MRA intracranial aneurysm detection method

Abstract

The invention relates to an MRA intracranial aneurysm detection method, which comprises the following steps: acquiring MRA image data to be detected and extracting a region of interest to obtain local image blocks around a blood vessel and a focus; the local image blocks are sent to a segmentation network, the trunk of the segmentation network adopts a UNet architecture of a residual structure, depth feature extraction is performed on the to-be-segmented aneurysm image blocks to obtain multi-scale features, and feature maps of different scales and an up-sampling result are fused through a context attention modulation module to obtain feature segmentation results of different scales; and the obtained segmentation result is sent to a classification network, the classification network adopts an ensemble learning device, various randomly enhanced lesion area images are respectively input into the ViT network to predict the aneurysm probability, and ensemble learning is carried out on output results of each ViT network to obtain the final prediction aneurysm probability. Compared with the prior art, the method has the advantages that the balance of high sensitivity and low false positive rate is realized, and the detection precision of the MRA intracranial aneurysm is effectively improved.

Description

Technical Field

[0001] The invention relates to the technical field of medical image processing, and in particular to an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control. Background Art

[0002] Intracranial aneurysms (IAs) are common cerebral vascular anomalies with an incidence of approximately 3%-7%, of which 93% of UIAs are <10 mm in diameter and 66% are <5 mm in diameter. Aneurysm rupture or mass effect can lead to serious morbidity and high mortality. Although the annual risk of rupture of unruptured intracranial aneurysms (UIA) is low, the potentially devastating sequelae of subarachnoid hemorrhage (SAH) emphasize the importance of appropriate diagnosis, treatment, and follow-up. Magnetic resonance angiography (MRA) is a non-invasive, radiation-free, and highly sensitive imaging technique that can detect asymptomatic occult aneurysms and is therefore often used for screening of unruptured aneurysms.

[0003] As people pay more attention to their health, the demand for MRA examinations has risen significantly. However, it usually takes doctors 20 minutes to read the film, and with the increasing workload, the diagnostic sensitivity of radiologists may decrease, leading to missed diagnoses and misdiagnoses. Therefore, it is particularly important to develop an automated method to detect aneurysms directly from the original scanned images to reduce the burden on doctors.

[0004] In recent years, research on aneurysm detection has mainly focused on traditional machine learning and deep learning methods. However, due to the small size and variable shape of aneurysms, it is difficult to extract their features. In order to improve detection accuracy, existing studies have introduced methods such as residual networks, multi-scale feature enhancement, or attention mechanisms to enhance the feature expression of small targets. Although these methods have made some progress in the automated detection of tiny aneurysms and have greatly reduced the workload of radiologists, the false positive rate is still high. In early screening scenarios, a high false positive rate will lead to a large number of misdiagnoses, and radiologists need to spend a lot of time to confirm whether it is an aneurysm, which increases the workload. Therefore, how to balance detection sensitivity and false positive rate in early screening of aneurysms is an urgent problem to be solved. Summary of the invention

[0005] The present invention aims to solve the problem of difficulty in balancing high sensitivity and low false positive rate indicators in current automatic aneurysm detection algorithms, and to provide an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control.

[0006] The purpose of the present invention can be achieved by the following technical solutions:

[0007] An MRA intracranial aneurysm detection method, comprising the following steps:

[0008] Acquire the MRA image data to be detected and extract the region of interest from the image data to obtain local image blocks around blood vessels and lesions;

[0009] The local image block is sent to the segmentation network, the backbone of which adopts the UNet architecture of the residual structure, and the deep feature extraction of the aneurysm image block to be segmented is performed to obtain multi-scale features, and the feature maps of different scales are fused with the upsampling results through the context attention modulation module to obtain the feature segmentation results of different scales;

[0010] The obtained segmentation results are sent to the classification network. The classification network uses an integrated learner to input multiple randomly enhanced lesion area images into the ViT network to predict the aneurysm probability, and integrate the output results of each ViT network to obtain the final predicted aneurysm probability.

