Interactive rapid cleaning method for whole-brain neuron three-dimensional image
By using a two-dimensional neuron image segmentation and classification network, combined with deep learning and manual proofreading, the problems of accuracy and completeness in the three-dimensional image classification of whole-brain neurons were solved, achieving efficient data cleaning and neuron reconstruction.
Patent Information
- Application Number
- CN202511609601.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-05
- Publication Date
- 2026-02-06
AI Technical Summary
Existing technologies for classifying three-dimensional images of whole-brain neurons suffer from insufficient classification accuracy, incomplete data, and excessive manual proofreading, especially in handling neuronal data effectively in complex contexts.
A two-dimensional neuron image segmentation and classification method is adopted, which combines deep learning networks and manual proofreading. The trained segmentation and classification networks predict and classify neuron data, and FIJI software is used for manual proofreading and redundant data removal to generate a cleaned dataset.
It improves the accuracy and completeness of neuron data classification, reduces the workload of manual proofreading, and enhances the efficiency and reliability of neuron reconstruction.
Smart Images
Figure CN121483510A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing, and more particularly to data cleaning of biomedical images. Background Technology
[0002] A deep understanding of the brain's fundamental functions is the prerequisite and foundation for effective treatment of brain diseases. At the neuronal level, research into key biological processes such as synaptic transmission mechanisms, axonal transport processes, and intracellular signal transduction not only helps to reveal the patterns of brain disease occurrence and development but also provides a theoretical basis for scientifically formulating intervention strategies. With the continuous development of high-resolution whole-brain imaging technologies (such as fluorescence microscopy with light section tomography (fMOST) and time-delayed integral imaging (TDI-fMOST), researchers can obtain three-dimensional image data of neurons covering the entire brain region. This type of data is of great value in analyzing synaptic transmission, axonal transport, intracellular signal transduction, and neural circuit connection patterns, and is an important foundation for a deeper understanding of brain function and the exploration of brain disease mechanisms.
[0003] In existing research, whole-brain imaging data is typically classified and screened to initially distinguish between data containing and without fiber signals, in order to facilitate subsequent neuron tracking and morphological reconstruction. However, these methods often rely on the linear or local intensity characteristics of image signals, and are limited by the complexity of the image background and the level of signal noise, resulting in significant uncertainty in the classification results.
[0004] The existing technology has the following problems and drawbacks:
[0005] 1. Complex image backgrounds reduce classification accuracy.
[0006] In the process of brain data labeling and imaging, images may have problems such as uneven signal, strong noise, or local signal loss. Existing technologies rely on the linear characteristics of fiber signals for classification. This complex background makes the classification results of three-dimensional images inaccurate.
[0007] 2. The classification results of neuron data lack completeness.
[0008] There are strong relationships between neuron data. If neuron data is missing at a certain point, it will lead to the loss of subsequent neuron reconstruction results. Existing classification methods cannot guarantee the completeness of neuron data and are prone to losing neuron data, which will affect the accuracy of subsequent neuron reconstruction.
[0009] 3. Manually calibrating neuronal data is extremely complex and time-consuming.
[0010] To ensure the integrity of neuron data, manual verification of the classified neuron data is required. However, the total number of neuron data sets is in the hundreds of thousands, and the workload of manually screening and identifying whether there are fiber signal data is enormous. Therefore, there is an urgent need for a boundary-effective neuron verification tool to address this problem. Summary of the Invention
[0011] Based on the aforementioned deficiencies in the existing technology, this invention provides an interactive and rapid cleaning method for three-dimensional images of whole-brain neurons. This method can achieve high-accuracy neuronal data classification in complex backgrounds, ensure data integrity, and reduce the workload of manual proofreading, thereby improving the efficiency and reliability of large-scale neuronal reconstruction.
[0012] To achieve the above objectives, the present invention provides an interactive and rapid cleaning method for three-dimensional images of whole-brain neurons, comprising the following steps:
[0013] Two-dimensional neuron image segmentation involves inputting the two-dimensional projection map of coronal neurons into a trained segmentation network for prediction, resulting in a neuron segmentation map.
