Morphological feature extraction method and morphological structure generation method of neurons

Through the self-supervised learning model of neuronal morphological features and the occlusion autocoding network, neuronal morphological features are extracted and generated, and the problem of relying on labeled information is solved, achieving high accuracy and widely applicable neuronal morphological features extraction and generation.

CN118918258APending Publication Date: 2024-11-08HUST SUZHOU INST FOR BRAINMATICS
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202410982846.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2024-07-22
Publication Date
2024-11-08

AI Technical Summary

Technical Problem

The existing neuronal morphological feature extraction methods rely on artificial labeling information, and are not effective when describing complex neuronal morphology, and cannot fully meet the needs of describing neuronal morphology.

Method used

The self-supervised learning model of neuronal morphological features is adopted to convert neuronal morphological data into standard point cloud data by blocking the self-coding network and self-attention mechanism, morphological features are extracted, and trained without labeling, combined with generation The model generates morphological structure.

Benefits of technology

It achieves highly accurate extraction of neuronal morphological characteristics without artificial labeling of information, and its performance is comparable to that of a supervised method. It is suitable for neuronal classification and morphological structure generation, expanding the scope of application of the model.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN118918258A_ABST
    Figure CN118918258A_ABST
Patent Text Reader

Abstract

The invention provides a morphological feature extraction method and a morphological structure generation method of neurons. The morphological feature extraction method of the neurons comprises the following steps: step 1, a preprocessing step: converting morphological data of the neurons into standard point cloudization data; step 2, a feature extraction step: inputting the standard point cloudization data into a pre-trained neuron morphological feature self-supervised learning model which adopts an occlusion self-encoding network as a skeleton and combines a self-attention mechanism to obtain a neuron morphological feature self-supervised learning model; and extracting morphological characteristics of neurons from the standard point cloudization data. According to the method, the information does not need to be labeled manually, and the accuracy is high.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention relates to the technical field of biological image data processing, and in particular to a method for extracting neuron morphological features and a method for generating a structure that can extract morphological features from neuron three-dimensional structure reconstruction data. Background Art

[0002] Neurons are the basic building blocks of the brain, with complex structures, diverse morphologies, and rich connection patterns, which are closely linked to the complex functions of the brain. In order to deeply study these complex and changeable neuronal structures, it is necessary to describe the neuronal morphology in detail, that is, to explore and summarize the unified morphological characteristics from the diversity of their morphology. The so-called morphological characteristics of neurons refer to the extraction of indicators that quantitatively describe the morphological structure of neurons by analyzing neuronal morphological data. These indicators cover the description of the morphological structure of neurons in multiple dimensions and angles. The key is to comprehensively and accurately quantify the morphology of neurons while losing as little information as possible. Therefore, accurately and comprehensively extracting the morphological characteristics of neurons, and then systematically and repeatably characterizing and analyzing neurons, is of great significance for promoting basic research in brain science such as neuron classification.

[0003] Unlike traditional image data, neuronal morphological data (biological image data) usually simplifies neuronal morphology into skeletonized tracings, i.e., a series of interconnected points that reveal the location, thickness, connectivity, and domain of neuronal cell bodies or processes, thus providing rich neuronal structural information. However, this representation method also brings challenges of structural complexity and morphological diversity. Currently, most studies on neuronal morphological characteristics are based on simple morphological parameters, aiming to fit experts' annotation and classification criteria for neuronal types. This traditional research method focuses on the design, selection, and combination of morphological features, but often requires the morphological features to be extracted to be determined in advance, and it is easy to ignore the relationship between morphological parameters. Therefore, this statistical method based on artificial definitions has limitations in describing neuronal structural information.

[0004] In the past decade or so, with the rapid development of deep learning technology, researchers have begun to use deep learning algorithms to automatically extract and analyze neuronal morphology and learn more abstract and high-level feature representations from data. This kind of learning-based morphological feature research has made significant progress in recent years, but most of it still relies on supervised learning frameworks and is mainly applicable to situations where label information is accurate and sufficient. When applied to other types, especially neurons with more complex morphology, its effect may be greatly reduced. In summary, the existing neuronal morphological feature extraction methods cannot fully meet the needs of describing neuronal morphology. Summary of the invention

[0005] Based on the above-mentioned defects of the prior art, the present invention provides a method for extracting morphological features of neurons and a method for generating morphological structures, which do not need to rely on artificially labeled information and have high accuracy.

