Intracranial artery lumen wall segmentation method based on HRMRI image

Abstract

The invention relates to the field of medical informatics, and provides an intracranial artery lumen wall segmentation method based on an HRMRI (High Resolution Magnetic Resonance Imaging) image, which comprises the following steps of: acquiring the HRMRI image and respectively marking a lumen and a wall of an intracranial artery to obtain a data set; extracting features of the data set through an encoder network to obtain high-dimensional features; establishing a double-branch decoder based on spatial position coding; and decoding the high-dimensional features through the double-branch decoder to obtain an intracranial artery segmentation image comprising a lumen segmentation result and a tube wall segmentation result. According to the invention, the efficiency and accuracy of intracranial artery lumen and tube wall segmentation are improved.

Description

Technical Field

[0001] The present invention relates to the technical field of medical informatics, and in particular, to a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images. Background Art

[0002] In clinical practice, quantitative indicators such as vascular wall thickness and plaque burden are important means for measuring plaques. As a modern imaging technology, HRMRI can obtain cross-sectional images of arteries and detect early abnormal changes in the vascular wall. The quantitative indicators of plaques are achieved by identifying the inner and outer wall boundaries of the vascular wall on HRMRI images. However, due to the complex signal characteristics near the vascular wall, layer-by-layer analysis of HRMRI images has certain complexity and difficulty in reproducibility, making the assessment of intracranial atherosclerotic plaque burden cumbersome. Clinicians usually need to spend a lot of time analyzing hundreds of frames of HRMRI images to accurately judge the condition. By visually identifying and manually outlining the intimal lumen, plaques, and the media-adventitia region, it is not only time-consuming and laborious but also prone to human errors and difficult to ensure accuracy. Existing vascular and lumen segmentation methods are mainly divided into two categories: one is the segmentation method based on classical machine learning, which relies on rule-based algorithms such as filtering enhancement, region growing, level set, and centerline guidance. These methods use the image features of the vascular lumen and wall for segmentation, but due to high manual dependence, it is difficult to be extended to clinical practice. The other is the method based on deep learning, which has received attention due to its higher scalability and accuracy. For example, fine features of intracranial vascular structures are learned through a feature extraction network. However, these methods usually ignore the correlation between the overall morphology of blood vessels and the plaque location, resulting in breaks in the segmentation results in relatively narrow vascular regions and unable to fully display the continuous structure of the lumen and wall. In traditional methods, the segmentation tasks of the lumen and the wall are usually separated and carried out independently, and two unrelated models are trained based on the original images respectively. Due to the lack of information sharing, the anatomical topological constraints of the vascular wall are difficult to be transmitted to the lumen segmentation process, limiting the model's understanding of the structural integrity. At the same time, the two models are prone to Feature Space Misalignment during the independent learning process. Even if the same image is input, there may be significant differences in the underlying feature expressions due to different focus areas (such as the lumen focusing on edge gradient changes and the wall focusing on texture features). Therefore, there is an urgent need in the art for a method for segmenting the lumen and wall of intracranial arteries that considers the continuous structure of the lumen and the wall. Summary of the Invention

[0003] The present invention provides a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images to solve the defects of the prior art.

[0004] The present invention provides a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images, including: S1: Collect HRMRI images and label the lumen and wall of intracranial arteries respectively to obtain a data set; S2: Extract the features of the data set through an encoder network to obtain high-dimensional features; S3: Establish a dual-branch decoder based on spatial position encoding; S4: Decode the high-dimensional features through the dual-branch decoder to obtain an intracranial artery segmentation image including the lumen segmentation result and the wall segmentation result.

[0005] According to the method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by the present invention, step S1 further includes: S11: Collect HRMRI images; S12: Preprocess the HRMRI images to obtain preprocessed images; S13: Register the preprocessed images through a reference image to obtain registered images; S14: Label the lumen of intracranial arteries and the wall of intracranial arteries in the registered images respectively to obtain a data set including multiple labeled registered images.

[0006] According to the method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by the present invention, the preprocessing of the HRMRI images in step S12 is to remove artifact images, remove signal loss images, and remove abnormal noise images.

