Cryoelectronic tomography few-sample classification method and system

Abstract

According to the small-sample classification method and system for the chilled electron tomography, discrete wavelet transform is introduced, the problem that noise of a chilled electron tomography scene is serious is solved, main information in data can be effectively reserved through the discrete wavelet transform, a high-frequency part containing noise can be removed, and preliminary noise reduction is achieved; self-distillation is improved, the model is iteratively optimized through KL divergence, when a large amount of complex noise exists in Cryo-ET data, the generalization ability of the model in a feature space can be enhanced through distillation of a multi-generation model, and the negative influence of the noise on feature extraction is effectively reduced. Besides, the model is gradually optimized in a self-iteration mode, so that the model can better adapt to learning tasks with few samples, and efficient classification performance is achieved under limited annotation data.

Description

Technical Field

[0001] The present invention belongs to the technical field related to cryo-electron tomography classification, and in particular relates to a cryo-electron tomography few-sample classification method and system. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] Cryo-electron tomography (Cryo-ET) has been widely used and developed in recent years as an important visualization technology for observing macromolecular structures in natural cellular environments in the biomedical field. However, in the process of subtomogram averaging (STA), the classification of macromolecules has always been a key factor restricting its accuracy and function prediction.

[0004] Currently, macromolecule classification methods are mainly divided into two categories: template matching methods and deep learning-based methods. Template matching methods classify macromolecules by calculating the constrained cross-correlation between data and known structures. However, such methods are highly dependent on templates, have reference-based biases, and are sensitive to noise, resulting in limited classification results. Although deep learning-based methods have achieved significant improvements in classification accuracy, when encountering macromolecule categories that have not been seen during training, data needs to be re-labeled and the model needs to be retrained, which not only brings extremely high time costs, but also limits its flexibility in practical applications.

[0005] In order to deal with the above problems, few-shot learning methods have emerged. These methods pre-train models on large-scale datasets and quickly adapt to new tasks using a small number of labeled samples, thereby achieving the recognition of new categories of macromolecules. However, most of the existing few-shot classification methods are based on metric learning, which learns image representations through similarity metrics. They mainly rely on statistical moments and have difficulty in effectively modeling high-dimensional feature distributions, which limits the performance of the model.

[0006] Specifically, existing few-shot classification methods, such as ProtoNet-CE, represent the category prototype by the mean vector in the embedding space and perform classification by calculating the Euclidean distance between the query sample and the prototype. However, this method only utilizes the marginal distribution information and ignores the joint distribution relationship between features, resulting in the model's inadequacy in capturing feature dependencies. In addition, although few-shot domain adaptation methods reduce distribution differences by comparing the covariance matrix between the source domain and the target domain, they also do not fully utilize the joint distribution information, limiting the generalization ability of the model.

[0007] In addition, the noise problem is particularly serious in cryo-electron tomography (Cryo-ET) scenarios. Cryo-ET data usually contain a large amount of high-frequency noise due to imaging under low temperature conditions and electron beam radiation. These noises are manifested as random shot noise and equipment noise, including noise introduced by electromagnetic interference during the imaging process of electron microscopes, randomness caused by low-dose radiation, and noise pollution caused by uneven sample quality under freezing conditions. These noises greatly affect the quality of the image, making feature extraction difficult, making the classification and identification of macromolecules more challenging. The impact of this high-frequency noise on Cryo-ET data is mainly reflected in the following aspects: Feature blurring: Noise will mask important details in the data, making the features of macromolecules blurred and difficult to be correctly extracted by neural networks. Increased classification error: Due to the interference of noise, the model may learn incorrect features from the noise, resulting in a decrease in the final classification accuracy. Extended training time: Noise interference will make it more difficult for the model to converge during training, resulting in increased training time, especially when the amount of data is limited. This effect is particularly significant.

[0008] In summary, there are at least the following problems in the classification of cryo-electron tomography images: 1. The joint distribution relationship between features is ignored, resulting in the model's inability to capture feature dependencies; 2. There is a large amount of high-frequency noise in the data, which causes problems such as feature blurring, increased classification errors, and prolonged model training. Summary of the invention

[0009] To overcome the shortcomings of the above-mentioned prior art, the present invention provides a cryo-electron tomography few-sample classification method and system, which enhances noise robustness by retaining only low-frequency components through DWT; the improved self-distillation technology can enhance the generalization ability of the model in the feature space and further reduce the negative impact of noise on feature extraction.

