Method, device, equipment and medium for automatically identifying and analyzing intracranial atherosclerotic plaque

Abstract

The invention discloses a method, a device, equipment and a medium for automatically identifying and analyzing intracranial atherosclerotic plaques, and relates to the technical field of medical image processing, and the key points of the technical scheme are that the method comprises the following steps: collecting an image of an intracranial atherosclerotic patient to be evaluated; importing into an intracranial atherosclerosis AI model preset by the detection equipment; automatically identifying and analyzing whether a feature image corresponding to an intracranial atherosclerosis plaque exists in the image or not by an intracranial atherosclerosis AI model; and if yes, generating a three-dimensional plaque model and displaying the three-dimensional plaque model in a visual window, and if not, outputting a detection result that no plaque exists. The device is used for realizing the method for automatically identifying and analyzing the intracranial atherosclerotic plaque. The apparatus comprises: a processor; computer readable instructions are stored in the storage, and when the processor executes the computer readable instructions, the method for automatically identifying and analyzing the intracranial atherosclerotic plaque is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical image processing, and more specifically, to a method, device, equipment and medium for automatically identifying and analyzing intracranial atherosclerotic plaques. Background Art

[0002] Intracranial atherosclerotic plaques are one of the important causes of cerebrovascular diseases, such as ischemic stroke. Existing imaging technologies, such as CT angiography, magnetic resonance imaging, and digital subtraction angiography, have certain effects in diagnosing intracranial atherosclerotic plaques, but these technologies still have problems such as strong subjective dependence, complex operation, and low diagnostic efficiency. Therefore, the medical community has gradually explored the use of artificial intelligence technology to assist in the analysis and identification of intracranial atherosclerotic plaques, thereby improving the accuracy and efficiency of diagnosis.

[0003] Although artificial intelligence has shown great potential in the identification of intracranial atherosclerotic plaques, the existing technology still has some shortcomings:

[0004] 1. Insufficient data quality and annotation: Training deep learning models requires a large amount of high-quality, accurately annotated imaging data. However, in actual medical scenarios, due to the scarcity and complexity of intracranial atherosclerotic plaque data, it is still challenging to obtain large-scale and accurately annotated datasets.

[0005] 2. Poor model interpretability: Although artificial intelligence models perform well in image recognition tasks, their complex internal mechanisms lead to poor model interpretability. It is difficult for doctors to fully understand the decision-making process of the model, which causes certain concerns in clinical applications.

[0006] 3. Applicability limitations: Most current AI models are trained under specific imaging devices or data formats. When applied to different devices and data in different formats, the performance of the model may decline, affecting its universal applicability. Summary of the invention

[0007] In view of the shortcomings of the prior art, the purpose of the present invention is to provide a method for automatically identifying and analyzing intracranial atherosclerotic plaques, which solves the problems of strong subjective dependence, low diagnostic efficiency, difficulty in quantifying plaque characteristics, and insufficient data and model adaptability in the prior art.

[0008] The present invention provides the following technical solution: a method for automatically identifying and analyzing intracranial atherosclerotic plaques, comprising the following steps:

[0009] Images of patients with intracranial atherosclerosis to be evaluated are collected and imported into the intracranial atherosclerosis AI model pre-set in the detection equipment; the intracranial atherosclerosis AI model automatically identifies and analyzes whether there are characteristic images corresponding to intracranial atherosclerotic plaques in the images; if so, a three-dimensional plaque model is generated and displayed in the visualization window; if not, a detection result of no plaque is output.

[0010] Furthermore, the method for generating the intracranial atherosclerosis AI model comprises:

[0011] Images of patients with intracranial atherosclerosis were collected, and an AI model of intracranial atherosclerosis was trained based on the network architecture built based on the deep learning model.

