A method for analyzing magnetic resonance images of the brain

Abstract

This invention relates to the field of three-dimensional magnetic resonance imaging (MRI) data processing technology for the brain, and more particularly to a method for analyzing MRI images of the brain. The method includes slicing 3D brain data into multiple views and extracting features from the 2D images; inputting an attention module to reflect the importance of different brain regions in the recognition contribution in the feature map; calculating cross-entropy classification loss to classify whether the subject has the disease; and simultaneously calculating Triplet Loss to distinguish whether multiple slices belong to the same subject. This invention not only helps reduce the number of network parameters and computational load, but also alleviates the overfitting problem caused by the small sample size of Alzheimer's disease datasets; it extracts interrelated features within different views, allowing for the use of diverse information to explain the intrinsic nature of Alzheimer's disease; it ensures consistent classification results for multiple 2D slice images of the same subject, while preserving and enhancing subtle intra- and inter-class feature information.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional magnetic resonance imaging (MRI) data processing technology for the brain, and more particularly to a method for analyzing MRI images of the brain. Background Technology

[0002] Existing processing methods for three-dimensional magnetic resonance imaging (MRI) data related to Alzheimer's disease include:

[0003] (1) Two-dimensional convolutional neural networks extract abstract feature representations from two spatial dimensions (x, y), and the output is a two-dimensional convolutional space, which can capture the inherent feature information in two-dimensional images.

[0004] (2) ShuffleNet is a computationally efficient convolutional neural network model that employs pointwise group convolution, depthwise separable convolution, and channel rearrangement methods, which significantly reduces the computational cost of the model while maintaining accuracy. It is an effective means of maintaining the best model accuracy under limited computing resources.

[0005] (3) Efficient Channel Attention (ECA) is a non-dimensionality-reducing local cross-channel interaction strategy and an adaptive method for selecting the size of a one-dimensional convolution kernel to determine the coverage of local cross-channel interactions. It enables the network to achieve significant performance gains with only a small increase in network parameters.

[0006] (4) Triplet Loss, which minimizes intra-class distance and maximizes inter-class distance while ensuring correct classification. It can make features of samples of the same class cluster together and features of samples of different classes move away from each other, thereby improving the robustness of the model.

[0007] However, the above-mentioned existing technologies have the following technical problems:

[0008] (1) Three-dimensional magnetic resonance data is complex. Considering the spatial complexity of structural magnetic resonance image data, one approach is to directly use a three-dimensional convolutional network to extract features from the spatial structure image. However, Alzheimer's disease magnetic resonance images are mainly composed of extremely subtle intra-class and inter-class features. Direct three-dimensional convolution to extract features would pose a significant challenge to the subsequent high-dimensional abstract feature learning of the network, and would also require high computational complexity and memory cost. Another approach is to slice the three-dimensional magnetic resonance image into two-dimensional images and then use a two-dimensional convolutional neural network to extract features. However, this approach ignores the inherent spatial advantages of magnetic resonance data, which would prevent the full utilization of the unique spatial feature information of the magnetic resonance data itself.

[0009] (2) Deep convolutional neural networks usually require a large number of samples to fit the model. However, compared with the existing million-sample natural image dataset, the Alzheimer's disease dataset has relatively few samples, which limits the ability of convolutional neural networks to effectively learn image features and leads to overfitting. In addition, studies have shown that in typical small-sample neuroimaging datasets, simple deep models often outperform more complex deep models by reducing overfitting.

[0010] (3) The magnetic resonance data of Alzheimer’s disease exhibits extremely subtle intra-class and inter-class differences. Using only a simple convolutional neural network cannot effectively focus on the feature space information, which will lead to insufficient feature extraction, noise, and task irrelevance. As a result, important subtle feature information in the data, especially intra-class features, is not enhanced, resulting in poor classification performance.

[0011] (4) Predicting Alzheimer’s disease by extracting two-dimensional slices from three-dimensional magnetic resonance data first ignores spatial dimension information; second, when the three-dimensional magnetic resonance image of the same subject is sliced ​​into multiple two-dimensional images and sent to the network for training, it is impossible to guarantee that the classification results of multiple two-dimensional slice images of the same subject are consistent, which leads to information leakage problem.

[0012] To address the aforementioned technical problems, we propose a method for analyzing magnetic resonance images of the brain according to the present invention. Summary of the Invention

[0013] The purpose of this invention is to provide a method for analyzing magnetic resonance images of the brain to solve the technical problems mentioned in the background art.

