Neuron morphological analysis method based on graph neural network

Abstract

The invention discloses a neuron morphology analysis method based on a graph neural network, and the method comprises the following steps: 1, constructing a data set comprising a plurality of neuron morphologies, the data set comprising original image data or three-dimensional voxel data of neurons; 2, converting the neuron morphological data into graph structure data by using a deep learning method; 3, inputting the neuron map data generated in the step 2 into a depth map neural network for feature learning and classification, introducing an attention mechanism to enhance the learning ability of a model for key features, and further improving the classification accuracy; 4, after feature learning and classification are completed, performing intelligent retrieval by using the generated neuron form hash codes; the beneficial effects of the invention are that the neural morphology is deeply learned and classified through the depth map neural network, and the classification accuracy is improved; a neuron data set with high geometric morphology similarity is selected for training, and the generalization ability of the method is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of neuron morphology analysis, and particularly relates to a neuron morphology analysis method based on a graph neural network. Background Art

[0002] In recent years, the research enthusiasm for graph neural networks in the field of deep learning has been increasing day by day, and graph neural networks have become a research hotspot in major deep learning conferences; the excellent ability of graph neural networks to process unstructured data has led to new breakthroughs in network data analysis, recommendation systems, physical modeling, natural language processing, and combinatorial optimization problems on graphs.

[0003] A graph neural network is a general term for algorithms that use neural networks to learn graph-structured data, extract and discover features and patterns in graph-structured data to meet the requirements of graph learning tasks such as clustering, classification, prediction, segmentation, and generation; in graph data, the data not only includes independent samples, but also includes the dependence or adjacency relationship between nodes, and a graph neural network can effectively capture this relationship and make predictions.

[0004] Neuron morphology analysis is of great significance for deeply understanding the structure and function of the nervous system; through morphology analysis, information such as the functional requirements, connection patterns, and signal transmission efficiency of neurons can be revealed, thereby providing strong support for the research and application of the nervous system.

[0005] A neuron morphology analysis method, device, and storage medium based on a graph neural network with a patent publication number of CN115187610B disclose receiving neuron morphology data to be analyzed; generating neuron graph data from the neuron morphology data through a morphology-aware deep hashing graph neural network trained in a contrastive learning manner in advance and extracting neuron morphology features from the neuron graph data; obtaining corresponding neuron morphology hash codes by binarizing the neuron morphology features extracted by the morphology-aware deep hashing graph neural network, and the neuron morphology hash codes are used for neuron morphology classification or neuron retrieval.

[0006] Traditional methods rely on feature extraction, often losing a large amount of useful information, which limits the accuracy and depth of analysis; with the development of deep learning technology, especially the ability of graph neural networks to process complex graph-structured data, it provides new possibilities for neuron morphology analysis. Summary of the Invention

[0007] The purpose of the present invention is to provide a neuron morphology analysis method based on a graph neural network, which improves the classification accuracy, reduces information loss, and realizes the accurate analysis and classification of neuron morphology.

[0008] To achieve the above object, the present invention provides the following technical solutions: A method for analyzing neuron morphology based on a graph neural network, comprising the following steps: Step 1: Construct a dataset containing various neuron morphologies, where the dataset includes the original image data or three-dimensional voxel data of neurons; Step 2: Use a deep learning method to convert the neuron morphology data into graph structure data; Step 3: Input the neuron graph data generated in Step 2 into a deep graph neural network for feature learning and classification, introduce an attention mechanism to enhance the model's learning ability for key features, and further improve the classification accuracy; Step 4: After completing feature learning and classification, use the generated neuron morphology hash code for intelligent retrieval.

[0009] As a preferred technical solution of the present invention, using a deep learning method to convert the neuron morphology data into graph structure data is specifically as follows: Generate neuron graph data from the neuron morphology data through a morphology-aware deep hashing graph neural network. The nodes in the graph represent specific parts of the neuron, and the edges represent the relationships between the nodes, thereby retaining the geometric morphology information of the neuron.

[0010] As a preferred technical solution of the present invention, the specific implementation method for constructing a dataset containing various neuron morphologies is as follows: Collect from neuron image data obtained from public databases, images in scientific literature, or laboratories; ensure that the collected data contains various types of neurons; Perform enhancement processing on the original image data to improve the quality and diversity of the data; standardize the image size and resolution to ensure the consistency of the input data; use image segmentation technology to separate the neurons from the background; Use morphological features or geometric feature descriptors to quantify the morphology of neurons; apply similarity measurement methods to identify neurons with high geometric morphological similarity; select a representative group of neurons with high geometric morphological similarity from the dataset to ensure the diversity and challenge of the dataset; store the preprocessed image or three-dimensional voxel data in TIFF format and assign labels to each data point to indicate its type and source information.

[0011] As a preferred technical solution of the present invention, the enhancement processing includes adjusting contrast, brightness, and applying filters.

[0012] As a preferred technical solution of the present invention, the image segmentation technology includes threshold segmentation, edge detection, and machine learning segmentation algorithms.

