Lumbosacral plexus nerve root analysis device and method based on magnetic resonance diffusion tensor imaging

Abstract

The invention discloses a lumbosacral plexus nerve root analysis device and method based on magnetic resonance diffusion tensor imaging. The device comprises a segmentation module which is used for preprocessing a diffusion weighted image, cutting a to-be-segmented region, and carrying out nerve root segmentation; the fiber bundle tracking module is used for tracking the nerve root fiber bundle based on the fitting information of the automatic segmentation result and the diffusion weighted image; the visualization and fusion module is used for converting the tracking result into a world coordinate system, then carrying out three-dimensional visualization, and carrying out multi-modal fusion display on the tracking result and other magnetic resonance image sequences; and the identification and quantitative analysis module is used for identifying nerve root overlapping and fracture conditions and generating an FA value distribution curve layer by layer for each separated nerve root for quantitative analysis. According to the method, the problems of long time consumption and large individual difference of manual sketching in current clinical application are solved, the lumbosacral plexus nerve root fiber bundle visualization result can be efficiently and stably obtained, and quantitative FA analysis is provided for each nerve root.

Description

Technical Field

[0001] This invention belongs to the field of medical image post-processing application technology, specifically relating to an analysis device and method for automatic segmentation, identification, quantitative analysis and multimodal fusion of lumbosacral plexus nerve roots based on magnetic resonance diffusion tensor imaging. Background Technology

[0002] The lumbosacral plexus, a key neural structure controlling movement and sensation in the lower limbs, can cause a range of serious clinical problems with abnormalities, including numbness, pain, and paralysis in the lower extremities. Therefore, precise localization, segmentation, and even quantitative assessment of the lumbosacral plexus roots are crucial for evaluating nerve damage, diagnosing diseases, and guiding spinal surgery.

[0003] Magnetic resonance diffusion tensor imaging (MRI), with its multiplanar imaging capabilities and high-resolution soft tissue characteristics, has become a key method for non-invasive visualization of the three-dimensional course and pathological changes of the lumbosacral plexus nerve roots. However, due to the complex structure of the lumbosacral plexus nerve roots, low image contrast, and differences in neuromorphology among patients, accurately segmenting the lumbosacral plexus nerve roots from images has always been a major challenge in medical image analysis.

[0004] Building upon precise segmentation, three-dimensional fiber tract tracking and quantitative analysis of nerve roots have become a crucial extension requirement for clinical decision-making. The reconstructed fiber tracts can directly provide anatomical information about the nerve root's course. Furthermore, multimodal fusion of fiber tract tracking results with high-resolution sequences such as T2-weighted images can significantly enhance the intuitiveness of lesion area visualization; while generating and quantitatively analyzing independent FA (Fraction Anisotropy) parameter maps for specific nerve roots provides critical data for accurate assessment of nerve damage. These technologies are essential for the diagnosis of neurological diseases, preoperative planning, and postoperative efficacy evaluation.

[0005] However, there are still gaps in the current clinical procedures for managing lumbosacral plexus nerve roots. On the one hand, current clinical management of nerve roots mainly relies on manual methods, including manually outlining the nerve root contours and tracing fiber tracts based on manually set seed points. These methods are not only time-consuming and labor-intensive, typically requiring a specialist physician to spend tens of minutes or even longer to complete, but also suffer from poor consistency among different operators, easily leading to significant subjective differences. On the other hand, existing methods struggle to automate the accurate identification of individual nerve roots, thus failing to generate independent FA maps for each nerve root for targeted quantitative analysis, limiting the level of precision in assessment. Furthermore, the inefficiency of manual processing when dealing with large-scale imaging data severely limits its widespread application in routine clinical practice.

[0006] In conclusion, developing an integrated method that can automatically perform lumbosacral plexus nerve root segmentation, single nerve identification, quantitative FA analysis, and multimodal fusion is not only a technological innovation requirement but also an urgent need to achieve precision medicine. Summary of the Invention

[0007] The purpose of this invention is to address the shortcomings of existing technologies by providing a fully automated and quantitative lumbosacral plexus nerve root analysis device and method based on magnetic resonance diffusion tensor imaging.