[0012] Furthermore, images of patients with intracranial atherosclerosis were collected, and an AI model of intracranial atherosclerosis was obtained through training based on the network architecture built by the deep learning model, including:

[0013] Obtain CT and / or MRI images of the sclerotic part of a patient with intracranial atherosclerosis, and perform diversified processing on the images to generate new samples;

[0014] Extracting characteristic images of the above images and samples to obtain a data set of characteristic images of intracranial atherosclerotic plaques, and dividing the data set of characteristic images of intracranial atherosclerotic plaques into a training set and a validation set;

[0015] The characteristic images of intracranial atherosclerotic plaques in the above training set are trained based on the deep learning model to generate an intracranial atherosclerosis AI model;

[0016] The intracranial atherosclerosis AI model was validated using the characteristic images of intracranial atherosclerotic plaques in the validation set;

[0017] After verification, the intracranial atherosclerosis AI model is obtained and set on the background server.

[0018] Furthermore, new samples are generated based on CT and / or MRI images of the sclerosis site of patients with intracranial atherosclerosis, including:

[0019] Interpolate between data samples through SMOTE to generate new synthetic samples;

[0020] And / or, generate new samples through data augmentation.

[0021] Furthermore, the data enhancement method includes:

[0022] Rotate, translate, flip, scale and crop existing CT and / or MRI image data of the sclerosis site of patients with intracranial atherosclerosis;

[0023] and / or changing the brightness, contrast, color and saturation of existing CT and / or MRI images of the sclerotic part of intracranial atherosclerosis patients to simulate the image performance under different equipment or imaging conditions;

[0024] and / or add Gaussian noise, blur, etc. to simulate imaging noise due to equipment or environmental factors, and simulate differences in image clarity and contrast caused by changes in equipment settings or patient status.

[0025] Furthermore, the verification method is to randomly divide the entire data set into K subsets of the same size, and then perform K trainings, using K-1 subsets for model training each time, and using the remaining 1 subset for verification, recording the evaluation results of the model after each verification, and finally evaluating the performance of the model by the average value of the K verifications.

[0026] Furthermore, it also includes fusing multimodal image data through a deep learning model, using the following method:

[0027] Image normalization: The grayscale value range and contrast of images of different modalities vary greatly. The normalization operation can adjust the grayscale or contrast of the image to the same range, so that data of different modalities can be processed uniformly.

[0028] Multi-scale feature extraction: Convolutional neural networks are used to extract different scale features of CT, MRI and other images, such as edge, texture and structural information. These features can represent different levels of information in the image. The extraction of multi-scale features enables the model to comprehensively utilize data from different modalities.

[0029] Fusion based on attention mechanism: In the fusion process, the attention mechanism is introduced to automatically assign weights of different modalities according to the distribution of important information in different modal images, so as to enhance the model's ability to capture key information in multimodal data.

[0030] Furthermore, it also includes an AI model of intracranial atherosclerosis to perform three-dimensional reconstruction and visualization of plaques. The method is: fusing data from different modalities such as CT and MRI, combining the output of the U-NET model with VTK's three-dimensional reconstruction, and embedding the generated three-dimensional model into the VTK visualization window.

[0031] The present invention provides a device for automatically identifying and analyzing intracranial atherosclerotic plaques. The device for automatically identifying and analyzing intracranial atherosclerotic plaques is used to implement the above-mentioned method for automatically identifying and analyzing intracranial atherosclerotic plaques, comprising:

[0032] A collection system that collects images of patients with intracranial atherosclerosis and uploads them to a backend server;

[0033] and a backend server, wherein the backend server pre-sets an intracranial atherosclerosis AI model and uses the AI ​​model to automatically identify and analyze collected images of patients with intracranial atherosclerosis;

[0034] If yes, the corresponding intracranial atherosclerotic plaque image or three-dimensional plaque model is output;

[0035] If not, the detection result without plaque is output.

