A magnetic resonance guided target region boundary extraction system and method

Abstract

This invention relates to the field of image processing and discloses a magnetic resonance-guided target region boundary extraction system and method, solving the problem that existing magnetic resonance-guided radiotherapy target region boundary extraction systems struggle to balance high-resolution dynamic image processing speed with target region delineation accuracy. This invention significantly reduces the computational data volume of the first boundary extraction model by compressing the original high-resolution 3D image into low-resolution compressed video information through a compressed video information acquisition module, thereby significantly improving the extraction speed of coarse-boundary target regions. Through multi-scale directional tensor analysis in the first boundary extraction model, this invention can identify the differences between the tubular structures of blood vessels and the structures of organs and bones, achieving adaptive gradient enhancement. Furthermore, this invention achieves an optimal balance between target region boundary extraction speed and accuracy by rapidly acquiring coarse boundaries from compressed video and performing fine-tuning in local high-resolution regions through a second boundary extraction model.

Description

Technical Field

[0001] This invention relates to the field of image processing, and in particular to a magnetic resonance-guided target region boundary extraction system. Background Technology

[0002] In magnetic resonance-guided affine therapy, magnetic resonance imaging (MRI) provides excellent soft tissue contrast, enabling real-time visualization of the anatomical structures of the treatment target area and surrounding organs at risk. To achieve precise radiotherapy dose delivery, the anatomical boundaries of the target organ must be accurately segmented from a sequence of continuously acquired three-dimensional MRI images.

[0003] Currently, deep learning-based convolutional neural network models have become the mainstream method for medical image segmentation. These models are trained under supervision using massive amounts of labeled medical image data, learning the end-to-end mapping relationship from voxels in the original image to organ boundary masks. After training convergence, the model can simultaneously segment multiple anatomical structures (such as liver, spinal cord, blood vessels, bones, etc.) in the input image and output the corresponding 3D masks, thereby achieving automated labeling of organ boundaries.

[0004] Magnetic resonance-guided radiotherapy systems typically acquire three-dimensional image sequences at high frame rates and high spatial resolution, with the amount of image data generated in a single treatment session reaching gigabytes. To ensure segmentation accuracy in complex anatomical contexts, existing convolutional neural network models often employ deep network structures and highly complex computational modules (such as dense attention mechanisms and multi-scale feature fusion). However, when processing high-resolution dynamic sequences, the computation time for a single frame of an existing convolutional neural network model increases significantly, making it difficult to match the boundary extraction speed with the real-time image acquisition rate.

[0005] If the resolution of the input image is reduced directly to increase speed, key anatomical details such as the edges of small blood vessels and the transition areas of organ boundaries will be lost, resulting in a decrease in the accuracy of target area delineation. If only local high-resolution areas are finely segmented, without the guidance of global spatial context, it is difficult to accurately determine the relative positional relationship between the target area and surrounding tissues. Summary of the Invention

[0006] The purpose of this invention is to solve the technical problem that existing magnetic resonance-guided radiotherapy target area boundary extraction systems cannot simultaneously achieve high-resolution dynamic image processing speed and target area delineation accuracy, and to provide a magnetic resonance-guided target area boundary extraction system and method.

[0007] To achieve the above-mentioned objectives, the embodiments of the present invention provide the following technical solutions:

[0008] A magnetic resonance-guided target region boundary extraction system includes a compressed video information acquisition module, an inverse feature extraction module, a first boundary extraction model, a supplementary information module, a global information integration module, and a second boundary extraction model.

[0009] The module for acquiring compressed video information acquires three-dimensional images through a magnetic resonance imaging device, integrates the three-dimensional images to obtain the original video information, and compresses the original video information to obtain compressed video information.

[0010] The reverse feature extraction module maps compressed video information and original video information to obtain the mapping relationship between compressed video information and original video information.

[0011] The first boundary extraction model processes compressed video information, combines a threshold to determine the edge points of the three-dimensional image in the compressed video information, and connects the edge points to generate the boundary of the coarse boundary target area.

[0012] The supplementary information module extracts the three-dimensional image region corresponding to the original video information of the coarse boundary target area by the mapping relationship between compressed video information and original video information, and uses the three-dimensional image region as supplementary information.

[0013] The global information integration module overlays the boundary of the coarse-boundary target area onto the three-dimensional image of the compressed video information in the form of annotations to form global information.

[0014] The second boundary extraction model uses global information to extract a global feature map, extracts a local feature map by supplementing information, obtains alignment features using the local feature map, fuses the alignment features and the global feature map to obtain a fused feature, and decodes the fused feature to obtain the boundary of the target region.

[0015] This invention addresses the technical problem in existing magnetic resonance-guided radiotherapy (MRI) target boundary extraction methods that directly reduce input image resolution to improve processing speed, leading to the loss of crucial anatomical details such as the edges of small blood vessels and the transition areas between organ boundaries, and a decrease in target delineation accuracy. The invention employs a compressed video information module to compress the original high-resolution 3D image into low-resolution compressed video information, significantly reducing the processing data volume of the first boundary extraction model and dramatically improving the extraction speed of coarse-boundary target areas. Simultaneously, a supplementary information module utilizes the mapping relationship between the compressed video information and the original video information to extract local 3D image regions corresponding to the coarse-boundary target areas from the original high-resolution video as supplementary information. This fully preserves the high-precision anatomical details of the edges of small blood vessels and the transition areas between organ boundaries, improving processing speed while avoiding the loss of crucial detail information.

[0016] This invention addresses the technical problem of accurately determining the relative positional relationship between the target area and surrounding tissues when performing fine segmentation only on local high-resolution regions due to the lack of global spatial context guidance. The invention uses a global information integration module to overlay the coarse target area boundary generated by the first boundary extraction model onto the 3D image of the compressed video information in labeled form, thus constructing global information carrying the approximate location and extent of the target area. The second boundary extraction model uses the global feature map extracted from this global information as spatial guidance, combining it with the local high-resolution feature map extracted from supplementary information for feature alignment and gating fusion. During the alignment process, affine transformation is used to compensate for overall displacement and deformation offset, modeling local elastic deformation, ultimately generating precise boundaries for the target area. This provides a reliable global spatial context reference for fine local segmentation, ensuring accurate determination of the relative positional relationship between the target area and surrounding tissues.

[0017] Compared with existing technologies, the beneficial effects of this invention are as follows: This invention compresses the original high-resolution 3D image into low-resolution compressed video information by acquiring compressed video information module, which greatly reduces the amount of computational data of the first boundary extraction model and significantly improves the extraction speed of coarse boundary target areas; This invention can identify the differences between the tubular structure of blood vessels and the structure of organs and bones through multi-scale directional tensor analysis in the first boundary extraction model, and realize adaptive gradient enhancement; This invention achieves the optimal balance between the speed and accuracy of target area boundary extraction by quickly acquiring coarse boundaries in compressed video and performing fine correction in local high-resolution areas through the second boundary extraction model.

