11.

Images