A time-of-flight magnetic resonance angiography cerebral vessel segmentation method, system, device and medium

Abstract

The application discloses a time-of-flight magnetic resonance angiography cerebral vessel segmentation method, system, device and medium, and relates to the field of medical image processing. The method comprises the following steps: acquiring a time-of-flight magnetic resonance angiography image to be processed; strengthening the blood vessel features in the time-of-flight magnetic resonance angiography image to be processed to obtain a preprocessed image; performing cerebral vessel segmentation on the preprocessed image by using a few-sample segmentation model to obtain a blood vessel probability map; the few-sample segmentation model is obtained by training a 3D U-Net by transferring the knowledge of a pre-trained video segmenter; and performing post-processing on the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed based on a human-computer interaction interface and a conditional random field to obtain a final segmentation image. The application can achieve high-precision and high-robustness segmentation effects only by using a small amount of samples, and the results can be optimized by an efficient human-computer interaction mode.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical image processing, in particular to a time-of-flight magnetic resonance angiography (TOF-MRA) brain vessel segmentation method, system, device and medium based on human-computer interaction. BACKGROUND

[0002] As a non-invasive neurovascular imaging technique without contrast agent, Time-of-Flight Magnetic Resonance Angiography (TOF-MRA) has important clinical value in the diagnosis and treatment of cerebral vascular diseases by taking advantage of the difference in magnetization between blood flow and static tissue. Accurate segmentation of the three-dimensional structure of cerebral vessels in its image is a key preprocessing step for precise quantitative analysis, surgical planning, and scientific research.

[0003] Despite the significant progress made in automated segmentation techniques, the inherent image quality problems of TOF-MRA remain a fundamental challenge to achieving accurate segmentation. The imaging principle of TOF-MRA is based on the inflow effect. While this mechanism can provide contrast between blood vessels and background tissue, it inevitably introduces complex and diverse image quality problems, which directly weaken the performance and reliability of automated segmentation algorithms.

[0004] To address the above challenges, cerebral vessel segmentation technology has undergone a development process from traditional image processing methods to modern deep learning models. However, the standard supervised learning paradigm still faces some inherent limitations when applied to TOF-MRA segmentation: first, it relies heavily on large amounts of high-quality labeled data, but labeling is time-consuming, expensive, and requires specialized knowledge, resulting in a lack of public datasets; second, the limited generalization ability of the model due to data scarcity makes it prone to overfitting, and its performance declines when processing data from different sources; third, the model is sensitive to image noise, artifacts, and other quality issues, and its robustness is insufficient to affect the segmentation quality; fourth, the segmentation capability of fine vascular systems is lacking, and there are challenges in outlining small blood vessels, maintaining blood vessel network topology, and processing pathological blood vessels, with some loss functions ignoring small blood vessels.

[0005] In addition, manual review and correction in clinical practice is necessary to ensure the accuracy of TOF-MRA cerebral vessel segmentation results. Existing interactive segmentation tools have manual and semi-automatic editing functions, but when used for this segmentation, users face challenges such as heavy workload, difficulty in learning, poor correction consistency, low efficiency in processing complex three-dimensional structures, and lack of intelligence.

[0006] In summary, current automated segmentation methods for TOF-MRA cerebral vessel images face challenges in dealing with data scarcity, image quality variability, small vessel segmentation accuracy, and interactive correction efficiency, among other issues. These problems limit their widespread application in clinical practice. SUMMARY

[0007] The purpose of the present application is to provide a time-of-flight magnetic resonance angiography brain vessel segmentation method, system, device and medium, which can reduce the dependence on large-scale labeled data and improve the robustness of segmentation, and efficiently obtain personalized segmentation results meeting clinical needs.

[0008] To achieve the above purpose, the present application provides the following scheme:

[0009] In a first aspect, the present application provides a time-of-flight magnetic resonance angiography brain vessel segmentation method, comprising:

[0010] Obtaining a time-of-flight magnetic resonance angiography image to be processed;

[0011] Strengthening the blood vessel features in the time-of-flight magnetic resonance angiography image to be processed to obtain a preprocessed image;

[0012] According to the preprocessed image, a few-sample segmentation model is used for brain vessel segmentation to obtain a blood vessel probability map; the few-sample segmentation model is obtained by training 3D U-Net by transferring the knowledge of a pre-trained video word segmenter;

[0013] Based on a human-computer interaction interface and a conditional random field, the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed are post-processed to obtain a final segmentation image.