[0014] To achieve the above objectives, the present invention adopts the following technical solution:

[0015] The present invention also provides a method for analyzing magnetic resonance images of the brain, the method comprising:

[0016] ShuffleNet V1 was used as the backbone network;

[0017] Multi-view angular slicing was performed on 3D brain data, and features were extracted from 2D images at different angles.

[0018] The extracted feature maps are input into the channel attention module, which reflects the importance of different brain regions in recognition.

[0019] The cross-entropy classification loss is calculated to classify whether a subject is ill; at the same time, the Triplet Loss is calculated to distinguish whether multiple slices belong to the same subject.

[0020] Furthermore, the multi-view angular slicing of 3D brain data mainly includes:

[0021] The 3D brain data was sliced ​​from three angles: coronal, sagittal, and transverse.

[0022] Furthermore, the method for constructing the channel attention module is as follows:

[0023] The results of the Shuffle basic unit in the last layer of the backbone network are globally averaged and pooled.

[0024] Local cross-channel interaction information is processed for each channel and the two adjacent channels to obtain a spatial attention feature map, which is then normalized using the sigmoid activation function.

[0025] Furthermore, it is characterized by assuming that the output of the Shuffle basic unit in the last layer of the backbone network is X∈R W×H×C W is the image width, H is the image height, and C is the number of channels. The channel attention is defined as follows:

[0026] ω=σ(C1D k (y))

[0027] Where σ represents the sigmoid activation function, C1D k This represents a one-dimensional convolution with a kernel size of k; y = g(X). This represents a global average pooling operation on the output result x.

[0028] Furthermore, the calculation of cross-entropy classification loss to classify whether a subject is ill involves the following steps:

[0029] The cross-entropy loss branch flattens and fully connects the result after the channel attention module;

[0030] The Triplet Loss calculation is used to distinguish whether multiple slices belong to the same subject. The steps are as follows:

[0031] The Triplet Loss branch first passes the ECA result through three 1×1 convolutional layers to reduce the number of channels and ReLU activation function layers, then flattens it, performs a fully connected layer, and finally performs L2 paradigm operations.

[0032] Furthermore, the cross-entropy classification loss formula is defined as follows:

[0033]

[0034] Where N represents the number of samples, The output of the sigmoid function represents the probability that the label of the current sample is 1. y represents the probability that the current sample label is 0, and y represents the label value 0 or 1.

[0035] The Triplet Loss formula is defined as follows:

[0036] L T =max(d(a,p)-d(a,n)+margin,0)

[0037] Where a is an anchor example, p is a positive example, n is a negative example, and d is the Euclidean distance metric;

[0038] The loss function of the backbone network is L = L C +L T .

[0039] The present invention also provides a magnetic resonance imaging analysis system for the brain, wherein the system selects ShuffleNet V1 as the backbone network;

[0040] The system includes:

[0041] The slicing module is used to slice 3D brain data from multiple view angles.

[0042] The feature extraction module is used to extract feature information from multiple views after slicing;

[0043] The channel attention module is used to reflect the importance of different brain regions in recognition in the feature map;

[0044] The Deep Metrics Learning module is used to calculate cross-entropy classification loss and Triplet Loss.

[0045] The present invention has at least the following beneficial effects:

[0046] (1) In order to reduce computational complexity and memory cost and avoid directly using three-dimensional convolution to extract high-dimensional features, this invention uses a two-dimensional convolutional neural network for feature extraction and uses ShuffleNet V1 as the backbone network. This not only helps to reduce the number of network parameters and computational cost, but also alleviates the overfitting problem caused by the small sample size of the Alzheimer's disease dataset while ensuring model performance.

[0047] (2) In order to improve the problem that convolutional neural networks cannot effectively focus on feature space information under small sample datasets, resulting in a limited amount of feature information extracted, this invention adopts a multi-view approach to slice 3D Alzheimer's disease image data from three angles: coronal, sagittal, and transverse, and extract 2D image features respectively, so that interrelated and complementary features can be extracted in different views, enabling the model to use sufficient and diverse information to explain the intrinsic nature of Alzheimer's disease.

[0048] (3) This invention adds an ECA module to the backbone network of ShuffleNet V1, which helps to highlight the differences in brain region features, unlike other attention mechanisms that can only focus on the feature distribution of a single channel. It avoids the negative impact of dimensionality reduction on channel attention classification prediction and effectively captures information from cross-channel interactions. Significant performance gains can be achieved with only a small increase in network parameters.