[0006] In order to achieve the above object, the present invention provides a method for extracting morphological features of neurons, comprising:

[0007] Step 1, preprocessing step: convert neuron morphological data into standard point cloud data;

[0008] Step 2, feature extraction step: input the standard point cloud data into a pre-trained neuron morphological feature self-supervised learning model, which uses an occluded autoencoder network as a skeleton and combines a self-attention mechanism to extract the morphological features of neurons from the standard point cloud data.

[0009] In one embodiment, in step 1, the process of converting neuron morphology data into standard point cloud data includes:

[0010] Extracting the three-dimensional coordinate points of neuron sampling points from the SWC file storing the neuron morphology data as point cloud raw data;

[0011] Normalizing the original point cloud data, mapping the coordinates to between -1 and 1 as point cloud normalized data;

[0012] The point cloud normalized data is sampled, and the farthest point sampling is adopted during sampling to achieve the unification of the number of points and keep the shape of the neuron unchanged, so as to obtain standard point cloud data.

[0013] In one embodiment, in step 2, the process of pre-training the neuron morphological feature self-supervised learning model includes:

[0014] Step S1, in each network training iteration, extracting N neuron morphological data from the training set according to the batch size and converting them into standard point cloud data;

[0015] Step S2, grouping the standard point cloud data to obtain a neuron point cloud group representation;

[0016] Step S3, randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud;

[0017] Step S4, mark embedding, mapping the occluded neuron point cloud group representation into a continuous numerical vector;

[0018] Step S5, encoding and decoding are performed through an encoding and decoding network; the encoding and decoding network includes an encoder and a decoder, and the decoder structure is similar to the encoder, both of which are composed of standard Transformer blocks; the encoder only encodes the visible part, and does not need to encode the occluded part; the decoder inputs the occluded part and the visible part encoding, outputs the decoded prediction block features, maps the prediction block features to a vector with the same dimension as the total number of coordinates of the midpoints of the occluded part through a fully connected layer, and then obtains the predicted morphological distribution of the occluded part through a reshape operation;

[0019] Step S6, using the chamfer distance loss function to calculate the reconstruction loss and train the neuron morphological feature self-supervised learning model;

[0020] Step S7, determining whether the set training end condition is met; if not, performing the next iterative training according to steps S1 to S6; if so, completing the training of the neuron morphological feature self-supervised learning model.

[0021] In one embodiment, in step S2, the process of grouping the standard point cloud data to obtain the neuron point cloud group representation includes: given a neuron standard point cloud data X∈R n×3 , where n represents the number of neuron points, and R represents the real number domain, which means that the input matrix set of size n*3 is real numbers. First, the center point of the neuron is selected, and p center points are sampled from the standard point cloud data through the farthest point sampling. Then, the K nearest neighbor algorithm is used to select k neighboring points for each center point, and the neuron point cloud grouping representation can be obtained.

[0022] In one embodiment, in step S3, the process of randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud includes: adopting a random occlusion design, setting the occlusion ratio to m, randomly selecting m center points, removing the corresponding groups, and obtaining the occluded group representation of the neuron point cloud.

[0023] In one embodiment, in step S4, the marker embedding, the process of mapping the occluded neuron point cloud group representation into a continuous numerical vector includes: for the occluded part, replacing the embedding of the occluded part with a weight-sharing learnable mask representation without performing other operations; for the visible part, using PointNet to implement group embedding, the main structure of PointNet is composed of MLP and maximum pooling layer, MLP maps the center coordinates to the embedding marker dimension, and provides position guidance for each neuron point cloud group representation.