[0008] To achieve the above objectives, the technical solution adopted by the device of the present invention is as follows:

[0009] A lumbosacral plexus nerve root analysis device based on magnetic resonance diffusion tensor imaging includes:

[0010] The segmentation module is used to preprocess the diffusion-weighted image, crop the region to be segmented according to the distribution characteristics of the nerve roots, and perform nerve root segmentation through the segmentation model.

[0011] The fiber tract tracking module is used to track nerve root fiber tracts based on the fitting information between the automatic segmentation results output by the segmentation module and the diffusion-weighted image.

[0012] The visualization and fusion module is used to convert the tracking results of the fiber bundle tracking module to the world coordinate system for three-dimensional visualization and to perform multimodal fusion display with other magnetic resonance image sequences.

[0013] The identification and quantitative analysis module is used to identify nerve root overlap and breakage, and to generate FA value distribution curves for each isolated nerve root layer by layer for quantitative analysis.

[0014] Furthermore, the segmentation model adopts the Attention U-Net model based on a three-axis attention convolution module. The specific structure is as follows: the model includes a multi-layer encoder and a corresponding decoder. Each encoder includes a three-axis attention convolution module and a downsampling module. In the deepest layer, a three-axis attention convolution module without changing the number of channels is used as a bottleneck layer to connect the encoder and the corresponding decoder. The features output by the last decoder are passed through a three-dimensional convolution layer to obtain the final segmentation result.

[0015] The present invention also provides an analysis method for the above-mentioned lumbosacral plexus nerve root analysis device based on magnetic resonance diffusion tensor imaging, the analysis method comprising the following steps:

[0016] After preprocessing the diffusion-weighted image, the region to be segmented is cropped according to the distribution characteristics of the nerve roots, and the nerve roots are segmented to obtain the automatic segmentation result.

[0017] The diffusion-weighted image is fitted, and the fitting information and automatic segmentation results are used to track nerve root fiber bundles.

[0018] The tracking results are transformed into the world coordinate system and then visualized in three dimensions. They are then fused with other magnetic resonance imaging sequences for multimodal display.

[0019] The study identifies overlapping and broken nerve roots, separates individual nerve roots, and generates FA value distribution curves for each separated nerve root layer by layer for quantitative analysis.

[0020] This invention first uses a deep learning model to automatically segment nerve root regions and uses these as seed points for fully automatic fiber tract tracking. Then, it uses directional information to remove parts with large deviations in the tracking results and generates an effective mask based on the coverage area. Through an identification algorithm (including the separation of overlapping nerve roots and the merging of broken areas), it achieves accurate identification of each individual nerve root. Finally, it generates an independent FA parameter map for each identified nerve root for quantitative analysis and fuses all results with high-resolution images for visualization. This invention solves at least the following problems of the prior art: (1) It solves the problems of long time consumption, strong subjectivity and poor consistency in the segmentation of lumbosacral plexus nerve roots in existing clinical diagnosis, which depends on manual operation; (2) It solves the problem of insufficient robustness and repeatability of the results due to the need to manually set seed points and thresholds in the fiber tracking process; (3) Most importantly, it solves the bottleneck problem that the existing methods cannot automatically and accurately distinguish and identify individual nerve roots, thus making it difficult to achieve independent quantitative assessment of specific nerve roots.