[0014] Neuron data classification involves inputting the neuron segmentation map into a trained classification network to classify the segmentation map, obtain the category of each data point, and determine the data containing fibers.
[0015] Manual verification involves drawing a grid on the projection map and combining it with the classification results. Locations that may contain signals are displayed based on correlation. Misidentified signal-containing images are manually supplemented using FIJI software.
[0016] Redundant data is removed by deleting invalid image blocks based on the index file, retaining only valid data containing neuron signals, and generating a cleaned dataset.
[0017] In one embodiment, the method for obtaining the two-dimensional projection map of the coronal neuron is as follows:
[0018] A coronal projection map is generated for the coronal plane of the neuron, and then the projection map is divided into small blocks to obtain a two-dimensional projection map of the coronal plane neuron.
[0019] In one embodiment, when the projection image is divided into small patches, the data block size of the small patch is 256×256 pixels. For images whose boundaries do not meet the 256-pixel size, zeros are padded to ensure the homogeneity of image size.
[0020] In one embodiment, when cropping data, 10 pixels of redundancy are retained to prevent obvious splicing marks.
[0021] In one embodiment, the training process of the segmentation network is as follows:
[0022] Data selection and cropping: Representative data blocks were selected from different locations in the whole brain and cropped to a size of 256×256×256 pixels for gold standard data preparation;
[0023] Two-dimensional data annotation first completes neuronal skeleton reconstruction through a semi-automatic method, then converts it into three-dimensional segmentation labels, and projects them into two-dimensional segmentation maps. The corresponding original three-dimensional images are projected into two-dimensional maps to form training pair datasets.
[0024] Hyperparameter optimization, with convolutional layers, pooling layers and upsampling layers as the core structure, gradually adjusts hyperparameters such as convolutional kernel size, learning rate, optimizer and number of layers to determine the optimal network structure;
[0025] Model training involves inputting the dataset into the network for training. The Adam optimizer is used with a learning rate of 0.0001, and the process is iterated 300 times to learn the mapping relationship between the two-dimensional image and the segmentation label, thereby obtaining a model that can accurately segment neuron signals.
[0026] In one embodiment, the training process of the classification network is as follows: a training dataset is established by segmenting small images and classification results, and then the classification network is trained based on these training data. The training is carried out for a total of 400 iterations, and each iteration completely traverses the entire training dataset.
[0027] In one embodiment, during manual calibration, grid lines are drawn at 256-pixel intervals across the entire projection image. Then, based on automatic classification, all signal-containing portions are reset to 0. Data potentially containing fibers is displayed based on fiber connectivity. Locations containing fiber signals, as determined by automatic classification, are not displayed. Incorrectly identified locations are marked using a multi-point marking function. Finally, FIJI's marker list generation function saves all marker points (X, Y), and index values for both directions are generated according to the following formula:
[0028] x ind ,y ind These are the index values for the x and y directions;
[0029] Based on the existing fiber signal index file, the data containing the signal is segmented and extracted from the coronal plane.
[0030] After manually marking one round, click "Next Round" to observe whether the data in the next connection contains fiber signals based on connectivity, until no fiber signal data is added.
[0031] In one embodiment, the method for clearing redundant data is as follows:
[0032] After obtaining all images containing neurons, the coordinates of each image are calculated and saved to a txt file. Image blocks not in the txt file are deleted by using C++ code to find and delete the corresponding files. Data containing neurons are retained, and finally, redundant neuron image data is obtained.
[0033] The interactive and rapid cleaning method for three-dimensional images of whole-brain neurons described in this invention can solve the following technical problems:
[0034] 1. Insufficient classification accuracy: During brain data labeling and imaging, images are often affected by factors such as signal inhomogeneity, noise interference, or local signal loss. Existing classification methods rely on the linear characteristics of fiber signals, making it difficult to effectively handle complex backgrounds, thus leading to inaccurate 3D image classification results.