[0007] According to the method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by the present invention, step S2 further includes: S21: Input the data set; S22: Convolve the images in the data set through a convolutional block to obtain a convolutional feature map; S23: Apply instance normalization to the convolutional feature map to obtain a normalized feature map; S24: Apply the ReLU activation function to the normalized feature map to obtain a non-linearly transformed feature map; S25: Pool the non-linearly transformed data to obtain a pooled feature map; S26: Stack the pooled feature maps output by multiple convolutional blocks to obtain high-dimensional features.

[0008] According to the method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by the present invention, the expression of the normalized feature map in step S23 is: ; Among them, is the sample index value, is the channel index value, is the height index value of the feature map, is the width index value of the feature map, is the th element of the -th channel of the -th input convolutional feature map with height is the th element of the -th channel of the -th output normalized feature map with height is the th mean value of the th channel of the th input convolutional feature map, is the variance of the

[0009] According to a method for intracranial artery lumen wall segmentation based on HRMRI images provided by the present invention, step S4 further includes: S41: Input the high-dimensional features; S42: Based on the self-attention mechanism, calculate the attention weights of the high-dimensional features to obtain a weighted feature map; S43: Decode the weighted feature map through the double-branch decoder and introduce spatial position encoding to obtain an intracranial artery segmentation image.

[0010] According to a method for intracranial artery lumen wall segmentation based on HRMRI images provided by the present invention, in step S43, the step of decoding the weighted feature map through the double-branch decoder further includes: S4311: Upsample the weighted feature map to obtain an upsampled feature map; S4312: Perform deconvolution on the upsampled feature map to obtain a deconvolution feature map; S4313: Add a bias term to the deconvolution feature map to obtain a transposed convolution feature map.

[0011] According to a method for intracranial artery lumen wall segmentation based on HRMRI images provided by the present invention, in step S43, the step of introducing spatial position encoding further includes: S4321: Calculate the spatial position encoding through a trigonometric function-based encoding function according to the coordinates of the high-dimensional features in the matrix; S4322: Add the spatial position encoding to each layer output feature map of the dual-branch decoder through a connection function.

[0012] According to a method for segmenting intracranial artery lumen wall based on HRMRI images provided by the present invention, the expression of the spatial position encoding in step S4321 is: ; in, To calculate the spatial position encoding, are the coordinates of the features in the matrix, is the batch size, is the first learning parameter of the encoding function, is the second learning parameter of the encoding function, is the third learned parameter of the encoding function.

[0013] According to a method for segmenting the lumen and wall of an intracranial artery based on HRMRI images provided by the present invention, the dual-branch decoder in step S3 is residually connected to the corresponding layer of the encoder network, and the structure of the dual-branch decoder specifically includes a parallel first branch decoder and a second branch decoder, the first branch decoder and the second branch decoder share a spatial position code, the first branch decoder is used for lumen segmentation, and the second branch decoder is used for wall segmentation.

[0014] The present invention provides a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images, which introduces a deep learning model with shared feature spatial position encoding and self-attention mechanism, and is used to automatically obtain the segmentation structure of the intracranial lumen and vascular wall based on the high-resolution magnetic resonance imaging of the intracranial arteries of patients. The present invention mainly includes the following steps: obtaining a high-resolution magnetic resonance imaging dataset of intracranial arteries; constructing an encoder network based on a modified U-net architecture to extract high-dimensional features of HRMRI images; establishing a dual-branch decoder for segmenting the lumen and wall of intracranial blood vessels respectively; ensuring the linkage of the segmentation of the lumen and wall of blood vessels based on shared spatial position encoding; introducing a self-attention mechanism to improve the ability to capture local information; and outputting the segmentation results of intracranial blood vessels and walls.