[0036] The present invention provides a detection device, the detection device comprising:

[0037] processor;

[0038] A memory having computer-readable instructions stored therein, wherein when the processor executes the computer-readable instructions, any of the above-mentioned methods for automatically identifying and analyzing intracranial atherosclerotic plaques is implemented.

[0039] The present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by a processor, any of the above-mentioned methods for automatically identifying and analyzing intracranial atherosclerotic plaques is implemented.

[0040] In summary, the present invention has the following beneficial effects:

[0041] The present invention adopts an optimized data processing and training strategy, generates diversified training samples, performs model training and evaluation on sample data, automatically extracts image features through deep learning algorithms, accurately quantifies key features such as volume, density, and composition of intracranial atherosclerotic plaques, and can fuse data from different imaging devices to perform three-dimensional reconstruction and display of identified plaques. The present invention solves the problems of strong subjective dependence, low diagnostic efficiency, difficulty in quantifying plaque features, and insufficient data and model adaptability in the prior art. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 is a flow chart of the method of the present invention;

[0043] Figure 2 It is the image detected by the method provided by the present invention. DETAILED DESCRIPTION

[0044] The present invention is further described in detail below with reference to the embodiments.

[0045] The embodiment of the present invention provides a method for automatically identifying and analyzing intracranial atherosclerotic plaques. The method can be implemented through a detection device port such as a terminal or a server.

[0046] A method for automatically identifying and analyzing intracranial atherosclerotic plaques comprises the following steps:

[0047] Images of patients with intracranial atherosclerosis to be evaluated are collected and imported into the intracranial atherosclerosis AI model pre-set in the detection equipment; the intracranial atherosclerosis AI model automatically identifies and analyzes whether there are characteristic images corresponding to intracranial atherosclerotic plaques in the images; if so, a three-dimensional plaque model is generated and displayed in the visualization window; if not, a detection result of no plaque is output.

[0048] The present invention provides a method for automatically identifying and analyzing intracranial atherosclerotic plaques through deep learning, adopts optimized data processing and training strategies, generates diversified training samples, performs model training and evaluation on sample data, automatically extracts image features through deep learning algorithms, accurately quantifies key features such as volume, density, and composition of intracranial atherosclerotic plaques, and can fuse data from different imaging devices to perform three-dimensional reconstruction and display of identified plaques. The present invention solves the problems of strong subjective dependence, low diagnostic efficiency, difficulty in quantifying plaque features, and insufficient data and model adaptability in the prior art.

[0049] The above steps of the present invention are further described below in conjunction with embodiments.

[0050] Preferably, the method for generating the intracranial atherosclerosis AI model comprises:

[0051] Images of patients with intracranial atherosclerosis are collected, and an AI model of intracranial atherosclerosis is obtained by training a network architecture built based on a deep learning model; the deep learning model may be a U-NET model.

[0052] To build an AI model that can automatically identify and analyze intracranial atherosclerotic plaques, it is necessary to collect a training data set for the model, and the training data set can be CT and / or MRI images.

[0053] The training data set includes CT and / or MRI images of the sclerosis site of several patients with intracranial atherosclerosis, and the images have plaque features, and the plaque features include morphology, density, texture, boundary, etc.

[0054] The plaque feature image is identified and extracted, and then the encoder of a deep learning model such as the U-NET model is used to extract the spatial features of the image. The decoder gradually restores the spatial resolution of the image to accurately locate the plaque area. At the same time, skip connections are used to transfer high-resolution features between the encoder and decoder, which helps the model retain detailed features during segmentation. The cross entropy loss function is used to evaluate the accuracy of segmentation. Automatic detection and classification of intracranial atherosclerotic plaques (ICAP) is achieved, reducing dependence on manual experience.