[0021] Compared with existing technologies, this invention achieves a technological leap from coarse-grained segmentation to precise identification of single nerve roots. It can not only greatly reduce manual intervention, improve processing efficiency and result consistency, but also provide clinical clinical quantitative analysis data of each lumbosacral plexus nerve root from L4 to S1, thereby providing more reliable and intuitive decision support for the precise localization of nerve compression lesions, severity assessment and preoperative planning. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is an overall frame diagram of the device of the present invention;

[0024] Figure 2 This is a schematic diagram illustrating the generation effect of the region to be segmented in an embodiment of the present invention;

[0025] Figure 3 This is a structural diagram of the three-axis attention convolution module in an embodiment of the present invention;

[0026] Figure 4 This is a structural diagram of the segmentation model in an embodiment of the present invention;

[0027] Figure 5 This is a schematic diagram of the segmentation results in an embodiment of the present invention;

[0028] Figure 6 This is a schematic diagram of the tracking results in an embodiment of the present invention;

[0029] Figure 7 This is a schematic diagram illustrating the visualization and fusion module effect in an embodiment of the present invention;

[0030] Figure 8 This is a schematic diagram of the FA curve in an embodiment of the present invention. Detailed Implementation

[0031] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0032] This invention aims to provide a device and method for analyzing lumbosacral plexus nerve roots based on magnetic resonance diffusion tensor imaging. It has functions of automatic segmentation, identification, quantitative analysis, and multimodal fusion. By automatically processing diffusion-weighted imaging (DWI) data, it can achieve automatic, efficient, and stable segmentation of L4-S1 lumbosacral plexus nerve roots and rapid tracking of corresponding fiber bundles. After removing deviating fibers, it can perform three-dimensional fusion visualization and further realize the identification and individual quantitative analysis of each nerve root.

[0033] like Figure 1As shown, the lumbosacral plexus nerve root analysis device based on magnetic resonance diffusion tensor imaging of the present invention includes: a segmentation module, a tracking module, a visualization and fusion module, and an identification and quantitative analysis module. The segmentation module preprocesses the DWI image and then crops the region to be segmented according to the distribution characteristics of the nerve roots, and performs nerve root segmentation by introducing an Attention U-Net model based on a triaxial attention convolution module. The fiber tract tracking module tracks nerve root fiber tracts based on the automatic segmentation results output by the segmentation module and the DTI fitting information, using seed point sampling after direction consistency filtering and a Streamline Tracking (STT) algorithm. The visualization and fusion module converts the tracking results to a world coordinate system for three-dimensional visualization and performs multimodal fusion with other magnetic resonance image sequences. The identification and quantitative analysis module identifies nerve root overlap and breakage, and generates FA value distribution curves layer by layer for quantitative analysis of each separated nerve root. The structure and workflow of each module are described in detail below.

[0034] 1. Segmentation Module

[0035] (1) Generate the region to be segmented

[0036] After preprocessing the acquired raw images, such as standardizing size and intensity, candidate regions containing the lumbosacral plexus nerve roots are extracted based on the anatomical distribution characteristics of the images to reduce irrelevant background interference. Data augmentation techniques such as random translation, random rotation, random flipping, and adding Gaussian noise can be used to improve the model's generalization performance.

[0037] In this embodiment, the specific implementation steps are as follows:

[0038] First, the original DWI image acquired using magnetic resonance diffusion tensor imaging technology is filled or cropped to a certain size (256×256×32 in this embodiment) and voxel intensity is normalized.

[0039] Next, based on prior knowledge of the location distribution of the lumbosacral plexus nerve roots, a sub-image to be segmented with a size of 96×96×32 centered at x=128, y=128 is cropped from the training set images. To improve the model's generalization performance, a random translation of 0-10 pixels is introduced to the center when cropping the training set images; that is, a random number between 0 and 10 is added to the x and y values ​​respectively as the final center used. The effect is as follows: Figure 2 As shown, the proportion of nerve roots in the image is significantly increased. This strategy can reduce irrelevant background interference, lower computational costs, and improve segmentation accuracy and inference efficiency.

[0040] (2) Target segmentation based on improved Attention U-Net

[0041] Based on the Attention U-Net structure, in order to better adapt to the anisotropic nature of data, this invention introduces an axial attention mechanism, using a triaxial attention convolution module to replace the double convolutional layer of the encoder in the original Attention U-Net, enabling the model to focus on high-resolution information and obtain accurate and stable segmentation results.