[0035] 2. Insufficient Data Completeness: Neuron data is highly interconnected; the absence of any partial neuron data will lead to incomplete neuron reconstruction results. However, existing classification methods struggle to guarantee the completeness of neuron data, easily resulting in the loss of valid data, thereby reducing the accuracy and reliability of reconstruction.
[0036] 3. Excessive workload for manual verification: To ensure the integrity of neuronal data, it is usually necessary to manually verify the classification results one by one. However, in whole-brain imaging data, the amount of neuronal data can easily reach hundreds of thousands. The process of manually screening and identifying fiber signals is not only tedious but also extremely time-consuming, which seriously restricts the processing and application of large-scale neuronal data.
[0037] The interactive rapid cleaning method for three-dimensional images of whole-brain neurons described in this invention has the following technical effects:
[0038] 1. To address the problem of insufficient classification accuracy caused by interference from complex neuronal signals and background signals, this invention proposes a classification method based on two-dimensional image segmentation. This method transforms a three-dimensional classification task into a two-dimensional segmentation task, which not only significantly shortens data processing time but also effectively reduces the interference of background signals on the classification results of neuronal data.
[0039] 2. To address the issues of incomplete neuronal data classification results and the complexity and time-consuming nature of manual verification, this invention proposes a manual verification method based on the connectivity of neuronal data. This method displays only the image data surrounding the fibers, allowing operators to process only a small amount of key data during verification, thus significantly reducing the workload and time cost of manual verification. Furthermore, through manual verification, incorrectly identified neuronal signals are added to the dataset, ensuring the completeness of the neuronal data. Attached Figure Description
[0040] The accompanying drawings described herein are for illustrative purposes only and are not intended to limit the scope of the invention in any way. Furthermore, the shapes and proportions of the components in the drawings are merely illustrative to aid in understanding the invention and do not specifically limit the shapes and proportions of the components. Those skilled in the art, guided by the teachings of this invention, can select various possible shapes and proportions to implement the invention according to specific circumstances.
[0041] Figure 1 A flowchart illustrating an interactive and rapid cleaning method for three-dimensional images of whole-brain neurons provided by this invention;
[0042] Figure 2 This is a schematic diagram of the process for creating segmented network tags in a specific embodiment of the present invention. Detailed Implementation
[0043] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of the embodiments of this invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this invention, and not all embodiments. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of this invention.
[0044] Please see Figure 1 As shown, the first embodiment of the present invention provides an interactive and rapid cleaning method for three-dimensional images of whole-brain neurons, comprising the following steps:
[0045] Two-dimensional neuron image segmentation involves inputting the two-dimensional projection map of coronal neurons into a trained segmentation network for prediction, resulting in a neuron segmentation map.
[0046] Neuron data classification involves inputting the neuron segmentation map into a trained classification network to classify the segmentation map, obtain the category of each data point, and determine the data containing fibers.
[0047] Manual verification involves drawing a grid on the projection map and combining it with the classification results. Locations that may contain signals are displayed based on correlation. Misidentified signal-containing images are manually supplemented using FIJI software.
[0048] Redundant data is removed by deleting invalid image blocks based on the index file, retaining only valid data containing neuron signals, and generating a cleaned dataset.
[0049] In one embodiment, the method for obtaining the two-dimensional projection map of the coronal neuron is as follows:
[0050] A coronal projection map is generated for the coronal plane of the neuron, and then the projection map is divided into small blocks to obtain a two-dimensional projection map of the coronal plane neuron.
[0051] In one embodiment, when the projection image is divided into small patches, the data block size of the small patch is 256×256 pixels. For images whose boundaries do not meet the 256-pixel size, zeros are padded to ensure the homogeneity of image size.
[0052] In one embodiment, when cropping data, 10 pixels of redundancy are retained to prevent obvious splicing marks.