[0055] Preferably, based on CT and / or MRI images of the sclerosis site of patients with intracranial atherosclerosis, an AI model of intracranial atherosclerosis is obtained by deep learning model training, including:

[0056] Obtain CT and / or MRI images of the sclerotic part of a patient with intracranial atherosclerosis, and perform diversified processing on the images to generate new samples;

[0057] Extracting the characteristic images of the above images and samples to obtain a data set of ICAP characteristic images, and dividing the data set of ICAP characteristic images into a training set and a validation set;

[0058] The characteristic images of intracranial atherosclerotic plaques in the above training set are trained based on the deep learning model to generate an intracranial atherosclerosis AI model;

[0059] The intracranial atherosclerosis AI model was validated using the characteristic images of the intracranial atherosclerotic plaques in the validation set;

[0060] After verification, the intracranial atherosclerosis AI model is obtained and set on the background server.

[0061] In actual medical scenarios, due to the scarcity and complexity of intracranial atherosclerotic plaque data, it is still challenging to obtain large-scale and accurately annotated datasets. To this end, the present invention increases the diversity of datasets through the following methods:

[0062] SMOTE is used to interpolate between data samples to generate new synthetic samples. These synthetic samples are not simply copied, but are randomly selected and linearly interpolated between samples according to the K-nearest neighbor algorithm to form new samples, thereby balancing the data distribution. By generating synthetic samples, the model can better learn the characteristics of rare diseases such as plaques and lesions, and improve classification accuracy.

[0063] Data enhancement: By transforming existing data in different ways, more diverse training samples can be generated to improve the generalization ability of the model. For example, existing data can be rotated, translated, flipped, scaled, and cropped. These transformations can simulate different shooting angles and changes in equipment parameters, allowing the model to learn more morphological changes. By changing the brightness, contrast, color, and saturation of the image, the image performance under different devices or imaging conditions can be simulated; Gaussian noise, blur, etc. can be added to simulate the imaging noise caused by equipment or environmental factors, and the image clarity and contrast differences caused by changes in equipment settings or patient status can be simulated. By simulating image transformations under different imaging devices and imaging conditions, the adaptability of the model under these conditions is enhanced, and the model's generalization ability for unknown data is improved, making the model more stable and reliable in different clinical scenarios. In addition, data enhancement increases the diversity of the training set, which helps the model avoid overfitting to specific training data.

[0064] After the data set is collected, feature training can be performed, and an AI model of intracranial atherosclerosis can be obtained through deep learning model training. The deep learning model of the present invention includes but is not limited to the U-NET model.

[0065] The present invention will not go into details about the principles and training process of the deep learning model.

[0066] The present invention uses the K-fold cross-validation method to verify the AI ​​model of intracranial atherosclerosis: first, the entire data set is randomly divided into K subsets of the same size, and then K trainings are performed, each time K-1 subsets are used for model training, and the remaining 1 subset is used for verification. The evaluation results of the model are recorded after each verification, and finally the performance of the model is evaluated by the average value of the K verifications. In order to make full use of limited data for model training and evaluation. It not only provides a stable performance evaluation method, but also effectively avoids overfitting of the model to a specific training set. It can also help determine the robustness and consistency of the model on different data subsets, discover the performance fluctuations of the model on certain data, and thus optimize the generalization ability of the model.

[0067] Different from the traditional single training / validation set division, K-fold cross-validation maximizes the efficiency of training set utilization by dividing the data multiple times, and can more comprehensively evaluate model performance, avoiding overfitting or underfitting problems caused by specific divisions, and helping to improve the generalization ability of the model in different clinical scenarios.

[0068] Preferably, in order to solve the problem of identifying image data of different devices and different formats on the intracranial atherosclerosis AI model of the present invention, enhance the adaptability of the intracranial atherosclerosis AI model to different types of data, and enable it to be stably applied in a variety of clinical scenarios, the present invention also fuses multimodal image data through a deep learning model, and adopts the following method:

[0069] Image normalization: The grayscale value range and contrast of images of different modalities vary greatly. The normalization operation can adjust the grayscale or contrast of the image to the same range, so that data of different modalities can be processed uniformly.