[0011] As a preferred technical solution, the extraction of the region of interest from the image data is specifically as follows:

[0012] Calculate the MRA image {X i} N Center point, based on the center position, remove irrelevant areas;

[0013] Set the image threshold σ and extract the partial mask {X′ i} N ;

[0014] Randomly set the image sampling block size s and the number of dense sampling times n, and for the MRA image {X i} N In a balanced way, {X′ i} N The foreground is upsampled n times to obtain an image block of size s

[0015] As a preferred technical solution, the image threshold is defined using the gray value histogram of the image and background, and the most suitable segmentation boundary is obtained by Otsu's method and used as the image threshold σ:

[0016]

[0017] Among them, ω1 and m1 are the weight and mean of the foreground, m T is the total mean, and the optimal image threshold is obtained by maximizing σ;

[0018] Based on the calculated threshold, the image is binarized to generate a mask:

[0019]

[0020] Where I(x,y,z) is the pixel value of the image at position (x,y,z) and σ is the image threshold.

[0021] As a preferred technical solution, multiple foreground points (x, y, z) are randomly selected from the mask Mask (x, y, z) based on threshold segmentation, which must include a Mask (x, y, z) = 1, and based on the size s of the image sampling block, the foreground points (x, y, z) are cut to obtain the area of ​​interest.

[0022] As a preferred technical solution, the size s of the image sampling block is set by the following formula:

[0023] s=D min ×2 L

[0024] Among them, D min is the minimum downsampling boundary of the network, and L is the number of network feature extraction layers.

[0025] As a preferred technical solution, the segmentation network is specifically as follows:

[0026] Mapped to feature maps of different scales through multi-layer feature extraction networks, each layer of feature extraction network contains a residual structure to retain the low-level semantic features of the target;

[0027] Based on the designed contextual attention modulation module,

[0028] Feature maps of different scales are obtained through the encoding module and the context attention modulation module, and then multi-scale segmentation results are obtained through the decoder;

[0029] Each upsampling layer is deeply supervised, and the loss is calculated by comparing the result of each upsampling layer with the mask of the corresponding scale to constrain the output results of feature maps of different scales.

[0030] As a preferred technical solution, the segmentation network adopts a context attention mechanism modulation module so that the context information of each spatial position is aggregated separately, as follows:

[0031] For a feature map with C channels of size G×W×D;

[0032] The global information guidance of low-semantic features is obtained through the top-down attention modulation module, and then the attention weight Attention(Q,K,V) is calculated;

[0033] The local information guidance of high-semantic features is obtained through a bottom-up point-channel attention modulation module, where the channel feature context of each spatial location is aggregated separately:

[0034]

[0035] Among them, the convolution kernel of Conv_1 is 1×1, and the number of channels is compressed to The convolution kernel of Conv_2 is 1×1, and the number of channels is reset to C;

[0036] The context-adjusted attention mechanism module is used to fuse the attention information in two directions:

[0037]

[0038] Among them, X is the low semantic feature of the feature map; Y is the high semantic feature of the feature map.

[0039] As a preferred technical solution, the segmentation network adopts the integrated combined loss Total Loss of CE and DICE:

[0040] Total Loss=α·CELOSS+β·DICELOSS

[0041]

[0042] Among them, α and β are weight factors; y i is the true label of the ith pixel / voxel, p i is the predicted probability of the ith pixel / voxel.

[0043] As a preferred technical solution, the classification network performs a variety of different random enhancements on the input to change the area around the lesion, and then inputs each of them into an independent Vit network for prediction probability, and integrates and learns the prediction results of each Vit network.

[0044] As a preferred technical solution, the method post-processes the aneurysm mask with a high probability of predicting aneurysm output by the classification network and displays it on the MIP map after the anterior and posterior circulation segmentation of the blood vessel, specifically:

[0045] The aneurysm lesion position is projected onto the MIP image after the anterior-posterior cycle segmentation. The projection angle is 15 degrees to the left and 15 degrees to the top and bottom. For the input image, the direction vector is V(x, y, z) and (d x ,d y ,d z ) is expressed as:

[0046]

[0047] Where t is the projected distance along the direction vector.