[0042] This embodiment presents a deep learning model that introduces an axial attention mechanism based on Attention U-Net.

[0043] Because DWI data has low resolution and low contrast, and the voxel spacing varies significantly along each axis, an axial attention mechanism is introduced to better utilize this anisotropy. This mechanism works by... Figure 3 The diagram shows the implementation of the Tri-Axial Attention Convolution (TAA Conv) module. This module includes an anisotropic feature extraction branch consisting of three parallel convolutional layers, a lightweight attention layer that generates corresponding weights for each feature path, and a final output part that normalizes the weights and weightedly fuses the features.

[0044] The specific workflow of the above-mentioned three-axis attention convolution module is as follows:

[0045] First, anisotropic convolutions are performed on the input data along the three axes to obtain feature information that is more focused on each axis. The specific feature transfer process is as follows:

[0046] featX = ConvX(IF)

[0047] featY = ConvY(IF)

[0048] featZ=ConvZ(IF)

[0049] In the formula, IF represents the input feature with dimensions [B,Cin,W,H,D], where B, Cin, W, H, and D are the batch number, the number of input channels, the feature map width, height, and depth, respectively. ConvX, ConvY, and ConvZ are three-axis convolutional layers, which in this embodiment are three-dimensional convolutional layers Conv3d with {kernel, padding} set to {(1,3,3),(0,1,1)}, {(3,1,3),(1,0,1)}, and {(3,3,1),(1,1,0)}, respectively. featX, featY, and featZ represent the feature information extracted in each axis, with dimensions [B,Cout,W,H,D], where Cout is the number of output channels.

[0050] Afterwards, through such Figure 3The lightweight attention module shown obtains the attention weights for each feature. The feature transfer process in the attention module is as follows:

[0051] AW=Conv3d2(ReLU(Conv3d1(GAP(feat))))

[0052] In the formula, feat represents the feature information along any axis, GAP is the global average pooling operation, Conv3d1 has a kernel size of 1, which transforms the number of channels from Cout to Cout / 4, that is, the number of channels after the feature passes through Conv3d1 is reduced to 1 / 4 of the number before input. ReLU is the activation function, Conv3d2 also has a kernel size of 1, which transforms the number of channels from Cout / 4 to 1, and AW is the attention weight of the final output with a size of [B,1,1,1,1].

[0053] Finally, the weighted summation yields the final output. The specific feature transfer process is as follows:

[0054] AWs = cat([Wx, Wy, Wz], dim = 1)

[0055] Wx′,Wy′,Wz′=Softmax(AWs)

[0056] OF=ReLU(BN(Wx′*featX+Wy′*featY+Wz′*featZ))

[0057] In the formula, Wx, Wy, and Wz represent the attention weights of each axis, respectively. The cat operation concatenates the weights in the first dimension for subsequent operations, resulting in AWs of size [B,3,1,1,1]. Wx', Wy', and Wz' represent the attention weights of each axis after being normalized by the Softmax operation. BN is the batch normalization operation, and OF is the final output feature of the three-axis attention convolution module with size [B,Cout,W,H,D].

[0058] This invention replaces the double convolutional layer of the encoder in Attention U-Net with a three-axis attention convolution module, which can effectively improve segmentation performance while only slightly increasing the computational cost.

[0059] The overall structure of the segmentation model in this embodiment is as follows: Figure 4As shown, the algorithm includes an encoder and a decoder. The double convolutional layer in each encoder is replaced with the aforementioned three-axis attention convolutional module, followed by a MaxPool3d layer for downsampling. The decoder directly uses the original structure of Attention U-Net. This embodiment has four encoders and four decoders. The deepest layer uses a three-axis attention convolutional module (without changing the number of channels) as a bottleneck layer connecting the encoder and the corresponding decoder. The features output by the last decoder are passed through a 3D convolutional layer to obtain the final segmentation result.