[0018] Furthermore, a magnetic resonance-guided target region boundary extraction system, wherein the compressed video information acquisition module includes an original video information acquisition submodule and an original video information compression submodule;

[0019] The original video information acquisition submodule acquires several three-dimensional images at high frame rate and high spatial resolution using a magnetic resonance imaging device, sorts the several three-dimensional images according to the acquisition time of each frame, and integrates the sorted three-dimensional images to obtain the original video information.

[0020] The original video information compression submodule compresses each frame of the original video information using the mean value to obtain compressed video information.

[0021] In the aforementioned solutions, addressing the technical challenges of massive amounts of high-resolution dynamic 3D sequence data and significant computational overhead in direct processing, this invention utilizes magnetic resonance imaging (MRI) equipment to dynamically acquire 3D image sequences of the treatment target area and surrounding organs at risk at high frame rates (1-10 frames per second) and high spatial resolution (voxel size ≤ 2 mm). These images are then integrated after being sorted according to the acquisition timestamp of each frame to obtain complete original video information recording the dynamic changes in anatomical structures. This invention employs a sub-module for compressing the original video information based on mean compression. The compression window is sized along the x, y, and z axes, dividing each frame of the original video information into several non-overlapping 3D blocks. The arithmetic mean of all voxel values ​​within each block is used as the voxel value of the corresponding voxel after compression. While preserving the overall contour features of key anatomical structures such as organ boundaries and blood vessel edges, the data volume of a single frame is compressed to one-thousandth of the original data volume, significantly reducing the computational load of the subsequent first boundary extraction model. This invention provides compressed video information with low data volume and high signal-to-noise ratio for the first boundary extraction model, ensuring that the extraction speed of coarse boundary target area can match the real-time acquisition rate of high frame rate dynamic images, while avoiding the problem of loss of key boundary details caused by direct downsampling.

[0022] Furthermore, a magnetic resonance-guided target region boundary extraction system is provided, wherein the first boundary extraction model includes a Gaussian filtering module, an edge intensity calculation module, a maximum value screening module, and a threshold connection module;

[0023] The Gaussian filtering module performs Gaussian smoothing on the three-dimensional image in the compressed video information to obtain smoothed data.

[0024] The edge intensity calculation module obtains the gradient vector at each scale by performing three-dimensional convolution on smoothed data and a multi-scale Gaussian kernel at each voxel location.

[0025] The maximum value filtering module obtains the anisotropy index of the orientation tensor matrix through eigenvalues, selects the total amplitude after enhancement at each voxel position using the anisotropy index, judges edge points based on the total amplitude after enhancement at each voxel position, and obtains the gradient direction through eigenvalues.

[0026] The threshold connection module determines the edge points based on the enhanced amplitude and uses the gradient direction and edge points to generate the boundary of the coarse boundary target area.

[0027] Furthermore, in a magnetic resonance-guided target region boundary extraction system, the total amplitude after enhancement at each voxel location is selected using anisotropy index, with the following formula:

[0028] ;

[0029] in, The enhanced total amplitude, For vascular enhancement amplitude, To enhance the amplitude at the edges of organs and bones, If, It is an exponential function with base e. It is a vascular enhancing factor. As a noise suppression factor, The anisotropy index, Let p be the p-th eigenvalue, where p = 1, 2, 3.

[0030] In the above scheme, addressing the technical problem that existing boundary extraction systems, using simple gradient operators, cannot distinguish between the tubular structures of blood vessels and the blocky / planar structures of organs and bones, leading to low accuracy in extreme point selection, this invention performs three-dimensional Gaussian smoothing on each frame of the three-dimensional image in the compressed video information through a Gaussian filtering module. This effectively suppresses blocky artifacts and high-frequency noise that may be introduced during compression, providing smooth data with a high signal-to-noise ratio for subsequent gradient calculations. This invention generates multi-scale Gaussian derivative kernels through an edge strength calculation module, performs three-dimensional convolution with the smoothed data to obtain gradient vectors at each scale, and fuses these multi-scale gradient vectors to construct a direction tensor matrix. This direction tensor matrix fully characterizes the structural strength distribution features in each direction within the neighborhood of each voxel. This invention performs eigenvalue decomposition on the direction tensor matrix through a maximum value selection module, obtaining eigenvalues ​​and corresponding eigenvectors arranged in descending order, and calculates the anisotropy index A based on the differences in eigenvalues. For tubular structures like blood vessels, the largest eigenvalue is significantly larger than other eigenvalues, triggering a vessel enhancement mechanism that exponentially amplifies the gradient amplitude along the principal feature vector, effectively compensating for gradient attenuation caused by the circular cross-section of small blood vessels. For planar structures like organs and bones, the edge response of the difference between the largest and second-largest eigenvalues ​​is enhanced, while noise suppression factors smooth organ boundaries. This invention uses a dual-threshold connection mechanism to trace and connect strong and weak edge points along the gradient direction, generating a continuous and complete coarse-boundary target region boundary. This invention provides reliable guidance for the approximate location and range of the target region in the subsequent second boundary extraction model. While rapidly completing the initial boundary localization in the low-resolution space of compressed video, it effectively distinguishes the differences between blood vessels, organs, bones, and background noise, significantly improving the accuracy and robustness of coarse-boundary target region extraction.

[0031] Furthermore, a magnetic resonance-guided target region boundary extraction system is provided, wherein the second boundary extraction model includes a global path processing module, a supplementary path processing module, a feature alignment module, and a feature fusion module;

[0032] The global path processing module obtains global network features by processing global information through convolution, and obtains a global feature map by processing global network features through spatial downsampling and residual learning.

[0033] The supplementary path processing module uses dilated convolution to process supplementary information to obtain supplementary convolutional features, and uses a visual transformer to enhance the supplementary convolutional features to obtain local feature maps.

[0034] The feature alignment module downsamples the local feature map to obtain a downsampled feature map. Using the downsampled feature map and the boundary of the coarse boundary target area, it obtains the normalized coordinates of the coarse boundary target area. It then concatenates the normalized coordinates of the coarse boundary target area with the global feature map. After concatenation, it uses a multilayer perceptron and a reshaping operation to obtain an affine matrix. It then uses the affine matrix to obtain affine coordinates. Finally, it combines the deformation offset and the affine coordinates to obtain the sampling coordinates. A three-dimensional deformation convolution is performed with the sampling coordinates as the center to obtain the alignment features.

[0035] The feature fusion module concatenates the aligned features and the global feature map to obtain the concatenated features, and then fuses the concatenated features with the gating weights to obtain the fused features. Finally, the fused features are decoded to obtain the boundary of the target region.