[0024] In one embodiment, in step S6, the chamfer distance loss function is:

[0025]

[0026] Among them, L represents the chamfer distance, which calculates the distance from each point in a point set a to the nearest point in another point set b, and takes the average or sum of these distances as the similarity measure of the two point sets, which is used to measure the similarity between the two point sets. P represents the grouped representation of neuronal morphology, gt represents the original input reference, and pre represents predicted reconstruction.

[0027] In one embodiment, step two also includes: using the morphological features of the neurons to distinguish different types of neurons: removing the decoder of the self-supervised learning model of the neuron morphological features and using pre-trained weights as initial weights for fine-tuning, extracting all local features of neuron morphology from the well-trained encoder, aggregating them into global features representing the entire shape through average pooling, and then inputting them into a fully connected network for fine-tuning and classification, so as to achieve the distinction between neuron types.

[0028] The present invention also provides a method for generating a neuron's morphological structure, comprising: using the neuron's morphological feature extraction method as described above and the following step three;

[0029] Step three, morphological structure generation step: using the morphological features extracted in step two to generate the morphological structure of neurons.

[0030] In one embodiment, in step 3, the process of using the morphological features to generate the morphological structure of neurons includes:

[0031] Based on the self-supervised learning model of neuron morphological features, the latent space description of the features is modified to represent the values ​​with a probability range rather than a numerical value. The samples similar to the given neurons are sampled and generated by means of a generative model. The method includes three components: encoding the latent distribution, and the output is the mean and standard deviation; reparameterization, in which the latent variables are generated by random sampling, and the sampling process itself is a non-differentiable operation. Reparameterization is performed to convert the non-differentiable operation into a differentiable process; KL divergence constraint, the objective function of VAE is the sum of the reconstruction loss and the KL divergence loss. The KL divergence loss is used to measure the difference between the latent distribution predicted by the encoder and the prior distribution, and the morphological structure of the neuron is obtained by sampling the latent space distribution and decoding.

[0032] The neuron morphological feature extraction method and morphological structure generation method described in the present invention adopt a neuron morphological feature self-supervised learning model. The neuron morphological feature self-supervised learning model adopts an occluded autoencoder network as a skeleton, combined with a self-attention mechanism, and can extract effective neuron morphological features without labeling, and the accuracy is comparable to the performance of supervised methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] The drawings described herein are only for explanation purposes and are not intended to limit the scope of the present invention in any way. In addition, the shapes and proportional dimensions of the components in the drawings are only for illustration purposes and are used to help understand the present invention. They do not specifically limit the shapes and proportional dimensions of the components of the present invention. Under the guidance of the present invention, those skilled in the art can select various possible shapes and proportional dimensions to implement the present invention according to specific circumstances. In the drawings:

[0034] Figure 1 A schematic flow chart of a method for generating a neuron morphological structure provided in the second embodiment of the present invention, wherein step 1 and step 2 included therein are schematic flow charts of a method for extracting a neuron morphological feature provided in the first embodiment of the present invention;

[0035] Figure 2 A schematic diagram of the process of pre-training the neuron morphological feature self-supervised learning model adopted by the present invention;

[0036] Figure 3 A comparison chart of the results of the present invention applied to the M1 pyramidal neuron generation task and the morphological data generated by MorphVAE;

[0037] Figure 4 An evaluation diagram showing the results of the present invention applied to the M1 pyramidal neuron generation task. DETAILED DESCRIPTION

[0038] In order to enable those skilled in the art to better understand the technical solutions in the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0039] See also Figure 1 As shown, the first embodiment of the present invention provides a method for extracting morphological features of neurons, comprising:

[0040] Step 1, preprocessing step: convert neuron morphological data into standard point cloud data;

[0041] Step 2, feature extraction step: input the standard point cloud data into a pre-trained neuron morphological feature self-supervised learning model, which uses an occluded autoencoder network as a skeleton and combines a self-attention mechanism to extract the morphological features of neurons from the standard point cloud data.