[0010] In order to achieve the above object, the present invention adopts the following technical solution: In a first aspect, the present invention provides a cryo-electron tomography few-sample classification method, comprising: Obtain a training set of cryo-electron tomography images; The student model is pre-trained by using the knowledge distillation algorithm, and the pre-trained student model is used as the teacher model. The teacher model and the student model are iteratively trained by using the self-distillation technology and the KL divergence to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; Decomposing the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; The low-frequency feature map is input into the trained student model to obtain the classification result of the cryo-electron tomography image to be classified.

[0011] In a second aspect, the present invention provides a cryo-electron tomography few-sample classification system, comprising: An acquisition module is configured to: acquire an image training set of cryo-electron tomography; A distillation module is configured to: pre-train a student model using a knowledge distillation algorithm, use the pre-trained student model as a teacher model, and iteratively train the teacher model and the student model using KL divergence through self-distillation technology to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; A decomposition module is configured to: decompose the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; A classification module is configured to: input the low-frequency feature map into a trained student model to obtain a classification result of the cryo-electron tomography image to be classified.

[0012] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0013] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0014] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0015] One or more of the above technical solutions have the following beneficial effects: In the present invention, discrete wavelet transform is introduced to solve the serious noise problem in cryo-electron tomography scenes. Discrete wavelet transform can effectively retain the main information (low-frequency part) in the data and remove the high-frequency part containing noise to achieve preliminary noise reduction. Compared with traditional convolution operations, the multi-scale characteristics of DWT can more fully describe the different levels of data features, making the model more robust when extracting features. DWT operations can also maintain the spatial consistency of the image well, thereby avoiding information loss and distortion.

[0016] In the present invention, self-distillation is improved, and the model is iteratively optimized through KL divergence. When faced with a large amount of complex noise in Cryo-ET data, the generalization ability of the model in the feature space can be enhanced through the distillation of multiple generations of models, effectively reducing the negative impact of noise on feature extraction. In addition, this method gradually optimizes the model through self-iteration, so that the model can better adapt to the learning task of a small number of samples, thereby achieving efficient classification performance under limited labeled data.

[0017] In the present invention, the model adopts a deep Brownian distance covariance-prototype network. The deep Brownian distance covariance module can solve the problem of incomplete feature modeling, improve the accuracy and stability of the classification model, and reduce the cost of retraining.

[0018] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 This is a flow chart of a macromolecule small sample classification test in Example 1 of the present invention; Figure 2 This is a flowchart of Cryo-ET few-sample classification in Example 1 of the present invention; Figure 3 Schematic diagram of the three-stage flow chart of the model in Embodiment 1 of the present invention. DETAILED DESCRIPTION

[0021] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0022] It should be noted that the terms used herein are for describing specific embodiments only and are not intended to be limiting of exemplary embodiments according to the present invention.

[0023] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0024] Embodiment 1 This embodiment discloses a cryo-electron tomography few-sample classification method, including: Obtain a training set of cryo-electron tomography images; The student model is pre-trained by using the knowledge distillation algorithm, and the pre-trained student model is used as the teacher model. The teacher model and the student model are iteratively trained by using the self-distillation technology and the KL divergence to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; Decomposing the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; The low-frequency feature map is input into the trained student model to obtain the classification result of the cryo-electron tomography image to be classified.

[0025] Combine the following Figure 1-Figure 3 The present embodiment is described in detail: In this embodiment, the model uses a deep Brownian distance covariance-prototype network, which aims to solve the problem of incomplete feature expression caused by relying only on marginal distribution in the prior art by introducing joint distribution modeling, thereby improving the accuracy and stability of the classification model. At the same time, the method of this embodiment can quickly adapt to new tasks with the support of a small amount of annotated data, reduce time costs, and enhance the model's ability to recognize new macromolecules.

[0026] This embodiment includes three main stages: a pre-training stage, a knowledge distillation stage, and an adaptation and classification stage.

[0027] In the pre-training stage, the classification model is pre-trained using a large-scale annotated macromolecular dataset. The feature extraction module uses DWT and Transformer to extract effective high-dimensional feature representations from Cryo-ET subvolume data. A deep Brownian distance covariance (BDC) module is added after feature extraction to model the joint distribution of features in the embedding space. The BDC module calculates the Euclidean distance matrix between features to obtain the Brownian distance covariance matrix, and after triangulation and vectorization, outputs the BDC vector for subsequent classification tasks.

