Method and system for processing intracranial large vessel image, electronic device, and medium

Abstract

The present application relates to the field of image processing, and discloses a method and system for processing an intracranial large vessel image, an electronic device, and a medium. The method comprises: acquiring an intracranial large vessel original image of an object to be identified and an intracranial large vessel original image of a sample subject; on the basis of the intracranial large vessel original images, using a cerebrovascular segmentation model to obtain cerebrovascular mask images; calculating regions of interest and bounding boxes of the cerebrovascular mask images; on the basis of the bounding boxes of the regions of interest, selecting corresponding regions of interest from the intracranial large vessel original images; respectively preprocessing mask image regions of interest and original image regions of interest to obtain images to be processed; marking a target area in the image to be processed of the sample subject; using a training set to train a convolutional neural network model to obtain a cerebrovascular occlusion classification model; and inputting the image to be processed of said object into the cerebrovascular occlusion classification model to obtain a target area identification result. The present application can improve the accuracy of target area identification.

Description

A method, system, electronic device and medium for processing intracranial large blood vessel images

[0001] This application claims priority to the Chinese patent application filed with the China Patent Office on February 29, 2024, with application number 202410223590.8 and invention name “A method, system, electronic device and medium for processing images of intracranial large blood vessels”, the entire contents of which are incorporated by reference into this application. Technical Field

[0002] The present application relates to the field of image processing technology, and in particular to a method, system, electronic equipment and medium for processing images of intracranial large blood vessels. Background Art

[0003] One common method for quickly confirming or ruling out the presence of intracranial large vessel occlusion is CT angiography. This rapid, three-minute examination can be easily performed following a non-contrast head CT scan. It provides a three-dimensional visualization of the intracranial vasculature, providing important diagnostic evidence for occlusive vascular lesions. Radiologists typically evaluate CTA images in the shortest possible time. However, the specific time depends on multiple factors, including image quality, the radiologist's experience and workload, and the hospital's workflow. Furthermore, the location and severity of the occlusion can also affect the evaluation time.

[0004] Deep learning is a type of machine learning. For convolutional neural networks, the features of three-dimensional CT angiography are very sparse. If three-dimensional images are input into a convolutional neural network, meaningful target recognition results may not be obtained. Therefore, it is very necessary to choose a suitable input.

[0005] In terms of preprocessing, maximum intensity projections (MIPs) have been used to facilitate the detection of sudden large vessel occlusions in CT angiography. Maximum intensity projection is a three-dimensional reconstruction technique for medical imaging information. It projects the voxel with the maximum CT value within a certain thickness (i.e., CT slice thickness) onto a background plane to reveal all or part of high-density blood vessels, bones, lung masses, and significantly enhancing soft tissue lesions. Alternatively, some studies have used a crude vessel segmentation algorithm to enhance the representation of sparse vascular features in the image, and then derive a maximum intensity projection image from the segmentation result.

[0006] Current methods for identifying and classifying large vessel occlusions in the brain generally rely on cumbersome preprocessing methods, which can result in information loss, such as when the segmentation network fails to identify vascular pixels or when the maximum intensity projection is performed, skull voxels are not stripped cleanly. Summary of the Invention

[0007] The purpose of this application is to provide a method, system, electronic equipment and medium for processing images of intracranial large blood vessels, which can improve the accuracy of target area recognition.

[0008] To achieve the above objectives, this application provides the following technical solutions:

[0009] A method for processing an image of an intracranial major blood vessel, the method comprising:

[0010] Obtaining original images of intracranial major blood vessels of the target to be identified and the sample subject; the original images of intracranial major blood vessels are 3D brain CT angiography images;

[0011] Applying a cerebral vascular segmentation model according to the original image of the intracranial large blood vessels to obtain a cerebral vascular mask image of the target to be identified and the sample subject;

[0012] Calculating a region of interest of the cerebral blood vessel mask image and a bounding box of the region of interest to obtain regions of interest of the mask image of the target to be identified and the sample subject;