[0053] The large image is split, and the coronal projection image (approximately 30,000×20,000 pixels) is cropped into a smaller 2D image. The insufficient boundary is filled with 0, and 10 pixels of redundancy are reserved to avoid stitching marks.
[0054] Segmentation prediction uses a trained network to predict the segmentation of a small image pixel by pixel, resulting in a corresponding two-dimensional segmentation result.
[0055] In one embodiment, the training process of the segmentation network is as follows:
[0056] Data selection and cropping: Representative data blocks were selected from different locations in the whole brain and cropped to a size of 256×256×256 pixels for gold standard data preparation;
[0057] Two-dimensional data annotation first completes neuronal skeleton reconstruction through a semi-automatic method, then converts it into three-dimensional segmentation labels, and projects them into two-dimensional segmentation maps. The corresponding original three-dimensional images are projected into two-dimensional maps to form training pair datasets.
[0058] Hyperparameter optimization, with convolutional layers, pooling layers and upsampling layers as the core structure, gradually adjusts hyperparameters such as convolutional kernel size, learning rate, optimizer and number of layers to determine the optimal network structure;
[0059] Model training involves inputting the dataset into the network for training. The Adam optimizer is used with a learning rate of 0.0001, and the process is iterated 300 times to learn the mapping relationship between the two-dimensional image and the segmentation label, thereby obtaining a model that can accurately segment neuron signals.
[0060] The training dataset generation primarily addresses the challenges and time-consuming nature of label creation. It involves the following two steps:
[0061] Data selection and cropping: Representative data blocks were selected from different locations throughout the brain and cropped to a size of 256×256×256 pixels for use in the preparation of gold standard data.
[0062] Two-dimensional data annotation is first performed by semi-automatically reconstructing the neuronal skeleton, then converting it into three-dimensional segmentation labels, and projecting it into a two-dimensional segmentation image B; the corresponding original three-dimensional image is projected into a two-dimensional image A, forming training pairs.
[0063] Deep learning network training primarily addresses the optimization of model parameters and the selection of optimal hyperparameters. It mainly involves two steps:
[0064] Hyperparameter optimization, with convolutional layers, pooling layers and upsampling layers as the core structure, gradually adjusts hyperparameters such as convolutional kernel size, learning rate, optimizer and number of layers to finally determine the optimal network structure.
[0065] For model training, datasets A and B are input into the network for training. The Adam optimizer (learning rate 0.0001) is used for 300 iterations to learn the mapping relationship between two-dimensional images and segmentation labels, thereby obtaining a model that can accurately segment neuron signals.
[0066] In one embodiment, the training process of the classification network is as follows: a training dataset is established by segmenting small images and classification results, and then the classification network is trained based on these training data. The training is carried out for a total of 400 iterations, and each iteration completely traverses the entire training dataset.
[0067] A classification network is trained using segmented images and classification results to classify segmented data into images containing fiber signals and images without fiber signals.
[0068] In one embodiment, during manual calibration, grid lines are drawn at 256-pixel intervals across the entire projection image. Then, based on automatic classification, all signal-containing portions are reset to 0. Data potentially containing fibers is displayed based on fiber connectivity. Locations containing fiber signals, as determined by automatic classification, are not displayed. Incorrectly identified locations are marked using a multi-point marking function. Finally, FIJI's marker list generation function saves all marker points (X, Y), and index values for both directions are generated according to the following formula:
[0069] x ind ,y ind These are the index values for the x and y directions;
[0070] Based on the existing fiber signal index file, the data containing the signal is segmented and extracted from the coronal plane.
[0071] After manually marking one round, click "Next Round" to observe whether the data in the next connection contains fiber signals based on connectivity, until no fiber signal data is added.
[0072] Manual correction involves drawing a grid on the projection map and combining it with the automatic classification results. The locations that may contain signals are displayed through correlation. The FIJI software is used to manually supplement the misidentified signal-containing images, which greatly reduces the amount of data that needs to be identified during manual correction. This achieves automated and manual collaborative data cleaning, balancing data cleaning efficiency and accuracy.
