Neuron-based classification methods, devices, media, and products based on improved classification networks

Abstract

This invention discloses a neuron classification method, apparatus, medium, and product based on an improved classification network, relating to the field of image data processing technology. The improved neuron point cloud classification network addresses the bias problem of manually designed features in traditional methods. Furthermore, in the improved neuron point cloud classification network, the Fourier attention module improves the lack of high-frequency information learning in coordinate-based point cloud networks, the channel attention module reduces information loss caused by max pooling in point cloud networks, and the multi-scale fusion module better integrates fine-grained and high-semantic information. Therefore, by adding a dual attention module and a multi-scale fusion module to the basic point cloud network, the classification network's ability to learn complex neuron features is improved, thereby increasing the accuracy of neuron classification.

Description

Technical Field

[0001] This invention relates to the field of image data processing, and in particular to a three-dimensional neuron classification method, apparatus, medium, and product based on attention and multi-scale point cloud networks. Background Technology

[0002] The study of neuronal morphology is a crucial topic in neuroscience because morphology plays a vital role in determining neuronal connectivity and functional characteristics (alterations in dendritic morphology have been observed in diseases such as Alzheimer's and schizophrenia). Rapid advancements in whole-brain optical microscopy and digital neuron tracing methods have significantly improved neuronal morphology research (e.g., the BigNeuron and NeuroMorpho.Org projects), enabling an increasing number of neuronal images to evolve from somatic cell digital reconstructions to data with three-dimensional spatial structures. These high-quality datasets provide new opportunities for neuronal morphological analysis.

[0003] Neuron classification studies typically involve morphological, physiological, and biochemical approaches. Because digitally reconstructed neuronal morphology is relatively simple to analyze and readily available, researchers tend to approach neuron classification by analyzing the geometric data of neurons.

[0004] Over the past few decades, researchers have proposed various methods and tools for calculating the statistical metrics and morphological features of neurons, which can be broadly categorized into traditional methods and deep learning-based methods. Especially considering the tree-like structure of 3D neuronal point cloud data, most traditional studies represent individual neurons using predefined quantitative metrics. A tool called L-Measure quantitatively represents neuronal morphology using 43 predefined statistical metrics (including total neuronal fiber length, number of segments, branching order, radial distance, and density map). To achieve more accurate comparisons, some researchers decompose neurons into different dendritic segments according to their bifurcation points before using morphological statistical features. However, designing morphological features based on neuroscientists' experience is not only cumbersome but also incomplete in utilizing morphological information, a deficiency that limits recognition capabilities.

[0005] Recently, some researchers have begun to leverage deep learning techniques to improve the performance of neuronal morphological representation. This involves using convolutional neural networks to extract features from neurons by projecting them along coordinate axes into two-dimensional images, or using recurrent neural networks or graph neural networks to capture higher-level representative features by starting with tree or graph concepts. However, the projection process can lead to the loss of considerable geometric information for neurons with complex spatial structures, often requiring supplementary information from human input. Therefore, a more impartial, data-driven approach to representing neurons is needed. Summary of the Invention

[0006] The purpose of this invention is to provide a neuron classification method, device, medium, and product based on an improved classification network, which can improve the accuracy of neuron classification.

[0007] To achieve the above objectives, the present invention provides the following solution:

[0008] A neuron classification method based on an improved classification network, the method comprising:

[0009] Obtain the 3D neuron point cloud data to be classified;

[0010] The three-dimensional neuron point cloud data to be classified is input into an improved neuron point cloud classification network to obtain the category information of the three-dimensional neuron point cloud. The improved neuron point cloud classification network includes three neuron representation modules, a multi-scale feature fusion module, and a fully connected layer connected in series. The outputs of the three neuron representation modules are used as the inputs of the multi-scale feature fusion module, the outputs of the multi-scale feature fusion module are used as the inputs of the fully connected layer, and the outputs of the fully connected layer are the morphological features of the three-dimensional neuron. The neuron representation module includes a first multilayer perceptron, a channel attention module, a second multilayer perceptron, a max pooling layer, and a Fourier attention module connected in series.