[0042] In one embodiment, in step 1, the process of converting neuron morphology data into standard point cloud data includes:

[0043] Extracting the three-dimensional coordinate points of neuron sampling points from the SWC file storing the neuron morphology data as point cloud raw data;

[0044] Normalizing the original point cloud data, mapping the coordinates to between -1 and 1 as point cloud normalized data;

[0045] The point cloud normalized data is sampled, and the farthest point sampling is adopted during sampling to achieve the unification of the number of points and keep the shape of the neuron unchanged, so as to obtain standard point cloud data.

[0046] In one embodiment, in step 2, the process of pre-training the neuron morphological feature self-supervised learning model includes:

[0047] Step S1, in each network training iteration, extracting N neuron morphological data from the training set according to the batch size and converting them into standard point cloud data;

[0048] Step S2, grouping the standard point cloud data to obtain a neuron point cloud group representation;

[0049] Step S3, randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud;

[0050] Step S4, mark embedding, mapping the occluded neuron point cloud group representation into a continuous numerical vector;

[0051] Step S5, encoding and decoding are performed through an encoding and decoding network; the encoding and decoding network includes an encoder and a decoder, and the decoder structure is similar to the encoder, both of which are composed of standard Transformer blocks; the encoder only encodes the visible part, and does not need to encode the occluded part; the decoder inputs the occluded part and the visible part encoding, outputs the decoded prediction block features, maps the prediction block features to a vector with the same dimension as the total number of coordinates of the midpoints of the occluded part through a fully connected layer, and then obtains the predicted morphological distribution of the occluded part through a reshape operation;

[0052] Step S6, using the chamfer distance loss function to calculate the reconstruction loss and train the neuron morphological feature self-supervised learning model;

[0053] Step S7, determining whether the set training end condition is met; if not, performing the next iterative training according to steps S1 to S6; if so, completing the training of the neuron morphological feature self-supervised learning model.

[0054] In one embodiment, in step S1, the conversion process is the same as the aforementioned process of converting neuron morphology data into standard point cloud data.

[0055] In one embodiment, in step S2, the process of grouping the standard point cloud data to obtain the neuron point cloud group representation includes: given a neuron standard point cloud data X∈R n×3 , where n represents the number of neuron points, and R represents the real number domain, which means that the input matrix set of size n*3 is real numbers. First, the center point of the neuron is selected, and p center points are sampled from the standard point cloud data through the farthest point sampling (FPS). Then, the K nearest neighbor algorithm (KNN) is used to select k neighboring points for each center point, and the neuron point cloud grouping representation can be obtained.

[0056] In one embodiment, in step S3, the process of randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud includes: adopting a simple random occlusion design, assuming that the occlusion ratio is m, randomly selecting m center points, removing their corresponding groups, and obtaining the occluded group representation of the neuron point cloud.

[0057] In one embodiment, in step S4, the marker embedding, the process of mapping the occluded neuron point cloud group representation to a continuous numerical vector includes: for the occluded part, the embedding of the occluded part is replaced with a weight-sharing learnable mask representation without other operations; for the visible part, group embedding is implemented using PointNet, the main structure of PointNet is composed of MLP and maximum pooling layers, and MLP maps the center coordinates to the embedding marker dimension, providing position guidance for each neuron point cloud group representation. Because the main structure of PointNet is composed of MLP and maximum pooling layers, it is not affected by the input order; and in the aforementioned step S2, the points of each group are standardized coordinates, so the center coordinates can be mapped to the embedding marker dimension through a simple learnable MLP, providing position guidance for each group.

[0058] In one embodiment, in step S6, the chamfer distance loss function is:

[0059]

[0060] Wherein, L represents the chamfer distance, which calculates the distance from each point in a point set a to the nearest point in another point set b, and takes the average or sum of these distances as the similarity measure of the two point sets, which is used to measure the similarity between the two point sets, P represents the grouping representation of neuron morphology, gt represents the original input reference, and pre represents the predicted reconstruction. In the present invention, a and b represent the predicted point set and the true reference point set, respectively.

[0061] In one embodiment, the step 2 further includes: using the morphological features of the neurons to distinguish different types of neurons: removing the decoder of the self-supervised learning model of the neuron morphological features and fine-tuning using the pre-trained weights as initial weights. Specifically, all neuron morphological local features are extracted from the well-trained encoder, aggregated into global features representing the entire shape through average pooling, and then input into a fully connected network for fine-tuning classification, so that simple network training can be used to achieve the distinction of neuron types.