[0015] The present invention extracts high-dimensional features of HRMRI images through an encoder network and combines a double-branch decoder based on spatial position encoding, which can more accurately segment the lumen and wall of intracranial arteries, fully utilize the spatial information and feature information of the images, and improve the accuracy of segmentation. Secondly, in the data preprocessing stage of the present invention, by eliminating artifact images, signal-loss images, and abnormal noise images, the noise and interference in the data can be effectively reduced, the robustness of the model can be enhanced, and it helps the model to maintain stable and accurate segmentation performance in complex and variable HRMRI images. In addition, the design of the double-branch decoder enables the segmentation of the lumen and wall to be carried out in parallel, thereby improving the processing efficiency. At the same time, due to the introduction of spatial position encoding, the model can better understand the spatial structure of the image, further improving the segmentation speed and accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] In order to more clearly illustrate the technical solutions in the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the following drawings are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0017] Figure 1 Schematic flowchart of a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by an embodiment of the present invention; Figure 2 Schematic diagram of the principle of a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images provided by an embodiment of the present invention; Figure 3 Schematic diagram of the effect of 3D vascular segmentation by the threshold filtering method and the region growing method provided by an embodiment of the present invention; Figure 4 Schematic diagram of the comparison of 3D vascular segmentation effects between the Unet model provided by an embodiment of the present invention and the model of the present invention; Figure 5 Schematic diagram of the comparison of the cross-sectional wall and lumen segmentation effects between the traditional Unet model and the model of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0018] To make the objectives, technical solutions and advantages of the present invention clearer, the technical solutions in the present invention will be clearly and completely described below with reference to the accompanying drawings in the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention, and they should not be construed as limiting the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the scope of protection of the present invention. In the description of the present invention, it should be understood that the terms used are only for the purpose of description and should not be construed as indicating or implying relative importance.

[0019] To better understand the embodiments of the present invention, the research background of the present invention will be explained in detail below.

[0020] Intracranial atherosclerosis is the most important pathological change leading to ischemic stroke, which is characterized by thickening of the vessel wall and plaque formation. Therefore, obtaining clear and detailed vessel wall and plaque structures is a prerequisite for accurate diagnosis and risk assessment of cerebrovascular diseases. Currently, traditional vascular examination techniques in clinical practice mainly include Transcranial Doppler (TCD), Computed Tomography Angiography (CTA), Magnetic Resonance Angiography (MRA), Digital Subtraction Angiography (DSA), etc. These examinations have low resolution and can only show the remaining lumen at the plaque, providing less information about the plaque itself and unable to accurately judge the condition of the vessel wall lesion itself. Moreover, CTA and DSA have certain radioactivity, and DSA is an invasive examination that may bring the risk of complications. High-resolution magnetic resonance imaging (HRMRI) can non-invasively and clearly display the vessel wall and its lesions, avoiding the possible invasive operations and complication risks in angiography. It can simultaneously provide multi-planar imaging and different contrast weightings (such as T1, T2, PDW, etc.), and can observe the morphology, size, location and thickness of the plaque in all directions and from multiple angles, improving the comprehensiveness of the assessment. These advantages make HRMRI an ideal tool for evaluating the intracranial arterial vessel wall and plaque structure.

[0021] In clinical practice, quantitative indicators such as vascular wall thickness and plaque burden are important means for measuring plaques. As a modern imaging technology, HRMRI can obtain cross-sectional images of arteries, detect early abnormal changes in the vascular wall, and the quantitative indicators of plaques are achieved by identifying the inner and outer wall boundaries of the vascular wall on HRMRI images. However, due to the complex signal characteristics near the vascular wall, layer-by-layer analysis of HRMRI images has certain complexity and reproducibility difficulties, making the assessment of intracranial atherosclerotic plaque burden cumbersome. Clinicians usually need to spend a lot of time analyzing hundreds of frames of HRMRI images to accurately judge the condition. By visually identifying and manually outlining the intimal lumen, plaque, and media-adventitia regions, this is not only time-consuming and laborious, but also prone to human errors and difficult to ensure accuracy.

[0022] In the field of high-resolution nuclear magnetic resonance imaging, in terms of the segmentation of intracranial blood vessels, artificial intelligence, especially deep learning technology, has shown great application value. Due to its unique advantages in vascular wall analysis, high-resolution nuclear magnetic resonance imaging has now become a research hotspot for vascular wall segmentation on this basis. Through deep learning technology, fine segmentation of the vascular wall can be achieved, providing an important basis for the early diagnosis and treatment of diseases.

[0023] Compared with the wall and lumen characteristics of large blood vessels such as the carotid and thoracic arteries, there are the following challenges in achieving intracranial vascular wall and vessel segmentation: High anatomical complexity: Compared with carotid and thoracic vessels, intracranial arteries have more bifurcation structures and more complex three-dimensional trajectories. Although existing methods have tried to introduce prior constraints such as region size and centroid position, these geometric information are too simple to accurately express the complex morphological characteristics of intracranial arteries, resulting in limited model generalization ability; Large interference from image quality and morphology: In intracranial HRMRI images, there are often problems such as blurred boundaries, low signal-to-noise ratio, and flow void artifacts. Especially in complex bifurcation regions or lesion regions, existing segmentation methods are prone to missed detections or misjudgments, and some models still need to rely on manual correction, limiting the reliability of automated applications; The wall structure is tiny and difficult to label: The intracranial vascular wall is thin and has low contrast, and existing models are difficult to effectively capture its weak characteristic signals. Since the labeling process highly depends on clinical experts and is time-consuming and laborious, existing deep learning methods for carotid and thoracic vessels face problems of insufficient generalization and limited training data in such tasks.