[0070] Multi-scale feature extraction: Convolutional neural networks are used to extract different scale features of CT, MRI and other images, such as edge, texture and structural information. These features can represent different levels of information in the image. The extraction of multi-scale features enables the model to comprehensively utilize data from different modalities.

[0071] Fusion based on attention mechanism: In the fusion process, the attention mechanism is introduced to automatically assign weights of different modalities according to the distribution of important information in different modal images, so as to enhance the model's ability to capture key information in multimodal data.

[0072] Compared with the existing registration methods based on rigid and non-rigid transformations, which may produce certain errors when processing organs or structures with more complex morphologies, image normalization can adjust the grayscale or contrast of the image to the same range, so that data of different modalities can be processed uniformly.

[0073] By using a deep learning model to fuse multimodal images, high-level features can be adaptively extracted, reducing the reliance on manually designed features. This method can automatically capture the key features of each modality image during the fusion process, reducing the information loss caused by insufficient feature extraction in traditional methods.

[0074] Through the use of deep learning fusion models, the present invention can capture the deep feature information in images of different modalities, and combine the attention mechanism to perform weight allocation, making the key features more prominent, thereby improving the recognition ability of the lesion area and the diagnostic accuracy.

[0075] More preferably, a deep convolutional neural network or a hybrid neural network architecture can be used to extract features from modalities such as CT and MRI respectively, and finally integrate these features through a fusion layer. The features of images of different modalities are learned simultaneously through a multi-layer neural network, and fused through joint optimization. For example, a two-stream network is used to process CT and MRI data streams respectively, and then feature fusion is performed through a specific fusion layer, such as weighted fusion, maximum pooling, attention mechanism, etc.

[0076] Preferably, in order to solve the problem that the existing models cannot provide three-dimensional information of plaques and have poor interpretability, the intracranial atherosclerosis AI model of the present invention can also perform three-dimensional reconstruction and visualization of plaques, using the following method:

[0077] (1) Data preprocessing

[0078] Obtaining segmentation results: First, the plaque area in medical images such as CT and MRI is segmented through the U-NET model to obtain the two-dimensional contour information of the plaque; the contour information may be represented in the form of a binary mask, where the plaque area is 1 and the background is 0.

[0079] Extraction of plaque contours: For each layer of two-dimensional segmentation results, the plaque boundary contour is extracted as the basic data for three-dimensional reconstruction; vtkContourFilter in VTK or similar contour extraction methods can be used.

[0080] (2) 3D reconstruction

[0081] Contour connection and volume reconstruction: By connecting the contours of the plaques in different slices layer by layer, the boundary information of the two-dimensional slices is combined to generate the preliminary three-dimensional structure of the plaque. Specifically, the vtkMarchingCubes algorithm provided by VTK can be used to convert the discrete points in the two-dimensional image data into a polygonal three-dimensional model.

[0082] Marching Cubes algorithm: This algorithm can gradually build the voxel data of CT or MRI images into a 3D polygonal mesh. The algorithm extracts isosurfaces from each 3D voxel block to form a network structure of polygons such as triangles. The final result is a visual and smooth 3D model.

[0083] (3) Surface smoothing and optimization

[0084] Mesh refinement and smoothing: In order to obtain a smoother and finer patch surface structure, mesh refinement and surface smoothing are used. vtkSmoothPolyDataFilter can be used to process and smooth the polygonal surface, reduce surface noise, and make the 3D model more consistent with the shape of the real patch.

[0085] Normal calculation: Surface normals are calculated by vtkPolyDataNormals, making the lighting and shadow effects of the 3D model more realistic, thereby enhancing the spatial visualization of the plaque and helping doctors better understand its morphological characteristics.

[0086] (4) Enhance the visualization of multi-dimensional features

[0087] Color mapping and transparency control: In order to allow doctors to more intuitively observe the different components or features of the plaque, different areas of the plaque can be color mapped, such as according to density, volume, surface roughness and other attributes. You can use vtkColorTransferFunction to color-code different features of the model, so that doctors can visually distinguish different plaque components.