[0036] In the aforementioned scheme, addressing the technical problems of lacking global spatial reference and difficulty in accurately determining the relative position of the target area and surrounding tissues in local fine segmentation, this invention extracts a global feature map carrying coarse boundary target area position information through a global path processing module, providing reliable spatial guidance for local fine segmentation; and extracts local high-resolution features from supplementary information through a supplementary path processing module, fully preserving key anatomical details such as the edges of small blood vessels and the transition zone between organ boundaries. This invention utilizes an affine transformation module to compensate for overall displacement and deformation offset, modeling local elastic deformation, and precisely calibrating the global and local feature maps in anatomical space to generate spatially consistent aligned features; and performs gated fusion of the aligned features and the global feature map through a feature fusion module, adaptively enhancing important features and suppressing irrelevant features, ultimately decoding to obtain the accurate target region boundary. This invention, through the dual-path collaborative processing of global guidance and local refinement, significantly reduces computation time while ensuring high-resolution boundary extraction accuracy, achieving an optimal balance between target region boundary extraction speed and accuracy.

[0037] Furthermore, a magnetic resonance-guided target region boundary extraction system, wherein the global path processing module includes a global convolution submodule and a residual extraction submodule;

[0038] The global convolution submodule performs convolution operations on global information, and batch normalizes and activates the global information after convolution to obtain global network features.

[0039] The residual extraction submodule performs spatial downsampling on the global network features. While performing spatial downsampling, it retains key information through residual blocks to obtain a global feature map.

[0040] In the aforementioned scheme, to address the technical problem of easily losing key spatial information of the target region edge during spatial downsampling, this invention performs convolution, batch normalization, and activation operations on global information through a global convolution submodule to obtain global network features representing the approximate location and range of the coarse-boundary target region. This invention, through a residual extraction submodule, retains important feature information from before sampling by utilizing the shortened paths of residual blocks during spatial downsampling, ensuring that the global feature map after downsampling still completely carries the spatial location guidance information of the coarse-boundary target region. The global feature map of this invention provides a stable and reliable global spatial context reference for subsequent feature alignment modules.

[0041] Furthermore, a magnetic resonance-guided target region boundary extraction system includes a supplementary path processing module comprising a dilated convolution submodule and a ViT enhancement submodule.

[0042] The dilated convolution submodule performs dilated convolution on the supplementary information to obtain supplementary convolution features;

[0043] The ViT enhancement submodule divides the supplementary convolutional features into three-dimensional slices and flattens them to obtain a slice sequence. The slice sequence is projected into a D-dimensional latent space to obtain an initial learnable positional encoding. The initial learnable positional encoding is processed by a Transformer encoding layer to obtain an attention feature sequence. The attention feature sequence is then projected back into the original spatial dimension for convolution operation to obtain a local feature map.

[0044] In the above scheme, this invention addresses the technical problem of limited receptive field and difficulty in modeling long-distance dependencies between anatomical landmarks that are far apart, as addressed by a dilated convolution submodule. This invention performs 3D dilated convolution on the supplementary information, expanding the receptive field while maintaining the spatial resolution of the feature map. This effectively captures the local contextual features of key anatomical structures such as small blood vessels and organ boundary transition zones in the supplementary information, obtaining supplementary convolutional features. The invention further divides the supplementary convolutional features into several non-overlapping 3D slices along the spatial dimension using a ViT enhancement submodule. Each slice is flattened into a one-dimensional vector and projected into a D-dimensional latent space, then added to a learnable positional encoding matrix to obtain an initial encoding sequence. This positional encoding mechanism fully preserves the absolute positional information of each slice in the original 3D space, avoiding spatial structure distortion caused by slice order confusion. This invention utilizes multiple Transformer coding layers to perform multi-head self-attention extraction layer by layer on the initial coding sequence. Through the self-attention mechanism, it explicitly models the global dependencies between different slices, effectively capturing the spatial association features between anatomical structures that are far apart, obtaining an attention feature sequence. This attention feature sequence is then mapped back to the original spatial dimension via linear projection, and reconstructed into a 3D spatial feature map. Finally, it undergoes convolution to fuse channel information, obtaining a local feature map with global context awareness. This invention provides the subsequent feature alignment module with a local feature map that retains both high-resolution local details and global spatial association information, effectively overcoming the limitations of dilated convolution in long-distance dependency modeling and improving the global consistency of target boundary localization.

[0045] Furthermore, a magnetic resonance-guided target region boundary extraction system includes a feature alignment module comprising a feature downsampling submodule, a bounding box mapping submodule, an affine transformation matrix mapping submodule, and a deformation convolution submodule.

[0046] The feature downsampling submodule downsamples the local feature map to the same spatial resolution as the global feature map to obtain a downsampled feature map.

[0047] The bounding box mapping submodule obtains the boundary of the coarse boundary target area, takes the maximum and minimum index coordinates in the width, height and depth directions, and converts them into normalized coordinates corresponding to the size of the downsampled feature map;

[0048] The affine transformation matrix mapping submodule performs a three-dimensional global average pooling operation on the global feature map, compressing it into a global feature vector. It then concatenates the global feature vector with the normalized coordinates of the coarse boundary target area. The concatenated features are mapped into an affine transformation parameter vector through a multilayer perceptron, and the affine transformation parameters are reshaped into affine moments.

[0049] The deformation convolution submodule constructs a reference network, obtains affine coordinates by transforming the reference network using an affine matrix, obtains deformation offsets by downsampling feature maps and global feature maps, combines deformation offsets and affine coordinates to obtain sampling coordinates, and performs three-dimensional deformation convolution centered on the sampling coordinates to obtain alignment features.

[0050] In the above solution, this invention addresses the technical problem of spatial misalignment between global feature maps derived from low-resolution compressed video and local feature maps derived from high-resolution supplementary information. This invention uses a feature downsampling submodule to adjust the spatial resolution of the local feature maps to match that of the global feature maps. Figure 1 This invention obtains downsampled feature maps, providing a unified scale benchmark for subsequent spatial alignment operations across feature maps. The bounding box mapping submodule obtains the maximum and minimum index coordinates of the coarse-boundary target region in the width, height, and depth directions, and converts them into normalized coordinates corresponding to the size of the downsampled feature map, establishing a coordinate mapping bridge between the global and local feature spaces for the coarse-boundary target region. The affine transformation matrix mapping submodule compresses the global feature map into a global feature vector through 3D global average pooling, concatenates it with normalized coordinates, and inputs it into a multilayer perceptron to generate an affine transformation parameter vector and reshape it into an affine matrix. This affine matrix can effectively compensate for overall translational and rotational displacements caused by respiratory movements or differences in scanning position. The invention constructs a normalized reference network through a deformation convolution submodule, transforms the reference network using an affine matrix to obtain affine coordinates, and simultaneously concatenates the downsampled feature map and the global feature map, generating a deformation offset through 3D convolution to model local non-rigid deformation caused by organ elastic deformation. The final sampled coordinates are obtained by superimposing affine coordinates and deformation offsets. A 3D deformation convolution is then performed on the downsampled feature map centered on these sampled coordinates to generate spatially consistent aligned features. This invention achieves precise calibration of global and local feature maps in anatomical space through the synergistic effect of affine transformation and deformation offsets, providing high-quality, spatially aligned feature representations for subsequent feature fusion modules.