[0024] To solve the above problems, the present invention proposes a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images.

[0025] The embodiments of the present invention will be described below with reference to the drawings.

[0026] As Figure 1 shown, the present invention provides a method for segmenting the lumen and wall of intracranial arteries based on HRMRI images, including: S1: Collect HRMRI images and label the lumen and wall of the intracranial arteries respectively to obtain a dataset.

[0027] Among them, step S1 further includes: S11: Collect HRMRI images.

[0028] In step S11, first select a magnetic resonance imaging device to ensure image quality, adjust the scanning parameters to obtain the best image effect, and ensure that the patient remains stationary during the scan to reduce the influence of motion artifacts.

[0029] S12: Preprocess the HRMRI images to obtain preprocessed images.

[0030] Among them, the preprocessing performed on the HRMRI images in step S12 is to remove artifact images, remove signal loss images, and remove abnormal noise images.

[0031] The preprocessing steps in step S12 are aimed at improving image quality, reducing noise and interference factors. Specifically, it includes removing artifact images: check whether there are artifacts in the images caused by equipment failures, patient movements or other factors, and remove these images; removing signal loss images: due to patient movement or equipment failure during the scan, signal loss may occur, so these images should also be removed to avoid interference with subsequent analysis; removing abnormal noise images: remove abnormal noise in the images, and the noise is caused by equipment noise, environmental interference, etc., and filters need to be used to reduce the noise.

[0032] S13: Register the preprocessed images with a reference image to obtain registered images.

[0033] Step S13 aims to ensure the spatial consistency of all images. First, select a high-quality image with clear anatomical structure as the reference image, and then use affine transformation to translate, rotate, and scale other preprocessed images to align with the reference image, so that all images are consistent in anatomical structure.

[0034] S14: Label the lumen of the intracranial artery and the wall of the intracranial artery in the registered images respectively to obtain a dataset including multiple labeled registered images.

[0035] In step S41, first, manually or using semi-automatic tools, mark the lumen and wall regions of the intracranial artery in the registered images. Subsequently, combine all the labeled registered images into a dataset, and the obtained dataset will be used for training machine learning.

[0036] S2: Extract the features of the dataset through an encoder network to obtain high-dimensional features.

[0037] Among them, step S2 further includes: S21: Input the said data set.

[0038] S22: Convolve the images in the said data set through a convolutional block to obtain a convolutional feature map.

[0039] S23: Apply instance normalization to the said convolutional feature map to obtain a normalized feature map.

[0040] Among them, the expression of the said normalized feature map in step S23 is: ; Among them, is the sample index value, is the channel index value, is the height index value of the feature map, is the width index value of the feature map, is the th element of the th channel of the -height and -width of the th input convolutional feature map, is the th element of the -height and -width of the th output normalized feature map, is the mean value of the th channel of the th input convolutional feature map, is the variance of the th channel of the th input convolutional feature map,

[0041] S24: Apply the ReLU activation function to the said normalized feature map to obtain a non-linearly transformed feature map.

[0042] S25: Pool the said non-linearly transformed data to obtain a pooled feature map.

[0043] Furthermore, the pooling layer in step S25 is used to reduce the dimension of the feature map while retaining the most important information. Pooling helps reduce the computational amount while improving the generalization ability of the model.

[0044] S26: Stack the pooled feature maps output by multiple convolutional blocks to obtain high-dimensional features.

[0045] Specifically, the main purpose of the encoder in step S2 is to extract high-dimensional features from the input intracranial artery HRMRI images, providing rich information for subsequent segmentation. The encoder is based on an improved U-Net structure, specifically including: Convolutional blocks: Each convolutional block contains a convolutional layer, an instance normalization layer (Instance Normalization), and an activation function (LeakyReLU). The convolutional layer uses 3D convolution (Convolutional3D) because 3D convolution can capture the hierarchical relationships of the images and is more suitable for processing 3D MRA images compared to 2D convolution; Convolution and pooling: During the convolution process, the image gradually reduces in size through multiple convolutional layers and pooling layers (Pooling Layer) to extract concentrated image features. Instance normalization helps maintain the stability of the features and speeds up the convergence rate.