[0073] In one embodiment, the method for clearing redundant data is as follows:
[0074] After obtaining all images containing neurons, the coordinates of each image are calculated and saved to a txt file. Image blocks not in the txt file are deleted by using C++ code to find and delete the corresponding files. Data containing neurons are retained, and finally, redundant neuron image data is obtained.
[0075] Data cleaning involves removing invalid image blocks from the index file and retaining only valid data containing neuronal signals to generate a cleaned dataset to support subsequent reconstruction and analysis.
[0076] Please refer to the following: Figure 2 The illustration shows a specific embodiment of an interactive and rapid cleaning method for three-dimensional images of whole-brain neurons according to the present invention, to facilitate the reader's better understanding of the invention.
[0077] S1, Training Dataset Generation Module
[0078] To address the complexity of neuron data and the difficulty of obtaining labeled data from two-dimensional data, a skeleton is obtained through three-dimensional reconstruction, and then a two-dimensional training dataset is generated. The steps include:
[0079] S11, Data Selection and Data Pruning Module
[0080] The coronal data was projected along the z-axis, converting 256 images into a single image. Based on the projection image and different data characteristics, 1000 two-dimensional images were selected and their locations were recorded. Then, the corresponding three-dimensional data was cropped from the whole-brain neuron dataset to obtain a 256×256×256 pixel data block, which was used for the subsequent preparation of the gold standard dataset.
[0081] S12, Two-dimensional dataset annotation preparation module
[0082] First, a combination of manual and semi-automatic methods was used to reconstruct the fiber skeleton. During reconstruction, efforts were made to ensure the fiber skeleton was located at the signal center to guarantee the accuracy of the subsequent training set. During fiber tracking, a combination of empirical judgment and software assistance was used to ensure relatively accurate reconstruction results even in weak signal regions and areas with irrelevant interference. After skeleton tracking was completed, an appropriate radius threshold was set, and a cylindrical structure was generated along the skeleton to obtain 3D binary neuron label data; this data was then projected into 2D training labels B. Finally, the original 3D image was projected into a 2D image to generate dataset A, as shown below. Figure 2 As shown.
[0083] S2, Deep Learning Network Training Module
[0084] Training a pre-built deep learning model using a pre-built training dataset involves the following two steps:
[0085] S21, Hyperparameter Optimization Module for Deep Learning Models
[0086] To effectively segment neuronal signals in 2D images and enable the designed deep learning network to learn image features more effectively during the segmentation process, a deep learning network structure consisting of convolutional layers, pooling layers, and upsampling layers was constructed. To ensure image segmentation accuracy, the network's hyperparameters were adjusted multiple times, with different numbers of max pooling operations, convolutional layers, and upsampling operations. The segmentation results were compared, and a deep learning network framework with 18 convolutional layers, 4 max pooling layers, and 4 upsampling layers was finally built. Pooling operations reduce the image size, while upsampling operations increase the image size. Simultaneously, images with the same scale as the max pooling layer (corresponding to the number of channels) are merged (i.e., stitched together). Upsampling ensures that the output image and input image have the same size.
[0087] S22, Parameter training module in deep learning models
[0088] The constructed datasets A and B are input into a deep learning network for training, enabling the network to segment neuron signals in a two-dimensional image.
[0089] The training process is a crucial step in deep learning models for feature learning and weight optimization. The model iteratively updates its parameters to gradually learn and extract fiber features from neuron images, achieving accurate neuron segmentation. Specifically, preprocessed data is first loaded into memory and then batch-fed into the network for training. Subsequently, the loss value is calculated based on the difference between the network's predictions and the labels, and the model parameters are continuously adjusted using an optimization function. This study employs the Adam optimizer with a learning rate of 0.0001 to effectively mitigate gradient explosion and vanishing problems. Training consists of 300 iterations, each iteration fully traversing the entire training set to ensure the model can fully learn the features of each sample. The experimental computing platform is configured with an NVIDIA Quadro RTX5000 GPU (32GB VRAM), dual Intel Xeon Gold6246R CPUs (3.39GHz), and 256GB of RAM.