[0013] Selecting a corresponding region of interest from the original image of the intracranial great vessels according to the bounding box of the region of interest, to obtain the region of interest of the original image of the target to be identified and the sample subject;

[0014] Preprocessing the mask image region of interest and the original image region of interest respectively to obtain the to-be-identified target and the to-be-processed image of the sample subject; the preprocessing includes scaling, normalization, and splicing;

[0015] Annotating a target area in the to-be-processed image of the sample subject to obtain an annotated image of the sample subject;

[0016] Using a training set to train a convolutional neural network model to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject;

[0017] The image to be processed of the target to be identified is input into a cerebral vascular occlusion classification model to obtain a target area identification result.

[0018] In one embodiment, the cerebral vascular segmentation model is a nnunet model.

[0019] In one embodiment, preprocessing the mask image region of interest and the original image region of interest to obtain the target to be identified and the processed image of the sample subject specifically includes:

[0020] Scaling the mask image region of interest and the original image region of interest to obtain mask image region of interest and original image region of interest with the same size;

[0021] Normalizing the mask image region of interest and the original image region of interest with the same pixel size to obtain the normalized mask image region of interest and the original image region of interest;

[0022] The normalized mask image region of interest and the original image region of interest are spliced ​​to obtain the target to be identified and the image to be processed of the sample subject.

[0023] In one embodiment, a maximum-minimum normalization method is applied to normalize the mask image ROI and the original image ROI with consistent pixel sizes.

[0024] In one embodiment, the processing method further comprises:

[0025] Data enhancement is performed on the training set; the data enhancement includes random scaling, random rotation, and random translation.

[0026] In one embodiment, the structure of the convolutional neural network model includes a feature extractor, two one-dimensional linear layers and a Sigmoid activation function connected in sequence;

[0027] The feature extractor includes a first extraction module, a second extraction module, a third extraction module, a fourth extraction module, a fifth extraction module and a sixth extraction module connected in sequence; the first extraction module includes a convolution unit and a residual block based on an attention mechanism connected in sequence; the second extraction module, the third extraction module, the fourth extraction module, the fifth extraction module and the sixth extraction module each include a maximum pooling layer, a convolution unit and a residual block based on an attention mechanism connected in sequence; the convolution unit includes a convolution layer, a group normalization layer and a ReLU activation function connected in sequence; the residual block based on the attention mechanism includes two convolution layers and a channel attention mechanism module connected in sequence.

[0028] In one embodiment, the training process of the convolutional neural network model specifically includes:

[0029] Applying a label smoothing regularization process to the labeled image of the sample subject to obtain a smoothed label;

[0030] Inputting the image to be processed of the sample subject into the convolutional neural network model to obtain an output label;

[0031] Calculating the cross entropy loss of multiple categories of labels based on the smoothed labels and the output labels;

[0032] The convolutional neural network model is trained according to the cross entropy loss of the multi-category labels.

[0033] A system for processing intracranial large vessel images, applying the above-mentioned method for processing intracranial large vessel images, the processing system comprises:

[0034] An acquisition module is used to acquire original images of intracranial large blood vessels of the target to be identified and the sample subject; the original images of intracranial large blood vessels are 3D brain CT angiography images;

[0035] a segmentation module, configured to apply a cerebral vascular segmentation model to the original image of the intracranial large blood vessels to obtain a cerebral vascular mask image of the target to be identified and the sample subject;

[0036] a calculation module, configured to calculate a region of interest of the cerebral vascular mask image and a bounding box of the region of interest, and obtain the mask image regions of interest of the target to be identified and the sample subject;

[0037] A selection module is configured to select a corresponding region of interest from the original image of the intracranial great vessels according to the bounding box of the region of interest, to obtain the region of interest of the original image of the target to be identified and the sample subject;

[0038] A preprocessing module, configured to preprocess the mask image region of interest and the original image region of interest, respectively, to obtain the to-be-identified target and the to-be-processed image of the sample subject; the preprocessing includes scaling, normalization, and splicing;

[0039] a labeling module, configured to label a target area in the image to be processed of the sample subject to obtain a labeled image of the sample subject;

[0040] A training module, configured to train a convolutional neural network model using a training set to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject;

[0041] The classification module is used to input the image to be processed of the target to be identified into the cerebral vascular occlusion classification model to obtain the target area recognition result.