[0011] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of the neuron classification method based on an improved classification network described above.

[0012] A computer-readable storage medium having a computer program stored thereon, characterized in that, when executed by a processor, the computer program implements the steps of the neuron classification method based on an improved classification network described above.

[0013] A computer program product includes a computer program, characterized in that, when executed by a processor, the computer program implements the steps of the neuron classification method based on an improved classification network described above.

[0014] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0015] This invention provides a neuron classification method, apparatus, medium, and product based on an improved classification network. The improved neuron point cloud classification network addresses the bias problem of manually designed features in traditional methods. Furthermore, in the improved neuron point cloud classification network, the Fourier attention module improves the lack of high-frequency information learning in coordinate-based point cloud networks, the channel attention module reduces information loss caused by max pooling in point cloud networks, and the multi-scale fusion module better integrates fine-grained and high-semantic information. Therefore, by adding a dual attention module and a multi-scale fusion module to the basic point cloud network, the classification network's ability to learn complex neuron features is improved, thereby increasing the accuracy of neuron classification. Attached Figure Description

[0016] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a schematic diagram of a neuron classification method based on an improved classification network provided in Embodiment 1 of the present invention;

[0018] Figure 2 A schematic diagram of the improved neuronal point cloud classification network structure provided in Embodiment 1 of the present invention;

[0019] Figure 3 This is a schematic diagram of the neuron representation module structure provided in Embodiment 1 of the present invention;

[0020] Figure 4 This is a schematic diagram of the channel attention module structure provided in Embodiment 1 of the present invention;

[0021] Figure 5 This is a schematic diagram of the Fourier attention module structure provided in Embodiment 1 of the present invention;

[0022] Figure 6 This is a schematic diagram of the multi-scale fusion module structure provided in Embodiment 1 of the present invention;

[0023] Figure 7 The training and testing process provided in Embodiment 1 of the present invention;

[0024] Figure 8 This is a diagram of the internal structure of a computer device. Detailed Implementation

[0025] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0026] This invention notes that reconstructed digital 3D neuron data exhibits a structure similar to point clouds, and deep learning models based on point cloud methods are widely used in various fields. Advances in reconstruction techniques have made obtaining representations of the original 3D coordinates of neurons more convenient, and these point cloud methods provide basic techniques for learning 3D spatial points. Therefore, point-level end-to-end feature representation of 3D neurons using point cloud methods becomes possible.

[0027] Existing point cloud feature learning networks typically use feature aggregation to learn global high-semantic features representing the entire input point cloud. When computing aggregated neighborhood features in point cloud networks, max pooling is often used. While max pooling satisfies the permutation invariance property, it only retains the maximum value of features within each neighborhood. This process may lead to a significant loss of non-maximum value information, and the maximum value does not necessarily represent the optimal feature for a neighborhood. Therefore, this invention uses a channel attention mechanism to weight the features of each neighborhood to better select the representation for each neighborhood. Furthermore, point cloud classification networks that directly extract features from points often use multilayer perceptrons as the backbone network, which may result in the loss of interactions between points and global information. Moreover, multilayer perceptron networks often struggle to learn high-frequency components in the features.

[0028] To address this issue, this invention proposes a neuronal classification network based on point cloud methods, incorporating attention and multi-scale fusion. This is the first time a point cloud method has been used to directly learn the morphological information of individual neurons. The comprehensive framework of this invention includes preprocessing, neuronal feature representation, and evaluation stages. This invention designs two attention modules and combines them with a multi-scale feature fusion approach to enhance the point cloud network's ability to capture complex neuronal morphological features. A data-driven approach is used to obtain point-level geometric features of neurons, reducing the bias of manually designed features and thus improving the performance of complex neuronal morphological representation.

[0029] The purpose of this invention is to provide a neuronal classification method, device, medium, and product based on an improved classification network.

[0030] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0031] Example 1

[0032] like Figure 1 As shown in this embodiment, a neuron classification method based on an improved classification network is described. The method includes:

[0033] S1: Obtain the 3D neuron point cloud data to be classified. As an example, the 3D neuron point cloud data to be classified can be 3D mouse brain single neuron point cloud data.