[0028] Specifically, in the pre-training stage, after the feature tensor is extracted by DWT and Transformer, it first enters the BDC module for dimensionality reduction. Then, the feature matrix is ​​flattened, and the Euclidean distance matrix between each feature is calculated to further obtain the Brown distance covariance matrix (BDC). The BDC matrix is ​​triangulated and vectorized, and finally the BDC vector is output. After the BDC vector is calculated, the similarity between the input data and its prototype is calculated.

[0029] In the feature extraction stage, the common maximum pooling, mean pooling and strided convolution operations are replaced by discrete wavelet transform (DWT) to decompose the feature map into low-frequency and high-frequency components. The low-frequency components contain the main information, such as the basic object structure, while the high-frequency components usually contain most of the noise. Therefore, only the low-frequency components are retained in the subsequent feature extraction to enhance noise robustness.

[0030] This embodiment combines the denoising and feature extraction methods of DWT and Transformer to improve the noise robustness and classification performance of the model. In the implementation process, the feature map is first decomposed into a low-frequency component and three high-frequency components (horizontal, vertical and diagonal high-frequency components) through DWT. In this process, DWT, as a multi-resolution analysis method, can effectively separate the structural information and noise information in the image. The low-frequency components mainly retain the basic structure and main features of the image, while the high-frequency components contain most of the noise and detail information. In the feature extraction stage of the present invention, we discard these high-frequency components and focus on the processing of low-frequency components, thereby reducing the interference of noise on the model.

[0031] Specifically, in the downsampling process, the traditional maximum pooling, mean pooling and strided convolution operations are replaced by DWT decomposition. Maximum pooling and mean pooling operations are prone to loss of feature information, especially in the presence of noise, these operations may further amplify the impact of noise on feature representation. DWT effectively retains the key structural features in the image by separating low-frequency and high-frequency information, while suppressing the propagation of noise. This alternative approach enables the feature extraction module to process noisy data more robustly.

[0032] The advantage of DWT is that it can decompose data at multiple resolution levels, providing a flexible feature representation method. In this way, DWT can effectively retain the main information (low-frequency part) in the data and remove the high-frequency part containing noise, thereby achieving a preliminary noise reduction effect at the input stage of the neural network. Compared with traditional convolution operations, the multi-scale characteristics of DWT can more fully describe the characteristics of different levels of data, making the model more robust when performing feature extraction. In addition, the decomposition and inverse transform (IDWT) operations of DWT can also maintain the spatial consistency of the image well, thereby avoiding information loss and distortion.

[0033] In order to further enhance the suppression of noise, the Transformer structure is combined, and the low-frequency features after noise reduction are input into the Transformer module for feature modeling. Transformer is good at capturing long-range dependencies and can globally model input features, thereby further improving the expressiveness of features and the robustness of the model to noise. In Cryo-ET data, macromolecular structures usually have complex spatial relationships, and Transformer can effectively model these complex relationships through the self-attention mechanism, so that the model can still accurately capture important features when facing noise interference.

[0034] Transformer's self-attention mechanism enables each feature point to pay attention to other feature points, which is particularly important for the complex interactions between features in Cryo-ET data. During the feature extraction process, the presence of noise may distort local features, and Transformer's global modeling capability can effectively integrate the information of the entire image, thereby suppressing the negative impact of local noise. This global modeling allows the model to extract macromolecular structures while avoiding noise interference as much as possible, thereby improving the overall classification performance of the model.

[0035] Specifically, the Multi-Head Self-Attention mechanism is used to enable the model to capture the relationship between low-frequency features from different "attention angles", thereby enhancing the diversity and expression of features. Combined with Positional Encoding, it can ensure that the Transformer retains spatial information when processing these features, so that the model can still accurately distinguish features at different positions in complex structure scenarios.

[0036] In addition, DWT can effectively reduce the impact of noise on the final classification results during the feature extraction process of this embodiment. By removing the high-frequency noise portion of the feature map and using Transformer to further model the global relationship of low-frequency features, the classification model of this embodiment is able to extract more stable and representative high-level features, thereby improving robustness to noise and classification accuracy. During the experiment, the performance of the classification model integrating DWT and Transformer on SHREC and synthetic datasets was verified. The results showed that this method can significantly improve the classification accuracy under noisy conditions, especially on test data with high noise interference.