[0042] An electronic device includes a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the above-mentioned method for processing images of intracranial major blood vessels.

[0043] A computer-readable storage medium stores a computer program, which implements the above-mentioned method for processing intracranial large blood vessel images when executed by a processor.

[0044] According to the specific embodiments provided in this application, this application has the following technical effects:

[0045] This application first uses a cerebral vascular segmentation model to segment the original three-dimensional CT angiography, then combines the original three-dimensional CT angiography and the segmented image information, and uses a training set to train a convolutional neural network model to obtain a cerebral vascular occlusion classification model, wherein the training set includes an image of a sample subject's original three-dimensional CT angiography and the segmented image information combined, as well as the corresponding labeled image; the image of the target to be identified, which is the combination of the original three-dimensional CT angiography and the segmented image information, is input into the cerebral vascular occlusion classification model to obtain a target area recognition result, thereby achieving the purpose of improving the accuracy of target area recognition.

[0046] Figures in the specification

[0047] The present application will be further described below with reference to the accompanying drawings:

[0048] FIG1 is a schematic diagram of the process of obtaining a mask result after processing a cerebral vascular CT angiography image of the present application;

[0049] FIG2 is a schematic diagram of the structure of the cerebral vascular occlusion classification model of the present application;

[0050] FIG3 is a schematic diagram of the CNN architecture designed based on the residual network of this application;

[0051] FIG4 is a flowchart of the practical application of the method for processing intracranial large vessel images of the present application;

[0052] FIG5 is a flow chart of the method for processing intracranial large vessel images of the present application. DETAILED DESCRIPTION

[0053] The following is a detailed description of the technical solutions in the embodiments of the present application in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, not all of the embodiments; based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative work shall fall within the scope of protection of this application.

[0054] The purpose of this application is to provide a method, system, electronic equipment and medium for processing images of intracranial large blood vessels, which can improve the accuracy of target area recognition.

[0055] The overall technical route of this application is: first, use a pre-trained and relatively accurate cerebral vascular segmentation model to segment the initial brain CT angiography image to obtain a cerebral vascular mask; then, calculate the bounding box of the region of interest based on the mask, use the bounding box to preprocess the cerebral vascular CT angiography image and the mask, and construct a classification network input; then, based on the multi-task multi-label classification learning method, use an appropriate loss function to train the classification model; finally, integrate the pre-trained cerebral vascular segmentation model and the cerebral vascular occlusion classification model to identify the target area of ​​the initial brain CT angiography image to be identified.

[0056] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application is further described in detail below with reference to the accompanying drawings and specific implementation methods.

[0057] Example 1

[0058] As shown in FIG5 , the present application provides a method for processing an intracranial large blood vessel image, the processing method comprising:

[0059] Step S1: obtaining original images of the intracranial major blood vessels of the target to be identified and the sample subject; the original images of the intracranial major blood vessels are 3D brain CT angiography images.

[0060] Step S2: applying a cerebral vascular segmentation model based on the original image of the intracranial large blood vessels to obtain a cerebral vascular mask image of the target to be identified and the sample subject.

[0061] Specifically, the cerebral blood vessel segmentation model is a nnunet model.

[0062] Step S3: Calculate the region of interest of the cerebral blood vessel mask image and the bounding box of the region of interest to obtain the mask image region of interest of the target to be identified and the sample subject.

[0063] Step S4: selecting a corresponding region of interest from the original image of the intracranial large blood vessels according to the bounding box of the region of interest, and obtaining the region of interest of the original image of the target to be identified and the sample subject.

[0064] Step S5: Preprocessing the mask image region of interest and the original image region of interest respectively to obtain the target to be identified and the processed image of the sample subject; the preprocessing includes scaling, normalization and splicing.