[0034] S2: Input the three-dimensional neuron point cloud data to be classified into the improved neuron point cloud classification network to obtain the category information of the three-dimensional neuron point cloud; such as Figure 2 As shown, the improved neuronal point cloud classification network includes three neuron representation modules, a multi-scale feature fusion module, and a fully connected layer connected in series. The outputs of the three neuron representation modules serve as the inputs to the multi-scale feature fusion module, the output of the multi-scale feature fusion module serves as the input to the fully connected layer, and the output of the fully connected layer represents the morphological features of the three-dimensional neurons. Figure 3 As shown, the neuron representation module includes a first multilayer perceptron, a channel attention module, a second multilayer perceptron, a max pooling layer, and a Fourier attention module connected in series.

[0035] As an optional implementation, step S2 specifically includes:

[0036] (21) Input the three-dimensional neuron point cloud data to be classified into the first neuron representation module to obtain the first feature map.

[0037] (22) Input the first feature map into the second neuron representation module to obtain the second feature map. The processing of the second neuron representation module and the third neuron representation module is the same as that of the first neuron representation module, and refer to the specific process of step (21).

[0038] (23) Input the second feature map into the third neuron representation module to obtain the third feature map.

[0039] (24) Input the first feature map, the second feature map and the third feature map into the multi-scale feature fusion module to obtain the feature representation of the neuron.

[0040] like Figure 3As shown, after the neuron point cloud data is input into the classification network, feature extraction is performed through the neuron representation module. First, the input feature dimensions are transformed using a first multilayer perceptron, increasing the feature dimensions. Next, the feature dimensions are weighted using an attention module to select the most representative features. Then, a second multilayer perceptron performs an identity mapping on the feature dimensions. Next, a max-pooling function is used to obtain features representing the neighborhood. Finally, a Fourier attention module performs frequency-domain weighting on the obtained neighborhood features. Therefore, combining... Figure 3 The specific process of step (21) is as follows:

[0041] (21-1) Input the three-dimensional neuron point cloud data to be classified into the first multilayer perceptron to obtain the feature dimension transformation feature map.

[0042] (21-2) Input the first feature dimension transformed feature map into the channel attention module to obtain the feature dimension weighted feature map.

[0043] (21-3) The weighted feature map of the feature dimensions is input into the second multilayer perceptron to obtain the feature dimension identity mapping feature map.

[0044] (21-4) Input the feature dimension identity mapping feature map into the max pooling layer to obtain the neighborhood feature map.

[0045] (21-5) Input the neighborhood feature map into the Fourier attention module to obtain the first feature map.

[0046] like Figure 4 As shown, the input to the channel attention module is set to x, L p M represents an average pooling filter. p X represents a max-pooling filter. l X represents the low-frequency portion of the input. h Since these represent high-frequency components, the core processing of the channel attention module can be described as follows:

[0047] X l =L p (x); X h =M p (x)-X l ;

[0048] FCAB(x)=MLP l X l +MLP h X h

[0049] =W l X l +Wh X h ;

[0050] Here, FCAB(x) represents the result of the core processing of the channel attention module.

[0051] Input through L p Obtain the low-frequency component X l Input will be passed through M p And by subtracting the low-frequency component, the opposite high-frequency component X is generated. h Learnable weights W are applied to the different frequency features obtained. l With W h Therefore, combining Figure 4 The specific process of step (21-2) can be described as follows:

[0052] (1) The first feature dimension transformation feature map is processed by an average pooling filter to obtain the low frequency component of the feature map.

[0053] (2) The low-frequency component is subtracted from the first feature dimension transformation feature map after processing by the max pooling filter to obtain the high-frequency component of the feature map.

[0054] (3) After setting learnable weights for the low-frequency component and the high-frequency component respectively, perform the addition operation.

[0055] (4) Multiply the feature map after the addition operation with the first feature dimension transformation feature map.

[0056] (5) Add the multiplied feature map to the first feature dimension transformed feature map to obtain the feature map after feature dimension weighting.