[0062] Please refer to Figure 1 As shown, the second embodiment of the present invention provides a method for generating a neuron morphological structure, including: using the neuron morphological feature extraction method as described in the first embodiment and the following step three;

[0063] Step three, morphological structure generation step: using the morphological features extracted in step two to generate the morphological structure of neurons.

[0064] In one embodiment, in step 3, the process of using the morphological features to generate the morphological structure of neurons includes:

[0065] Based on the self-supervised learning model of neuron morphological features, the latent space description of the features is modified, and the values ​​are expressed in probability ranges rather than numerical values, and samples similar to given neurons are generated by generative models. It includes three components: encoding potential distribution, and the output is the mean and standard deviation; reparameterization, the latent variables are generated by random sampling, and the sampling process itself is a non-differentiable operation. Reparameterization is performed to convert the non-differentiable operation into a differentiable process; KL divergence constraint, the objective function of VAE is the sum of reconstruction loss and KL divergence loss, and KL divergence loss is used to measure the difference between the potential distribution predicted by the encoder and the prior distribution, and the morphological structure of the neuron is obtained by sampling the potential space distribution and decoding. Because the sampling process itself is a non-differentiable operation, in order to be able to back propagate the error, it is necessary to reparameterize the non-differentiable operation into a differentiable process. The prior distribution can be, for example, a standard normal distribution, and the KL divergence loss is used to measure the difference between the potential distribution predicted by the encoder and the prior distribution, which forces the data encoding learned by the encoder to be as close to the standard normal distribution as possible, thereby realizing the learning of the potential space distribution.

[0066] The following is a specific embodiment of the present invention.

[0067] In a specific implementation, the present invention uses training data mainly from NeuroMorpho.Org, and selects seven typical mouse neuron morphology data, a total of 2595 neuron data. For the convenience of description, it is named Neuron7 data set.

[0068] In order to evaluate the self-supervised learning method adopted by the present invention, a linear evaluation method is adopted, as follows:

[0069] Linear evaluation refers to fixing the feature representation obtained by self-supervised learning, training a simple linear classifier on this basis, and then testing the classification accuracy to illustrate the effectiveness of the method. The comparison method uses the classic method MorphVAE in the field of neuronal morphology for illustration. The linear evaluation method uses LiearSVM classification accuracy to characterize feature quality. Table 1 is the evaluation results, which show that the linear evaluation accuracy of the self-supervised model of the present invention reaches 75.43%, which is nearly 10% higher than the prior art.

[0070] Table 1 Linear evaluation table of self-supervised methods

[0071]

[0072] In a specific embodiment, in step 2, the self-supervised learning method is modified for the neuron type classification task in the following manner:

[0073] The decoder of the self-supervised learning method is removed and the pre-trained weights are used as the initial weights for fine-tuning. Specifically, all local features of neuron morphology are extracted from the well-trained encoder, aggregated into global features representing the entire shape through average pooling, and then input into the fully connected network for fine-tuning classification. Considering the fairness of the comparison dataset, an additional public dataset M1EXC provided in MorphVAE is added, which contains 275 excitatory pyramidal neuron dendrites from the primary motor cortex M1 of adult mice.

[0074] Table 2 Classification performance evaluation table of different classification methods

[0075]

[0076]

[0077] Table 2 is a table of classification performance evaluation of different classification methods. Among them, the feature extraction method based on the statistical distribution of traditional morphological measurements is combined with linear SVM for classification, which is used as the basic method in the traditional field. MorphoGNN, a neuron morphological representation method based on graph neural network, is an effective method in the field of neuron classification, which is used as a basic method in the supervision field. MorphVAE, a neuron morphological feature learning method based on self-supervised learning, describes the neuron structure from the perspective of branches, and learns the embedding of complex morphological features of neurons based on the generative model, which is used as a basic method in the field of self-supervision. The automatic encoder scheme based on the self-attention mechanism, that is, the present invention will illustrate the model performance through two experiments in the classification task. One is to use the representation of the trained encoder Transformer to train the linear support vector machine classifier, and to explain whether the self-supervised learning method has learned an effective feature representation through linear evaluation. The second is supervised fine-tuning, which mainly explores the ability of the self-supervised method to migrate to downstream tasks. By freezing the weights of the encoder and removing the decoder, and adding a simple fully connected layer to achieve the classification task, in the specific implementation, only one linear layer is added for supervised fine-tuning. The results show the good performance of the present invention in migration and expansion to the neuron classification task, with a classification accuracy of more than 79%.