[0060] Figure 5 The image shows a comparison between the automatic segmentation results of the segmentation module and the original annotations. It can be seen that the segmentation module has segmented the main part of the nerve root quite well, providing a good foundation for subsequent tracking and visualization.

[0061] 2. Tracking Module

[0062] (1) DTI fitting

[0063] DTI data were obtained by fitting DWI data using the weighted least squares method.

[0064] The specific implementation steps of this embodiment are as follows (all functions are implemented in the Dipy library):

[0065] First, the bval and bvec data, i.e., the b value and b vector information, are loaded through the read_bvals_bvecs function;

[0066] Next, the gradient table is generated based on the bval and bvec data using the gradient_table function;

[0067] Then initialize the TensorModel using the gradient table;

[0068] Finally, the weighted least squares fitting is achieved using the fit function of TensorModel.

[0069] (2) Fiber bundle tracking

[0070] After filtering for directional consistency in the automatic segmentation results output by the segmentation model, uniformly sampled seed points are used as the starting point for the tracking algorithm. The STT algorithm is used to gradually track and obtain complete fiber bundles, and consistent tracking results that conform to physiological structure are obtained through constraint conditions. No manual intervention is required throughout the process, ensuring the robustness of the results.

[0071] This embodiment first filters valid points from the segmentation results based on directional consistency. The specific steps are as follows:

[0072] Step 1: Use the label function in the skimage.measure library to label all connected regions in the segmentation result, and then use the regionprops function to calculate the size of each connected region. Only retain regions with a volume greater than a preset threshold to remove minor noise.

[0073] Step 2: Convert the segmentation result after removing minor noise into a Boolean array to determine the initial region for seed point generation. The set of 3D voxel coordinates of the initial region is represented as:

[0074] {p=(p x ,p y ,p z )|segmentation(p x ,p y ,p z )=1}

[0075] p x ,p y ,p z These are the coordinates along the x, y, and z axes, respectively; Segmentation represents the segmentation result array, where points equal to 1 represent those belonging to neurons, and points equal to zero otherwise.

[0076] Step 3: For all voxels within each retained connected region, calculate the principal eigenvector direction based on the DTI data and then obtain the average value.

[0077] Step 4: For each candidate seed point p within the region, if the principal feature vector direction v of its voxel... p Its connected domain If the included angle is greater than a preset threshold (45° in actual use), the seed point is discarded to ensure that the tracking starting direction conforms to the overall diffusion characteristics of the region. The filtering criteria are:

[0078]

[0079] Since the fitted direction is neither positive nor negative, it is necessary to unify the z-component to be positive before calculating the included angle.

[0080] After screening, the retained seed points are considered effective voxels. Then, adjustable density uniform sampling is used as the starting point for the tracking algorithm. The specific steps are as follows:

[0081] Step 5, set the sampling density for each of the x, y, and z axes as follows:

[0082] d=(d x ,d y ,d z )

[0083] Where, d x,d y ,d z is a positive integer representing the number of samples along each axis within each voxel.

[0084] Step 6: Construct a locally uniformly distributed sampling grid. The coordinates (relative to the origin of the voxel) of each sub-sampling point within the grid are:

[0085]

[0086] Where i = 0, 1, ..., d x -1; j = 0, 1, ..., d y -1; k = 0, 1, ..., d z -1.

[0087] Step 7: For each effective voxel p, sum it with all local offsets within the sampling grid to obtain the actual seed point coordinates:

[0088] s = p + g

[0089] The above steps allow for flexible adjustment of the number of seed points based on the set density while ensuring uniform sampling. This facilitates the subsequent fiber tracking process and eliminates the need for manual setting of the starting point, thus ensuring the robustness of the tracking results.