[0048] Compared with the prior art, the present invention has the following beneficial effects:

[0049] 1) The segmentation model of the present invention adopts the UNET network under the residual structure of the aneurysm segmentation network, and designs a contextual attention modulation module. The multi-scale context is exchanged through the top-down global attention mechanism module and the bottom-up local attention mechanism module. The contextual attention mechanism modulation module supplements the most advanced top-up high-level semantic feedback path and adopts the reverse bottom-up context modulation path to encode smaller-scale visual details to a deeper level, thereby providing semantic information and spatial details for microaneurysm detection. Richer encoding improves the sensitivity of the model.

[0050] 2) In order to reduce the false positive rate of aneurysms, the present invention designs an aneurysm classification network guided by global information knowledge. The classification network part designs an integrated learner to receive the input observation area around the lesion after random enhancement to adapt to different suspicious lesion properties, thereby solving the impact of differences in aneurysm size, shape and position on diagnosis. The classification model adopts the ViT architecture to better combine the global information around the aneurysm, obtain the characteristic differences between aneurysms and non-aneurysms, and send the prediction results of the previous segmentation network to the classification network to obtain the probability of predicting it as an aneurysm, thereby eliminating some aneurysms with low probability.

[0051] 3) By introducing a region of interest extraction module before the traditional feature extraction module to reduce the large-area sampling of background negative classes while enhancing the target context area samples, the model can focus more on the aneurysm part. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 It is a flow chart of the MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control of the present invention;

[0053] Figure 2 A schematic diagram of region of interest extraction of an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control according to an embodiment of the present invention;

[0054] Figure 3 A segmentation network structure diagram of an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control according to an embodiment of the present invention;

[0055] Figure 4 A schematic diagram of a context attention modulation module according to an embodiment of the present invention;

[0056] Figure 5 A classification network structure diagram of an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control according to an embodiment of the present invention;

[0057] Figure 6This is a visualization diagram of the results of an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control described in an embodiment of the present invention. DETAILED DESCRIPTION

[0058] The present invention is described in detail below in conjunction with the accompanying drawings and specific embodiments. This embodiment is implemented based on the technical solution of the present invention, and provides a detailed implementation method and specific operation process, but the protection scope of the present invention is not limited to the following embodiments.

[0059] Example 1

[0060] The present invention aims to provide an MRA intracranial aneurysm detection method to solve the problem that it is difficult to balance high sensitivity and low false positive rate indicators in the current automatic aneurysm detection algorithm. First of all, by introducing a region of interest extraction module before the traditional feature extraction module to reduce the large-area sampling of the background negative class while enhancing the target context area samples, the model can pay more attention to the aneurysm part. Secondly, a context attention modulation module is designed to exchange multi-scale contexts, enrich the spatial semantic information of the aneurysm, and improve the sensitivity of the model. On the other hand, in order to reduce the false positive rate of the aneurysm obtained, an aneurysm classification model guided by global information knowledge is designed, and the prediction result of the segmentation network in the previous step is sent to the classification network to obtain the probability of predicting it as an aneurysm, thereby eliminating some aneurysms with low probability.

[0061] like Figure 1 The figure shows one embodiment of an MRA intracranial aneurysm detection method based on sensitivity optimization and false positive rate control of the present invention, comprising the following steps:

[0062] Step 1: The collected original MRA data are used to delineate the aneurysm lesion area and divided into a training set and a test set, both of which contain one-to-one corresponding original images and binary label maps.

[0063] Obtain the MRA image data to be detected and its binary label map, and divide it into a training set and a test set according to the sequence. i} N And the corresponding binary label map {Y i} N , where N is the total number of training samples.

[0064] Step 2: After the region of interest extraction module is used to obtain the blocks around the blood vessels and lesions, the original images and label images in the training set are placed in the region of interest extraction module. Since MRA images have the characteristic of highlighting the blood vessel area, the approximate range of the region of interest can be obtained by setting a threshold.