[0037] DWT can effectively improve the performance of convolutional neural networks under noise interference. By removing the high-frequency noise part in the feature map, DWT can retain the main low-frequency structural features, making the model more robust and improving the classification accuracy. At the same time, Transformer can globally model the features through its self-attention mechanism, which is particularly suitable for processing complex macromolecular structural relationships. Experimental results also show that the convolutional network combining DWT and Transformer has achieved significant performance improvements on multiple noise interference data sets (such as ImageNet-C). Therefore, this method is used to enhance the performance of this embodiment in a noisy environment, especially in a low signal-to-noise ratio scenario such as cryo-electron tomography (Cryo-ET). The introduction of DWT and Transformer greatly improves the robustness and accuracy of classification.

[0038] In the knowledge distillation stage, an improved self-distillation technology is introduced to further enhance the generalization ability of the model. In this embodiment, the traditional knowledge distillation method is improved, and the same network architecture is used for self-distillation, that is, the teacher model and the student model share the same architecture during the training process. Specifically, the student model is first pre-trained, and then in the self-distillation stage, the pre-trained student model is used as the teacher model, and the student model is optimized by a weighted combination of Kullback-Leibler (KL) divergence and traditional loss. During multiple distillation processes, the student model is gradually improved, and the robustness to complex noise is continuously improved.

[0039] The advantage of self-distillation in this embodiment is that the performance of the model can be improved without changing the model structure, especially when facing a large amount of complex noise in Cryo-ET data. Through the distillation of multiple generations of models, the generalization ability of the model in the feature space can be enhanced, and the negative impact of noise on feature extraction can be effectively reduced. In addition, this method can also gradually optimize the model through self-iteration, so that the model can better adapt to the learning task of a small number of samples, thereby achieving efficient classification performance under limited labeled data.

[0040] First, the introduction of self-distillation technology makes the model no longer rely on a single teacher-student model structure, but by transferring knowledge from the same model at different stages, the model is continuously enhanced in the process of self-optimization. Through the iteration of each generation of distillation, the feature extraction ability of the model has been significantly improved, so that it can more stably extract effective features when facing high-noise data. This generation-by-generation distillation strategy allows the model to better adapt to the complexity and uncertainty of Cryo-ET data, especially in low signal-to-noise ratio scenarios. The model gradually enhances its robustness to noise through continuous self-learning and improvement.

[0041] Secondly, during the distillation process, by using KL divergence as part of the loss function, it is ensured that in each generation of knowledge transfer, the student model not only needs to imitate the output of the teacher model, but also needs to learn the teacher model's understanding of the feature distribution. This dual constraint allows the model to maintain the accuracy of the overall output while capturing deeper relationships at the level of feature representation during feature learning. This is particularly important for modeling complex spatial relationships in Cryo-ET data, as it helps the model better distinguish different macromolecular structures, even in the presence of high noise.

[0042] In addition, the advantages of self-distillation technology are also reflected in its adaptability to few-sample learning. Since labeled data is usually very scarce in Cryo-ET tasks, the model needs to be able to maintain high classification performance with a small amount of labeled data. Through multiple generations of self-distillation, the model of this embodiment can gradually accumulate and strengthen the understanding of macromolecular characteristics, reduce dependence on large amounts of labeled data, and enable it to still achieve better classification results in few-sample scenarios. Compared with traditional knowledge distillation methods, the self-distillation process of this embodiment does not need to rely on external complex teacher models, but simplifies the training process and improves efficiency through iterative optimization of the model itself.

[0043] The improved self-distillation technology enhances the generalization ability of the model and the depth of feature learning through multiple iterations, enabling it to better cope with the noise and complexity in Cryo-ET data. This process combines the loss constraint of KL divergence and the strategy of generation-by-generation optimization, so that the model can significantly improve the robustness to noise and the ability to distinguish complex macromolecular structures without adding additional network complexity. In this way, the model of this embodiment can still maintain efficient and stable classification performance in low signal-to-noise ratio and high complexity scenarios.

[0044] In the adaptation and classification phase, a small amount of labeled data of new categories is used for fine-tuning. A logistic regression classifier is used to classify the low-dimensional features of the support set data, thereby achieving prediction of the query set data. In the adaptation and validation phase, a meta-learning algorithm is used for model adaptation. The model after G-generation distillation is used to calculate the Brownian distance covariance (BDC) vectors of the support set and query set, and the support set data and its labels are used to train the logistic regression (LR) classifier, which is then used to classify the query set data.