[0065] S5 specifically includes:

[0066] Step S51: scaling the mask image ROI and the original image ROI to obtain mask image ROI and original image ROI with the same size.

[0067] Step S52: normalizing the mask image ROI and the original image ROI with the same pixel size to obtain the normalized mask image ROI and the original image ROI.

[0068] Specifically, a maximum-minimum normalization method is applied to normalize the mask image ROI and the original image ROI with consistent pixel sizes.

[0069] Step S53: splicing the normalized mask image region of interest and the original image region of interest to obtain the target to be identified and the image to be processed of the sample subject.

[0070] Step S6: marking the target area in the to-be-processed image of the sample subject to obtain a marked image of the sample subject.

[0071] Step S7: Using the training set to train the convolutional neural network model to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject.

[0072] Before training the convolutional neural network model, the processing method further includes:

[0073] Data enhancement is performed on the training set; the data enhancement includes random scaling, random rotation, and random translation.

[0074] In practical applications, considering that in real situations, due to differences in human bodies and the imaging process, image inconsistencies may occur, this application uses three basic data augmentation methods: random scaling, random rotation, and random translation. To ensure that only minor image changes are caused, the ranges of scaling, rotation angle, and translation are set to 0.9 to 1.2, -30° to 30°, and -10 to 10, respectively. Note that in order to ensure a one-to-one correspondence between the contrast image and the mask, the same data augmentation processing should be performed on both during training.

[0075] Furthermore, due to factors such as the machine, environment, and contrast agent injection time, the cerebral vascular region in the image may appear dim or blurred, making it difficult for the model to focus on the region of interest. To enhance the model's adaptability to this phenomenon, this application uses a mask to adjust the grayscale value of the cerebral vascular region in the original image, with an adjustment factor set between 0.9 and 1.2.

[0076] Enhancing the diversity of datasets is a common data processing method in deep learning. It can improve the performance of computer vision models and help the models better generalize to different image changes.

[0077] As a specific implementation, as shown in FIG2 , the structure of the convolutional neural network model includes a feature extractor, two one-dimensional linear layers and a Sigmoid activation function connected in sequence.

[0078] The feature extractor includes a first extraction module, a second extraction module, a third extraction module, a fourth extraction module, a fifth extraction module, and a sixth extraction module, which are connected in sequence. The first extraction module includes a convolution unit and a residual block based on an attention mechanism, which are connected in sequence. The second, third, fourth, fifth, and sixth extraction modules each include a maximum pooling layer, a convolution unit, and a residual block based on the attention mechanism, which are connected in sequence. The convolution unit includes a convolution layer, a group normalization layer, and a ReLU activation function, which are connected in sequence. The residual block based on the attention mechanism includes two convolution layers and a channel attention mechanism module, which are connected in sequence. Specifically, the residual block based on the attention mechanism includes two identical convolution layers and a channel attention mechanism module, which are connected in sequence. In this application, the first extraction module is stage 1 in Figure 3; the second extraction module is stage 2 in Figure 3; the third extraction module is stage 3 in Figure 3; the fourth extraction module is stage 4 in Figure 3; the fifth extraction module is stage 5 in Figure 3; and the sixth extraction module is stage 6 in Figure 3. The channel attention mechanism module is the SE module in Figure 3; the residual block based on the attention mechanism is the residual SE module in Figure 3; a convolution unit is the 3×3×3 3D convolution kernel in Figure 3.

[0079] In practice, the cerebral vascular occlusion classification model in this application uses a fixed-size 3D cerebral CT angiography image and a correspondingly sized cerebral vascular mask as input. These two images are concatenated along the channel direction and directly input into the feature extractor. The extracted feature map is then vectorized into a feature vector. The final two occlusion classification heads each use a one-dimensional linear layer with a sigmoid activation function, ultimately outputting vectors of probability values ​​for each class, i.e., vectors of length 2 and 4, respectively.