[0051] Furthermore, a magnetic resonance-guided target region boundary extraction system, wherein the feature fusion module includes a feature splicing submodule, a gating weight submodule, and a decoder;

[0052] The feature splicing submodule splices the aligned features and the global feature map to obtain spliced ​​features;

[0053] The gate weight submodule performs convolution and activation operations on the spliced ​​features to obtain gate weights, and then performs element-wise multiplication of the gate weights and spliced ​​features to obtain fused features.

[0054] The decoder generates a boundary probability map by decoding the fused features, and obtains the boundary of the target region using threshold transformation.

[0055] In the above scheme, to address the technical problem of feature redundancy or mutual interference between global coarse-grained information and local fine-grained information in the stitched features, this invention uses a feature stitching submodule to stitch the aligned features and the global feature map along the channel dimension, obtaining stitched features that simultaneously carry global spatial guidance information and local high-resolution details. This invention uses a gating weight submodule to perform convolution operations and sigmoid activation on the stitched features, generating a gating weight matrix with values ​​in the [0,1] interval. Then, the gating weights are multiplied element-wise with the stitched features, which can adaptively enhance important feature channels related to the target region boundary while suppressing the response of irrelevant background regions or redundant features, obtaining more discriminative fused features. This invention uses a decoder to perform multi-layer upsampling and convolution processing on the fused features, gradually restoring the spatial resolution of the feature map to be consistent with the original video information. Finally, a boundary probability map is output after convolutional layers and a sigmoid activation function. The value at each voxel position represents the probability that the voxel belongs to the target region boundary. By setting a binarization threshold, the boundary probability map is converted into the final target region boundary. This invention, through gated adaptive fusion and multi-layer progressive upsampling, effectively suppresses background noise and feature redundancy while fully preserving the fine details of the target area boundary, significantly improving the positioning accuracy and continuity of the target area boundary.

[0056] A method for extracting the boundary of a target region guided by magnetic resonance imaging includes the following sub-steps:

[0057] Three-dimensional images are acquired using magnetic resonance imaging equipment, the three-dimensional images are integrated to obtain the original video information, and the original video information is compressed to obtain compressed video information.

[0058] By mapping the compressed video information to the original video information, the mapping relationship between the compressed video information and the original video information is obtained.

[0059] Process compressed video information, combine threshold to determine the edge points of the 3D image in the compressed video information, and connect the edge points to generate the boundary of the coarse boundary target area;

[0060] By leveraging the mapping relationship between compressed video information and original video information, the three-dimensional image region corresponding to the original video information of the coarse boundary target area is extracted, and the three-dimensional image region is used as supplementary information.

[0061] The boundaries of the coarse-boundary target area are superimposed on the 3D image of the compressed video information in the form of annotations to form global information;

[0062] Global feature maps are extracted using global information, local feature maps are extracted using supplementary information, alignment features are obtained using local feature maps, and fused features are obtained by fusing alignment features and global feature maps. Finally, the fused features are decoded to obtain the boundary of the target region. Attached Figure Description

[0063] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly described below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a structural diagram of a magnetic resonance-guided target region boundary extraction system.

[0065] Figure 2 Extract the structural diagram of the model for the second boundary. Detailed Implementation

[0066] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0067] It should be noted that similar reference numerals and letters in the following figures denote similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, the terms "first," "second," etc., are used only for distinguishing descriptions and should not be construed as indicating or implying relative importance, or suggesting any such actual relationship or order between these entities or operations. Additionally, the terms "connected," "linked," etc., can refer to a direct connection between elements or an indirect connection via other elements.

[0068] like Figure 1 As shown, a magnetic resonance-guided target region boundary extraction system includes a compressed video information acquisition module, an inverse feature extraction module, a first boundary extraction model, a supplementary information module, a global information integration module, and a second boundary extraction model.

[0069] The module for acquiring compressed video information acquires three-dimensional images through a magnetic resonance imaging device, integrates the three-dimensional images to obtain the original video information, and compresses the original video information to obtain compressed video information.

[0070] The module for acquiring compressed video information includes a sub-module for acquiring original video information and a sub-module for compressing original video information.

[0071] The original video information acquisition submodule acquires several three-dimensional images at high frame rate and high spatial resolution using a magnetic resonance imaging (MRI) device, sorts the several three-dimensional images according to the acquisition time of each frame, and integrates the sorted three-dimensional images to obtain the original video information.

[0072] The three-dimensional image is an anatomical image of the treatment target area (e.g., tumor) and surrounding organs at risk (e.g., spinal cord, intestines).

[0073] The high frame rate is 1-10 frames per second, and the high spatial resolution is voxel size ≤ 2mm.

[0074] The original video information compression submodule compresses each frame of the original video information using the mean value to obtain compressed video information.

[0075] The submodule for compressing the original video information specifically involves acquiring each frame of the original video information as a 3D image, setting the size of the compression window in the x, y, and z axes, dividing the 3D image into several non-overlapping 3D blocks through the compression window, using the average of all voxel values ​​within each 3D block as the voxel value of each frame of the compressed 3D image, traversing all 3D blocks to complete single-frame compression, and obtaining the compressed video information. The formula is:

[0076] ;

[0077] in, To compress the voxel values ​​of the 3D image in the video information, i is the index along the x-axis, j is the index along the y-axis, and k is the index along the z-axis. The window size is along the x-axis. This represents the window size along the y-axis. This represents the window size along the z-axis. This represents the offset along the x-axis. This represents the offset along the y-axis. This is the offset in the z-axis direction. For the 3D images in the original video information The voxel value of the coordinates.

[0078] In the embodiment, the compressed window's Set to 10×10×10.

[0079] In the above-mentioned scheme, the neural network model has multiple convolutional kernels processing the data. The convolutional kernels essentially learn how to extract key feature information through constantly changing parameters. However, this application fixes the way the convolutional kernels process the data, that is, compressing the original video information into compressed video information and removing non-critical information from the 3D image (for example, most organs, muscles, internal cavities, etc. in the 3D image, where voxels mainly show texture features, which are of no use in delineating the target area without tumor diagnosis tasks and are considered non-critical information). The key information of this invention is the area related to the boundary of specific target areas such as blood vessels, organs, bones, and tumors. Because this area has boundary features, the voxel value will change drastically. By compressing the video information, the constantly changing parameters of the convolutional kernel are reduced, avoiding the waste of computing power.

[0080] The reverse feature extraction module maps compressed video information to original video information to obtain the mapping relationship between compressed video information and original video information.

[0081] The inverse feature extraction module specifically obtains the dimensions of the compressed window in the x, y, and z axes, based on the voxel values ​​of the three-dimensional image in the compressed video information. voxel position Obtain the voxel coordinate range of the 3D image in the original video information, and obtain the mapping relationship:

[0082] The voxel coordinate range of the 3D image in the x-axis direction in the original video information is: ;

[0083] The voxel coordinate range of the 3D image in the original video information along the y-axis is: ;

[0084] The voxel coordinate range of the 3D image in the z-axis direction in the original video information is: .

