Rare brain lesion distinguishing method based on meta learning

Abstract

The invention provides a rare brain lesion distinguishing method based on meta-learning, belongs to the technical field of medical information intelligent diagnosis, and solves the problems that traditional artificial intelligence training needs a large amount of clinical diagnosis information, and the traditional artificial intelligence training needs a large amount of clinical diagnosis information from the data level. And the number of samples of some disease patients cannot meet the requirements of traditional neural network training. According to the technical scheme, a data set of a brain nuclear magnetic resonance image is divided into a plurality of tasks, and each task comprises a support set and a query set; training and testing the CNN model on a pre-training task, and obtaining a globally optimal initialization parameter theta through back propagation; fine-tuning the trained model by using a support set in the new category data set; and testing by adopting the query set to obtain evaluation on the model. The method has the beneficial effects that the prior experience of distinguishing common brain diseases is utilized to generate diagnosis opinions for rare diseases, and the generalization ability of the model is remarkably improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical information intelligent diagnosis, and in particular to a rare brain lesion identification method based on meta-learning. Background Art

[0002] As a common method of representing data relationships, network structures are widely used in various studies. The human brain can also be abstracted into a complex network structure by defining nodes and connecting edges. This network is usually referred to as a brain network, and its structure is closely related to human cognitive functions and brain diseases. Analyzing and studying brain networks can provide powerful tools for exploring how the brain works, studying the pathological mechanisms of neurodegenerative diseases, and improving the diagnosis and treatment of mental illnesses and brain damage.

[0003] In recent years, with the rapid development of artificial intelligence technology, especially deep learning, more and more researchers are paying attention to the use of artificial intelligence technology to assist doctors in disease diagnosis. At present, many studies have shown that in disease identification based on imaging data, artificial intelligence technology can reach or exceed the level of human experts in the identification of some diseases. Therefore, using artificial intelligence for analysis in the field of brain networks is an effective way to improve the efficiency of brain disease diagnosis, but traditional artificial intelligence training requires a large amount of clinical diagnostic information. At the data level, the number of samples of patients with some diseases may not meet the needs of traditional neural network training. Wang Kang mentioned in the article "Research on Image Classification Methods Based on Small Sample Learning" that small sample learning does not rely on a large number of training samples, avoids the high cost of labeled data, learns from limited labeled samples and quickly generalizes to new tasks containing a small number of labeled samples, and overcomes the limitation of deep learning relying on large-scale labeled data. Summary of the invention

[0004] The purpose of the present invention is to provide a rare brain lesion identification method based on meta-learning. Therefore, the present invention proposes a brain network analysis method based on meta-learning, which can effectively use small sample brain magnetic resonance imaging to perform brain network analysis, thereby training a model that can accurately judge the patient's disease and improve the efficiency of medical diagnosis.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is specifically: a rare brain lesion identification method based on meta-learning, comprising the following steps:

[0006] Step 1: Build a traditional CNN convolutional neural network model. The model includes 4 convolution blocks as feature extraction modules and 1 fully connected layer as a classifier. Each convolution block includes 1 convolution layer Convld, 1 BatchNormld layer, 1 maximum pooling layer and 1 non-linear activation function ReLU. Then input the data of the support set in each task into the CNN model, fine-tune the parameters, and then use the query set for testing to obtain the classification results.

[0007] Step 2: For the extraction of functional brain network features, let A F ∈R N×N is the Pearson correlation coefficient analysis matrix of the functional brain network. After the convolutional neural network extracts the features, the feature map C is obtained. c ∈R n×n .

[0008] The step 2 comprises the following steps:

[0009] (1) First, the functional activity of the brain region of interest (ROI) is reflected by the signal fluctuation of rs-fMRI. For each sample, use x i =(x i1 ,x i2 ,...,x iM ), i = 1, 2, ..., N to represent the original rs-fMRI signal sequence of the i-th ROI, where M represents the total number of voxels and N represents the number of all ROIs.

[0010] (2) The Pearson correlation coefficient is used to characterize the functional connection strength between two signal sequences. Assume that the functional connection between the i-th ROI and the j-th ROI is represented by ρ ij Indicates, then

[0011] ρ ij =corr(x i , x j ) (3)

[0012] The algorithm for Pearson correlation is as follows:

[0013]

[0014] in, and They all represent the corresponding sample averages.

[0015] (3) For each sample, the functional connections between the paired ROIs are calculated. The static brain network can be expressed in data structure using a symmetric matrix C:

[0016] C=ρ ij, where 1≤i,j≤N and ρ ij =ρ ji ; (5)

[0017] (4) Perform convolution calculation and pooling on each matrix C. After completion, the feature map obtained is C c ∈R n×n .