[0080] The input to the cerebrovascular occlusion classification model in this application consists of a cerebral CT angiography image and a cerebrovascular mask obtained after running it through the cerebrovascular segmentation model. The cerebrovascular segmentation network is trained using the existing NNUNet framework without any modifications. Simply preprocessing the dataset and then training it within the framework yields excellent segmentation results.

[0081] As shown in Figure 1, combining the original image and the mask can restore the texture of the cerebral blood vessels. In Figure 1, grayscale is used to clearly demonstrate the texture. The overlay comparison image, obtained by overlaying and comparing the grayscale texture images of the original image and the mask, shows that the mask accurately represents the blood vessels in the original image.

[0082] The 3D brain image and the cerebrovascular mask are concatenated in the channel dimension to form the final network input. Furthermore, assuming the image shape is represented as (W, H, D, C) --> (image width, image height, image depth, number of channels), a 3D brain image has a shape of (W, H, D, 1). Then, the cerebrovascular mask obtained after the aforementioned cerebrovascular segmentation model also has a shape of (W, H, D, 1). Overlaying the two in the final dimension creates an image of (W, H, D, 2). This channel image stacking method is primarily intended to facilitate data input into the network for training.

[0083] The feature extractor can learn and extract information from input data through convolution, pooling, and other operations. This application designs a CNN architecture based on the residual network as shown in Figure 3. The architecture consists of six similar stages, each of which contains a convolution unit and a customized residual block (Res-SE block) based on the attention mechanism. The convolution unit contains a 3×3×3 convolution layer with a stride of 1, a 3D Group Normalization (GN) layer, and a ReLU activation function. In addition, starting from the second stage, a 2×2×2 maximum pooling layer with a stride of 2 is added before the convolution unit. In the first convolution block, the number of feature channels is set to 4 and doubled after entering the next convolution block to infer a sufficiently rich representation of vascular information. The Res-SE block first performs two identical convolution operations. The subsequent SE block can calculate a set of weights. This set of weights can be applied to the input to enhance the importance of features between channels. The SE block first performs global average pooling to obtain the global compressed feature vector of the current feature map. It then uses two layers of full connection to obtain the weight of each channel. The feature map is weighted using the weight value. Finally, a residual connection is used to add the weighted feature map to the input feature map of the Res-SE block to form a new feature map to continue the calculation of the next stage. This design can explicitly model the interdependence between the network's evolving feature channels to improve the quality of the network's generated representations, while also alleviating the problems of vanishing and exploding gradients and making training more stable.

[0084] Furthermore, the training process of the convolutional neural network model specifically includes:

[0085] A label smoothing regularization process is applied to the labeled image of the sample subject to obtain a smoothed label.

[0086] The image to be processed of the sample subject is input into the convolutional neural network model to obtain an output label.

[0087] Calculate the cross entropy loss of multiple categories of labels based on the smoothed labels and the output labels.

[0088] The convolutional neural network model is trained according to the cross entropy loss of the multi-category labels.

[0089] In practical applications, the loss function used in this application has made some improvements based on the multi-class cross-entropy loss (Multiclass Cross-Entropy Loss). Before calculating the loss function, label smoothing regularization is used for the ground truth labels. It is often used for classification problems, especially for problems with hard labels (one-hot encoding). The core idea of ​​label smoothing is not to set the target labels to strictly 0 or 1, but to adjust them to a probability distribution between 0 and 1. The purpose of this is to reduce the overfitting of the model to the training data and make it more robust. This application uses the following formula for label smoothing to obtain soft labels:

[0090] Among them, y true is the original hard label (i.e., ground truth label), which is the target label annotated by professional radiologists to the sample, set to strictly 0 and 1 (1 means that the corresponding blood vessel segment is occluded, and 0 means no occlusion), ε is the smoothing factor (the value is 0.1 in this application), and V represents the number of categories in the current classification task. Furthermore, if the radiology department diagnoses that a sample has L-ICA and R-ICA occlusion, then y in the binary classification task true is [0,1], y in the four-category task true is [1,1,0,0]. After label smoothing, the two classification tasks They are [0.05,0.95], [0.925,0.925,0.025,0.025] respectively.