[0088] Transparency and clipping planes: VTK can provide transparency control and clipping plane functions to enhance interpretability; doctors can observe the internal structure of the plaque through the clipping plane, or view the surface and internal features at the same time by setting transparency. This multi-angle display makes the model interpretation more comprehensive.

[0089] (5) Interactive 3D visualization

[0090] Visualization window: The generated 3D model is embedded in the VTK visualization window to provide interactive operation functions. Doctors can observe different angles and positions of plaques by rotating, zooming in and out of the model. VTK supports interaction with 3D models, allowing doctors to observe and analyze from different angles.

[0091] Measurement function: Through VTK's measurement tools, doctors can also measure the specific size, volume and surface area of ​​plaques on the 3D model to enhance the accuracy of diagnosis. This provides doctors with a more quantitative means of plaque analysis.

[0092] By combining the output of the U-NET model with VTK's 3D reconstruction, the present invention realizes a closed-loop system from AI-assisted segmentation to intuitive visualization. Doctors can not only see the segmentation results of the AI ​​model, but also understand the lesion structure more intuitively through the 3D model. The present invention fuses data from different modalities such as CT and MRI, and realizes joint display through 3D reconstruction, which enhances the adaptability of the AI ​​model to multimodal data.

[0093] The generated three-dimensional plaque model provides an intuitive display of the lesion structure, helping doctors better understand the decision-making process of the artificial intelligence model. This transparent processing can enhance doctors' trust and acceptance of AI-assisted diagnosis results. Doctors can rotate, scale, and measure plaque models in the interactive window provided by VTK to obtain more accurate diagnostic information. This real-time interactive function enhances the doctor's understanding of the model and makes the AI ​​diagnosis results more clinically valuable. Three-dimensional reconstruction combined with AI quantitative analysis provides doctors with richer diagnostic information and reduces the error of relying on two-dimensional images for inference. This integration method improves the accurate assessment of plaque morphology and composition, which helps more accurate clinical diagnosis.

[0094] In one aspect, the present invention provides a device for automatically identifying and analyzing intracranial atherosclerotic plaques, wherein the device is used to implement the above-mentioned method for automatically identifying and analyzing intracranial atherosclerotic plaques, comprising:

[0095] A collection system that collects images of patients with intracranial atherosclerosis and uploads them to a backend server;

[0096] and a backend server, wherein the backend server pre-sets an intracranial atherosclerosis AI model and uses the AI ​​model to automatically identify and analyze collected images of patients with intracranial atherosclerosis;

[0097] If yes, the corresponding intracranial atherosclerotic plaque image or three-dimensional plaque model is output;

[0098] If not, the detection result without plaque is output.

[0099] In one aspect, the present invention provides a detection device, comprising:

[0100] processor;

[0101] A memory having computer-readable instructions stored therein, wherein when the processor executes the computer-readable instructions, any of the above-mentioned methods for automatically identifying and analyzing intracranial atherosclerotic plaques is implemented.

[0102] On the other hand, the present invention provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor, implement any of the above-mentioned methods for automatically identifying and analyzing intracranial atherosclerotic plaques.

[0103] The technical features of the above embodiments may be combined arbitrarily. To make the description concise, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0104] This specific embodiment is merely an explanation of the present invention and is not a limitation of the present invention. After reading this specification, those skilled in the art may make non-creative modifications to the present embodiment as needed. However, as long as they are within the scope of the claims of the present invention, they are protected by the patent law.

Claims

1. A method for automatically identifying and analyzing intracranial atherosclerotic plaques, characterized in that: The following steps are involved: Images of patients with intracranial atherosclerosis to be evaluated are collected and imported into the intracranial atherosclerosis AI model pre-set in the detection equipment; the intracranial atherosclerosis AI model automatically identifies and analyzes whether there are characteristic images corresponding to intracranial atherosclerotic plaques in the images; if so, a three-dimensional plaque model is generated and displayed in the visualization window; if not, a detection result of no plaque is output.