[0078] In a specific embodiment, in step 3, the self-supervised learning method is modified and expanded for the neuron type generation task in the following manner, including:

[0079] Step S31, encoding potential distribution: modifying the latent space description of features based on the self-supervised learning method, expressing them in terms of probability ranges of values ​​rather than numerical values, and outputting them as mean and standard deviation;

[0080] Step S32, reparameterization: The latent variables are generated by random sampling, but the sampling process itself is a non-differentiable operation. In order to be able to back-propagate the error, reparameterization is required to convert the non-differentiable operation into a differentiable process;

[0081] Step S33, calculate the loss function: the loss function is the sum of the reconstruction loss and the KL divergence loss, where the KL divergence loss is used to measure the difference between the potential distribution predicted by the encoder and the prior distribution (such as the standard normal distribution), which forces the data encoding learned by the encoder to be as close to the standard normal distribution as possible, thereby achieving learning of the latent space distribution.

[0082] In a specific implementation, the M1 EXC dataset is used to conduct actual experiments, and finally the trained generative model is used to sample and generate sample results similar to a given neuron.

[0083] When evaluating the results, common morphological features based on specific aspects of neurons were considered, as shown in Table 3.

[0084] Table 1 Summary of traditional morphological characteristics of dendrites

[0085]

[0086] Finally, the results of the present invention in the task of generating the neuron morphological structure are as follows: Figure 3 As shown, where:

[0087] Figure 3 From the perspective of subjective qualitative evaluation, the samples generated by the present invention show relatively rich morphological features as a whole, can show the complex branching structure of neurons, and produce a large number of neuronal morphologies of the same type but different shapes. Compared with the morphological data generated by MorphVAE, the neuronal morphological structure generated by the present invention has certain advantages in visual inspection by neuron reconstruction experts.

[0088] Figure 4 24 groups of commonly used traditional morphological parameters were measured, and the comparison of morphological statistical distribution among reference samples, MorphVAE, the present invention and training sets was shown.

[0089] The above results show the good scalability of the present invention in the task of generating the morphological structure of neurons.

[0090] In summary, the present invention focuses on the local fine structure of neurons, and starting from the neuronal point structure, designs the main framework for neuron morphological feature extraction based on point cloud data and unlabeled data. In the actual neuron type differentiation task, it is better than the existing self-supervised learning method, and the classification performance is comparable to that of the supervised method, which illustrates the advantage of the present invention in extracting complex morphological neuron features. When further generating neuron morphological structure, the present invention realizes the generation display on the dendrites of M1 excitatory neurons. Compared with the existing methods, this method has better morphological parameter rationality in the distribution of distance-based and topology-based statistical indicators.

[0091] The present invention proposes a dendrite morphological feature extraction method suitable for neuron point data and independent of annotation. This method takes into account the complex multi-branching of neurons, converts neuron morphological data into point cloud form, simplifies neuron morphological representation through grouping strategy, and further combines with occlusion autoencoder network to make up for the shortcomings of insufficient description in feature design and selection based on traditional feature extraction methods. In order to solve the problem that learning-based neuron morphological features rely on annotation information, a self-supervised learning method that can be used for neuron morphological feature extraction is designed, which effectively expands the scope of application of the model. In addition, combined with occlusion design and self-attention mechanism, a refined representation of the local structure of neurons is achieved, further satisfying the feature extraction of complex morphological neurons.