[0065] The image {X i} N The center point is used to remove irrelevant areas based on the center position. Then, a threshold σ is set to extract the partial mask {X' i} N and {Y' i} N Finally, we randomly set the block size s and the number of dense sampling times n to perform i} N In a balanced way in {X' i} N The prospect or {Y i} N The foreground is upsampled n times to obtain an image block of size s and local binary label map

[0066] Step 3: By combining the region of interest and the label map, and by setting the foreground sampling probability and the sampling block size, we obtain positive samples that contain the foreground part and negative samples that contain the region of interest but not the foreground part, as well as the corresponding binary label map, and send them together to the segmentation network for segmentation. The input of the segmentation network is the entire MRA image, and the image will pass through a sliding window with a window size of the sampling block to obtain the segmentation result of the entire MRA image.

[0067] The local image blocks are sent to the segmentation network, and the backbone of the network adopts the UNet architecture under the residual structure. The segmentation network has the following two characteristics: first, the top-down global attention mechanism module and the bottom-up local attention mechanism module are designed to enrich the generation of feature maps; second, by adding deep supervision to each upsampling layer, the output results of feature maps of different scales are constrained, thereby avoiding the gradient disappearance caused by the small shape of the aneurysm. Specifically, the aneurysm image block to be segmented is passed through the encoding module and the context attention modulation module to obtain feature maps of different scales, and then the multi-scale segmentation results are obtained through the decoder. The loss of each result is calculated with the corresponding mask of different scales.

[0068] The specific process of segmentation network training is as follows: the training image will be mapped to feature maps of different scales through a multi-layer feature extraction network. Each layer of the feature extraction network contains a residual structure to better retain the low-level semantic features of the target (including rich textures and details). Then, based on the designed contextual attention modulation module, the feature maps of different scales are fused with the results of decoder upsampling to obtain the contextual semantic enhanced feature map. Finally, the target segmentation result of the network output image block is matched with the label block for similarity matching measurement as the loss feedback gradient.

[0069] Step 4: The blocks predicted by the segmentation network as aneurysms are passed through the classification model to obtain the probability of each block being an aneurysm. Since aneurysms are local protrusions in the location of blood vessels, aneurysm detection relies heavily on the extraction of global image feature information. Therefore, the ViT architecture that can combine global information is selected as the backbone of the classification network to better combine the global information around the aneurysm and obtain the feature differences between aneurysms and non-aneurysms. This classification network has the following two characteristics:

[0070] First, because the image input to the segmentation model is obtained by extracting the region of interest (not the entire image), the segmentation model will also output results of the same scale. Therefore, it is necessary to infer all possible candidate regions for each image input to the segmentation network. Then, according to the predicted recommended areas, the observation areas are extracted from around different candidate areas for further classification to reduce false positives; second, an integrated learner is designed for the classification network part, which receives the input observation areas around the lesions after random enhancement to adapt to different suspicious lesion properties, thereby solving the impact of differences in aneurysm size, shape and position on diagnosis. The global block information of the input image is used, combined with a multi-scale feature integrated learner to adapt to the judgment of suspicious lesions of different sizes and positions, and the probability of the network predicting an aneurysm is obtained, further reducing the false positive rate.

[0071] The classification network training process is as follows: the image with the aneurysm location output by the segmentation network in the previous step is sent to the network as input, where the image containing the real aneurysm is a positive sample, and the image without the real aneurysm is a negative sample. The training image is extracted from the aneurysm feature by the ViT architecture that combines the global image information, and the training set is densely inferred to finally output the probability of whether the network predicts whether it contains a real aneurysm, and the difference with the real label is used as the loss to return the gradient.

[0072] Step 5: Eliminate aneurysms with low prediction probability and map aneurysms with high prediction probability to MIP images for easy observation by radiologists. Since aneurysms are usually very small in size, in order to better prompt radiologists, the output aneurysm mask is post-processed to be displayed on the MIP image after the anterior and posterior circulation segmentation of the blood vessels, so as to more clearly show the output of the network.