[0018] Step 3: In the meta-training set, use episode to represent each task. According to the N-way k-shot few-shot learning principle, randomly select N categories from the data set, and then select k samples from each category as the support set from the selected N categories. Samples are used as target sets, totaling m S =N×k; then select b samples from the remaining samples in each of the N categories as the query set Total m Q =N×b, each episode simulates an N-way k-shot few-shot learning task; where N represents the number of classification categories in a few-shot classification task, and k represents the number of samples in each category. On the meta-validation set and meta-test set, each task consists of a support set and a target set, and the selection method is the same as above.

[0019] Step 4: Use MAML to optimize this model. The update formula is as follows:

[0020]

[0021]

[0022] Among them, θ is the meta-learning model parameter, and the adjustment method is gradient descent to find the partial derivative.

[0023] First, the parameters θ are randomly initialized. Then, the parameters are train ,T test} is updated iteratively. i It is T train The parameters after the support set gradient update in each task, is the loss value of the query set in task i. Finally, according to the final parameter θ meta , in T test The model is fine-tuned on the support set in and evaluated on the query set. In addition, a and β are the learning rates when each task is executed and when all tasks are completed, respectively.

[0024] Compared with the prior art, the present invention has the following beneficial effects:

[0025] 1. The small sample brain disease diagnosis method proposed in the present invention utilizes the information in the small sample more accurately than the traditional model. In some scenarios, even when the accuracy is equivalent, the amount of data required by the present invention is greatly reduced compared with the traditional convolutional neural network (CNN) feature extraction and classification method, which not only helps to improve the diagnostic efficiency, but also reduces the burden of data collection and processing.

[0026] 2. This method cleverly combines the diagnostic experience of other common brain diseases as prior information. This comprehensive use of the correlation characteristics between different diseases enables the method to capture the higher-dimensional features produced by brain lesions. By introducing prior information, the present invention effectively reduces the training time of the neural network and improves the overall training efficiency. This feature has significant practical significance for accelerating the diagnosis process of brain diseases, especially in resource-limited or time-sensitive situations. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention.

[0028] Figure 1 It is a structural diagram of the CNN and MAML models in the present invention.

[0029] Figure 2 This is the overall framework diagram of the rare brain lesion identification method based on meta-learning in the present invention.

[0030] Figure 3 Comparison chart of the accuracy of the three evaluation indicators of MAML+CNN in the present invention. DETAILED DESCRIPTION

[0031] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. Of course, the specific embodiments described here are only used to explain the present invention and are not used to limit the present invention.

[0032] Example 1

[0033] See also Figure 1 , Figure 2 The technical solution provided in this embodiment is a rare brain lesion identification method based on meta-learning, which specifically includes the following steps:

[0034] S1. In the meta-training set, each task is represented by episode. According to the N-way k-shot few-shot learning principle, N categories are randomly selected from the data set, and then k samples are selected from each category as the support set. Samples are used as target sets, totaling mS =N×k; then select b samples from the remaining samples in each of the N categories as the query set Total m Q =N×b, each episode simulates an N-way k-shot few-shot learning task; where N represents the number of classification categories in a few-shot classification task, and k represents the number of samples in each category of images. On the meta-validation set and meta-test set, each task consists of a support set and a target set, and the selection method is the same as above.

[0035] S2. Extract the functional brain network features from the original data in the dataset. F ∈R N×N Pearson correlation coefficient analysis matrix for functional brain networks.

[0036] S3, using MAML to train the parameters of the convolutional neural network in task T = {T train ,T test} to obtain the global optimal initialization parameter θ. The θ update formula is as follows:

[0037]

[0038]

[0039] Among them, θ is the meta-learning model parameter, and the adjustment method is gradient descent to find the partial derivative. First, the parameter θ is randomly initialized, and then the parameter is in this batch of tasks T = {T train ,T test} to perform update iterations; θ i 'It's T train The parameters after the support set gradient update in each task, is the loss value of the query set in task i; finally, according to the final parameter θ meta , in T test The model is fine-tuned on the support set in and its query set is used for model evaluation; in addition, α and β are the learning rates when each task is executed and when all tasks are completed, respectively.

[0040] S4: First, input the data of the support set in each task into the CNN learner to fine-tune the parameters. Then use the query set for testing to obtain the classification results.

[0041] Step S2.1 first reflects the functional activity of the brain region of interest (ROI) through the signal fluctuation of rs-fMRI. For each sample, use x i =(x i1 ,x i2 ,...,x iM), i = 1, 2, ..., N to represent the original rs-fMRI signal sequence of the i-th ROI, where M represents the total number of voxels and N represents the number of all ROIs.

[0042] Step S2.2 uses the Pearson correlation coefficient to characterize the functional connection strength between two signal sequences. Assume that the functional connection between the i-th ROI and the j-th ROI is represented by ρ ij Indicates, then

[0043] ρ ij =corr(x i , x j ) (3)

[0044] The specific algorithm of Pearson correlation is as follows:

[0045]

[0046] in and They all represent the corresponding sample averages.