[0046] Furthermore, the encoder in step S2 consists of 4 consecutive convolutional blocks. After 4 times of convolution and pooling, a concentrated feature map is finally obtained, which contains the high-dimensional feature information of the image, forming a high-dimensional feature representation that increases the capacity of the model and enables it to learn more complex data structures.

[0047] S3: Establish a dual-branch decoder based on spatial position encoding.

[0048] Among them, the dual-branch decoder in step S3 is residually connected to the corresponding layers of the encoder network. The structure of the dual-branch decoder specifically includes a parallel first-branch decoder and a second-branch decoder. The first-branch decoder and the second-branch decoder share the spatial position encoding. The first-branch decoder is used for lumen segmentation, and the second-branch decoder is used for wall segmentation.

[0049] Furthermore, the dual-branch decoder is for simultaneously processing the segmentation tasks of the lumen and the wall. Since the lumen and the wall are closely related in anatomical structure but have their own unique features, establishing two parallel branches in step S3 to process them separately can more effectively capture and utilize this information.

[0050] The dual-branch decoder consists of two parallel decoder branches. The established dual-branch decoder shares the same spatial position encoding. Spatial position encoding is a method of embedding spatial position information into the feature representation, which can ensure the spatial consistency of the model during the decoding process, thereby more accurately locating the lumen and the wall.

[0051] In addition, residual connections are established between the corresponding layers of the dual-branch decoder and the encoder network. By directly adding the input to the output, the residual connections can alleviate the problem of vanishing gradients in deep networks and accelerate the training process. The residual connections enable the features extracted by the encoder network to be directly transmitted to the corresponding layers of the decoder, thereby enhancing the feature reusability and information flow.

[0052] In the decoder stage of step S3 and subsequent step S4, the present invention constructs two parallel decoder branch structures for intracranial vascular lumen segmentation and intracranial vascular wall segmentation respectively, and the spatial position encoding of the two branches is shared. Since, from a physiological perspective, where there is a vascular lumen, it must be surrounded by a vascular wall, there is a mutually dependent relationship between the intracranial vascular lumen and the intracranial vascular wall, thereby enhancing the linkage effect during the segmentation of the intracranial vascular lumen and the intracranial vascular wall.

[0053] To solve the problem of feature disappearance during multiple convolutions of blood vessels in the decoder part, residual connections are introduced to directly transmit the features in the encoder to the decoder. Each layer of the decoder is connected corresponding to a certain layer of the encoder. The residual link can be simply understood as a linear function, and the expression is: ; where, represents the input of this layer, represents the output of this layer for the input after a series of transformations (such as convolution, activation function, etc.), represents the output after the residual connection. This formula embodies the core idea of the residual connection, that is, directly adding the input to the output of this layer to form the output after the residual connection.

[0054] S4: Decode the high-dimensional features through the dual-branch decoder to obtain an intracranial artery segmentation image including the lumen segmentation result and the wall segmentation result.

[0055] Among them, step S4 further includes: S41: Input the high-dimensional features.

[0056] S42: Calculate the attention weights of the high-dimensional features based on the self-attention mechanism to obtain a weighted feature map.

[0057] Due to the design of the dual-branch decoder structure and the introduction of spatial position encoding as the communication channel between the dual-branch decoders, the decoder cannot obtain the correct features when performing the corresponding segmentation tasks (i.e., the unique features of the lumen will affect the results during wall segmentation, and the unique features of the wall will affect the results during lumen segmentation). Therefore, in step S42 of the present invention, a self-attention mechanism is introduced into the decoder. Through the self-attention mechanism, when the two branch decoders pay more attention to the unique features of their own tasks, they can also fuse the spatial position relationship. The obtained weighted feature map emphasizes the parts of the input features that are more important for the segmentation task, enabling the model to more accurately capture the features of the intracranial artery lumen and wall.

[0058] S43: Decode the weighted feature map through the dual-branch decoder, and introduce spatial position encoding to obtain an intracranial artery segmentation image.