[0014] In a second aspect, the present application provides a time-of-flight magnetic resonance angiography brain vessel segmentation system, comprising:

[0015] An image acquisition module is configured to acquire a time-of-flight magnetic resonance angiography image to be processed;

[0016] A preprocessing module is configured to strengthen the blood vessel features in the time-of-flight magnetic resonance angiography image to be processed to obtain a preprocessed image;

[0017] A preliminary segmentation module is configured to use a few-sample segmentation model to perform brain vessel segmentation according to the preprocessed image to obtain a blood vessel probability map; the few-sample segmentation model is obtained by training 3D U-Net by transferring the knowledge of a pre-trained video word segmenter;

[0018] A post-processing module is configured to post-process the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed based on a human-computer interaction interface and a conditional random field to obtain a final segmentation image.

[0019] In a third aspect, the present application provides a computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the time-of-flight magnetic resonance angiography brain vessel segmentation method.

[0020] In a fourth aspect, the present application provides a computer readable storage medium, having a computer program stored thereon, wherein the computer program is executed by a processor to implement the time-of-flight magnetic resonance angiography brain vessel segmentation method.

[0021] According to the embodiments provided in the present application, the present application has the following technical effects:

[0022] The present application provides a time-of-flight magnetic resonance angiography brain vessel segmentation method, system, device and medium, which strengthens the blood vessel features by introducing a preprocessing process, and uses the knowledge transfer of a pre-trained video segmenter, so that a small sample segmentation model only needs a small amount of complete labeled time-of-flight magnetic resonance angiography to complete the training and achieve robust segmentation performance, significantly reducing the dependence on large-scale labeled data. An interactive post-processing process based on a human-computer interaction interface and a conditional random field provides an intuitive visual interface and a convenient modification tool, and users can quickly locate and correct segmentation errors to efficiently obtain personalized segmentation results that meet clinical needs. In summary, the present application can achieve high-precision and high-robustness segmentation effect with only a small amount of samples, and the result can be optimized through an efficient human-computer interaction mode, which has important research value and broad clinical application prospect. BRIEF DESCRIPTION OF DRAWINGS

[0023] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the drawings needed in the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present application, and other drawings can be obtained by those skilled in the art without creative labor.

[0024] Figure 1 An application environment diagram of a time-of-flight magnetic resonance angiography brain vessel segmentation method in an embodiment of the present application.

[0025] Figure 2 A whole process schematic diagram of a time-of-flight magnetic resonance angiography brain vessel segmentation method provided in an embodiment of the present application.

[0026] Figure 3 A detailed process schematic diagram of a time-of-flight magnetic resonance angiography brain vessel segmentation method provided in an embodiment of the present application.

[0027] Figure 4A flowchart of an image preprocessing process in an embodiment of the present application.

[0028] Figure 5 A schematic diagram of a network structure of a residual encoder U-Net in an embodiment of the present application.

[0029] Figure 6 A schematic diagram of a fusion process of features extracted by a pre-trained video segmenter and feature maps extracted by a residual encoder U-Net in an embodiment of the present application.

[0030] Figure 7 A schematic diagram of functional modules of a time-of-flight magnetic resonance angiography cerebral vessel segmentation system provided in an embodiment of the present application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only some of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative efforts fall within the scope of protection of the present application.

[0032] The present application aims to solve the problems of poor robustness and low accuracy of existing TOF-MRA cerebral vessel segmentation methods under a few-sample condition, and the problems of complex operation and low efficiency of existing interactive segmentation tools. Specifically, the present application aims to provide a cerebral vessel image segmentation process that can obtain robust segmentation results by training a model with only a small amount of complete labeled TOF-MRA images, and can optimize and verify the results through efficient human-computer interaction.

[0033] The core goal of few-shot learning (FSL) is to enable a model to learn and effectively generalize from a small amount of labeled samples. Common implementation strategies include meta-learning, model fine-tuning, data augmentation, and metric learning.

[0034] Transfer learning (TL) migrates and applies the knowledge learned by a model pre-trained on a large-scale source domain dataset (such as ImageNet in the natural image field or Kinetics in the video field) to a target medical image task.

[0035] To make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the present application will be further described in detail below with reference to the drawings and specific embodiments.

[0036] The time-of-flight magnetic resonance angiography cerebral vessel segmentation method provided in the embodiments of the present application can be applied to, for example, Figure 1The application environment shown. Among them, the terminal 102 communicates with the server 104 through the network. The data storage system can store the data required by the server 104 to process. The data storage system can be set up separately, or integrated on the server 104, or placed on the cloud or other servers. The terminal 102 can send the time-of-flight magnetic resonance vascular image to be processed to the server 104, and the server 104 receives the time-of-flight magnetic resonance vascular image to be processed, and strengthens the vascular features in the time-of-flight magnetic resonance vascular image to be processed to obtain a preprocessed image; according to the preprocessed image, a few sample segmentation model is used for cerebral vascular segmentation to obtain a vascular probability map; based on the human-computer interaction interface and the conditional random field, the vascular probability map and the time-of-flight magnetic resonance vascular image to be processed are post-processed to obtain the final segmentation image. The server 104 can feed back the final segmentation image obtained to the terminal 102. In addition, in some embodiments, the time-of-flight magnetic resonance vascular image cerebral vascular segmentation method can also be implemented by the server 104 or the terminal 102 alone.