[0090] DTI data uses a symmetric positive definite second-order tensor D to quantitatively represent the diffusion characteristics of each voxel, specifically in the form of:

[0091]

[0092] By performing eigenvalue decomposition on the fitted diffusion tensor, the main diffusion direction and diffusion intensity of water molecules in each voxel can be obtained. Furthermore, a series of quantitative indicators representing diffusion characteristics can be calculated based on eigenvalues ​​and eigenvectors. Among them, the fractional anisotropy (FA) is used to measure the directional consistency of diffusion, and its calculation formula is shown below:

[0093]

[0094] Where λ1, λ2, and λ3 are the eigenvalues ​​arranged from largest to smallest.

[0095] The complete fiber bundle is obtained by tracing step by step from the starting point based on the STT algorithm. The specific steps are as follows:

[0096] Step 8: For each seed point, use it as the starting point for tracking. Calculate the principal feature vector direction w based on the DTI data of that point. This direction is in the physical coordinate system and needs to be transformed back to the voxel coordinate system using an affine matrix. The specific formula is as follows:

[0097] v = R -1 ·w

[0098] Where R is the linear part of the affine matrix corresponding to the DTI data (in this embodiment, it is a 3×3 rotation and scaling matrix, obtained by reading it using the read function of the nrrd library), R -1 It is its inverse matrix.

[0099] Step 9: Step a fixed length along the transformed principal eigenvector direction to reach the next point of the fiber bundle. Determine the next step direction based on the DTI data at that point. Continuously execute the above steps to track and obtain a complete and continuous fiber bundle. The specific update formula is shown below:

[0100]

[0101] Where x t ∈R 3 R represents the spatial coordinates of the t-th point in the trajectory. 3 Represents a three-dimensional real vector, v t It is the main diffusion direction vector calculated based on the DTI data at the current position. Δs is the preset step length, which is usually smaller than the voxel spacing to ensure the continuity and accuracy of the trajectory.

[0102] In practice, using As the stepping direction to obtain smoother fibers, v t-1 This refers to the step direction from the previous step. Since the step length is less than the voxel spacing, trilinear interpolation is needed to obtain the DTI value at the current position to ensure the continuity and accuracy of tracking.

[0103] During trajectory generation in step 9, if the angle of directional change between two steps exceeds a preset threshold θ... max Or the current voxel FA value (also obtained through trilinear interpolation) is lower than the threshold τ. FA If the path is not found to be valid, the tracking result will be terminated immediately. The judgment formula is as follows:

[0104]

[0105] FA(x t )<τ FA

[0106] After tracking each independent region is completed, further path post-processing is performed to remove fibers that are too short, thereby enhancing the accuracy and clinical applicability of the results. Simultaneously, the tracking results are screened again for directional consistency; the specific steps are as follows:

[0107] Step 10: Fit the main extension direction vector for each fiber bundle using principal component analysis (PCA), calculate the average value of the main extension direction vectors of all fiber bundles, and remove the entire fiber if the angle between the direction and the mean is greater than a set threshold.

[0108] For each fiber bundle, five consecutive points are taken as a segment. The main extension direction vector is also fitted to the segment. The angle between the main extension direction vector and the main extension direction vector of the current fiber bundle is calculated for each segment. If it is less than a preset threshold, it is retained. Once a segment exceeds the threshold, the tracking is terminated.

[0109] 3. Visualization and Fusion Module

[0110] (1) Fiber bundle visualization

[0111] After the tracked fiber bundles are converted to the world coordinate system, they are visualized in three dimensions. Each voxel is colored according to its corresponding quantitative index to intuitively show the compression of the nerve root.

[0112] In this embodiment, the fiber bundles are transformed to the world coordinate system using an affine matrix, and then three-dimensional rendering of the fiber bundles is achieved using the FURY visualization library. The tracked fiber bundles are displayed in three-dimensional space as curves, which can intuitively reflect the direction and spatial distribution of nerve fibers.

[0113] To enhance the clinical significance of visualization, each voxel point is colored in conjunction with its corresponding quantitative indicators, such as the anisotropy index FA, to visually represent the compression of the nerve root through the depth of color.