[0091] Then, the smoothed ground truth label is calculated using the following formula and the network output label y pred Multi-class soft label cross entropy loss.

[0092] Among them, n mask and p mask are two Boolean masks, representing Less than or equal to The value is greater than or equal to The value of . It is a decimal very close to zero and can be set to 1e-7 or 1e-6 to avoid numerical instability caused by division in deep learning and filter invalid values ​​in labels. ⊙ is the Hadamard product (element-wise product, a binary operation). LSE() represents the mean of the elements in the tensor after performing element-by-element logsumexp operations. is the cross entropy loss.

[0093] This application adopts a multi-task learning structure with simultaneous training of two-class and four-class classifications, and the final loss is defined as the following formula. They are the two-category soft label cross entropy loss and the four-category soft label cross entropy loss, respectively, where α and β are the adjustment factors of the loss weights of different classification tasks (the values ​​in this application are 0.25 and 0.75).

[0094] Step S8: inputting the image to be processed of the target to be identified into the cerebral vascular occlusion classification model to obtain a target area recognition result.

[0095] As shown in FIG4 , the main workflow of the method for processing intracranial large vessel images provided in this application is as follows:

[0096] (1) Obtaining an original image. Specifically, the original image is a 3D brain CT angiography image.

[0097] (2) The original image is input into a pre-trained cerebral vascular segmentation model to obtain a cerebral vascular mask.

[0098] Specifically, a relatively accurate cerebral vascular segmentation network is used to segment the context-sensitive image, generating a cerebral vascular segmentation mask for the original 3D brain CT angiography image. Training is performed using the existing nnunet framework without any modifications. Simply preprocessing the dataset and then training it within the framework yields excellent segmentation results.

[0099] (3) Calculate the bounding box of the region of interest using the above cerebral vascular mask.

[0100] Specifically, a bounding box is used to identify and define the location of the region of interest in the visual data. Using the mask information above, a minimal 3D bounding box is found that fully encompasses the cerebral vessels. To reduce the segmentation model's boundary errors while retaining the vascular edge information in the original image, the bounding box is expanded by 10 voxels along each of the x, y, and z axes.

[0101] (4) Build a classification network input image.

[0102] Specifically, preprocessing the original image and cerebral vascular mask can improve the prediction accuracy of deep learning methods for brain age, and speed up the processing and analysis, making it efficient and easy to use. The above preprocessing includes:

[0103] (4.1) Use the bounding box to select the original image and the cerebral vascular mask, and then scale the shapes of both to 256×256×128.

[0104] (4.2) Window width normalization: To better display vascular details, the window width is set to [-1024, 2048]. That is, the CT value less than -1024 and the CT value greater than 2048 are set as the upper and lower limits of the window width, respectively. Then, the maximum and minimum value normalization (formula below) is used to scale all pixels of the image to between 0 and 1. Where x i is the grayscale value of a pixel in the 3D image x, min(x) and max(x) represent the minimum and maximum pixel values ​​of the image x, respectively.

[0105] (4.3) The preprocessed brain image and cerebrovascular mask are concatenated in the channel dimension to form a network input with a dimension of [256, 256, 128, 2].

[0106] (5) Construct a classification model for cerebral vascular occlusion.

[0107] (6) Using image and mask data, combined with doctor annotation information, the classification model is trained using a multi-task multi-label classification loss function, and the model with high classification accuracy and strong generalization performance is selected for preservation.

[0108] (7) Using the trained cerebral vascular occlusion classification model, combined with the aforementioned pre-trained cerebral vascular segmentation model, patients with intracranial large vessel occlusion are classified and identified to provide support for subsequent clinical treatment.

[0109] This application has the following advantages:

[0110] 1. This application uses a deep learning approach to identify and classify intracranial large vessel occlusions using CT angiography alone, without the need for complex preprocessing steps.

[0111] 2. This application proposes a multi-task learning classification network architecture, which performs two-class and four-class classification tasks simultaneously, and can improve the accuracy of the classification results.