[0037] Among them, the terminal 102 can be, but not limited to, various desktop computers, notebook computers, smart phones, tablet computers, Internet of Things devices and portable wearable devices. The server 104 can be implemented by an independent server or a server cluster composed of multiple servers, and can also be a cloud server.

[0038] In an exemplary embodiment, as shown in Figure 2 and Figure 3 , a time-of-flight magnetic resonance vascular image cerebral vascular segmentation method is provided, which is executed by a computer device, specifically can be executed by a terminal or a server alone, or can be executed by a terminal and a server together. In the embodiment of the present application, the method is applied to the server 104 in Figure 1 , which includes the following steps 201 to 204.

[0039] Step 201, obtaining a time-of-flight magnetic resonance vascular image to be processed.

[0040] Step 202, strengthening the vascular features in the time-of-flight magnetic resonance vascular image to be processed to obtain a preprocessed image.

[0041] In a specific application example, the preprocessed image includes a time-of-flight magnetic resonance vascular image after skull removal and a contrast-enhanced time-of-flight magnetic resonance vascular image. As shown in Figure 4 , the present application designs a set of preprocessing procedures including multi-modal image registration, skull removal and histogram standardization to significantly strengthen the vascular features in the image, laying a foundation for subsequent accurate segmentation. Specifically, step 202 includes the following steps 21 to 23.

[0042] Step 21, register the time-of-flight magnetic resonance angiography image to be processed with the T1 weighted image to obtain a registered time-of-flight magnetic resonance angiography image and a registered T1 weighted image. The T1 weighted image and the time-of-flight magnetic resonance angiography image to be processed are images of the same subject.

[0043] Specifically, rigid and non-rigid registration of the TOF-MRA image and the T1 weighted image is completed based on the Advanced Normalization Tools (ANTs), and the spatial correspondence between different modal images is accurately established to ensure that subsequent processing is performed in a unified spatial coordinate system.

[0044] The goal of image registration is to find a spatial transformation such that the floating image (e.g. the TOF-MRA image) is aligned with the reference image (e.g. the T1 weighted image) after transformation. This is usually achieved by optimizing a cost function, which can be expressed as: ; where is the cost function value, is a similarity measure function (e.g. mutual information, normalized cross correlation), represents applying the spatial transformation to the floating image, is a regularization term used to constrain the smoothness or reasonableness of the transformation, is the regularization coefficient.

[0045] For rigid registration, the spatial transformation includes rotation and translation; for non-rigid registration, the spatial transformation is usually a more complex deformation field.

[0046] Step 22, skull stripping the registered T1 weighted image to generate a binary brain tissue mask, and applying the binary brain tissue mask to the time-of-flight magnetic resonance angiography image to be processed to obtain a skull-stripped time-of-flight magnetic resonance angiography image.

[0047] Specifically, the Brain Extraction Tool 2 (BET2) is used to skull strip the registered T1 weighted image (automatically extract the brain tissue region) to generate a binary brain tissue mask. This binary brain tissue mask is applied to the TOF-MRA image in the same space to remove the interference of non-brain tissue such as skull and scalp, and only retain the blood vessel and brain parenchyma information to obtain a TOF-MRA image containing only brain parenchyma and blood vessels.

[0048] Step 23: Histogram normalization is performed on the time-of-flight magnetic resonance angiography images after skull removal to obtain contrast-enhanced time-of-flight magnetic resonance angiography images.

[0049] Specifically, the histogram normalization method implemented in the Torchio medical image processing library is used to globally match and normalize the grayscale distribution of TOF-MRA images after skull removal, aiming to significantly improve the contrast of vascular details. By mapping the grayscale intensity of the images to a predefined standard range or reference histogram, the grayscale distribution between different images is normalized, and the contrast between blood vessels and background tissues is enhanced.

[0050] The goal of histogram normalization is to normalize the input image Histogram of TOF-MRA image after craniotomy Transform into a target histogram Thus, a standardized image is obtained. (Contrast-enhanced time-of-flight magnetic resonance angiography). This can be achieved using the Cumulative Distribution Function (CDF). Let... and The input image and target histogram are respectively displayed at gray levels. The cumulative distribution function value at that location. For each pixel grayscale value in the input image. Its standardized grayscale value You can find it in the following ways: ;in, It is the inverse function of the target cumulative distribution function. Furthermore, the implementation in Torchio may involve more complex techniques such as landmark-based histogram matching.

