Classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information

Abstract

The present invention discloses a classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information, which mainly relates to the technical field of medical image classification. The method comprises the following steps: S1, collecting data images and performing maximum density projection operation to convert the original image format data into image format data; S2, preprocessing the image; S3, sending the left hemisphere image and the right hemisphere image into the main network respectively to extract the image features FM of the corresponding hemisphere. left ,FM r‑flip ; S4, the FM obtained left ,FM r‑flip Send it to the symmetric information processing module to obtain the classification loss #imgabs0#S5, and the obtained FM left ,FM r‑flip The deep supervision module is sent to obtain the contrast loss #imgabs1#S6, and the model is updated using #imgabs2# and #imgabs3# until the effect converges. The present invention can solve the problem of poor automatic identification effect of the responsible blood vessels for intracranial large blood vessel blockage in the existing system.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image classification, and in particular to a classification method for identifying intracranial large vessel blockage based on bilateral contrast difference information. Background Art

[0002] Stroke is a serious brain disease that poses a serious threat to human life and health. Clinically, ischemic stroke (IS) accounts for the largest proportion of cases, with acute ischemic stroke (AIS) caused by large vessel occlusion (LVO) being the most severe. However, if patients receive accurate diagnosis and treatment within the first few days of onset, the effects of the disease can be effectively mitigated. This not only facilitates recovery but also significantly reduces the impact of complications. Therefore, rapid diagnosis of LVO and appropriate treatment based on symptoms are imperative.

[0003] Computed tomographic angiograms (CTA) are the primary method for physicians to determine whether a patient has a major intracranial vascular obstruction and to locate the offending vessel. However, a single cranial CTA examination can generate hundreds of high-resolution images for review, posing a challenge to radiologists and clinical imaging workflows and potentially delaying patient treatment. Therefore, helping physicians quickly and accurately identify the specific location of LVO is a key step in ensuring timely surgical treatment for patients.

[0004] In recent years, although relevant researchers have carried out a lot of work on the task of identifying LVO and achieved good results, some problems still exist. For example, most of the current research is limited to binary classification tasks and does not identify the specific location of blocked blood vessels. In addition, they all rely on existing network models to complete training and classification, and do not adjust the network's own structure according to the actual data. Summary of the Invention

[0005] The purpose of the present invention is to solve the problems existing in the prior art and provide a classification method for identifying intracranial large vessel blockage based on bilateral contrast difference information. It is a method that achieves high-precision identification and classification by comparing the difference information of corresponding positions on both sides of the brain, which can solve the problem of poor automatic identification effect of the responsible blood vessels for intracranial large vessel blockage.

[0006] To achieve the above-mentioned purpose, the present invention is implemented through the following technical solutions:

[0007] A classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information includes the following steps:

[0008] S1. Acquire data images and perform maximum intensity projection operations to convert raw image format data into image format data for subsequent use;

[0009] S2, preprocess the image, first correct the input image, and then divide the input image into the left hemisphere image X left and right hemisphere image X right , then X right The image is flipped horizontally to obtain X r-flip ;

[0010] S3, send the left brain image and the right brain image to the main network to extract the image features FM of the corresponding brain left ,FM r-flip ;

[0011] S4. Obtain FM left ,FM r-flip Send it to the symmetric information processing module to obtain the classification loss

[0012] S5. Obtain FM left ,FM r-flip Send it to the deep supervision module to obtain the contrast loss

[0013] S6. Utilization and Update the model until the effect converges.

[0014] Preferably, in step S1, the original silhouette data is copied from the workstation, and then projected from the slice stacking direction using the maximum intensity projection technology to obtain a MIP projection image, and the classification label is annotated.

[0015] Preferably, in step S2, the maximum outline of the brain is first extracted by thresholding, and then the minimum circumscribed rectangular frame of the brain region is obtained using the outline mask. Then, the perspective transformation is used to correct the image of the brain region, and finally the left hemisphere image X is obtained by segmenting from the midline. left and right hemisphere image X right , and X right The image is flipped horizontally to get X r-flip .

[0016] Preferably, in step S3, the input image X left 、X r-flip It will first pass through the same backbone network to extract the same low-level features F left 、F r-flip , then F left 、F r-flipEach hemisphere will pass through its own unique attention convolution branch to obtain high-level semantic features FM left ,FM r-flip ;

[0017] Among them, the attention convolution branch is composed of multiple attention convolution blocks, which can complete the attention enhancement in spatial and channel dimensions.