[0085] It should be noted that if the mapping relationship cannot be obtained through inverse feature extraction, after selecting several calibration points for alignment, the mapping relationship between compressed video information and original video information can be obtained. The calibration points are points where the compressed video information and original video information are obviously in the same position.

[0086] The first boundary extraction model processes compressed video information, combines a threshold to determine the edge points of the three-dimensional image in the compressed video information, and connects the edge points to generate the boundary of the coarse boundary target area.

[0087] The first boundary extraction model includes a Gaussian filtering module, an edge intensity calculation module, a maximum value filtering module, and a threshold connection module.

[0088] The Gaussian filtering module performs Gaussian smoothing on the 3D image in the compressed video information to obtain smoothed data, using the following formula:

[0089] ;

[0090] in, voxel position Smoothed data, Let the radius be the Gaussian kernel. To compress voxel locations in video information voxel values, It is a three-dimensional Gaussian smoothing kernel. is the standard deviation of the three-dimensional Gaussian smoothing kernel.

[0091] In the embodiments, .

[0092] The edge intensity calculation module obtains the gradient vector at each scale by performing three-dimensional convolution on smoothed data and a multi-scale Gaussian kernel at each voxel location.

[0093] The edge strength calculation module specifically generates m scale parameters, and calculates the edge strength at the m-th scale respectively. Taking the partial derivative along the index direction yields the multi-scale Gaussian kernel for partial derivative calculation, as shown in the formula:

[0094] ;

[0095] in, To obtain the multi-scale Gaussian kernel with respect to the partial derivative with respect to index i, To obtain the multi-scale Gaussian kernel with respect to the partial derivative with respect to index j, To obtain the multi-scale Gaussian kernel with partial derivatives with respect to index k, To find the partial derivative with respect to index i, To find the partial derivative with respect to index j, To find the partial derivative with respect to index k, Let m be the m-th Gaussian kernel scaling parameter, where m is the scaling index, m = 1, 2, 3...M, and M is the total number of scaling parameters. This is the preset minimum scale parameter.

[0096] In the embodiments, .

[0097] By performing a 3D convolution on the smoothed data and the multi-scale Gaussian kernel with partial derivatives, the gradient vector at the m-th scale is obtained, as shown in the formula:

[0098] ;

[0099] in, The position of the voxel at the m-th scale gradient vector, This is a 3D convolution operation.

[0100] The gradient vectors at various scales are fused at multiple scales to construct the orientation tensor matrix, as shown in the formula:

[0101] ;

[0102] in, voxel position The direction tensor matrix, where T is the matrix transpose. The numerical stability coefficient, is the weight coefficient for the m-th scale.

[0103] In the embodiments, .

[0104] The maximum value filtering module obtains the anisotropy index of the direction tensor matrix through eigenvalues, selects the total amplitude after enhancement at each voxel position using the anisotropy index, judges edge points based on the total amplitude after enhancement at each voxel position, and obtains the gradient direction through eigenvalues.

[0105] The maximum value filtering module specifically performs eigenvalue decomposition on the direction tensor matrix to obtain eigenvalues ​​arranged in descending order of size. The anisotropy index of the direction tensor matrix is ​​then obtained from these eigenvalues, using the following formula:

[0106] ;

[0107] in, Let be the direction tensor matrix. Let p be the eigenvector corresponding to the p-th eigenvalue. For the p-th eigenvalue, , The anisotropy index.

[0108] The enhancement amplitudes of blood vessels and organ / skeletal margins are obtained based on the anisotropy index. The total enhancement amplitude is selected by determining the magnitude of the anisotropy index, using the following formula:

[0109] ;

[0110] in, The enhanced total amplitude, For vascular enhancement amplitude, To enhance the amplitude at the edges of organs and bones, If, It is an exponential function with base e. It is a vascular enhancing factor. It is a noise suppression factor.

[0111] It should be noted that if Therefore, there is no need to increase the total amplitude.

[0112] In the embodiments, , .

[0113] The gradient direction is chosen by selecting the eigenvector corresponding to the first eigenvalue, as shown in the formula:

[0114] ;

[0115] in, The gradient direction.

[0116] The threshold connection module determines the edge points based on the enhanced amplitude and uses the gradient direction and edge points to generate the boundary of the coarse boundary target area.

[0117] The threshold connection module is specifically set to a high threshold. and low threshold Determine edge points:

[0118] like If so, then it is considered a strong edge point;

[0119] like If it is, then it is considered a weak edge point;

[0120] in, voxel position The enhanced total amplitude.

[0121] In the embodiments, , .

[0122] Starting from strong edge points, along the gradient direction Tracing and connecting adjacent weak edge points generates the boundary of the coarse-boundary target area.

[0123] In the above scheme, a simple and unlabeled first boundary extraction model is used to process the compressed video information to obtain the boundary of the coarse boundary target area, saving the computing power spent on non-critical areas. In existing technologies, extraction models often use gradient operators to calculate all gradients and amplitudes in the 3D image to filter out extreme points. This filtering method cannot distinguish complex organ information, resulting in low accuracy of extreme point extraction. However, this application uses multi-scale directional tensor matrix analysis to identify the differences between blood vessels (tubular) and organ bones (block / area-like) to achieve adaptive gradient enhancement. Specifically, for tubular structures like blood vessels, which are cylindrical, the gradient is strongest along the long axis of the blood vessel and decays rapidly in the radial direction. Therefore, for blood vessel structures, the gradient will exhibit... The characteristic of rapid reduction. As for organs and bones, their volume is larger and more uniform, exhibiting... This characteristic. middle, Used to describe the concentration in the principal direction. Used to describe attenuation in the sub-direction, the marking characteristics of organ and bone edges and blood vessel edges in A can accurately capture the difference between the two, achieving powerful edge capture.

[0124] The supplementary information module extracts the three-dimensional image region corresponding to the original video information of the coarse boundary target area by mapping the compressed video information and the original video information, and uses the three-dimensional image region as supplementary information.

[0125] In the above scheme, inputting supplementary information into the second boundary extraction model can increase the accuracy of the second boundary extraction model in extracting the target area. The supplementary information is used to refine the boundary correction of global information through feature fusion.

[0126] The global information integration module overlays the boundaries of the coarse-boundary target area onto the 3D image of the compressed video information in the form of annotations to form global information.

[0127] The second boundary extraction model uses global information to extract a global feature map, extracts a local feature map by supplementing information, obtains alignment features using the local feature map, fuses the alignment features and the global feature map to obtain a fused feature, and decodes the fused feature to obtain the boundary of the target region.

[0128] like Figure 2 As shown, the second boundary extraction model includes a global path processing module, a supplementary path processing module, a feature alignment module, and a feature fusion module.

[0129] The global path processing module includes a global convolution submodule and a residual extraction submodule.

[0130] The global path processing module obtains global network features by processing global information through convolution, and obtains a global feature map by processing global network features through spatial downsampling and residual learning.