[0013] As a preferred technical solution of the present invention, the similarity measurement methods include Euclidean distance, Manhattan distance, and cosine similarity.

[0014] As a preferred technical solution of the present invention, the specific implementation method for inputting neuron graph data into a deep graph neural network for feature learning and classification is as follows: Define specific parts of neurons as graph nodes; define the edges between nodes according to the connectivity of the neuron structure, and use spatial distance and morphological connection as the weights of the edges; Select or design a neural network architecture suitable for graph data, input the graph structure data into the model, and perform feature extraction and node representation learning through graph convolution or graph attention mechanism; In the feature learning stage, the model extracts deep features of graph nodes through multi-layer graph convolution or graph attention operations; In the classification stage, map the node features to class labels.

[0015] As a preferred technical solution of the present invention, in the classification stage, use a fully connected layer or a softmax layer to map the node features to class labels.

[0016] Compared with the prior art, the beneficial effects of the present invention are: Deep learning and classification of neuron morphology through a deep graph neural network to improve the classification accuracy; Select a neuron dataset with high geometric morphological similarity for training to enhance the generalization ability of the method; Use a graph neural network to convert neuron morphology data into graph structure data, retaining the geometric morphological information of neurons; Perform intelligent retrieval through the generated neuron morphology hash code to improve the retrieval efficiency and provide a powerful tool for neuroscience research. Description of the Drawings

[0017] Figure 1 It is a flowchart of the neuron morphology analysis method of the present invention. Detailed Embodiments

[0018] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0019] Embodiment 1 Please refer to Figure 1, which is the first embodiment of the present invention. This embodiment provides a method for neuron morphology analysis based on graph neural networks, including the following steps: Step 1: Construct a dataset containing various neuron morphologies, where the dataset includes the original image data or three-dimensional voxel data of neurons; Step 2: Use deep learning methods to convert neuron morphology data into graph structure data, specifically as follows: Generate neuron graph data from neuron morphology data through a morphology-aware deep hashing graph neural network. The nodes in the graph represent specific parts of the neuron, and the edges represent the relationships between the nodes, thereby preserving the geometric morphology information of the neuron; Step 3: Input the neuron graph data generated in Step 2 into a deep graph neural network for feature learning and classification. Introduce an attention mechanism to enhance the model's learning ability for key features and further improve the classification accuracy; Step 4: After completing feature learning and classification, perform intelligent retrieval using the generated neuron morphology hash code.

[0020] In this embodiment, preferably, the specific implementation method for constructing a dataset containing various neuron morphologies is as follows: Collect from publicly available databases (such as the neuron morphology database of the Allen Institute for Brain Science), images in scientific literature, or neuron image data obtained from laboratories; ensure that the collected data contains various types of neurons, including neurons from different brain regions, different species, and different developmental stages; Perform enhancement processing on the original image data, such as adjusting contrast, brightness, applying filters, etc., to improve the quality and diversity of the data; standardize the image size and resolution to ensure the consistency of the input data; use image segmentation techniques (such as threshold segmentation, edge detection, machine learning segmentation algorithms) to separate neurons from the background; Quantify the morphology of neurons using morphological features (such as length, number of branches, volume, etc.) or more advanced geometric feature descriptors; apply similarity measurement methods (such as Euclidean distance, Manhattan distance, cosine similarity, etc.) to identify neurons with high geometric morphological similarity; select a set of representative neurons with high geometric morphological similarity from the dataset to ensure the diversity and challenge of the dataset; Store the preprocessed images or three-dimensional voxel data in TIFF format and assign labels to each data point to indicate its type and source information.

[0021] In this embodiment, preferably, the specific implementation method for inputting neuron graph data into a deep graph neural network for feature learning and classification is as follows: Define specific parts of the neuron (such as cell body, dendrite, axon, etc.) as graph nodes; Edges between nodes can be defined according to the connectivity of the neuron structure, and spatial distance, morphological connection, etc. can be used as the weights of the edges; Select or design a neural network architecture suitable for graph data, such as Graph Convolutional Network (GCN), Graph Attention Network (GAT), etc.; input the graph structure data into the model, and perform feature extraction and node representation learning through graph convolution or graph attention mechanism; In the feature learning stage, the model extracts deep features of graph nodes through multiple layers of graph convolution or graph attention operations; In the classification stage, use a fully connected layer or a softmax layer to map the node features to class labels.

[0022] Embodiment 2 Please refer to Figure 1 , which is the second embodiment of the present invention. This embodiment is based on the previous embodiment, and the difference is: The specific implementation method of introducing an attention mechanism to enhance the model's learning ability of key features and further improve the classification accuracy is as follows: Introduce an attention layer in the graph neural network, such as the attention mechanism in the graph attention network; design the attention mechanism to calculate the relative importance between nodes, so as to dynamically adjust the weights of node features; Use trainable attention coefficients to calculate the correlation or importance between node pairs; Apply the softmax function to normalize the attention coefficients into a probability distribution to ensure that the sum of the weights is 1; Weight the node features according to the calculated attention weights; Aggregate the weighted features to generate new node representations, which contain richer context information and key features; Train the model using the cross-entropy loss function; apply optimization algorithms (such as Adam, SGD) to minimize the loss function, thereby updating the model parameters.