[0045] Therefore, by combining the deep Brownian distance covariance (BDC) module, DWT denoising and Transformer global modeling, this embodiment effectively improves the model's robustness to noise and classification stability in the task of classifying macromolecules with a small number of samples. In this way, the method of this embodiment not only achieves accurate modeling of joint distribution in embedded feature modeling, but also ensures efficient and stable classification performance under noise interference.

[0046] In addition, the present invention also proposes alternative classifier solutions, such as support vector machines or random forests, to adapt to different application scenarios.

[0047] Embodiment 2 The purpose of this embodiment is to provide a cryo-electron tomography few-sample classification system, including: An acquisition module is configured to: acquire an image training set of cryo-electron tomography; A distillation module is configured to: pre-train a student model using a knowledge distillation algorithm, use the pre-trained student model as a teacher model, and iteratively train the teacher model and the student model using KL divergence through self-distillation technology to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; A decomposition module is configured to: decompose the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; A classification module is configured to: input the low-frequency feature map into a trained student model to obtain a classification result of the cryo-electron tomography image to be classified.

[0048] In further embodiments, there is also provided: An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, no further description is given here.

[0049] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0050] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0051] A computer-readable storage medium is used to store computer instructions. When the computer instructions are executed by a processor, the method described in embodiment 1 is completed.

[0052] The method in the first embodiment can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it will not be described in detail here.

[0053] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in the first embodiment is implemented.

[0054] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0055] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0056] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, etc. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, etc.

[0057] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0058] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A cryo-electron tomography few-sample classification method, characterized in that: include: Obtain a training set of cryo-electron tomography images; The student model is pre-trained by using the knowledge distillation algorithm, and the pre-trained student model is used as the teacher model. The teacher model and the student model are iteratively trained by using the self-distillation technology and the KL divergence to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; Decomposing the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; The low-frequency feature map is input into the trained student model to obtain the classification result of the cryo-electron tomography image to be classified.

2. A cryo-electron tomography few-sample classification method according to claim 1, characterized in that: Before inputting the low-frequency feature map into the trained student model, it also includes using Transformer to perform global feature modeling on the low-frequency feature map, and inputting the modeled global features into the trained student model.

3. A cryo-electron tomography few-sample classification method according to claim 1, characterized in that: The pre-training process of the deep Brownian distance covariance-prototype network is: Use the deep Brownian distance covariance module to reduce the dimension of the input features and calculate the Euclidean distance between each feature to obtain the Euclidean distance matrix; The calculated Euclidean distance matrix is ​​triangulated and quantized to obtain the deep Brownian distance covariance vector; The similarity between the deep Brownian distance covariance vector and the deep Brownian distance covariance-prototype network input feature is calculated, and the deep Brownian distance covariance-prototype network is trained with a small sample according to the calculated similarity.

4. The cryo-electron tomography few-sample classification method according to claim 1, characterized in that: A meta-learning framework is adopted to calculate the Brownian distance covariance vectors of the support set and the query set for the deep Brownian distance covariance-prototype network after self-distillation, and the logistic regression classifier is trained using the support set data and its corresponding labels for classification of the query set data.

5. The cryo-electron tomography few-sample classification method according to claim 1, characterized in that: The student model is optimized by a weighted combination of KL divergence and traditional loss functions.

6. A cryo-electron tomography few-sample classification system, characterized in that: include: An acquisition module is configured to: acquire an image training set of cryo-electron tomography; A distillation module is configured to: pre-train a student model using a knowledge distillation algorithm, use the pre-trained student model as a teacher model, and iteratively train the teacher model and the student model using KL divergence through self-distillation technology to obtain a trained student model; wherein the student model adopts a deep Brownian distance covariance-prototype network; A decomposition module is configured to: decompose the cryo-electron tomography image to be classified by discrete wavelet transform to obtain a low-frequency feature map; A classification module is configured to: input the low-frequency feature map into a trained student model to obtain a classification result of the cryo-electron tomography image to be classified.

7. A cryo-electron tomography few-sample classification system as claimed in claim 6, characterized in that: Also includes: Features A modeling module is configured to use Transformer to perform global feature modeling on the low-frequency feature map.

8. An electronic device, characterized in that: The method comprises a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 5 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 5.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 5.

Images