[0131] The global convolution submodule performs convolution operations on the global information, and then batch normalizes and activates the global information after the convolution operation to obtain the global network features, as shown in the formula:

[0132] ;

[0133] in, As a global network feature, For global information, This is a 3×3 convolution operation. For batch normalization, For the ReLU function;

[0134] The residual extraction submodule performs spatial downsampling on the global network features. While performing spatial downsampling, it retains key information through residual blocks to obtain a global feature map.

[0135] The residual extraction submodule specifically performs convolution, batch normalization, and activation on the input residual block group data in the input layer, and outputs it to the intermediate layer for convolution, batch normalization, and activation to obtain the backbone path. The formula is as follows:

[0136] ;

[0137] in, To input the data for the residual block group, This refers to the 3×3 convolution operation of the input layer in the residual block group. For the 3×3 convolution operation in the middle layer of the residual block group, Main path;

[0138] The number of input layer channels and the number of output layer channels are compared to obtain the short-circuit path, using the following formula:

[0139] ;

[0140] in, Otherwise, This represents the number of input layer channels in the residual block group. This represents the number of output layer channels in the residual block group. For the 1×1 convolution operation of the output layer in the residual block group, This is a short-circuit path;

[0141] By shortening the path and the main path, the key information of each residual block is obtained, as shown in the formula:

[0142] ;

[0143] in, Key information for residual blocks;

[0144] Residual extraction is performed on the global network features to obtain the global feature map, using the following formula:

[0145] ;

[0146] in, For global feature maps, For residual extraction, For spatial downsampling, This refers to the composition of functions. For the residual block processing operation of the two stacks at the fourth level, This is a residual block processing operation for the two stacks of the first level.

[0147] The supplementary path processing module includes a dilated convolution submodule and a ViT enhancement submodule.

[0148] The supplementary path processing module uses dilated convolution to process supplementary information to obtain supplementary convolutional features, and uses a visual transformer (ViT) to enhance the supplementary convolutional features to obtain a local feature map.

[0149] The dilated convolution submodule performs dilated convolution on the supplementary information to obtain supplementary convolution features, as shown in the formula:

[0150] ;

[0151] in, For location index Supplementary convolutional features at the location, For activation function, For the dilated convolution kernel bias term, The dilated convolution kernel weight matrix is... The kernel width is half the width. The height-direction offset of the dilated convolution kernel. This is an offset in the width direction of the dilated convolution kernel. This represents the depth-direction offset of the dilated convolution kernel. The dilation rate of the hollow convolution kernel. For dilated convolution kernels at offset The weight value, To supplement the height coordinates of the convolutional features, To supplement the width coordinates of the convolutional features, To supplement the depth coordinates of the convolutional features, To supplement information In location index The voxel value.

[0152] The ViT enhancement submodule divides the supplementary convolutional features into three-dimensional slices and flattens them to obtain a slice sequence. The slice sequence is projected into a D-dimensional latent space to obtain an initial learnable positional encoding. The initial learnable positional encoding is processed by a Transformer encoding layer to obtain an attention feature sequence. The attention feature sequence is then projected back into the original spatial dimension for convolution operation to obtain a local feature map.

[0153] The ViT enhancement submodule specifically defines the slice size as follows: The supplementary convolutional features are divided into N non-overlapping 3D slices along the spatial dimension. These 3D slices are then flattened into 1D vectors to obtain the slice sequence, as shown in the formula:

[0154] ;

[0155] in, It is a slice sequence. , For the set of real numbers, This is a three-dimensional spatial slicing operation. The dimension of the slice in the height direction. The dimension of the slice in the width direction. The dimension of the slice in the depth direction. To supplement convolutional features;

[0156] The slice sequence is projected into the D-dimensional latent space using a slice embedding weight matrix. After projection, it is combined with the position encoding matrix to obtain the initial learnable position encoding, as shown in the formula:

[0157] ;

[0158] in, For the initial learnable position encoding, Embed the weight matrix for the slice. This is the position encoding matrix;

[0159] The L-layer Transformer encoding layer processes the initial learnable positional encodings and extracts attention information layer by layer. Attention feature sequence is obtained. The formula is:

[0160] ;

[0161] in, For the first The Transformer of the layer encodes the intermediate sequence of the layer. For multi-head self-attention operation, For layer normalization operation, For the first The input sequence of the Transformer encoding layer, For the index of the Transformer encoding layer, , This is for multilayer perceptron operation.

[0162] The attention feature sequence is projected back to the original spatial dimension through a linear projection matrix. After returning to the original spatial dimension, it is reconstructed into a spatial feature map. The spatial feature map is then convolved to obtain a local feature map, as shown in the formula:

[0163] ;

[0164] in, It is a linear projection matrix. For the reconstruction operation, For spatial feature maps, This is a 1×1 convolution operation. This is a local feature map.

[0165] The feature alignment module downsamples the local feature map to obtain a downsampled feature map. Using the downsampled feature map and the boundary of the coarse boundary target area, it obtains the normalized coordinates of the coarse boundary target area. It then concatenates the normalized coordinates of the coarse boundary target area with the global feature map. After concatenation, it uses a multilayer perceptron and a reshaping operation to obtain an affine matrix. It then uses the affine matrix to obtain affine coordinates. Finally, it combines the deformation offset and the affine coordinates to obtain the sampling coordinates. A three-dimensional deformation convolution is performed with the sampling coordinates as the center to obtain the alignment features.

[0166] The feature alignment module includes a feature downsampling submodule, a bounding box mapping submodule, an affine transformation matrix mapping submodule, and a deformation convolution submodule.

[0167] The feature downsampling submodule downsamples the local feature map to the same spatial resolution as the global feature map to obtain a downsampled feature map.

[0168] The bounding box mapping submodule obtains the boundary of the coarse-boundary target region, takes the maximum and minimum index coordinates in the width, height, and depth directions, and converts them into normalized coordinates corresponding to the size of the downsampled feature map, using the following formula:

[0169] ;

[0170] in, Normalized coordinates for the coarse-boundary target region. This is the minimum width index for normalizing the coarse-boundary target region. This is the minimum height index for normalizing the coarse-boundary target region. This is the minimum depth index for normalizing the coarse-boundary target region. This is the maximum width index for normalized coarse-boundary target regions. This is the maximum height index for normalized coarse-boundary target regions. For the maximum depth index normalized to the coarse-boundary target region, The minimum width index for the coarse-boundary target region. This is the minimum height index for the coarse-boundary target region. Minimum depth index for coarse-boundary target regions. Index of the maximum width of the coarse-boundary target region. The maximum height index for the coarse-boundary target area. The maximum depth index for the coarse-boundary target region. To reduce the width of the downsampled feature map, To reduce the height of the sampled feature map, This is to reduce the depth of the downsampled feature map.

[0171] The affine transformation matrix mapping submodule performs a 3D global average pooling operation on the global feature map, compressing it into a global feature vector. It then concatenates the global feature vector with the normalized coordinates of the coarse-boundary target region. The concatenated features are mapped to an affine transformation parameter vector through a multilayer perceptron. The affine transformation parameters are then reshaped into an affine matrix, as shown in the formula:

[0172] ;

[0173] in, Let be the affine transformation parameter vector. For splicing, This is a three-dimensional global average pooling operation. For global feature vectors, For global feature maps, To reshape the matrix into a 3×4 matrix, It is an affine matrix.