[0073] It should be noted that the acquired image is an image scanned by a TOF-MRA sequence without any original processing. In the embodiment of the present application, the data set during training is divided into the size and location area of ​​the reference aneurysm. The training and test are divided in a ratio of 4:1.

[0074] It should be noted that since MRA images are not the gold standard for aneurysm diagnosis, this embodiment is based on the experience of three senior radiologists, and the back-to-back voting principle is used to include cases that are unanimously reviewed as aneurysms into the data set.

[0075] It should be noted that for step 2, the extraction of the region of interest is as follows: Figure 2 The image threshold is defined using the gray value histogram of the image and background, and the most suitable segmentation boundary is obtained by Otsu's method and used as the threshold σ.

[0076]

[0077] Among them, ω1 and m1 are the weight and mean of the foreground, m T is the total mean. The optimal threshold is obtained by maximizing σ.

[0078] Using the calculated threshold σ, the image is binarized to generate a mask:

[0079]

[0080] Where I(x,y,z) is the pixel value of the image at position (x,y,z), and σ is the threshold obtained in the above steps.

[0081] It should be noted that the size s of the image sampling block is set by the following formula:

[0082] s=D min ×2 L (3)

[0083] Among them, D min is the minimum downsampling boundary of the network, and L is the number of network feature extraction layers.

[0084] In this embodiment, since the shapes of the segmentation targets are basically small, the corresponding feature information may be lost as the number of downsampling increases. Therefore, the segmentation network structure designed in the embodiment of the present application reduces the number of network feature extraction layers L, and also increases the network's lowest downsampling boundary to prevent the disappearance of the feature layer.

[0085] Randomly select multiple foreground points (x, y, z) from the threshold segmentation Mask (x, y, z) obtained above, which must contain a Mask (x, y, z) = 1, and cut the foreground points (x, y, z) based on the size of the sampling block s. Get the region of interest.

[0086] It should be noted that for step 3, the main body segmentation network structure of this embodiment is based on UNet. Figure 3As shown in FIG, this embodiment adjusts the existing framework based on the application of the existing UNet technology to make it more suitable for the aneurysm detection processing framework. In general, as the network goes deeper, the target is easily overwhelmed by the number of network layers. This embodiment designs a context attention mechanism modulation module so that the context information of each spatial position is aggregated separately. For details, see Figure 4 .

[0087] Given a low-semantic feature X and a high-semantic feature Y (features at the lower layer) with C channels and a feature map of size H×W×D, the top-down attention adjustment obtains the global information guidance of the low-semantic feature, which can be formulated as:

[0088] Q=XW Q ,K=XW K ,V=XW V (4)

[0089] Among them, W Q , W K , W V is a trainable weight matrix.

[0090] The attention weight is then calculated and can be formulated as:

[0091]

[0092] Among them, d k is the dimension of the key, used to scale the dot product to prevent the values ​​from being too large.

[0093] The bottom-up point channel attention modulation module obtains local information guidance of high semantic features, where the channel feature context of each spatial position is aggregated separately, which can be formulated as:

[0094]

[0095] Among them, the convolution kernel of Connv_1 is 1×1, and the number of channels is compressed to The convolution kernel of Conv_2 is 1×1, and the number of channels is reset to C.

[0096] The attention information of the two directions is fused using the subsequent context-adjusted attention mechanism module, which can be formulated as:

[0097]

[0098] It should be noted that the contextual attention mechanism modulation module invented in this embodiment supplements the most advanced top-up high-level semantic feedback path, and adopts a reverse bottom-up contextual modulation path to encode smaller-scale visual details to a deeper level, thereby providing richer encoding of semantic information and spatial details for microaneurysm detection.

[0099] This embodiment is also modified based on the original UNet architecture to an integrated combination loss of CE and DICE to better adapt to the task of aneurysm detection. The specific CE function formula is as follows:

[0100]

[0101] In order to make the generated target closer to the foreground, DICE LOSS is increased. The DICE function formula is as follows:

[0102]

[0103] The final combination loss is calculated as follows:

[0104] Total Loss=α·CE LOSS+β·DICE LOSS (10)

[0105] Among them, α and β are weight factors used to balance the two losses; y i is the true label of the ith pixel / voxel, p i is the predicted probability of the ith pixel / voxel.