[0059] Among them, in step S43, the step of decoding the weighted feature map through the dual-branch decoder further includes: S4311: Upsample the weighted feature map to obtain an upsampled feature map.

[0060] During the decoding process, since the encoder network usually includes multiple downsampling (such as pooling) steps, resulting in a gradual decrease in the resolution of the feature map, in order to restore the resolution of the original image in step S43, the decoder network first needs to perform an upsampling operation. The purpose is to increase the resolution of the weighted feature map to a level close to that of the original image or the target resolution. Finally, the feature map obtained after upsampling is called the upsampled feature map, which retains the important information in the original weighted feature map but has a higher resolution.

[0061] S4312: Perform deconvolution on the upsampled feature map to obtain a deconvolved feature map.

[0062] The purpose of deconvolution is the opposite of convolution, that is, to increase the resolution of the feature map. In the deconvolution operation of step S4312, specifically, a convolution kernel (or filter) is used to slide on the upsampled feature map, but the size of the output feature map will be larger than that of the input feature map. The feature map obtained after the deconvolution operation is called the deconvolved feature map, whose resolution is further increased while retaining the key information in the upsampled feature map.

[0063] S4313: Add a bias term to the deconvolved feature map to obtain a transposed convolution feature map.

[0064] In step S4313, after the deconvolution operation, a bias term needs to be added to the feature map. The bias term allows the model to add a learnable constant at each position of the output feature map, which helps the model better fit the training data and may improve the accuracy of segmentation.

[0065] The bias term is usually a vector with the same number of channels as the output feature map, and each element corresponds to the bias value of a channel. Before passing the transposed convolution feature map to the next layer, this bias vector needs to be added to each channel of the feature map. The feature map obtained after adding the bias term is called the transposed convolution feature map, which contains higher-resolution segmentation information for the subsequent final segmentation output.

[0066] Among them, in step S43, the step of introducing spatial position encoding further includes: S4321: Calculate and obtain the spatial position encoding according to the coordinates of the high-dimensional feature in the matrix through a trigonometric function-based encoding function.

[0067] The purpose of step S4321 is to calculate and obtain the spatial position encoding according to the coordinates of the high-dimensional feature in the matrix to ensure spatial consistency during the decoding process. The specific implementation method is to use a trigonometric function-based encoding function to calculate the spatial position encoding, which can capture information of different frequencies when processing periodic data and is more suitable for the separate encoding of the lumen wall. The finally calculated spatial position encoding is a vector with the same dimension as the feature, containing information about the spatial position of the feature.

[0068] Among them, the expression of the spatial position encoding in step S4321 is: ; Among them, is the calculated spatial position encoding, are the coordinates of the feature in the matrix, is the batch size, is the first learning parameter of the encoding function, is the second learning parameter of the encoding function, is the third learning parameter of the encoding function.

[0069] S4322: Add the spatial position encoding to the output feature map of each layer of the double-branch decoder through a connection function.

[0070] The purpose of step S4322 is to add the calculated spatial position encoding to the output feature map of each layer of the double-branch decoder to enhance the spatial information of the feature map. Using the spatial position encoding as the channel between the double-branch decoders enables the two decoders to transmit information to each other.

[0071] Specifically, in step S4322, spatial position encoding is added to the output feature map of each layer of the dual-branch decoder. That is, at each stage of the decoding process, the feature map is combined with spatial position information to maintain spatial consistency. By introducing spatial position encoding, the model can better understand the spatial relationships in the feature map and improve the accuracy of segmentation. Especially when dealing with the lumen and wall segmentation of intracranial arteries with complex spatial structures in medical images in the present invention, it can better ensure the correspondence of spatial position information.

[0072] The spatial position encoding in steps S4321 to S4322 is a technique to enhance the model's spatial understanding, especially suitable for segmenting structures such as intracranial blood vessels, and helps the model more accurately locate features in the image.

[0073] The process of introducing spatial position encoding is as follows: In each layer of the decoder, multi-level and multi-dimensional features extracted by the encoder are fused to generate a highly refined feature representation, and spatial position encoding is embedded therein. The position encoding is used to mark the coordinates of the features in the image to help the model identify different image regions during the segmentation process.

[0074] Subsequently, the two branches share the same position encoding to ensure the spatial correlation of lumen and wall segmentation. Specifically, in each layer of decoding, the position encoding is embedded into the feature map so that the lumen and the wall can cooperate during segmentation.