[0174] In the embodiments, Affine transformation parameter vector of length 12.

[0175] The deformation convolution submodule constructs a reference network, obtains affine coordinates by transforming the reference network using an affine matrix, obtains deformation offsets by downsampling feature maps and global feature maps, combines deformation offsets and affine coordinates to obtain sampling coordinates, and performs three-dimensional deformation convolution centered on the sampling coordinates to obtain alignment features.

[0176] The deformation convolution submodule specifically defines the feature dimensions of the alignment features and constructs a normalized reference network based on these feature dimensions, using the following formula:

[0177] ;

[0178] in, For reference network Floating-point numbers in the index direction, , For the width of the alignment feature, For the height of the alignment feature, To align the depth of the feature, For reference network width index, For reference network height indexing, For reference network depth index;

[0179] In the above scheme, a normalized reference network is constructed based on the defined feature dimension of the alignment feature. The reference network is related to the accuracy of the target area (for example, when the accuracy range of the target area boundary needs to be without deviation from the original video information, the feature dimension of the alignment feature is consistent with the resolution of the original video information). If the accuracy requirement of the target area is reduced, the feature dimension of the alignment feature is also reduced. By controlling the feature dimension of the alignment feature, the amount of computation can be controlled, and the relationship between the amount of computation and the efficiency of computation can be balanced.

[0180] Affine coordinates are obtained by performing an affine transformation on the reference network using an affine matrix, as shown in the formula:

[0181] ;

[0182] in, Affine coordinates;

[0183] The downsampled feature map and the global feature map are concatenated, and the deformation offset is obtained by a 3D convolution operation. The formula is as follows:

[0184] ;

[0185] in, For downsampling feature maps, This is a 3×3×3 3D convolution operation. This is the deformation offset.

[0186] Combining the deformation offset and affine coordinates, the sampling coordinates are obtained using the following formula:

[0187] ;

[0188] in, These are the sampling coordinates;

[0189] Using the sampling coordinates as the center, perform 3D deformation convolution sampling on the downsampled feature map to obtain the alignment features, as shown in the formula:

[0190] ;

[0191] in, for Alignment features at coordinate indices This is the channel index of the downsampled feature map. To reduce the number of channels in the downsampled feature map, The half-width of the deformable convolution kernel, This represents the spatial offset of the deformable convolution kernel. For deformable convolution kernels in downsampled feature map channels Spatial offset of deformable convolution kernel The weight of the position, For spatial transformation sampling operators, This represents the width offset of the deformable convolution kernel. This represents the height offset of the deformable convolution kernel. Depth offset of the deformable convolution kernel.

[0192] In the above scheme, the alignment mechanism uses the synergistic effect of affine transformation (global rigid alignment) and deformation offset (local elastic adaptation) to accurately calibrate feature maps from different slices, phases, or imaging devices in anatomical space. The affine coordinates can then compensate for the overall displacement caused by respiratory motion or differences in scanning position. Furthermore, the local deformation field learned by 3D convolution models the elastic deformation of organs (such as myocardial contraction and tumor morphological changes). Through trilinear interpolation convolution with coordinate offset, spatially consistent feature representations are generated while maintaining the continuity of the anatomical structure. This allows the global feature map and the supplementary feature map to be accurately aligned together. During training, it can accurately capture the error effects caused by the elastic deformation of different organs in different samples.

[0193] The feature fusion module concatenates the aligned features and the global feature map to obtain the concatenated features, and then fuses the concatenated features with the gating weights to obtain the fused features. Finally, the fused features are decoded to obtain the boundary of the target region.

[0194] The feature fusion module includes a feature splicing submodule, a gating weight submodule, and a decoder;

[0195] The feature stitching submodule stitches together the alignment features and the global feature map to obtain the stitched features, as shown in the formula:

[0196] ;

[0197] in, For alignment features, This is a splicing feature.

[0198] The gate weight submodule performs convolution and activation operations on the concatenated features to obtain gate weights. The gate weights and concatenated features are then multiplied element-wise to obtain the fused features, as shown in the formula:

[0199] ;

[0200] in, For convolution operations, For gating weights, As a feature of fusion, For element-wise multiplication, This is the activation function.

[0201] The decoder generates a boundary probability map by decoding the fused features, and obtains the boundary of the target region using threshold transformation.

[0202] The decoder specifically samples and convolves the fused features through multiple upsampling layers to gradually restore the spatial resolution. Then, after convolution and activation, a boundary probability map is obtained, as shown in the formula:

[0203] ;

[0204] in, Let S be the feature map after processing by the s-th upsampling layer, where s = 1, 2, ..., S, and s is the index of the upsampling layer, and S is the number of upsampling layers. This is for upsampling layer sampling operations. For 1×1×1 3D convolution operations, The input features are those of the first upsampling layer. It is the Sigmoid activation function. This is a boundary probability diagram.

[0205] The boundary of the target region is obtained by thresholding the boundary probability map. The formula is as follows:

[0206] ;

[0207] in, The voxel positions of the 3D image in the original video information The boundary of the target area, The boundary of the target area, The three-dimensional image in the original video information The probability that the coordinates belong to the boundary of the target region. This is the binarization threshold.

[0208] In the embodiments, .

[0209] It should be noted that the specific ways in which each module performs its operations in the above embodiments have been described in detail in the embodiments of the system, and will not be elaborated here.

[0210] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A magnetic resonance-guided target region boundary extraction system, characterized in that, This includes a module for acquiring compressed video information, a reverse feature extraction module, a first boundary extraction model, a supplementary information module, a global information integration module, and a second boundary extraction model; The module for acquiring compressed video information acquires three-dimensional images through a magnetic resonance imaging device, integrates the three-dimensional images to obtain the original video information, and compresses the original video information to obtain compressed video information. The reverse feature extraction module maps compressed video information and original video information to obtain the mapping relationship between compressed video information and original video information. The first boundary extraction model processes compressed video information, combines a threshold to determine the edge points of the three-dimensional image in the compressed video information, and connects the edge points to generate the boundary of the coarse boundary target area. The supplementary information module extracts the three-dimensional image region corresponding to the original video information of the coarse boundary target area by the mapping relationship between compressed video information and original video information, and uses the three-dimensional image region as supplementary information. The global information integration module overlays the boundary of the coarse-boundary target area onto the three-dimensional image of the compressed video information in the form of annotations to form global information. The second boundary extraction model uses global information to extract a global feature map, extracts a local feature map by supplementing information, obtains alignment features using the local feature map, fuses the alignment features and the global feature map to obtain a fused feature, and decodes the fused feature to obtain the boundary of the target region.