[0049] (4) This invention combines cross-entropy classification loss with Triplet Loss to form the loss function of the network model. It can classify whether a two-dimensional slice image is diseased, ensuring that the classification results of multiple two-dimensional slice images of the same subject are consistent, and can maximize the inter-class distance, minimize the intra-class distance, and retain and enhance the corresponding extremely subtle intra-class and inter-class feature information. Attached Figure Description

[0050] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic diagram of feature extraction from magnetic resonance images using a lightweight network based on ShuffleNet V1.

[0052] Figure 2 Diagram of the ECA attention mechanism structure;

[0053] Figure 3 Here is a structural diagram of the deep metric learning module;

[0054] Figure 4 This is a diagram of the overall network structure. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0056] Example 1:

[0057] Reference Figure 1-4 This invention discloses a method and system for analyzing magnetic resonance images of the brain, comprising:

[0058] ShuffleNet V1 was used as the backbone network;

[0059] Multi-view angular slicing of 3D brain data and feature extraction of 2D images were performed separately.

[0060] The extracted feature maps are input into the channel attention module, which reflects the importance of different brain regions in recognition.

[0061] The cross-entropy classification loss is calculated to classify whether a subject is ill; at the same time, the Triplet Loss is calculated to distinguish whether multiple slices belong to the same subject.

[0062] Specifically:

[0063] (1) Feature extraction of magnetic resonance images based on ShuffleNet V1 multi-view lightweight convolutional neural network

[0064] See Figure 1 In the lightweight convolutional neural network design: a two-dimensional ShuffleNet V1 is used as the backbone network, which mainly consists of multiple Shuffle downsampling units and Shuffle basic units. The basic units include 1×1 group convolutions, 3×3 depthwise separable convolutions, and Channel Shuffle. The downsampling unit consists of two branches: the basic unit and a 3*3 average pooling layer with a stride of 2. The number of channels is 24, 144, 288, and 576, respectively. After each convolutional layer, a BatchNorm layer and a ReLU activation function layer are set. In order to avoid overfitting and increase the receptive field, a downsampling operation with a stride of 2×2 is added after all other parts except the basic unit.

[0065] In the design of the multi-view feature extraction module: the feature extraction module has three branches, each composed of the same sub-network. It adopts a multi-view approach, slicing the 3D Alzheimer's disease image data from three angles: coronal, sagittal, and transverse. Then, the 2D images of the three planes are input into the three branches for feature extraction, so that interrelated and complementary features are extracted in different views, thereby improving the classification effect on the dataset. The model can use sufficient and diverse information to explain the intrinsic nature of Alzheimer's disease.

[0066] (2) ECA attention mechanism

[0067] See Figure 2Alzheimer's disease is highly correlated with lesions in certain brain regions. Therefore, locating the disease-related brain regions can improve the robustness of the model. Thus, a channel attention module is introduced. The main approach is to perform global average pooling on the output of the Shuffle basic unit in the last layer of the backbone network, and then perform local cross-channel interaction information for each channel and its two adjacent channels to obtain a spatial attention feature map. This map is then normalized using the sigmoid activation function, reflecting the importance of different brain regions in the recognition process. Disease-sensitive brain regions are often given greater weight, while relatively less important brain regions are given smaller weights. Assume the output of the Shuffle basic unit in the last layer of the backbone network is X∈R. W×H×C Let W be the image width, H be the image height, and C be the number of channels. Then, ECA is defined as follows:

[0068] ω=σ(C1D k (y))

[0069] Where σ represents the sigmoid activation function, C1Dk Let y = g(X) represent a one-dimensional convolution with kernel size k. This represents a global average pooling operation on the output result X.

[0070] (3) The deep metric learning module performs classification optimization.

[0071] See Figure 3 The ECA attention module is designed with two branches. One branch calculates the cross-entropy classification loss to classify whether a subject is ill. The cross-entropy classification loss formula is defined as follows:

[0072]

[0073] Where N represents the number of samples, The output of the sigmoid function represents the probability that the label of the current sample is 1. y represents the probability that the current sample label is 0, and y represents the label value 0 or 1.

[0074] Another branch calculates Triplet Loss to distinguish whether multiple slices belong to the same subject. Specifically, the cross-entropy loss branch flattens and fully connects the ECA-processed results; the Triplet Loss branch first passes the ECA-processed results through three 1×1 convolutional layers to reduce the number of channels and ReLU activation function layers, then flattens and fully connects them, and finally performs L2 paradigm operations. The Triplet Loss formula is:

[0075] L T=max(d(a,p)-d(a,n)+m arg in,0)

[0076] Where a is an anchor example, p is a positive example, n is a negative example, and d is the Euclidean distance metric;

[0077] Cross-entropy classification loss and Triplet Loss together constitute the loss function of the network model:

[0078] L = L C +L T .