[0090] S3, Two-dimensional coronal neuron projection map segmentation prediction module
[0091] The two-dimensional projection image of the coronal neurons is input into the trained network for prediction to obtain the neuron segmentation image. This involves the following two steps:
[0092] S31, split the coronal projection image into smaller image modules.
[0093] To input whole-brain neuron data into the network for prediction, the large two-dimensional projection image needs to be divided into 256×256 pixel data blocks. Since a coronal projection image can reach approximately 30,000 pixels in the x-direction and 20,000 pixels in the y-direction, images whose boundaries do not meet the 256-pixel size requirement are padded with zeros to ensure homogeneity of image size before being input into the network for prediction. Furthermore, during data cropping, a 10-pixel redundancy is retained to prevent noticeable stitching artifacts.
[0094] S32 uses a trained deep network model to make predictions and obtain a two-dimensional segmented image for each input small image.
[0095] After acquiring the small image obtained from S31, the small image is input into a deep learning network. Using the pre-trained network model parameters, pixel-by-pixel prediction is performed on the image to obtain the segmentation of each pixel position in the small image, resulting in the final neuron segmentation structure diagram, as shown below. Figure 2 As shown.
[0096] S4, a neuronal image segmentation and classification network
[0097] To accurately distinguish between images containing neuronal signals and those without, training data was built using the segmentation images obtained from S32 and the classification results. The classification network was then trained based on this data, undergoing 400 iterations, each iterating through the entire training set. This process established connections between the original image and the segmentation image, and between the segmentation image and the classification result. By using the segmentation network before classification, interference from background signals in the original image was eliminated during the classification process.
[0098] After training the network, a segmentation network is used to obtain segmentation maps for all data. Then, a classification network is used to classify the segmentation maps to obtain the category of each data point and determine the data containing fibers.
[0099] S5, Neuron Data Manual Correction Module
[0100] In manual correction, we draw grid lines at 256-pixel intervals across the entire projection image. Then, based on automatic classification, we reset all signal-containing areas to 0. Next, based on fiber connectivity, we display data that may contain fibers. This minimizes the number of data points identified as containing fiber signals. Locations with fiber signals identified by automatic classification are not displayed, thus clearly showing misidentified locations (identifying signal-containing areas as non-signal-containing areas). After locating these locations, we mark them using multi-point marking. Finally, we save all marked points (X, Y) using FIJI's marked point list generation function. Then, we generate the following formula...
[0101]
[0102] x ind ,y ind These are the index values for the x and y directions. This step supplements the existing fiber signal index file, with the ultimate goal of segmenting and extracting data containing signals from the coronal plane. Since the proportion of data containing fiber signals in the mouse whole-brain data is relatively small, and the number of data incorrectly identified as having no signal is further reduced after classification, the manpower required for the manual supplementation stage is very limited, fully meeting efficiency requirements. Furthermore, the small proportion of data with no signal that is misidentified as having signal is negligible and requires no additional manual removal, thus not affecting the accuracy of reconstruction and having a negligible impact on the data volume.
[0103] After manually marking one round, click "Next Round" to observe whether the data in the next connection contains fiber signals based on connectivity, until no fiber signal data is added.
[0104] S6, Neuron Data Cleaning Module
[0105] After obtaining all images containing neurons, we calculate the coordinates of each image and save them to a txt file. For image blocks not in the txt file, we delete them. We use C++ code to find the corresponding file and delete it. For data containing neurons, we retain it. Finally, we obtain cleaned and redundant neuron image data, which greatly alleviates the pressure of storing and computing large amounts of data.