[0023] The specific implementation method of using the generated neuron morphological hash code for intelligent retrieval is as follows: After the feature learning stage, use a hash function to map the deep features of neurons to compact binary hash codes; local sensitive hashing (LSH), iterative quantization (ITQ) or other hash methods can be used to achieve this; Store the generated hash codes in a hash table for quick retrieval; the hash table can be implemented using data structures such as hash maps or Bloom filters; When retrieving a morphology similar to a given query neuron, first calculate the hash code of the query neuron; look up entries similar to the query hash code in the hash table (such as using the Hamming distance as a similarity metric); Return a list of morphologies that are most similar to the query neuron, which have similar hash codes and deep features; Evaluate the quality of the retrieval results using evaluation metrics such as accuracy, recall, and F1-score; adjust the hash function and retrieval strategy according to the evaluation results to optimize the performance.

[0024] Although the embodiments of the present invention have been shown and described, see the above detailed description, for those of ordinary skill in the art, it can be understood that various changes, modifications, substitutions, and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A neuron morphology analysis method based on graph neural network, characterized by: The steps include: Step 1: Construct a data set containing various neuron morphologies, including the original image data or three-dimensional voxel data of neurons; Step 2: Use deep learning methods to convert neuron morphological data into graph structure data; Step 3: Input the neuron graph data generated in step 2 into the deep graph neural network for feature learning and classification. Introduce the attention mechanism to enhance the model's ability to learn key features and further improve the classification accuracy. Step 4: After completing feature learning and classification, use the generated neuron morphological hash code for intelligent retrieval.

2. The neuron morphology analysis method based on graph neural network according to claim 1, characterized in that: Using deep learning methods, neuron morphological data are converted into graph structured data, as follows: neuron morphological data are generated into neuron graph data through a morphology-aware deep hash graph neural network, in which nodes represent specific parts of neurons and edges represent the relationships between nodes, thereby retaining the geometric morphological information of neurons.

3. The neuron morphology analysis method based on graph neural network according to claim 1, characterized in that: The specific implementation method of constructing a data set containing various neuron morphologies is as follows: Collect neuron image data from public databases, images in the scientific literature, or laboratory-acquired data; ensure that the collected data contains a variety of neuron types; Enhance the original image data to improve the quality and diversity of the data; standardize the image size and resolution to ensure the consistency of the input data; use image segmentation technology to separate neurons from the background; Use morphological features or geometric feature descriptors to quantify the morphology of neurons; apply similarity metrics to identify neurons with high geometric morphology similarity; select a set of representative neurons with high geometric morphology similarity from the dataset to ensure the diversity and challenge of the dataset; The preprocessed images or three-dimensional voxel data are stored in TIFF format, and a label is assigned to each data point to indicate its type and source information.

4. The neuron morphology analysis method based on graph neural network according to claim 3, characterized in that: Enhancement processing includes adjusting contrast, brightness, and applying filters.

5. The neuron morphology analysis method based on graph neural network according to claim 3, characterized in that: The image segmentation technology includes threshold segmentation, edge detection, and machine learning segmentation algorithm.

6. The neuron morphology analysis method based on graph neural network according to claim 3, characterized in that: Similarity measurement methods include Euclidean distance, Manhattan distance, and cosine similarity.

7. The neuron morphology analysis method based on graph neural network according to claim 1, characterized in that: The specific implementation method of inputting neuron graph data into the deep graph neural network for feature learning and classification is as follows: Define specific parts of neurons as graph nodes; define edges between nodes based on the connectivity of neuron structures, and use spatial distance and morphological connections as edge weights; Select or design a neural network architecture suitable for graph data, input graph structure data into the model, and perform feature extraction and node representation learning through graph convolution or graph attention mechanism; In the feature learning stage, the model extracts deep features of graph nodes through multi-layer graph convolution or graph attention operations; In the classification stage, node features are mapped to class labels.

8. The neuron morphology analysis method based on graph neural network according to claim 7, characterized in that: In the classification stage, a fully connected layer or a softmax layer is used to map node features to category labels.

9. The neuron morphology analysis method based on graph neural network according to claim 1, characterized in that: The specific implementation method of intelligent retrieval using the generated neuron morphology hash code is as follows: After the feature learning phase, a hash function is used to map the deep features of neurons into compact binary hash codes; Store the generated hash codes in a hash table for fast retrieval; When it is necessary to retrieve morphologies similar to a given query neuron, first the hash code of the query neuron is calculated; entries similar to the query hash code are searched in the hash table; Returns a list of morphologies most similar to the query neuron, those with similar hash codes and deep features.

10. The neuron morphology analysis method based on graph neural network according to claim 9, characterized in that: A hash function is used to map the deep features of neurons into compact binary hash codes, which are implemented using locality sensitive hashing and iterative quantization.

Images