[0092] The present invention provides an accurate and comprehensive method for extracting complex morphological features of neuronal morphological data, and has the following advantages:

[0093] 1. Starting from the point structure of neurons, the present invention simplifies the neuron morphological representation through a grouping strategy and introduces the local branching factors of neuron morphology into the network model, thus making up for the problem of insufficient representation of traditional morphological features.

[0094] 2. The present invention combines the occluded autoencoder network structure to design a self-supervised learning method for extracting neuron morphological features. When learning the complex morphological features of neurons, there is no need to manually annotate type information, but the structure and distribution characteristics of the data itself are used to guide the learning process. By designing appropriate auxiliary tasks, it is possible to effectively learn and understand neuron representations, thereby solving the problem that the neuron morphological feature extraction method relies on annotated information.

[0095] 3. The present invention has good scalability and wide adaptability, is independent of neuron labeling information, has good generalizability to different neuron morphology tasks, and can be used to generate more similar and different neuron morphology structures based on the feature extraction of a certain type of neurons.

[0096] The present invention can extract effective neuron morphological features without labeling, and has a significant effect on the extraction of complex morphological features of neuron morphological data. Without the need for neuron labeling, the performance is comparable to that of supervised methods. The present invention can be widely used in different neuron morphological tasks, and to a certain extent solves the problem that neuron morphological features are difficult to generalize and apply to other neuron tasks. For neuron classification tasks, a simple fully connected network can be trained on the basis of feature extraction to achieve good classification results; on the other hand, for neuron morphology generation tasks, the mean and standard deviation of the encoded potential distribution and the re-parameterization technique can be modified and added on the basis of the self-supervised learning method to achieve sampling of neuron morphological features and generate neuron morphological data that conforms to the reference sample morphological distribution.

[0097] It should be understood that the above description is for illustration and not for limitation. Many embodiments and many applications beyond the examples provided will be apparent to those skilled in the art upon reading the above description. Therefore, the scope of the present teachings should not be determined with reference to the above description, but rather with reference to the foregoing claims and the full scope of equivalents to which such claims are entitled. For the purpose of comprehensiveness, all articles and references, including disclosures of 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 to be a waiver of such subject matter, nor should it be considered that the applicant has not considered such subject matter to be part of the disclosed application subject matter.

Claims

1. A method for extracting morphological features of neurons, characterized in that: include: Step 1, preprocessing step: convert neuron morphological data into standard point cloud data; Step 2, feature extraction step: input the standard point cloud data into a pre-trained neuron morphological feature self-supervised learning model, which uses an occluded autoencoder network as a skeleton and combines a self-attention mechanism to extract the morphological features of neurons from the standard point cloud data.

2. The method for extracting neuron morphological features according to claim 1, characterized in that: In the step 1, the process of converting the neuron morphology data into standard point cloud data includes: Extracting the three-dimensional coordinate points of neuron sampling points from the SWC file storing the neuron morphology data as point cloud raw data; Normalizing the original point cloud data, mapping the coordinates to between -1 and 1 as point cloud normalized data; The point cloud normalized data is sampled, and the farthest point sampling is adopted during sampling to achieve the unification of the number of points and keep the shape of the neuron unchanged, so as to obtain standard point cloud data.

3. The method for extracting neuron morphological features according to claim 2, characterized in that: In the step 2, the process of pre-training the neuron morphological feature self-supervised learning model includes: Step S1, in each network training iteration, extracting N neuron morphological data from the training set according to the batch size and converting them into standard point cloud data; Step S2, grouping the standard point cloud data to obtain a neuron point cloud group representation; Step S3, randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud; Step S4, mark embedding, mapping the occluded neuron point cloud group representation into a continuous numerical vector; Step S5, encoding and decoding are performed through an encoding and decoding network; the encoding and decoding network includes an encoder and a decoder, and the decoder structure is similar to the encoder, both of which are composed of standard Transformer blocks; the encoder only encodes the visible part, and does not need to encode the occluded part; the decoder inputs the occluded part and the visible part encoding, outputs the decoded prediction block features, maps the prediction block features to a vector with the same dimension as the total number of coordinates of the midpoints of the occluded part through a fully connected layer, and then obtains the predicted morphological distribution of the occluded part through a reshape operation; Step S6, using the chamfer distance loss function to calculate the reconstruction loss and train the neuron morphological feature self-supervised learning model; Step S7, determining whether the set training end condition is met; if not, performing the next iterative training according to steps S1 to S6; if so, completing the training of the neuron morphological feature self-supervised learning model.