[0057] like Figure 5 As shown, the input to the Fourier attention module is set to x. It is a two-dimensional positive Fourier transform. For the corresponding inverse transform, the Fourier attention module can be described as:

[0058]

[0059] X'=X×W

[0060]

[0061] The input x undergoes a Fourier transform to obtain the spectrum X of x. Then, the spectrum X is multiplied by a frequency weight matrix W with learnable parameters to obtain a weighted X′. Finally, the weighted X′ is inversely transformed to update the input x. Therefore, combining... Figure 5 The specific process of step (21-5) is as follows:

[0062] (1) Perform a two-dimensional Fourier forward transform on the neighborhood feature map to obtain the forward transform result.

[0063] (2) Multiply the positive transformation result with a frequency weight matrix of a learnable parameter to obtain a weighted positive transformation result.

[0064] (3) Perform a two-dimensional inverse Fourier transform on the weighted forward transform result to obtain the inverse transform result.

[0065] (4) Add the inverse transformation result to the neighborhood feature map to obtain the first feature map.

[0066] like Figure 6 As shown, the first feature map is set as x0, the second feature map as x1, and the third feature map as x2. The multi-scale feature fusion process first uses three multilayer perceptrons to change the input feature dimensions of x0, x1, and x2 to the same dimension. Next, a batch normalization (BN) layer is used to prevent overfitting. x0 is pooled to obtain a size identical to x1. Then, x0 and x1 are added in the same dimension. The result is then pooled again to obtain a size identical to x2. This is then passed through another BN layer to prevent overfitting. Finally, the result is concatenated with x2 in the same feature dimension to obtain the final output. Therefore, combining... Figure 6 The specific process of step (24) is as follows:

[0067] (24-1) The feature dimensions of the first feature map, the second feature map and the third feature map are transformed to the same feature dimension using three third multilayer perceptrons respectively.

[0068] (24-2) Use three bn layers to perform overfitting on three feature maps with the same feature dimensions.

[0069] (24-3) Perform pooling operation on the feature map corresponding to the first feature map after overfitting to obtain the first pooled feature map.

[0070] (24-4) Add the first pooled feature map and the overfitted feature map corresponding to the second feature map in the dimensional direction.

[0071] (24-5) Perform pooling operation on the feature map after addition to obtain the second pooled feature map.

[0072] (24-6) The feature map after the second pooling is processed by the bn layer and then concatenated with the feature map after the overfitting processing corresponding to the third feature map to obtain the feature representation of the neuron.

[0073] (25) The feature representation of the neuron is input into the fully connected layer to obtain the category information of the three-dimensional neuron point cloud.

[0074] Before executing step S2, the improved neuronal point cloud classification network needs to be trained and tested to obtain a well-trained classification network.

[0075] Combination Figure 7 The specific process of training and testing is as follows:

[0076] Acquisition and preprocessing of 3D single neuron dataset: 3D mouse brain single neuron data in SWC format were acquired. For each sample in the dataset, the largest connected component containing the neuron cell body was retained. The coordinates of each neuron were normalized to 0-1. The number of fixed point clouds between samples was aligned using interpolation and sampling algorithms. Finally, the processed data was divided into training and test sets in a 7:3 ratio.

[0077] Construction of a neuronal point cloud classification network: Building a deep learning network model using Python based on the PyTorch framework. Figures 2 to 6 The network model structure is defined, and the loss function and optimization function are set. The output of the multi-scale fusion module is set to a 2048-dimensional vector, and the loss function of the classification network is the taxonomic loss between the class output by the fully connected layer and the true class label of the vector.

[0078] Training the classification network: The number of iterations for training the classification network is set, and the network is trained using the training set. During training, the input to the classification network consists of N*3 single neurons and their corresponding classification labels. Here, N represents the number of points in the point cloud corresponding to a single neuron; 3 indicates three dimensions.

[0079] The training set is input into the classification network in batches, and the first feature map is obtained through the first neuron representation module. The first feature map is then input into the second neuron representation module to obtain the second feature map. The second feature map is then input into the third neuron representation module to obtain the third feature map. Finally, the first, second, and third feature maps are all input into the multi-scale fusion module to obtain the feature representation of the neuron. The feature representation of the neuron is then backpropagated through the output of the fully connected layer and the ground truth label using the cross-entropy loss function.