[0106] It should be noted that for step 4, in order to obtain global image features, the subject classification network structure of this embodiment is based on ViT. Figure 5 As shown, the classification network proposed in this embodiment accepts the area around the lesion after three different random enhancement changes (including affine transformation, contrast adjustment, size scaling, and random flipping, etc.) as input, and performs integrated learning on the network results, thereby reducing the influence of different lesion positions and sizes and increasing the robustness of the model.

[0107] It should be noted that for step 5, in order to provide a clearer position for the radiologist, the present invention combines previous work and projects the display result of the aneurysm lesion position onto the MIP map after the anterior and posterior circulation segmentation. Figure 6 As shown in the figure, the segmentation visualization of three different locations and sizes of aneurysms on the original image and MIP image. The projection angle of the present invention is 15 degrees to the left and 15 degrees to the top and bottom. Assuming that the input image is V (x, y, z) and the direction vector is the formula (d x ,d y ,d z ), the formula can be expressed as follows:

[0108]

[0109] Where t is the projected distance along the direction vector.

[0110] Experimental details and parameter settings: The model is implemented in Python and the code uses the Pytorch framework. The experiment uses the SGD optimizer and the weight decay is empirically determined to be 3×10 -5 , which are uniformly applied to all experiments. The batch size is set to 2. The initial learning rate is set to 2×10 -2 , PolyLR will decay the learning rate after each Epoch. A total of 1000 rounds of training are performed.

[0111] Comparison method: In order to verify the effectiveness of the proposed method, the basic framework UNet architecture was selected as a comparison. In addition, it was compared with the most advanced SwinUNet and MedNeXT methods.

[0112] Evaluation indicators: In order to compare with other methods, dice, sensitivity (also known as sensitivity), and FPs / Case (also known as the number of false positives per case) are used as evaluation indicators for detection. All aneurysms are considered positive. Since the result of the system is a cubic area, if more than 30% of the aneurysms are in this area, it is considered to be discovered. If all aneurysms in a case are discovered, it is considered to be TP (true positive). On the other hand, if there are undiscovered aneurysms in the case, it is considered to be FN (false negative). If there is a false detection in this case, that is, there is no aneurysm at the detection site, it is considered to be FP (false positive). There may be multiple FPs in a case. According to the above classification, we can get four sample situations, as shown in the following table:

[0113] Table 1 Sample situation

[0114] Actual situation\predicted situation Positive Negative Positive TP FP Negative TN FN

[0115] According to the above classification, the following sensitivity and FPs / Case expression formula can be obtained:

[0116]

[0117] Where N is the total number of lesions in aneurysms.

[0118] The experimental results are shown in the following table:

[0119] Table 2 Experimental results

[0120]

[0121]

[0122] The preferred specific embodiments of the present invention are described in detail above. It should be understood that a person skilled in the art can make many modifications and changes based on the concept of the present invention without creative work. Therefore, any technical solution that can be obtained by a person skilled in the art through logical analysis, reasoning or limited experiments based on the concept of the present invention on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A method for detecting intracranial aneurysms by MRA, characterized in that the steps include: Acquire the MRA image data to be detected and extract the region of interest from the image data to obtain local image blocks around blood vessels and lesions; The local image block is sent to the segmentation network, the backbone of which adopts the UNet architecture of the residual structure, and the deep feature extraction of the aneurysm image block to be segmented is performed to obtain multi-scale features, and the feature maps of different scales are fused with the upsampling results through the context attention modulation module to obtain the feature segmentation results of different scales; The obtained segmentation results are sent to the classification network. The classification network uses an integrated learner to input multiple randomly enhanced lesion area images into the ViT network to predict the aneurysm probability, and integrate the output results of each ViT network to obtain the final predicted aneurysm probability.