[0051] Step 203: Based on the preprocessed image, perform cerebral blood vessel segmentation using a few-sample segmentation model to obtain a blood vessel probability map.

[0052] The few-shot segmentation model is trained on a 3D U-Net by transferring knowledge from a pre-trained video segmenter. The few-shot segmentation model is based on the U-Net architecture and incorporates features extracted at different scales by the pre-trained video segmenter.

[0053] In a specific application example, the pre-trained video segmenter is trained using natural scene video streams and features corresponding to different resolutions. Efficient few-shot learning is achieved by embedding the pre-trained knowledge into the natural scene video. The 3D U-Net is a residual encoder U-Net.

[0054] The training process of the few-shot segmentation model is as follows: a small amount of TOF-MRA image data set with complete pixel-level cerebral vessel label is obtained and preprocessed. In order to expand the training set, data enhancement techniques including random rotation, flipping, elastic deformation, intensity transformation and the like are adopted to enhance the TOF-MRA image data set. All input images (whether single modality or multi-modality) are standardized and preprocessed by the Z-Score method before being input into the few-shot segmentation model to obtain the few-shot training sample set. Based on the few-shot training sample set and the pre-trained video segmenter, the parameters of the residual encoder U-Net are optimized.

[0055] During the training process of the few-shot segmentation model, the parameters of the pre-trained video segmenter are frozen, the features extracted by the pre-trained video segmenter are injected into a specific level of the encoder or decoder of the residual encoder U-Net, and only the parameters of the residual encoder U-Net and possibly the parameters of the feature adaptation layer are trained. This strategy helps to efficiently transfer the general visual features learned by the pre-trained video segmenter to the medical image segmentation task, especially in the case of a small amount of training samples, to ensure efficient use of pre-trained features and guide the network to quickly train, reduce the interference of different network parameter initialization on pre-trained features, and improve the robustness of the model.

[0056] In the selection of the loss function, for example, the combination of the Dice loss and the cross-entropy loss can be used as the total loss function. The total loss function L can be the weighted sum of the two: ; wherein, is the Dice loss, is the cross-entropy loss, is the weight coefficient of the Dice loss, is the weight coefficient of the cross-entropy loss.

[0057] The Dice loss is used to measure the overlap between the predicted segmentation result and the true label . For a binary classification problem, it is defined as: .

[0058] Or the more commonly used form: ; wherein, is the probability of the pixel predicted by the few-shot segmentation model as foreground (vessel), is the true label (0 or 1, 1 represents a vessel and 0 represents a non-vessel) of the pixel , and is the total number of pixels, is a small smoothing constant to prevent the denominator from being zero.

[0059] ​For binary classification problems (e.g., each pixel is classified as vascular or non-vascular), the cross-entropy loss takes the form of a binary cross-entropy loss: .

[0060] The optimizer can be an Adaptive Moment Estimation (Adam) optimizer or an Adaptive Moment Estimation Weight (AdamW) optimizer that introduces weight decay. The Adam optimizer combines the advantages of the Adaptive Gradient (AdaGrad) and Root Mean Square Propagation (RMSProp) algorithms.

[0061] Its parameter update rule is: at the first... In the next iteration, for the parameters Perform the following steps (1) to (6).

[0062] (1) Calculate the gradient: .

[0063] (2) Update the first moment estimate (momentum): .

[0064] (3) Update the second moment estimate (velocity): .

[0065] (4) Calculate the first moment estimate after bias correction: .

[0066] (5) Calculate the second-order moment estimate after bias correction: .

[0067] (6) Update parameters: .

[0068] in, It is the first The loss function value at the next iteration It is the first The parameters at the next iteration It is the first The parameters at the next iteration It is the first The gradient at the next iteration It is the first First-order moment estimation at the next iteration It is the first First-order moment estimation at the next iteration It is the first Second-order moment estimation at the next iteration It is the first the second moment estimation at the i-th iteration, and is the decay rate (usually close to 1, e.g. 0.9 and 0.999), is the i-th iteration, is the first moment estimation after bias correction at the i-th iteration, is the i-th iteration, is the second moment estimation after bias correction at the i-th iteration, is the learning rate.

[0069] Regarding the training parameters, for example, the input 3D data block size is [112, 256, 256] voxels, and the batch size can be set to 3. The learning rate, training rounds and other parameters are adjusted according to the specific hardware resources.

[0070] In this application, step 203 includes the following steps 31 and 32.

[0071] Step 31, using a pre-trained video tokenizer to encode the features of the pre-processed image, to obtain a plurality of features of different resolutions.