[0018] Preferably, the attention convolution branch is:

[0019] For the input convolution feature F input , it will complete the weighted operation on the channel and spatial position to obtain the output feature F out , in terms of channel weighting, F input First, global maximum pooling and global average pooling operations are performed in the width and height dimensions to obtain F MP With F AP , then F MP With F AP After passing through the MLP structure, the corresponding position value addition operation will be completed, and finally the final channel weight W will be generated under the action of the Sigmoid activation function. cha , described by the formula:

[0020] W cha =σ(MLP(AvgPool(F input ))+MLP(MaxPool(F input )))

[0021] Where σ(·) is the Sigmoid activation function, and MLP is the multi-layer perceptron structure;

[0022] For the weighting of spatial information, F input The maximum pooling and average pooling operations will also be used first, but in the channel direction, which will give the same result as F input Features with the same width and height but channel 1 and Then these two features are concatenated and convolved, and the spatial weight W is generated by the Sigmoid activation function after the convolution operation. spa , described by the formula:

[0023] W spa =σ(conv(concat[AvgPool(F input ), MaxPool(F input )]))

[0024] Where conv is a 3×3 convolution operation, concat is a concatenation operation, and σ(·) is a Sigmoid activation function.

[0025] Preferably, in step S4, firstly, the feature information of both hemispheres is subjected to a wise-subtract operation to obtain the difference feature F u , then the difference feature F u The network is sent to the continuous convolution layer for information extraction, then flattened according to the batch size dimension, and finally sent to the fully connected layer to complete the classification operation, and the classification loss is obtained. The loss function for classification is the cross entropy loss function;

[0026] Described in formula language:

[0027]

[0028] Among them, Conv(·) represents continuous convolution operation, View(·) represents flattening operation, FC represents fully connected layer, and F is obtained. f After that, a Softmax operation is required to obtain the final prediction probability of each category. The classification loss function used in the end is is the cross entropy loss function, and the formula is expressed as:

[0029]

[0030] Among them, N represents the number of samples, M represents the number of classification categories, and y ic If the true category of sample i is equal to c, it takes 1, otherwise it takes 0, p ic represents the predicted probability that the observed sample i belongs to category c.

[0031] Preferably, in step S5, a half-brain label is set for the half-brain of the input sample, 0 indicates that the half-brain is normal, 1 indicates that the middle cerebral artery of the half-brain is blocked, and 2 indicates that the cervical artery of the half-brain is blocked. Then, the supervised contrast loss function is used to complete the deep supervision function according to the set label to obtain the contrast loss function. Described by the formula:

[0032]

[0033] In the formula, N represents the number of samples, Indicates that the half-brain label in the sample is y i the number of Indicates finding samples with the same half-brain label except itself. If satisfied, then is 1, otherwise it is 0; s i，j is the cosine similarity function, which is used to measure the similarity between two features. Indicates that sample k and sample i belong to different half-brain labels. t is the temperature coefficient, which is a scalar used to control the degree of discrimination of the model for samples and is set to 0.5.

[0034] Preferably, in step S6, the final loss function is: Among them, α and β are the proportional weights of the loss of the classification loss module and the deep supervision module, respectively. α is set to 15 and β is set to 1.

[0035] Compared with the prior art, the beneficial effects of the present invention are:

[0036] The present invention redesigns the network structure by taking advantage of the fact that blockage of blood vessels on one side of the brain can lead to information differences between the two sides of the brain. The network model also uses a dual attention mechanism of space and channel to enhance the model's information extraction capabilities. Furthermore, the present invention introduces a deep supervised contrast loss function, which helps the model learn better by narrowing the distance between the same class and widening the distance between different classes during training. This ultimately achieves accurate classification of the vessels responsible for LVO, resolving the existing problem of poor automated identification of vessels responsible for large intracranial vessel blockage. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 It is a schematic diagram of the structure of the bilateral comparison difference information network model of the present invention;

[0038] Figure 2 is a flow chart of the present invention for dividing an input image into left and right hemispheres;

[0039] Figure 3 It is a schematic diagram of converting original silhouette data into image format data in the present invention;

[0040] Figure 4 It is a schematic diagram of the main network structure;

[0041] Figure 5 This is a schematic diagram of the structure of the attention convolution module;

[0042] Figure 6 It is a structural diagram of the symmetric information processing module;

[0043] Figure 7 This is a schematic diagram of the effect that the deep supervision module aims to achieve. DETAILED DESCRIPTION

[0044] Below in conjunction with specific embodiment, further set forth the present invention.Should be understood that these embodiments are only used to illustrate the present invention and are not used in limiting the scope of the present invention.In addition, should be understood that after reading the content taught by the present invention, those skilled in the art can make various changes or modifications to the present invention, and these equivalent forms fall within the scope limited by the application equally.

[0045] Example: As shown in the attached Figure 1-7 As shown, the present invention is a classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information, comprising the steps of:

[0046] S1: The specific conversion method is as follows Figure 1 As shown, this example is based on data collected by the First Affiliated Hospital of Chongqing Medical University. A total of 435 samples were collected, including 276 healthy individuals, 28 with right carotid artery occlusion, 64 with right middle artery occlusion, 20 with left carotid artery occlusion, and 47 with left middle artery occlusion. From the original data, 61 healthy individuals, 18 with right middle artery occlusion, 9 with right carotid artery occlusion, 15 with left middle artery occlusion, and 6 with left carotid artery occlusion were randomly selected as an independent test set. The remaining samples served as the training set, and data augmentation was performed on the training set.