[0047] Step S2.3 calculates the functional connections between pairs of ROIs for each sample. The static brain network can be expressed in data structure using a symmetric matrix C:

[0048] C=ρ ij , where 1≤i,j≤N and ρ ij =ρ ji ;; (5)

[0049] The brain MRI datasets used in the experiment are selected from the ADNI Alzheimer's disease dataset, COBRE schizophrenia dataset, UAIS laboratory public depression research dataset, ABIDE autism dataset, and NITRC ADHD dataset. The ADNI Alzheimer's disease dataset and COBRE schizophrenia dataset are used in the 2-way 5-shot experiment, and the above five major categories of datasets are used in the 5-way 1-shot and 5-way 5-shot experiments.

[0050] The raw data was preprocessed, including time layer correction, head motion correction, registration, segmentation, and normalization. The processed data was imported into the 2-way 5-shot, 5-way 1-shot, and 5-way 5-shot experiments, as shown in the following table.

[0051]

[0052] The method of this embodiment achieved accuracies of 0.87, 0.6, and 0.7 in 2-way 5-shot, 5-way 1-shot, and 5-way 5-shot experiments, respectively. Compared with traditional CNN, the accuracy of this method is improved in various cases.

[0053] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principle of the present invention should be included in the protection scope of the present invention.

Claims

1. A rare brain lesion identification method based on meta-learning, characterized in that: The following steps are involved: S1. On the meta-training set and the meta-test set, use episode to represent each task, and extract the support set and target set of each task from the original data; S2. For the extraction of functional brain network features, let A F ∈R N×N is the Pearson correlation coefficient analysis matrix of the functional brain network. After the convolutional neural network extracts the features, the feature map C is obtained. c ∈R n×n ; S3, construct a CNN convolutional neural network model, and use CNN to extract local spatial features from the feature map obtained in S2; S4. Use MAML to optimize this model.

2. The rare brain lesion identification method based on meta-learning according to claim 1, characterized in that: In the step S1, specifically: According to the N-way k-shot few-shot learning principle, N categories are randomly selected from the data set, and then k samples are selected from each category as the support set. Samples are used as target sets, totaling m S = N × k, select b samples from the remaining samples in each of the N categories as the query set Total m Q =N×b, each episode simulates an N-way k-shot few-shot learning task; where N represents the number of classification categories in a few-shot classification task, and k represents the number of samples in each category of images. On the meta-validation set and meta-test set, each task consists of a support set and a target set, and the selection method is the same as above.

3. The rare brain lesion identification method based on meta-learning according to claim 1, characterized in that: The step S2 includes the following steps: S21, the functional activity of the brain region of interest is reflected by the signal fluctuation of rs-fMRI. For each sample, x i =(x i1 ,x i2 ,...,x iM ), i = 1, 2, ..., N to represent the original rs-fMRI signal sequence of the i-th ROI, where M represents the total number of voxels and N represents the number of all ROIs; S22. The Pearson correlation coefficient is used to characterize the functional connection strength between two signal sequences. Assume that the functional connection between the i-th ROI and the j-th ROI is represented by ρ ij express: r ij =corr(x i ,x j ) (3) The algorithm for Pearson correlation is as follows: in, and All represent the corresponding sample averages; S23. For each sample, the functional connections between the paired ROIs are calculated. The static brain network is expressed in the data structure using a symmetric matrix C: C = ρ ij , where 1 ≤ i, j ≤ N and ρ ij = ρ ji ; (5) S24, perform convolution calculation and pooling on each matrix C, and the feature map obtained after completion is C c ∈R n×n .

4. The rare brain lesion identification method based on meta-learning according to claim 1, characterized in that: In the step S3, specifically: The model consists of multiple convolution blocks as feature extraction modules and a single fully connected layer as a classifier. Each convolution block consists of a convolution layer Convld, a BatchNormld layer, a maximum pooling layer and a non-linear activation function ReLU. Each convolution block is connected by a maximum pooling layer.

5. The rare brain lesion identification method based on meta-learning according to claim 1, characterized in that: In step S4, the optimization of the model using MAML is specifically as follows: The update formula is as follows: Among them, θ is the meta-learning model parameter, and the adjustment method is gradient descent to find the partial derivative. The parameter θ is randomly initialized. The parameter is in the batch of tasks T = {T train ,T test }, θ i ' is the meta-training set T train The parameters after the support set gradient update in each task, is the loss value of the query set in task i, in T train After the above update iteration is completed, we get θ meta , and then the parameter θ meta In T test , and the query set is used for model evaluation. α and β are the learning rates when each task is executed and when all tasks are completed, respectively.

Images