[0072] The pre-processed image (for example, the time-of-flight magnetic resonance angiography image after skull removal and the contrast-enhanced time-of-flight magnetic resonance angiography image, which can be input as a multi-channel or processed separately and fused) is regarded as a video stream, and a pre-trained video tokenizer (Vidtok) is used, for example, vidtok_kl_causal_ 488 _ 16 chn_v 1_1and vidtok_kl_causal_ 41616_16 chn_v 1_1, the input pre-processed image is encoded into a hidden space at different compression ratios (for example, 4x8x8 and 4x16x16), forming features of different resolutions. Specifically, the pre-trained video tokenizer includes at least two encoders of different compression ratios, which respectively encode the input 3D TOF-MRA image into a hidden space, and fuse with the feature maps of different depths of the residual encoder U-Net.

[0073] For example, vidtok_kl_causal_ 488 _ 16 chn_v 1_1, the pre-processed image is encoded into a hidden space at a compression ratio of 4x8x8, and the output feature is spliced and fused with the feature map of the second-to-last layer (or the corresponding deeper layer of the encoder) of the decoder of the residual encoder U-Net. vidtok_kl_causal_ 41616_16 chn_v1_1 The preprocessed image is compressed and encoded into the latent space at a ratio of 4×16×16. Its output features are stitched and fused with the feature map of the penultimate layer of the residual encoder U-Net decoder (or the deeper layer corresponding to the encoder, i.e. the first level of upsampling after the bottleneck layer).

[0074] Step 32: Use the pre-trained 3D U-Net to segment the cerebral blood vessels in the preprocessed image, and fuse features of multiple different resolutions into the feature maps of different downsampling stages of 3D U-Net (e.g., the penultimate and penultimate layers of the network) to obtain a blood vessel probability map.

[0075] like Figure 5 As shown, the U-Net residual encoder architecture comprises seven processing stages, with inputs such as 3D TOF-MRA image blocks of size [112, 256, 256]. The number of feature maps (channels) in each stage are 32, 64, 128, 256, 256, 256, and 256 (bottleneck layer), respectively. The number of convolutional blocks varies with depth: 1 in the first stage, 3 in the second stage, 4 in the third stage, and 6 in each of the fourth to seventh stages. All convolutional operations are based on 3×3×3 three-dimensional convolutional kernels (torch.nn.modules.conv.Conv3d) with a bias term enabled (conv_bias:true). Regarding the downsampling strategy, isotropic downsampling is achieved through convolutions with a stride of [2,2,2] during the transition from stage 1 to stage 5; anisotropic downsampling is achieved through convolutions with a stride of [1,2,2] during the transition from stage 5 to stage 6 and from stage 6 to stage 7. This means maintaining resolution along the Z-axis while performing a 2x downsampling on the X and Y axes. The initial convolution (i.e., the first convolutional block in stage 1) has a stride of [1,1,1]. Residual connections are introduced between feature extraction modules in the encoder path (e.g., within or between each convolutional block) to facilitate training of deep networks and improve gradient flow.

[0076] The decoder structure of the residual encoder U-Net consists of six upsampling stages, which are symmetrical to the encoder structure. Each decoder stage upsamples the feature map by transposing convolution or interpolation combined with convolution to restore the resolution of the feature map, and then concatenates it with the feature map passed from the corresponding level of the encoder through skip connections. The concatenated feature map is then processed by convolutional blocks.

[0077] The general layer component in the residual encoder U-Net follows the following design: after each convolution operation, a three-dimensional instance normalization (InstanceNorm3d, eps:1e-05, affine:true) is connected, followed by a LeakyReLU activation function (torch.nn.LeakyReLU, inplace:true), and no Dropout layer is configured.

[0078] As shown in Figure 6 , to is the feature map obtained by each processing stage of the residual encoder U-Net encoder, to is the feature map obtained by each processing stage of the residual encoder U-Net decoder, S is the feature map obtained by the encoder.

[0079] Step 204, based on the human-computer interaction interface and the conditional random field, the blood vessel probability map and the time-of-flight magnetic resonance blood vessel image to be processed are post-processed to obtain a final segmentation image.

[0080] In one specific application example, the human-computer interaction interface displays the current segmentation result by using a 3D Slicer plug-in and receives the user's correction instruction. The blood vessel probability map is used as the initial input, and through the interactive interface integrated in the three-dimensional visualization platform (such as 3D Slicer), the user is allowed to visually check and interactively correct the segmentation result, and the conditional random field is used to optimize and generate a globally optimal segmentation mask according to the user's correction. Step 204 includes steps 41 to step 45.