[0075] The core of the spatial position encoding mechanism introduced in the present invention is that at the end stage of the encoder processing flow, through an efficient feature aggregation process, all the extracted multi-level and multi-dimensional features are concentrated and integrated to form a highly refined and information-rich feature representation. Subsequently, this concentrated feature set is further fed into a specific position encoding module to perform precise spatial position encoding on regions with significant features. Subsequently, the present invention adds the spatial encoding to the feature map of each layer of the decoder using the concat function and introduces the self-attention mechanism in each layer.

[0076] Finally, after the processing of steps S1 to S4, the dual-branch decoder can output two segmentation result images, corresponding to the lumen and wall of the intracranial blood vessels respectively. By combining the segmentation results of the lumen and the wall, a complete intracranial artery segmentation image is obtained, which is convenient for subsequent analysis and diagnosis of intracranial vascular lesions.

[0077] Such as Figure 2As shown in the figure, it is a schematic diagram of the network principle of the intracranial artery lumen and wall segmentation method based on HRMRI images of the present invention. First, the original HRMRI image dataset is input. Specifically, for the original HRMRI image dataset, it is the intracranial artery image data obtained by high-resolution magnetic resonance imaging scanning. This dataset contains the intracranial artery HRMRI images of multiple patients, with high image resolution, which can clearly present the structure and details of the intracranial arteries. For each image, a professional medical image annotation tool is used for manual annotation. The annotation content includes: the artery lumen area and the vascular wall area, to ensure the integrity and accuracy of the image information and provide high-quality annotation data for subsequent model training. All annotations are completed by radiologists to ensure the consistency and reliability of the annotation data.

[0078] Subsequently, a dual-branch decoder structure is established through an encoder network and a decoder network with residual connections. Sharing the same encoder for the two tasks can enable the model to obtain the same feature representation. Then, by establishing channel transmission of spatial feature encoding between the two branch decoders, the lumen model can know the segmentation position of the wall model, so that the two models can guide each other in segmentation, thereby improving the segmentation effect. Finally, the results of segmenting the internal structure of the intracranial blood vessels are output, that is, the intracranial blood vessel lumen segmentation result and the intracranial blood vessel wall segmentation result. The obtained segmentation results are of great significance for subsequent research on vascular lesions and evaluation of vascular health status, etc.

[0079] The following combines Figures 3 to 5 to describe the result comparison between the intracranial artery lumen and wall segmentation method based on HRMRI images provided by the present invention and traditional segmentation methods.

[0080] As Figure 3 shown, it is the effect of 3D blood vessel segmentation by traditional threshold filtering method and region growing method. It can be seen from Figure 3 that due to the blood vessels passing through the middle of the bones in the intracranial basal segment, the traditional segmentation effect is extremely poor. Figure 4 It is the 3D blood vessel segmentation effect diagram of the traditional Unet model and the improved model of the present invention. Among them, the green part is the segmentation result of the traditional Unet model, and the corresponding dice coefficient is 0.852. The red part is the segmentation result of the improved model of the present invention, and the corresponding dice coefficient is 0.878. Figure 5 It is the cross-sectional wall and lumen segmentation effect diagram of the traditional Unet model and the improved model of the present invention. Among them, Figure 5 the left figure in it is the result schematic of the traditional Unet model segmentation. The yellow inner surrounded shape is the lumen, and the green outer surrounded shape is the wall. The lumen dice coefficient is 0.868, and the wall dice coefficient is 0.876; Figure 5The right figure in the middle shows the result of segmentation by the improved model of the present invention. The lumen Dice coefficient is 0.900, and the wall Dice coefficient is 0.902. The Dice coefficient is a geometric similarity metric function used to measure the similarity between the model prediction and the true target. Its value range is from 0 to 1. The closer it is to 1, the better the model construction effect. It can be seen from the above results that the improved model established by the method for segmenting the intracranial artery lumen and wall based on HRMRI images of the present invention has significantly better effect in lumen and wall segmentation than the traditional method, with high accuracy.