[0080] The formula for the cross-entropy loss function is as follows:

[0081]

[0082] Where output is the network output, with a size of 1*n, where n represents the types of neurons in the dataset; label is the label of each neuron represented by a one-hot vector encoding; and m represents the batch size set when training the network.

[0083] Testing the classification network: The trained classification network is tested using the training set. Finally, the tested classification network is used to classify the 3D neuron point cloud data to be classified.

[0084] In this embodiment, a point cloud method is used for the first time to learn the morphological information of individual neurons end-to-end. The comprehensive framework of this invention includes preprocessing, neuron feature representation, and evaluation stages, reducing the bias of manually designed features in traditional methods. Furthermore, two attention modules are designed to address the problems of point cloud networks lacking high-frequency information and max pooling losing non-maximum information. A multi-scale method utilizes high semantic information while compensating for the lack of fine-grained information, ultimately enabling the classification network to learn better morphological features of neurons and improving classification accuracy.

[0085] Example 2

[0086] A computer device includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the processor executing the computer program to implement the steps of a neuron classification method based on an improved classification network as described in Embodiment 1.

[0087] Example 3

[0088] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a neuron classification method based on an improved classification network as described in Embodiment 1.

[0089] Example 4

[0090] A computer program product includes a computer program that, when executed by a processor, implements the steps of a neuron classification method based on an improved classification network as described in Example 1.

[0091] Example 5

[0092] A computer device, which may be a database, may have an internal structure diagram as shown below. Figure 8As shown, the computer device includes a processor, memory, input / output (I / O) interfaces, and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is also connected to the system bus via the I / O interfaces. The processor provides computational and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores the operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores pending transactions. The I / O interfaces are used for exchanging information between the processor and external devices. The communication interface is used for communicating with external terminals via a network connection. When the computer program is executed by the processor, it implements a neuron classification method based on an improved classification network as described in Embodiment 1.

[0093] It should be noted that the object information (including but not limited to object device information, object personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this invention are all information and data authorized by the object or fully authorized by all parties, and the collection, use and processing of related data must comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0094] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided by this invention can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided by this invention may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided by this invention may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0095] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0096] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A neuron classification method based on an improved classification network, characterized in that, The method includes: Obtain the 3D neuron point cloud data to be classified; The three-dimensional neuron point cloud data to be classified is input into an improved neuron point cloud classification network to obtain the category information of the three-dimensional neuron point cloud. The improved neuron point cloud classification network includes three neuron representation modules, a multi-scale feature fusion module, and a fully connected layer connected in series. The outputs of the three neuron representation modules are used as the inputs of the multi-scale feature fusion module, the outputs of the multi-scale feature fusion module are used as the inputs of the fully connected layer, and the outputs of the fully connected layer are the morphological features of the three-dimensional neuron. The neuron representation module includes a first multilayer perceptron, a channel attention module, a second multilayer perceptron, a max pooling layer, and a Fourier attention module connected in series. The input to the channel attention module is set as follows: , L p This represents an average pooling filter. M p This represents a max-pooling filter. X l This represents the low-frequency portion of the input to the channel attention module. X h If the high-frequency components represent the input to the channel attention module, then the processing procedure of the channel attention module can be described as follows: ； ； ； In the formula, FCAB( () represents the processing result of the channel attention module; W l and W h As weight; The input to the Fourier attention module is set as follows: , It is a two-dimensional positive Fourier transform. For a two-dimensional inverse Fourier transform, the processing procedure of the Fourier attention module can be described as follows: ； ； ； In the formula, This is the frequency weight matrix; Here, the first feature map is set as x0, the second feature map as x1, and the third feature map as x2; the multi-scale feature fusion process of the multi-scale feature fusion module is as follows: First, the input feature dimensions of x0, x1, and x2 are changed to the same dimension using three multilayer perceptrons. Second, a batch normalization (BN) layer is used to prevent overfitting. x0 is pooled to obtain a size with the same dimensions as x1. The pooled x0 and x1 are added together in the same dimension. Then, the result of the addition is pooled again to obtain a size with the same dimensions as x2. The pooled and added result is then passed through a BN layer to prevent overfitting. Finally, the result of the addition after the BN layer is concatenated with x2 in the same feature dimension to obtain the final output.