[0081] Step 41, generating a three-dimensional reconstruction model of cerebral blood vessels based on the blood vessel probability map.

[0082] Step 42, generating a two-dimensional view based on the time-of-flight magnetic resonance blood vessel image to be processed. Specifically, it includes the two-dimensional view of the time-of-flight magnetic resonance blood vessel image to be processed on the sagittal plane, coronal plane and horizontal plane.

[0083] Specifically, the TOF-MRA image to be processed and the blood vessel probability map are loaded in the 3D Slicer environment, and the blood vessel probability map is used as the initial input of the unary potential of the conditional random field.

[0084] Step 43, for any round of interaction, according to the current three-dimensional reconstruction model of cerebral blood vessels and the current two-dimensional view, the human-computer interaction interface is used to display the current segmentation result and receive the user's correction instruction.

[0085] Specifically, the current segmentation result is displayed on the current 3D reconstruction model and the current 2D view (for example, in the form of a color contour or a semi-transparent mask). The user can check the segmentation result from different angles and levels by zooming, panning, rotating, etc., and quickly locate the over-segmentation (false positive) or under-segmentation (false negative) area.

[0086] Step 44, based on the correction instruction, the current cerebral vascular 3D reconstruction model and / or the current 2D view are corrected to obtain a modified blood vessel probability map.

[0087] The user can use the interactive tools provided by 3D Slicer (for example, a brush, an eraser, a region selection, etc., which can be reused or customized by the present application) to make corrections on the 2D view or the cerebral vascular 3D reconstruction model: for the under-segmentation area (blood vessels are not detected), the user can mark the area as "blood vessels", and the plug-in will increase the probability value of the corresponding voxel (for example, set a higher probability value close to 1); for the over-segmentation area (non-blood vessels are misdetected as blood vessels), the user can mark the area as "background" or "non-blood vessels", and the plug-in will reduce the probability value of the corresponding voxel (for example, set a lower probability value close to 0). The interactive points or regions provided by the user will be used as hard constraints or strong priors in the conditional random field, directly modifying the unary potential.

[0088] Step 45, the blood vessel probability map is used as the initial input of the unary potential of the conditional random field, the user's correction instruction is used as the hard constraint or strong prior of the conditional random field, the global optimal segmentation mask is calculated according to the modified blood vessel probability map by using the conditional random field, and the global optimal segmentation mask is fed back to the cerebral vascular 3D reconstruction model and the 2D view for the next round of interaction. After multiple rounds of interaction, the final segmentation image is obtained.

[0089] When the user completes a round of modification, the optimization process of the conditional random field is triggered. The conditional random field considers the blood vessel probability map, the user-provided correction instruction, and the spatial continuity between pixels by minimizing an energy function containing unary potential and binary potential, calculates a global optimal segmentation mask, and updates and feeds back the global optimal segmentation mask to each view of 3D Slicer in real time for the user to evaluate.

[0090] Specifically, the goal of the conditional random field is to find a label configuration that minimizes the Gibbs energy function : where is the label of pixel , is the unary potential, indicating that pixel takes label The cost is usually derived from the predicted probability of the few-sample segmentation model, and the correction of user interaction can directly affect the unary potential. It is a binary potential, representing adjacent pixels. and pixels ( , It is a pixel Take the labels from the neighborhood of each and The time cost is often used to encourage smoothness in segmentation results, for example, by using the Potts model or models based on image features (such as intensity, gradient): ;in, It is an indicator function that returns 1 when the condition is true and 0 otherwise. It is a weight, which can depend on the pixel. and pixels The similarity between them.

[0091] Users can perform multiple rounds of interactive correction and optimization based on the feedback results of the conditional random field until a satisfactory final segmentation mask is obtained. The final result can be saved in standard medical image formats, such as Neuroimaging Informatics Technology Initiative (NIfTI) and Nearly Raw Raster Data (NRRD).

[0092] Through the above steps, this application realizes a complete TOF-MRA cerebral blood vessel segmentation process, from image preprocessing and automatic segmentation based on few-shot learning to efficient human-computer interaction post-processing.

[0093] Compared with the prior art, this application has the following beneficial effects.

[0094] (1) Highly efficient few-shot learning capability: By introducing a preprocessing process to enhance vascular features and utilizing the knowledge transfer of a pre-trained video segmenter, the few-shot segmentation model can be trained with only a very small number of fully annotated TOF-MRA images and achieve robust segmentation performance, significantly reducing the dependence on large-scale annotated data.

[0095] (2) Improved segmentation accuracy and robustness: The preprocessing process and deep learning network architecture, especially the effective integration of residual connections and pre-trained features, help the few-shot segmentation model capture more subtle vascular structures, improving the accuracy of segmentation and robustness to different data quality.