[0106] It should be understood that the above description is for illustrative purposes and not for limitation. Many embodiments and applications beyond the provided examples will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of this teaching should not be determined by reference to the above description, but rather by reference to the foregoing claims and the full scope of their equivalents. For purposes of completeness, all articles and references, including patent applications and publications, are incorporated herein by reference. The omission of any aspect of the subject matter disclosed herein in the foregoing claims is not intended as a waiver of that subject matter, nor should it be construed as an indication that the applicant has not considered that subject matter as part of the disclosed application subject matter.
Claims
1. An interactive and rapid cleaning method for three-dimensional images of whole-brain neurons, characterized in that, Includes the following steps: Two-dimensional neuron image segmentation involves inputting the two-dimensional projection map of coronal neurons into a trained segmentation network for prediction, resulting in a neuron segmentation map. Neuron data classification involves inputting the neuron segmentation map into a trained classification network to classify the segmentation map, obtain the category of each data point, and determine the data containing fibers. Manual verification involves drawing a grid on the projection map and combining it with the classification results. Locations that may contain signals are displayed based on correlation. Misidentified signal-containing images are manually supplemented using FIJI software. Redundant data is removed by deleting invalid image blocks based on the index file, retaining only valid data containing neuron signals, and generating a cleaned dataset.
2. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 1, characterized in that, The method for obtaining the two-dimensional projection map of the coronal facial neurons is as follows: A coronal projection map is generated for the coronal plane of the neuron, and then the projection map is divided into small blocks to obtain a two-dimensional projection map of the coronal plane neuron.
3. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 2, characterized in that, When the projection image is divided into smaller images, the data block size of the smaller image is 256×256 pixels. For images whose boundaries do not meet the 256-pixel size, zeros are used to ensure the homogeneity of image size.
4. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 3, characterized in that, When cropping data, retain 10 pixels of redundancy to prevent obvious splicing marks.
5. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 1, characterized in that, The training process of the segmentation network is as follows: Data selection and cropping: Representative data blocks were selected from different locations in the whole brain and cropped to a size of 256×256×256 pixels for gold standard data preparation; Two-dimensional data annotation first completes neuronal skeleton reconstruction through a semi-automatic method, then converts it into three-dimensional segmentation labels, and projects them into two-dimensional segmentation maps. The corresponding original three-dimensional images are projected into two-dimensional maps to form training pair datasets. Hyperparameter optimization, with convolutional layers, pooling layers and upsampling layers as the core structure, gradually adjusts hyperparameters such as convolutional kernel size, learning rate, optimizer and number of layers to determine the optimal network structure; Model training involves inputting the dataset into the network for training. The Adam optimizer is used with a learning rate of 0.0001, and the process is iterated 300 times to learn the mapping relationship between the two-dimensional image and the segmentation label, thereby obtaining a model that can accurately segment neuron signals.
6. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 1, characterized in that, The training process of the classification network is as follows: a training dataset is established by segmenting small images and classification results, and then the classification network is trained based on these training data. The training is carried out for a total of 400 iterations, and each iteration completely traverses the entire training dataset.
7. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 1, characterized in that, In manual calibration, grid lines are drawn at 256-pixel intervals across the entire projection image. Then, based on automatic classification, all signal-containing areas are reset to 0. Data potentially containing fibers is displayed based on fiber connectivity. Locations containing fiber signals identified automatically are not displayed. Incorrectly identified locations are marked using a multi-point marking function. Finally, FIJI's marker list generation function saves all marker points (X, Y), and then index values for both directions are generated using the following formula: x ind ,y ind These are the index values for the x and y directions; Based on the existing fiber signal index file, the data containing the signal is segmented and extracted from the coronal plane. After manually marking one round, click "Next Round" to observe whether the data in the next connection contains fiber signals based on connectivity, until no fiber signal data is added.
8. The interactive rapid cleaning method for three-dimensional images of whole-brain neurons as described in claim 1, characterized in that, The methods for clearing redundant data are as follows: After obtaining all images containing neurons, the coordinates of each image are calculated and saved to a txt file. Image blocks not in the txt file are deleted by using C++ code to find and delete the corresponding files. Data containing neurons are retained, and finally, redundant neuron image data is obtained.