[0047] S2: As attached Figure 2 As shown in the figure, the maximum outline of the brain is first extracted by thresholding, and then the minimum circumscribed rectangular frame of the brain area is obtained using the outline mask. Then, the perspective transformation is used to correct the image of the brain area, and finally the left and right hemisphere images are obtained by dividing them from the midline. left 、X right , and X right The image is flipped horizontally to get X r-flip .

[0048] S3: As attached Figure 4 As shown, the input image X left 、X r-flip It will first pass through the same backbone network to extract the same low-level features F left 、F r-flip , then F left 、F r-flip Each hemisphere will pass through its own unique attention convolution branch to obtain high-level semantic features FM left ,FM r-flip The attention convolution branch is composed of multiple attention convolution blocks, which can complete the attention in the spatial and channel dimensions. Figure 5 As shown, for the input convolution feature F input , it will complete the weighted operation on the channel and spatial position to obtain the output feature F out , in terms of channel weighting, F input First, global maximum pooling and global average pooling operations are performed in the width and height dimensions to obtain F MP With F AP , then F MP With F AP After passing through the MLP structure, the corresponding position value addition operation will be completed, and finally the final channel weight W will be generated under the action of the Sigmoid activation function. cha , described by the formula:

[0049] W cha=σ(MLP(AvgPool(F input ))+MLP(MaxPool(F input )))

[0050] Where σ(·) is the Sigmoid activation function, and MLP is the multi-layer perceptron structure. For the weighting of spatial information, F input The maximum pooling and average pooling operations will also be used first, but in the channel direction, which will give the same result as F input Features with the same width and height but channel 1 and Then these two features are concatenated and convolved, and the spatial weight W is generated by the Sigmoid activation function after the convolution operation. spa , described by the formula:

[0051] W spa =σ(conv(concat[AvgPool(F input ), MaxPool(F input )]))

[0052] Where conv is a 3×3 convolution operation, concat is a concatenation operation, and σ(·) is a Sigmoid activation function.

[0053] S4: The specific process is as shown in the attached Figure 6 As shown, firstly, the feature information of both hemispheres is subjected to a wise-subtract operation (i.e., the corresponding element positions of the corresponding channels are subtracted) to obtain the difference feature F u , then the difference feature F u The network is fed into successive convolutional layers for information extraction, then flattened by the batch size dimension, and finally fed into a fully connected layer to complete the classification operation, which can be described in formula language as:

[0054]

[0055] Where Conv(·) represents continuous convolution operation, View(·) represents flattening operation, FC represents fully connected layer, and F is obtained. f After that, a Softmax operation is required to obtain the final prediction probability of each category. The classification loss function used in the end is is the cross entropy loss function, and the formula is expressed as:

[0056]

[0057] In the formula, N represents the number of samples, M represents the number of classification categories, and y ic If the true category of sample i is equal to c, it takes 1, otherwise it takes 0, p icrepresents the predicted probability that the observed sample i belongs to category c.

[0058] S5: The purpose of the deep supervision module is to make the network learn better and faster by shortening the distance between the same label classes and alienating the distance between different label classes in the feature space, thereby reducing the learning difficulty of the classifier. The final effect is as shown in the attached figure. Figure 7 As shown, the same hemi-brain labels will be close to each other in the feature space, and different hemi-brain labels will be far away from each other. For the setting of hemi-brain labels, 0 means that the hemi-brain is normal, 1 means that the middle cerebral artery of the hemi-brain is blocked, and 2 means that the cervical artery of the hemi-brain is blocked. Then, according to the set labels, the supervised contrast loss function is used to complete the deep supervision function to obtain the contrast loss. The specific formula is described as:

[0059]

[0060] Where N represents the number of samples, Indicates that the half-brain label in the sample is y i the number of Indicates finding samples with the same half-brain label except itself. If satisfied, then is 1, otherwise it is 0; s i,j is the cosine similarity function, which is used to measure the similarity between two features. Indicates that sample k and sample i belong to different half-brain labels. t is the temperature coefficient, which is a scalar used to control the degree of discrimination of the model for samples. In this study, it is set to 0.5.

[0061] S6: The final loss function is: Here, α and β are the proportional weights of the loss of the classification loss module and the deep supervision module, respectively, and are set to 15 and 1. Regarding network training details, the input size of each half-brain image is 512×256, the batch size is 16, SGD is selected as the network optimizer, the momentum is set to 0.9, the initial learning rate is 0.001, and the learning rate is adjusted using a warm-up training strategy.