[0096] (3) Convenient and efficient human-computer interaction: the interactive post-processing process based on the 3D Slicer plug-in provides an intuitive visual interface and convenient modification tools, and users can quickly locate and correct segmentation errors, combined with conditional random field optimization, to efficiently obtain personalized segmentation results that meet clinical needs.

[0097] (4) Ease of use and integration: integrating the entire process (preprocessing, segmentation, post-processing), especially the post-processing part as a 3D Slicer plug-in, provides an out-of-the-box solution for clinicians and researchers to use.

[0098] Based on the same inventive concept, the embodiments of the present application also provide a time-of-flight magnetic resonance angiography brain vessel segmentation system for implementing the time-of-flight magnetic resonance angiography brain vessel segmentation method described above. The implementation scheme for solving the problem provided by the system is similar to the implementation scheme described in the above method, so the specific limitations in one or more time-of-flight magnetic resonance angiography brain vessel segmentation system embodiments provided below can be referred to the limitations of the time-of-flight magnetic resonance angiography brain vessel segmentation method described above, which will not be repeated here.

[0099] In one exemplary embodiment, as shown in Figure 7 , a time-of-flight magnetic resonance angiography brain vessel segmentation system is provided, which includes an image acquisition module 701, a preprocessing module 702, a preliminary segmentation module 703, and a post-processing module 704.

[0100] The image acquisition module 701 is configured to acquire a time-of-flight magnetic resonance angiography image to be processed.

[0101] The preprocessing module 702 is configured to enhance the blood vessel features in the time-of-flight magnetic resonance angiography image to be processed to obtain a preprocessed image.

[0102] The preliminary segmentation module 703 is configured to perform brain vessel segmentation on the preprocessed image using a few-shot segmentation model to obtain a blood vessel probability map. The few-shot segmentation model is obtained by training a 3D U-Net by transferring the knowledge of a pre-trained video segmenter.

[0103] The post-processing module 704 is configured to perform post-processing on the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed based on a human-computer interaction interface and a conditional random field to obtain a final segmentation image.

[0104] In one exemplary embodiment, a computer device is provided, which includes a memory and a processor, and the memory stores a computer program. When the processor executes the computer program, the steps in the above method embodiments are implemented.

[0105] In an exemplary embodiment, a computer readable storage medium storing a computer program is provided, the computer program, when executed by a processor, implements the steps in the above method embodiments.

[0106] In an exemplary embodiment, a computer program product is provided, comprising a computer program, the computer program, when executed by a processor, implements the steps in the above method embodiments.

[0107] It should be noted that the user information (including but not limited to user equipment information, user personal information, etc.) and data (including but not limited to data for analysis, stored data, displayed data, etc.) involved in the present application are all information and data authorized by the user or authorized by all parties, and the collection, use and processing of related data need to comply with relevant regulations.

[0108] In the present application, all actions of obtaining signals, information or data are carried out in compliance with the corresponding data protection regulations and policies of the country where the device is located, and with the authorization given by the owner of the corresponding device.

[0109] Those of ordinary skill in the art can understand that all or part of the processes in the above method embodiments can be completed by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer readable storage medium. When the computer program is executed, it can include the processes of the above method embodiments. In the embodiments provided by the present application, any reference to a memory, database or other medium can include at least one of a non-volatile and volatile memory. The non-volatile memory can include a read-only memory (ROM), a magnetic tape, a floppy disk, a flash memory, an optical storage, a high-density embedded non-volatile memory, a resistive memory (ReRAM), a magnetoresistive random access memory (MRAM), a ferroelectric memory (FRAM), a phase change memory (PCM), a graphene memory, etc. The volatile memory can include a random access memory (RAM) or an external cache memory, etc. As an illustration but not limitation, the RAM can be in various forms, such as a static random access memory (SRAM) or a dynamic random access memory (DRAM), etc.

[0110] The database involved in each of the embodiments provided in the present application can include at least one of a relational database and a non-relational database. The non-relational database can include a distributed database based on a blockchain, and the like, without being limited thereto. The processor involved in each of the embodiments provided in the present application can be a general-purpose processor, a central processing unit, a graphics processing unit, a digital signal processor, a programmable logic device, a data processing logic device based on quantum computing, and the like, without being limited thereto.

[0111] The technical features of the above embodiments can be combined in any manner. To make the description concise, all possible combinations of the technical features in the above embodiments are not described, but it should be considered that any combination of the technical features is within the scope of the present disclosure, as long as there is no contradiction.

[0112] The principles and implementation manners of the present application are described by using specific examples herein, and the above embodiments are only used to help understand the method of the present application and its core idea. Meanwhile, for those skilled in the art, the specific implementation manners and application ranges can be changed according to the idea of the present application. In summary, the content of the present description should not be understood as a limitation of the present application.