[0079] The overall network structure diagram of this invention is as follows: Figure 3 .

[0080] To this end, we constructed a relevant system, and the overall network structure diagram of this method is as follows: Figure 4 .

[0081] The system includes:

[0082] The slicing module is used to slice 3D brain data from multiple view angles.

[0083] The feature extraction module is used to extract feature maps from the sliced ​​multi-view images.

[0084] The channel attention module is used to reflect the importance of different brain regions in recognition in the feature map;

[0085] The Deep Metrics Learning module is used to calculate cross-entropy classification loss and Triplet Loss.

[0086] thus:

[0087] (1) This invention uses ShuffleNet V1 as the backbone network to reduce the number of network parameters and computational cost; the original three-dimensional magnetic resonance image is taken as a two-dimensional slice image, and then a two-dimensional convolutional neural network is used to extract features, thereby reducing the computational complexity and memory cost while ensuring the model performance remains unchanged.

[0088] (2) The present invention adopts a multi-view approach to extract 2D image features from different angles, thereby extracting interrelated and complementary features in different views.

[0089] (3) Add an ECA module to the ShuffleNet V1 backbone network to strengthen the weights of task-related features and suppress noise and task-irrelevant features.

[0090] (4) Use Triplet Loss to ensure consistent classification results for multiple two-dimensional slice images of the same subject.

[0091] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A method for analyzing magnetic resonance imaging of the brain, characterized in that, The method includes: ShuffleNet V1 was used as the backbone network; Multi-view angular slicing was performed on 3D brain data, and feature extraction was performed on 2D images from different angles to obtain feature maps; The extracted feature maps are input into the channel attention module to reflect the importance of different brain regions in the feature maps to recognition. The cross-entropy classification loss is calculated to classify whether a subject is ill; at the same time, the triplet loss is calculated to distinguish whether multiple slices belong to the same subject. The method for constructing the channel attention module is as follows: The results of the Shuffle basic unit in the last layer of the backbone network are globally averaged and pooled. Local cross-channel interaction information is obtained for each channel and the two adjacent channels to obtain a spatial attention feature map, which is then normalized using the sigmoid activation function; The calculation of cross-entropy classification loss to classify whether a subject is diseased involves the following steps: The cross-entropy loss branch flattens and fully connects the result after the channel attention module; The Triplet Loss calculation is used to distinguish whether multiple slices belong to the same subject. The steps are as follows: The Triplet Loss branch first passes the ECA-processed result through three 1×1 convolutional layers to reduce the number of channels and ReLU activation function layers, then flattens it, performs a fully connected layer, and finally performs L2 paradigm operations.

2. The method for analyzing magnetic resonance images of the brain according to claim 1, characterized in that, in, Multi-view angular slicing of 3D brain data mainly includes: The 3D brain data was sliced ​​from three angles: coronal, sagittal, and transverse.

3. The method for analyzing magnetic resonance images of the brain according to claim 1, characterized in that, Suppose the output of the shuffle basic unit in the last layer of the backbone network is , Image width, Image height, Given the number of channels, the channel attention of the channel attention module is defined as follows: ； in, This represents the sigmoid activation function. Indicates the kernel size as One-dimensional convolution; y=g(X), g(X)= , representing the output result Perform global average pooling.

4. The method for analyzing magnetic resonance images of the brain according to claim 1, characterized in that, The cross-entropy classification loss formula used to calculate the cross-entropy classification loss is defined as follows: ； Where N represents the number of samples, The output of the sigmoid function represents the probability that the label of the current sample is 1. This represents the probability that the current sample label is 0, where y represents the label value 0 or 1; The Triplet Loss formula used to calculate the Triplet Loss is defined as follows: ； Where a is an anchor example, p is a positive example, n is a negative example, and d is the Euclidean distance metric; The loss function of the backbone network .

5. A magnetic resonance imaging (MRI) system for the brain used in the brain MRI image analysis method according to any one of claims 1-4, characterized in that, The system selects ShuffleNet V1 as the backbone network; The system includes: The slicing module is used to slice 3D brain data from multiple view angles. The feature extraction module is used to extract feature information from multiple views after slicing; The channel attention module is used to reflect the importance of different brain regions in recognition in the feature map; The Deep Metrics Learning module is used to calculate cross-entropy classification loss and Triplet Loss.

Images