2. The method according to claim 1, characterized in that The method for generating the intracranial atherosclerosis AI model comprises: Images of patients with intracranial atherosclerosis were collected, and an AI model of intracranial atherosclerosis was trained based on the network architecture built based on the deep learning model.

3. The method according to claim 2, characterized in that Collect images of patients with intracranial atherosclerosis, and train the network architecture built based on the deep learning model to obtain an AI model for intracranial atherosclerosis, including: Obtain CT and / or MRI images of the sclerotic part of a patient with intracranial atherosclerosis, and perform diversified processing on the images to generate new samples; Extracting characteristic images of the above images and samples to obtain a data set of characteristic images of intracranial atherosclerotic plaques, and dividing the data set of characteristic images of intracranial atherosclerotic plaques into a training set and a validation set; The characteristic images of intracranial atherosclerotic plaques in the above training set are trained based on the deep learning model to generate an intracranial atherosclerosis AI model; The intracranial atherosclerosis AI model was validated using the characteristic images of intracranial atherosclerotic plaques in the validation set; After verification, the intracranial atherosclerosis AI model is obtained and set on the background server.

4. The method according to claim 3, characterized in that Generate new samples based on CT and / or MRI images of the sclerosis site in patients with intracranial atherosclerosis, including: Interpolate between data samples through SMOTE to generate new synthetic samples; And / or, generate new samples through data augmentation.

5. The method according to claim 4, characterized in that Data augmentation methods include: Rotate, translate, flip, scale and crop existing CT and / or MRI image data of the sclerosis site of patients with intracranial atherosclerosis; and / or changing the brightness, contrast, color and saturation of existing CT and / or MRI images of the sclerotic part of intracranial atherosclerosis patients to simulate the image performance under different equipment or imaging conditions; and / or add Gaussian noise, blur, etc. to simulate imaging noise due to equipment or environmental factors, and simulate differences in image clarity and contrast caused by changes in equipment settings or patient status.

6. The method according to claim 3, characterized in that The verification method is to randomly divide the entire data set into K subsets of the same size, and then perform K trainings, each time using K-1 subsets for model training and using the remaining 1 subset for verification, recording the evaluation results of the model after each verification, and finally evaluating the performance of the model by the average value of the K verifications.

7. The method according to any one of claims 1 to 6, characterized in that: It also includes the fusion of multimodal image data through a deep learning model, using the following method: Image normalization: The grayscale value range and contrast of images of different modalities vary greatly. The normalization operation can adjust the grayscale or contrast of the image to the same range, so that data of different modalities can be processed uniformly. Multi-scale feature extraction: Convolutional neural networks are used to extract different scale features of CT, MRI and other images, such as edge, texture and structural information. These features can represent different levels of information in the image. The extraction of multi-scale features enables the model to comprehensively utilize data from different modalities. Fusion based on attention mechanism: In the fusion process, the attention mechanism is introduced to automatically assign weights of different modalities according to the distribution of important information in different modal images, so as to enhance the model's ability to capture key information in multimodal data.

8. The method according to claim 7, characterized in that It also includes an AI model for intracranial atherosclerosis to perform three-dimensional reconstruction and visualization of plaques, specifically: The data of different modalities such as CT and MRI are fused, the output of the U-NET model is combined with the 3D reconstruction of VTK, and the generated 3D model is embedded in the visualization window of VTK.

9. A device for automatically identifying and analyzing intracranial atherosclerotic plaques, the device being used to implement the method described in any one of claims 1 to 8.

10. A detection device, comprising: processor; A memory having computer-readable instructions stored thereon, wherein when the processor executes the computer-readable instructions, the method according to any one of claims 1 to 8 is implemented.

Images