2. The MRA intracranial aneurysm detection method according to claim 1, characterized in that: The extraction of the region of interest from the image data is specifically as follows: Calculate MRA image {X i } N Center point, based on the center position, remove irrelevant areas; Set the image threshold σ and extract the partial mask {X′ i } N ; Randomly set the image sampling block size s and the number of dense sampling times n, and for the MRA image {X i } N In a balanced way, {X′ i } N The foreground is upsampled n times to obtain an image block of size s 3. The MRA intracranial aneurysm detection method according to claim 2, characterized in that: The image threshold is defined using the gray value histogram of the image and background, and the most suitable segmentation boundary is obtained by Otsu's method and used as the image threshold σ: Among them, ω1 and m1 are the weight and mean of the foreground, m T is the total mean, and the optimal image threshold is obtained by maximizing σ; Based on the calculated threshold, the image is binarized to generate a mask: Where I(x,y,z) is the pixel value of the image at position (x,y,z) and σ is the image threshold.

4. The MRA intracranial aneurysm detection method according to claim 3, characterized in that: Randomly select multiple foreground points (x, y, z) from the mask Mask(x, y, z) based on threshold segmentation, which must contain a Mask(x, y, z) = 1, and cut the foreground points (x, y, z) based on the size s of the image sampling block to obtain the region of interest.

5. The MRA intracranial aneurysm detection method according to claim 3, characterized in that: The size s of the image sampling block is set by the following formula: s=D min ×2 L Among them, D min is the minimum downsampling boundary of the network, and L is the number of network feature extraction layers.

6. The MRA intracranial aneurysm detection method according to claim 1, characterized in that: The segmentation network is specifically as follows: Mapped to feature maps of different scales through multi-layer feature extraction networks, each layer of feature extraction network contains a residual structure to retain the low-level semantic features of the target; Based on the designed contextual attention modulation module, Feature maps of different scales are obtained through the encoding module and the context attention modulation module, and then multi-scale segmentation results are obtained through the decoder; Each upsampling layer is deeply supervised, and the loss is calculated by comparing the result of each upsampling layer with the mask of the corresponding scale to constrain the output results of feature maps of different scales.

7. The MRA intracranial aneurysm detection method according to claim 6, characterized in that: The segmentation network adopts the context attention mechanism modulation module so that the context information of each spatial position is aggregated separately, as follows: For a feature map with C channels of size H×W×D; The global information guidance of low-semantic features is obtained through the top-down attention modulation module, and then the attention weight Attention(Q,K,V) is calculated; The local information guidance of high-semantic features is obtained through a bottom-up point-channel attention modulation module, where the channel feature context of each spatial location is aggregated separately: Among them, the convolution kernel of Conv_1 is 1×1, and the number of channels is compressed to The convolution kernel of Conv_2 is 1×1, and the number of channels is reset to C; The context-adjusted attention mechanism module is used to fuse the attention information in two directions: Among them, X is the low semantic feature of the feature map; Y is the high semantic feature of the feature map.

8. The MRA intracranial aneurysm detection method according to claim 1, characterized in that: The segmentation network adopts the integrated combination loss of CE and DICE Total Loss: Total Loss=α·CE LOSS+β·DICE LOSS Among them, α and β are weight factors; y i is the true label of the ith pixel / voxel, p i is the predicted probability of the ith pixel / voxel.

9. The MRA intracranial aneurysm detection method according to claim 1, characterized in that: The classification network performs a variety of different random enhancements on the input to change the area around the lesion, and then inputs each of them into an independent Vit network to predict the probability, and integrates the prediction results of each Vit network for learning.

10. The MRA intracranial aneurysm detection method according to claim 1, characterized in that: The method performs post-processing on the aneurysm mask with a high probability of predicting aneurysm output by the classification network, and displays it on the MIP map after the anterior and posterior circulation segmentation of the blood vessel, specifically: The aneurysm lesion position is projected onto the MIP image after the anterior-posterior cycle segmentation. The projection angle is 15 degrees to the left and 15 degrees to the top and bottom. For the input image, the direction vector is V(x, y, z) and (d x ,d y ,d z ) is expressed as: Where t is the projected distance along the direction vector.

Images