[0081] A method for segmenting the intracranial artery lumen and wall based on HRMRI images provided by the present invention significantly improves the segmentation accuracy of the intracranial artery lumen and wall through a dual-branch decoder structure, shared feature space position encoding, and self-attention mechanism, avoids the common fracture phenomenon in traditional methods, and solves the problem of the fragmentation of the traditional lumen and wall segmentation tasks. The use of the dual-branch decoder structure and residual connection ensures the spatial continuity and consistency of the segmentation, making the segmentation result more in line with the anatomical structure of the blood vessel and helping to present the global morphology completely. In summary, the present invention is not only superior to the traditional method in terms of segmentation accuracy and stability, but also provides strong technical support for the accurate diagnosis of intracranial artery diseases.

[0082] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. An intracranial artery lumen and wall segmentation method based on HRMRI images, characterized in that, Including: S1: Collect HRMRI images and separately label the lumen and wall of intracranial arteries to obtain a dataset; S2: Extract the features of the dataset through an encoder network to obtain high-dimensional features; S3: Establish a dual-branch decoder based on spatial position encoding; S4: Decode the high-dimensional features through the dual-branch decoder to obtain an intracranial artery segmentation image including lumen segmentation results and wall segmentation results.

2. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 1, characterized in that Step S1 further includes: S11: Collect HRMRI images; S12: Preprocess the HRMRI images to obtain preprocessed images; S13: Register the preprocessed images with a reference image to obtain registered images; S14: Separate label the lumen of intracranial arteries and the wall of intracranial arteries in the registered images to obtain a dataset including multiple labeled registered images.

3. The intracranial artery lumen and wall segmentation method based on HRMRI images according to claim 2, characterized in that, The preprocessing of the HRMRI images in step S12 is to remove artifact images, remove signal loss images, and remove abnormal noise images.

4. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 1, characterized in that, Step S2 further includes: S21: Input the dataset; S22: Convolve the images in the dataset through a convolutional block to obtain a convolutional feature map; S23: Apply instance normalization to the convolutional feature map to obtain a normalized feature map; S24: Apply the ReLU activation function to the normalized feature map to obtain a non-linearly transformed feature map; S25: Pool the non-linearly transformed data to obtain a pooled feature map; S26: Stack the pooled feature maps output by multiple convolutional blocks to obtain high-dimensional features.

5. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 4, characterized in that, The expression of the normalized feature map in step S23 is: ； Among them, is the sample index value, is the channel index value, is the height index value of the feature map, is the width index value of the feature map, is the th element of the th channel of the -height and -width convolutional feature map of the th input, is the th element of the -height and -width normalized feature map of the th output, is the mean value of the th channel of the th input convolutional feature map, is the variance of the th channel of the th input convolutional feature map, is the anti-zero constant.

6. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 1, characterized in that, Step S4 further includes: S41: Input the high-dimensional features; S42: Calculate the attention weights of the high-dimensional features based on the self-attention mechanism to obtain a weighted feature map; S43: Decode the weighted feature map through the dual-branch decoder and introduce spatial position encoding to obtain an intracranial artery segmentation image.

7. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 6, characterized in that, In step S43, the step of decoding the weighted feature map through the dual-branch decoder further includes: S4311: Upsample the weighted feature map to obtain an upsampled feature map; S4312: Perform deconvolution on the upsampled feature map to obtain a deconvolution feature map; S4313: Add a bias term to the deconvolution feature map to obtain a transposed convolution feature map.

8. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 6, characterized in that, In step S43, the step of introducing spatial position encoding further includes: S4321: Calculate the spatial position encoding according to the coordinates of the high-dimensional features in the matrix through a trigonometric-based encoding function; S4322: Add the spatial position encoding to each layer of output feature map of the dual-branch decoder through a connection function.

9. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 8, characterized in that, The expression of the spatial position encoding in step S4321 is: ； Among them, is the calculated spatial position encoding, is the coordinate of the feature in the matrix, is the batch size, is the first learning parameter of the encoding function, is the second learning parameter of the encoding function, is the third learning parameter of the encoding function.

10. A method for segmenting the lumen and wall of intracranial arteries based on HRMRI images according to claim 1, characterized in that The dual-branch decoder in step S3 is residually connected to the corresponding layer of the encoder network. The structure of the dual-branch decoder specifically includes a parallel first-branch decoder and a second-branch decoder. The first-branch decoder and the second-branch decoder share spatial position encoding. The first-branch decoder is used for lumen segmentation, and the second-branch decoder is used for wall segmentation.

Images