4. The method for extracting neuron morphological features according to claim 3, characterized in that: In step S2, the process of grouping the standard point cloud data to obtain the neuron point cloud group representation includes: given a neuron standard point cloud data X∈R n×3 , where n represents the number of neuron points, and R represents the real number domain, which means that the input matrix set of size n*3 is real numbers. First, the center point of the neuron is selected, and p center points are sampled from the standard point cloud data through the farthest point sampling. Then, the K nearest neighbor algorithm is used to select k neighboring points for each center point, and the neuron point cloud grouping representation can be obtained.

5. The method for extracting neuron morphological features according to claim 3, characterized in that: In the step S3, the process of randomly occluding the group representation of the neuron point cloud to obtain the occluded group representation of the neuron point cloud includes: adopting a random occlusion design, setting the occlusion ratio to m, randomly selecting m center points, removing the corresponding groups, and obtaining the occluded group representation of the neuron point cloud.

6. The method for extracting neuron morphological features according to claim 3, characterized in that: In step S4, the marker embedding, the process of mapping the occluded neuron point cloud group representation into a continuous numerical vector includes: for the occluded part, replacing the embedding of the occluded part with a weight-sharing learnable mask representation without performing other operations; for the visible part, using PointNet to implement group embedding, the main structure of PointNet consists of MLP and maximum pooling layers, and MLP maps the center coordinates to the embedding marker dimension to provide position guidance for each neuron point cloud group representation.

7. The method for extracting neuron morphological features as claimed in claim 3, characterized in that: In step S6, the chamfer distance loss function is: Among them, L represents the chamfer distance, which calculates the distance from each point in a point set a to the nearest point in another point set b, and takes the average or sum of these distances as the similarity measure of the two point sets, which is used to measure the similarity between the two point sets. P represents the grouped representation of neuronal morphology, gt represents the original input reference, and pre represents predicted reconstruction.

8. The method for extracting neuron morphological features as claimed in claim 3, characterized in that: The step 2 also includes: using the morphological features of the neurons to distinguish different types of neurons: removing the decoder of the neuron morphological feature self-supervised learning model and using pre-trained weights as initial weights for fine-tuning, extracting all neuron morphological local features from the well-trained encoder, aggregating them into global features representing the entire shape through average pooling, and then inputting them into a fully connected network for fine-tuning and classification, so as to achieve the distinction between neuron types.

9. A method for generating a neuron morphological structure, characterized in that: include: Using the neuron morphological feature extraction method as claimed in claim 1 and the following step 3; Step three, morphological structure generation step: using the morphological features extracted in step two to generate the morphological structure of neurons.

10. The method for generating the neuron morphological structure according to claim 9, characterized in that: In the step 3, the process of using the morphological features to generate the morphological structure of neurons includes: Based on the self-supervised learning model of neuron morphological features, the latent space description of the features is modified to represent the values ​​with a probability range rather than a numerical value. The samples similar to the given neurons are sampled and generated by means of a generative model. The method includes three components: encoding the latent distribution, and the output is the mean and standard deviation; reparameterization, in which the latent variables are generated by random sampling, and the sampling process itself is a non-differentiable operation. Reparameterization is performed to convert the non-differentiable operation into a differentiable process; KL divergence constraint, the objective function of VAE is the sum of the reconstruction loss and the KL divergence loss. The KL divergence loss is used to measure the difference between the latent distribution predicted by the encoder and the prior distribution, and the morphological structure of the neuron is obtained by sampling the latent space distribution and decoding.

Citation Information

Cited By

  • Local perception point cloud representation learning pre-training method and device based on attention mechanism

    CN119418143A

  • Local perception point cloud representation learning pre-training method and device based on attention mechanism

    CN119418143B