2. The neuron classification method based on an improved classification network according to claim 1, characterized in that, The three-dimensional neuron point cloud data to be classified is input into an improved neuron point cloud classification network to obtain the category information of the three-dimensional neuron point cloud, specifically including: The three-dimensional neuron point cloud data to be classified is input into the first neuron representation module to obtain the first feature map; The first feature map is input into the second neuron representation module to obtain the second feature map; The second feature map is input into the third neuron representation module to obtain the third feature map; The first feature map, the second feature map, and the third feature map are all input into the multi-scale feature fusion module to obtain the feature representation of the neuron; The feature representation of the neuron is input into the fully connected layer to obtain the category information of the three-dimensional neuron point cloud.

3. The neuron classification method based on an improved classification network according to claim 2, characterized in that, The three-dimensional neuron point cloud data to be classified is input into the first neuron representation module to obtain the first feature map, specifically including: The three-dimensional neuron point cloud data to be classified is input into the first multilayer perceptron to obtain the first feature dimension transformation feature map; The first feature dimension transformed feature map is input into the channel attention module to obtain the feature map after feature dimension weighting; The weighted feature map of the feature dimensions is input into the second multilayer perceptron to obtain the feature dimension identity mapping feature map; The feature dimension identity mapping feature map is input into the max pooling layer to obtain the neighborhood feature map; The neighborhood feature map is input into the Fourier attention module to obtain the first feature map.

4. The neuron classification method based on an improved classification network according to claim 3, characterized in that, The first feature dimension transformed feature map is input into the channel attention module to obtain the feature dimension weighted feature map, specifically including: The low-frequency component of the feature map is obtained by processing the first feature dimension transformation feature map through an average pooling filter. The high-frequency component of the feature map is obtained by subtracting the low-frequency component after processing the first feature dimension transformation feature map through a max pooling filter. After setting learnable weights for the low-frequency components and the high-frequency components respectively, an addition operation is performed. Multiply the feature map after the addition operation with the feature map transformed by the first feature dimension; The feature map after multiplication is added to the feature map after the first feature dimension transformation to obtain the feature map after feature dimension weighting.

5. The neuron classification method based on an improved classification network according to claim 3, characterized in that, The neighborhood feature map is input into the Fourier attention module to obtain the first feature map, specifically including: Perform a two-dimensional forward Fourier transform on the neighborhood feature map to obtain the forward transform result; The forward transformation result is multiplied by a frequency weight matrix with a learnable parameter to obtain a weighted forward transformation result; Perform a two-dimensional inverse Fourier transform on the weighted forward transform result to obtain the inverse transform result; The inverse transformation result is added to the neighborhood feature map to obtain the first feature map.

6. The neuron classification method based on an improved classification network according to claim 2, characterized in that, The first feature map, the second feature map, and the third feature map are all input into the multi-scale feature fusion module to obtain the feature representation of the neuron, specifically including: Three third-layer perceptrons are used to transform the feature dimensions of the first feature map, the second feature map, and the third feature map to the same feature dimension, respectively. Three batch normalization (BN) layers are used to perform overfitting on three feature maps with the same feature dimension. The overfitted feature map corresponding to the first feature map is pooled to obtain the first pooled feature map. The first pooled feature map and the overfitted feature map corresponding to the second feature map are added together in the dimensional direction. The feature maps after addition are pooled to obtain the second pooled feature map. The feature map after the second pooling is processed by the bn layer and then concatenated with the feature map after the overfitting treatment corresponding to the third feature map to obtain the feature representation of the neuron.

7. A computer device, comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that the processor executes the computer program to implement the steps of a neuronal classification method based on an improved classification network according to any one of claims 1-6.

8. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of a neuron classification method based on an improved classification network as described in any one of claims 1-6.

9. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of a neuron classification method based on an improved classification network as described in any one of claims 1-6.

Images