[0112] 3. This application uses deep learning methods to identify and classify intracranial large vessel occlusions through CT angiography, which can assist clinicians in interpreting imaging diagnoses for preliminary screening of occlusions. Specifically, deep learning technology is used to identify the specific blood vessels at the time of large vessel occlusion: the left internal carotid artery (L-ICA), the right internal carotid artery (R-ICA), the left branch of the M1 segment of the middle cerebral artery (L-M1), and the right branch (R-M1).

[0113] 4. The method proposed in this application does not require complex image preprocessing steps. The entire process only requires the original CT angiography data, which is less likely to cause information loss and improves the accuracy of the recognition results.

[0114] Example 2

[0115] In order to execute the method corresponding to the above embodiment 1 and achieve the corresponding functions and technical effects, a system for processing intracranial large blood vessel images is provided below. The processing system includes:

[0116] The acquisition module is used to obtain the original images of the intracranial large blood vessels of the target to be identified and the sample subject; the original images of the intracranial large blood vessels are 3D brain CT angiography images.

[0117] The segmentation module is used to apply a cerebral vascular segmentation model to the original image of the intracranial large blood vessels to obtain the cerebral vascular mask image of the target to be identified and the sample subject.

[0118] A calculation module is used to calculate the region of interest of the cerebral blood vessel mask image and the bounding box of the region of interest to obtain the mask image region of interest of the target to be identified and the sample subject.

[0119] A selection module is configured to select a corresponding region of interest from the original image of the intracranial great vessels according to a bounding box of the region of interest, and obtain the region of interest of the original image of the target to be identified and the sample subject.

[0120] A preprocessing module is used to preprocess the mask image region of interest and the original image region of interest respectively to obtain the target to be identified and the processed image of the sample subject; the preprocessing includes scaling, normalization and splicing.

[0121] The labeling module is used to label the target area in the image to be processed of the sample subject to obtain the labeled image of the sample subject.

[0122] A training module is used to train a convolutional neural network model using a training set to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject.

[0123] The classification module is used to input the image to be processed of the target to be identified into the cerebral vascular occlusion classification model to obtain the target area recognition result.

[0124] Example 3

[0125] An embodiment of the present application provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for processing intracranial large blood vessel images of embodiment 1.

[0126] As a specific implementation, the electronic device may be a server.

[0127] In addition, an embodiment of the present application further provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the method for processing intracranial large blood vessel images of embodiment 1.

[0128] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0129] This document uses specific examples to illustrate the principles and implementation methods of this application. The description of the above examples is only intended to help understand the method and core concept of this application. At the same time, for those skilled in the art, based on the concept of this application, there may be changes in the specific implementation methods and application scope. In summary, the content of this specification should not be understood as limiting this application.

Claims

1. A method for processing intracranial large blood vessel images, characterized in that: The processing method comprises: Obtaining original images of intracranial major blood vessels of the target to be identified and the sample subject; the original images of intracranial major blood vessels are 3D brain CT angiography images; Applying a cerebral vascular segmentation model according to the original image of the intracranial large blood vessels to obtain a cerebral vascular mask image of the target to be identified and the sample subject; Calculating a region of interest of the cerebral blood vessel mask image and a bounding box of the region of interest to obtain regions of interest of the mask image of the target to be identified and the sample subject; Selecting a corresponding region of interest from the original image of the intracranial great vessels according to the bounding box of the region of interest, to obtain the region of interest of the original image of the target to be identified and the sample subject; Preprocessing the mask image region of interest and the original image region of interest respectively to obtain the to-be-identified target and the to-be-processed image of the sample subject; the preprocessing includes scaling, normalization, and splicing; Annotating a target area in the to-be-processed image of the sample subject to obtain an annotated image of the sample subject; Using a training set to train a convolutional neural network model to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject; The image to be processed of the target to be identified is input into a cerebral vascular occlusion classification model to obtain a target area identification result.