2. The magnetic resonance-guided target region boundary extraction system according to claim 1, characterized in that, The module for acquiring compressed video information includes a sub-module for acquiring original video information and a sub-module for compressing original video information. The original video information acquisition submodule acquires several three-dimensional images at high frame rate and high spatial resolution using a magnetic resonance imaging device, sorts the several three-dimensional images according to the acquisition time of each frame, and integrates the sorted three-dimensional images to obtain the original video information. The original video information compression submodule compresses each frame of the original video information using the mean value to obtain compressed video information.

3. The magnetic resonance-guided target region boundary extraction system according to claim 1, characterized in that, The first boundary extraction model includes a Gaussian filtering module, an edge strength calculation module, a maximum value filtering module, and a threshold connection module; The Gaussian filtering module performs Gaussian smoothing on the three-dimensional image in the compressed video information to obtain smoothed data. The edge intensity calculation module obtains the gradient vector at each scale by performing three-dimensional convolution on smoothed data and a multi-scale Gaussian kernel at each voxel location. The maximum value filtering module obtains the anisotropy index of the orientation tensor matrix through eigenvalues, selects the total amplitude after enhancement at each voxel position using the anisotropy index, judges edge points based on the total amplitude after enhancement at each voxel position, and obtains the gradient direction through eigenvalues. The threshold connection module determines the edge points based on the enhanced amplitude and uses the gradient direction and edge points to generate the boundary of the coarse boundary target area.

4. The magnetic resonance-guided target region boundary extraction system according to claim 3, characterized in that, The formula for selecting the total amplitude of enhancement at each voxel location using the anisotropy index is as follows: ； in, The enhanced total amplitude, For vascular enhancement amplitude, To enhance the amplitude at the edges of organs and bones, If, It is an exponential function with base e. It is a vascular enhancing factor. As a noise suppression factor, The anisotropy index, Let p be the p-th eigenvalue, where p = 1, 2, 3.

5. The magnetic resonance-guided target region boundary extraction system according to claim 1, characterized in that, The second boundary extraction model includes a global path processing module, a supplementary path processing module, a feature alignment module, and a feature fusion module; The global path processing module obtains global network features by processing global information through convolution, and obtains a global feature map by processing global network features through spatial downsampling and residual learning. The supplementary path processing module uses dilated convolution to process supplementary information to obtain supplementary convolutional features, and uses a visual transformer to enhance the supplementary convolutional features to obtain local feature maps. The feature alignment module downsamples the local feature map to obtain a downsampled feature map. Using the downsampled feature map and the boundary of the coarse boundary target area, it obtains the normalized coordinates of the coarse boundary target area. It then concatenates the normalized coordinates of the coarse boundary target area with the global feature map. After concatenation, it uses a multilayer perceptron and a reshaping operation to obtain an affine matrix. It then uses the affine matrix to obtain affine coordinates. Finally, it combines the deformation offset and the affine coordinates to obtain the sampling coordinates. A three-dimensional deformation convolution is performed with the sampling coordinates as the center to obtain the alignment features. The feature fusion module concatenates the aligned features and the global feature map to obtain the concatenated features, and then fuses the concatenated features with the gating weights to obtain the fused features. Finally, the fused features are decoded to obtain the boundary of the target region.

6. The magnetic resonance-guided target region boundary extraction system according to claim 5, characterized in that, The global path processing module includes a global convolution submodule and a residual extraction submodule; The global convolution submodule performs convolution operations on global information, and batch normalizes and activates the global information after convolution to obtain global network features. The residual extraction submodule performs spatial downsampling on the global network features. While performing spatial downsampling, it retains key information through residual blocks to obtain a global feature map.

7. The magnetic resonance-guided target region boundary extraction system according to claim 5, characterized in that, The supplementary path processing module includes a dilated convolution submodule and a ViT enhancement submodule: The dilated convolution submodule performs dilated convolution on the supplementary information to obtain supplementary convolution features; The ViT enhancement submodule divides the supplementary convolutional features into three-dimensional slices and flattens them to obtain a slice sequence. The slice sequence is projected into a D-dimensional latent space to obtain an initial learnable positional encoding. The initial learnable positional encoding is processed by a Transformer encoding layer to obtain an attention feature sequence. The attention feature sequence is then projected back into the original spatial dimension for convolution operation to obtain a local feature map.

8. The magnetic resonance-guided target region boundary extraction system according to claim 5, characterized in that, The feature alignment module includes a feature downsampling submodule, a bounding box mapping submodule, an affine transformation matrix mapping submodule, and a deformation convolution submodule. The feature downsampling submodule downsamples the local feature map to the same spatial resolution as the global feature map to obtain a downsampled feature map. The bounding box mapping submodule obtains the boundary of the coarse boundary target area, takes the maximum and minimum index coordinates in the width, height and depth directions, and converts them into normalized coordinates corresponding to the size of the downsampled feature map; The affine transformation matrix mapping submodule performs a three-dimensional global average pooling operation on the global feature map, compressing it into a global feature vector. It then concatenates the global feature vector with the normalized coordinates of the coarse boundary target area. The concatenated features are mapped into an affine transformation parameter vector through a multilayer perceptron, and the affine transformation parameters are reshaped into an affine matrix. The deformation convolution submodule constructs a reference network, obtains affine coordinates by transforming the reference network using an affine matrix, obtains deformation offsets by downsampling feature maps and global feature maps, combines deformation offsets and affine coordinates to obtain sampling coordinates, and performs three-dimensional deformation convolution centered on the sampling coordinates to obtain alignment features.

9. The magnetic resonance-guided target region boundary extraction system according to claim 5, characterized in that, The feature fusion module includes a feature splicing submodule, a gating weight submodule, and a decoder; The feature splicing submodule splices the aligned features and the global feature map to obtain spliced ​​features; The gate weight submodule performs convolution and activation operations on the spliced ​​features to obtain gate weights, and then performs element-wise multiplication of the gate weights and spliced ​​features to obtain fused features. The decoder generates a boundary probability map by decoding the fused features, and obtains the boundary of the target region using threshold transformation.

10. A method for extracting the boundary of a target region guided by magnetic resonance imaging (MRI), used to implement the magnetic resonance-guided target region boundary extraction system according to any one of claims 1-9, characterized in that, Includes the following sub-steps: Three-dimensional images are acquired using magnetic resonance imaging equipment, the three-dimensional images are integrated to obtain the original video information, and the original video information is compressed to obtain compressed video information. By mapping the compressed video information to the original video information, the mapping relationship between the compressed video information and the original video information is obtained. Process compressed video information, combine threshold to determine the edge points of the 3D image in the compressed video information, and connect the edge points to generate the boundary of the coarse boundary target area; By leveraging the mapping relationship between compressed video information and original video information, the three-dimensional image region corresponding to the original video information of the coarse boundary target area is extracted, and the three-dimensional image region is used as supplementary information. The boundaries of the coarse-boundary target area are superimposed on the 3D image of the compressed video information in the form of annotations to form global information; Global feature maps are extracted using global information, local feature maps are extracted using supplementary information, alignment features are obtained using local feature maps, and fused features are obtained by fusing alignment features and global feature maps. Finally, the fused features are decoded to obtain the boundary of the target region.

Images