Claims

1. A classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information, characterized in that: Including steps: S1. Acquire data images and perform maximum intensity projection operations to convert original image format data into image format data for subsequent use; S2, preprocess the image, first correct the input image, and then divide the input image into the left hemisphere image X left and right hemisphere image X right , then X right The image is flipped horizontally to get X r-flip ; S3, send the left and right hemisphere images to the main network to extract the image features FM of the corresponding hemispheres left ,FM r-flip ; Input image X left 、X r-flip It will first pass through the same backbone network to extract the same low-level features F left 、F r-flip , then F left 、F r-flip Each hemisphere will pass through its own unique attention convolution branch to obtain high-level semantic features FM left ,FM r-flip ; Among them, the attention convolution branch is composed of multiple attention convolution blocks, which can complete the attention blessing in the spatial and channel dimensions; The attention convolution branch is: For the input convolution feature F input , it will complete the weighted operation on the channel and spatial position to obtain the output feature F out , in terms of channel weighting, F input First, global maximum pooling and global average pooling operations are performed in the width and height dimensions to obtain F MP With F AP , then F MP With F AP After passing through the MLP structure, the corresponding position value addition operation will be completed, and finally the final channel weight W will be generated under the action of the Sigmoid activation function. cha , described by the formula: W cha =σ(MLP(AvgPool(F input ))+MLP(MaxPool(F input ))) Where σ(·) is the Sigmoid activation function, and MLP is the multi-layer perceptron structure; For the weighting of spatial information, F input The maximum pooling and average pooling operations will also be used first, but in the channel direction, which will give the same result as F input Features with the same width and height but channel 1 and Then these two features are concatenated and convolved, and the spatial weight W is generated by the Sigmoid activation function after the convolution operation. spa , described by the formula: W spa =σ(conv(concat[AvgPool(F input ),MaxPool(F input )])) Among them, conv is a 3×3 convolution operation, concat is a concatenation operation, and σ(·) is a Sigmoid activation function; S4. Obtain FM left ,FM r-flip Send it to the symmetric information processing module to obtain the classification loss First, perform a wise-subtract operation on the feature information of both hemispheres to obtain the difference feature F u , then the difference feature F u The network is sent to the continuous convolution layer for information extraction, then flattened according to the batch size dimension, and finally sent to the fully connected layer to complete the classification operation, and the classification loss is obtained. The loss function for classification is the cross entropy loss function; Described in formula language: Among them, Conv(·) represents continuous convolution operation, View(·) represents flattening operation, FC represents fully connected layer, and F is obtained. f After that, a Softmax operation is required to obtain the final prediction probability of each category. The classification loss function used in the end is is the cross entropy loss function, and the formula is expressed as: Among them, N represents the number of samples, M represents the number of classification categories, and y ic If the true category of sample i is equal to c, it takes 1, otherwise it takes 0, p ic represents the predicted probability that the observed sample i belongs to category c; S5. Obtain FM left ,FM r-flip Send it to the deep supervision module to obtain the contrast loss S6. Utilization and Update the model until the effect converges.

2. The classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information according to claim 1, characterized in that: In step S1, the original silhouette data is copied from the workstation, and then projected from the slice stacking direction using the maximum intensity projection technology to obtain the MIP projection image, and the classification label is annotated.

3. The classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information according to claim 1, characterized in that: In step S2, the maximum outline of the brain is first extracted by thresholding, and then the minimum circumscribed rectangular frame of the brain area is obtained using the outline mask. Then, the perspective transformation is used to correct the image of the brain area, and finally the left hemisphere image X is obtained by segmenting from the midline. left and right hemisphere image X right , and X right The image is flipped horizontally to get X r-flip .

4. The classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information according to claim 1, characterized in that: In step S5, a half-brain label is set for the half-brain of the input sample, 0 indicates that the half-brain is normal, 1 indicates that the middle cerebral artery of the half-brain is blocked, and 2 indicates that the cervical artery of the half-brain is blocked. Then, the supervised contrast loss function is used to complete the deep supervision function according to the set label to obtain the contrast loss function. Described by formula: In the formula, N represents the number of samples, Indicates that the half-brain label in the sample is y i the number of Indicates finding samples with the same half-brain label except itself. If satisfied, then is 1, otherwise it is 0; s i，j is the cosine similarity function, which is used to measure the similarity between two features. Indicates that sample k and sample i belong to different half-brain labels. t is the temperature coefficient, which is a scalar used to control the degree of discrimination of the model for samples and is set to 0.

5.

5. The classification method for identifying intracranial large vessel occlusion based on bilateral contrast difference information according to claim 1, characterized in that: In step S6, the final loss function is: Among them, α and β are the proportional weights of the loss of the classification loss module and the deep supervision module, respectively. α is set to 15 and β is set to 1.

Images