Claims

1. A method for segmenting cerebral blood vessels using time-of-flight magnetic resonance angiography, characterized in that, The method comprises: acquiring a time-of-flight magnetic resonance angiography image to be processed; after preprocessing the image, a small sample segmentation model is used for brain vessel segmentation to obtain a blood vessel probability map, specifically comprising: using a pre-trained video segmenter to encode the features of the preprocessed image to obtain multiple features at different resolutions; using a pre-trained 3D U-Net to segment the brain vessels of the preprocessed image, and simultaneously fusing the multiple features at different resolutions into the feature maps at different down-sampling stages of the 3D U-Net to obtain the blood vessel probability map; the small sample segmentation model is obtained by transferring the knowledge of the pre-trained video segmenter to train the 3D U-Net; the pre-trained video segmenter is pre-trained using natural scene video streams and corresponding features at different resolutions; before segmentation, the features of the blood vessels in the time-of-flight magnetic resonance angiography image to be processed are strengthened to obtain a preprocessed image, specifically comprising: registering the time-of-flight magnetic resonance angiography image to be processed with a T1 weighted image to obtain a registered time-of-flight magnetic resonance angiography image and a registered T1 weighted image; wherein the rigid and non-rigid registration of the TOF-MRA image and the T1 weighted image is completed based on the advanced normalization tool, and the T1 weighted image and the time-of-flight magnetic resonance angiography image to be processed are images of the same subject; the skull is removed from the registered T1 weighted image and the time-of-flight magnetic resonance angiography image using the brain extraction tool, a binary brain tissue mask retaining brain vessels is generated by merging, and the binary brain tissue mask is applied to the time-of-flight magnetic resonance angiography image to be processed to obtain a skull-removed time-of-flight magnetic resonance angiography image; the skull-removed time-of-flight magnetic resonance angiography image is histogram normalized using the histogram normalization method implemented in the Torchio medical image processing library to obtain a contrast-enhanced time-of-flight magnetic resonance angiography image; the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed are post-processed based on a human-computer interaction interface and a conditional random field to obtain a final segmentation image, specifically comprising: generating a brain vessel three-dimensional reconstruction model based on the blood vessel probability map; generating two-dimensional views on the sagittal plane, coronal plane and horizontal plane based on the time-of-flight magnetic resonance angiography image to be processed; for any round of interaction, according to the current brain vessel three-dimensional reconstruction model and the current two-dimensional view, the current segmentation result is displayed using the human-computer interaction interface and the user's correction instruction is received; based on the correction instruction, the current brain vessel three-dimensional reconstruction model and / or the current two-dimensional view are corrected to obtain a modified blood vessel probability map; the blood vessel probability map is used as the initial input of the unary potential of the conditional random field, the user's correction instruction is used as the hard constraint or strong prior of the conditional random field, the globally optimal segmentation mask is calculated using the conditional random field according to the modified blood vessel probability map, and the globally optimal segmentation mask is fed back to the brain vessel three-dimensional reconstruction model and the two-dimensional view for the next round of interaction; after multiple rounds of interaction, a final segmentation image is obtained.

2. The method of claim 1, wherein the method further comprises: The 3D U-Net is a residual encoder U-Net.

3. The method of claim 1, wherein the method further comprises: The human-computer interaction interface displays the current segmentation result by using a 3D Slicer plug-in and receives the correction instruction of the user.

4. A time-of-flight magnetic resonance angiography cerebral vessel segmentation system, applied to the time-of-flight magnetic resonance angiography cerebral vessel segmentation method of any one of claims 1-3, characterized in that, The system comprises: An image acquisition module is configured to acquire a time-of-flight magnetic resonance angiography image to be processed. A preprocessing module is configured to strengthen a blood vessel feature in the time-of-flight magnetic resonance angiography image to be processed to obtain a preprocessed image. A preliminary segmentation module is configured to perform brain blood vessel segmentation on the preprocessed image by using a few-shot segmentation model to obtain a blood vessel probability map, wherein the few-shot segmentation model is obtained by training a 3D U-Net by transferring knowledge of a pre-trained video segmenter. A post-processing module is configured to perform post-processing on the blood vessel probability map and the time-of-flight magnetic resonance angiography image to be processed based on a human-computer interaction interface and a conditional random field to obtain a final segmentation image.

5. A computer device comprising: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the time-of-flight magnetic resonance angiography brain blood vessel segmentation method of any one of claims 1-3.

6. A computer-readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the time-of-flight magnetic resonance angiography brain blood vessel segmentation method of any one of claims 1-3.

Images