2. The method for processing intracranial large vessel images according to claim 1, characterized in that: The cerebral vascular segmentation model is a nnunet model.

3. The method for processing intracranial large vessel images according to claim 1, characterized in that: Preprocessing the mask image region of interest and the original image region of interest respectively to obtain the target to be identified and the processed image of the sample subject specifically includes: Scaling the mask image region of interest and the original image region of interest to obtain mask image region of interest and original image region of interest with the same size; Normalizing the mask image region of interest and the original image region of interest with the same pixel size to obtain the normalized mask image region of interest and the original image region of interest; The normalized mask image region of interest and the original image region of interest are spliced ​​to obtain the target to be identified and the image to be processed of the sample subject.

4. The method for processing intracranial large vessel images according to claim 3, characterized in that: A maximum-minimum normalization method is applied to normalize the mask image region of interest and the original image region of interest with consistent pixel sizes.

5. The method for processing intracranial large vessel images according to claim 1, characterized in that: The processing method further comprises: Data enhancement is performed on the training set; the data enhancement includes random scaling, random rotation, and random translation.

6. The method for processing intracranial large vessel images according to claim 1, characterized in that: The structure of the convolutional neural network model includes a feature extractor, two one-dimensional linear layers and a Sigmoid activation function connected in sequence; The feature extractor includes a first extraction module, a second extraction module, a third extraction module, a fourth extraction module, a fifth extraction module and a sixth extraction module connected in sequence; the first extraction module includes a convolution unit and a residual block based on an attention mechanism connected in sequence; the second extraction module, the third extraction module, the fourth extraction module, the fifth extraction module and the sixth extraction module each include a maximum pooling layer, a convolution unit and a residual block based on an attention mechanism connected in sequence; the convolution unit includes a convolution layer, a group normalization layer and a ReLU activation function connected in sequence; the residual block based on the attention mechanism includes two convolution layers and a channel attention mechanism module connected in sequence.

7. The method for processing intracranial large vessel images according to claim 1, characterized in that: The training process of the convolutional neural network model specifically includes: Applying a label smoothing regularization process to the labeled image of the sample subject to obtain a smoothed label; Inputting the image to be processed of the sample subject into the convolutional neural network model to obtain an output label; Calculating the cross entropy loss of multiple categories of labels based on the smoothed labels and the output labels; The convolutional neural network model is trained according to the cross entropy loss of the multi-category labels.

8. A system for processing intracranial large blood vessel images, characterized in that: The processing system comprises: An acquisition module is used to acquire original images of intracranial large blood vessels of the target to be identified and the sample subject; the original images of intracranial large blood vessels are 3D brain CT angiography images; a segmentation module, configured to apply a cerebral vascular segmentation model to the original image of the intracranial large blood vessels to obtain a cerebral vascular mask image of the target to be identified and the sample subject; a calculation module, configured to calculate a region of interest of the cerebral vascular mask image and a bounding box of the region of interest, and obtain the mask image regions of interest of the target to be identified and the sample subject; A selection module is configured to select a corresponding region of interest from the original image of the intracranial great vessels according to the bounding box of the region of interest, to obtain the region of interest of the original image of the target to be identified and the sample subject; A preprocessing module, configured to preprocess the mask image region of interest and the original image region of interest, respectively, to obtain the to-be-identified target and the to-be-processed image of the sample subject; the preprocessing includes scaling, normalization, and splicing; a labeling module, configured to label a target area in the image to be processed of the sample subject to obtain a labeled image of the sample subject; A training module, configured to train a convolutional neural network model using a training set to obtain a cerebral vascular occlusion classification model; the training set includes the to-be-processed image of the sample subject and the corresponding annotated image of the sample subject; The classification module is used to input the image to be processed of the target to be identified into the cerebral vascular occlusion classification model to obtain the target area recognition result.

9. An electronic device, characterized in that: The electronic device comprises a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to execute the method for processing an intracranial major blood vessel image according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that The device stores a computer program, which, when executed by a processor, implements the method for processing an intracranial great vessel image according to any one of claims 1 to 7.

Images