[0114] This module not only improves the interpretability of fiber bundle results, making them easier for doctors to understand and analyze, but also supports joint display with other medical images, assisting in clinical diagnosis and treatment planning.

[0115] Figure 6 A visualization of the traced fiber bundles is presented, showing that the fiber range is well covered and the course is continuous, which has clinical applicability.

[0116] (2) Integration

[0117] Other high-resolution magnetic resonance images are converted to the world coordinate system and then overlaid and fused with fiber tract tracking results to improve the intuitiveness and accuracy of clinical diagnosis by utilizing multimodal information.

[0118] This embodiment converts other high-resolution images to a world coordinate system and renders them as a semi-transparent layer, while simultaneously presenting fiber bundles in a linear form, achieving a visual fusion effect. The effect is as follows... Figure 7 As shown.

[0119] 4. Identification and Quantitative Analysis Module

[0120] (1) Effective mask generation

[0121] For each independent connected region, fiber bundle tracing is performed, and each point on the fiber bundle is mapped back to the voxel coordinate system. The covered three-dimensional voxel region is used as an effective mask, which can more accurately reflect the actual area through which nerve fibers pass compared to the original segmentation result.

[0122] (2) Independent nerve root recognition

[0123] Based on the effective mask, a post-processing algorithm is used to identify overlapping areas between different nerve roots and multiple possible fracture areas within the same nerve root, thus separating individual nerve roots from the whole. Combining the anatomical priors of the lumbosacral plexus nerve roots, the effective mask is separated into six independent nerve root regions—L4, L5, and S1—using spatial relationships.

[0124] In this embodiment, for each effective mask, the first step is to determine whether there are multiple overlapping nerve roots using morphological methods and then separate them. The specific steps are as follows:

[0125] 1) To effectively identify connected regions layer by layer using a mask, if there are multiple independent connected regions with an area greater than a preset value (e.g., set to 5), it indicates that there is overlap.

[0126] 2) Record the slice index with multiple connected regions as the resolvable range, extract independent connected regions from the resolvable part in the effective mask to separate the non-overlapping parts of each nerve root;

[0127] 3) Fit the average principal feature vector to each of the separated independent connected regions. For each voxel in the overlapping part, assign it to the nearest connected region according to the cosine similarity to achieve the separation effect.

[0128] After completing the separation of overlapping nerve roots, determine whether there are multiple effective masks belonging to the same nerve root. The specific steps are as follows:

[0129] 1) For each effective mask, determine whether there are other effective masks that are connected to it after expansion;

[0130] 2) If two connected effective mask regions exist, further determine whether they belong to the same nerve root based on the angle between the vector connecting the centroids of the two regions and the mean of the average principal eigenvectors of the two regions. The determination formula is as follows:

[0131]

[0132] Where c1 and c2 are the centroid coordinates of the two regions, v1 and v2 are the average principal feature vectors of the two regions, and θ is a preset angle threshold, which is set to 25° in this embodiment.

[0133] If the above judgment formula is met, then the two regions are considered to belong to the same nerve root, and the two regions are merged.

[0134] After processing, the regions are divided into two groups, left and right, based on the x-coordinate of their centroids, and then labeled L4, L5, and S1 respectively based on the vertical relationship of their z-coordinates.

[0135] (3) FA diagram generation

[0136] Based on the separated independent nerve root masks, the average FA value within each mask region is calculated layer by layer along the z-axis, and a distribution curve of the FA value along the nerve root is generated accordingly. Figure 8 Line graphs are used to visually display the changing trends. This method transforms traditional voxel-level FA information into more clinically interpretable spatial distribution features, which can intuitively reflect the continuous changes in nerve fibers. It is particularly helpful in locating specific segments with abnormally low FA values, providing key quantitative evidence for the precise localization and severity assessment of nerve compression or lesions.

[0137] The analysis method provided in this embodiment first performs size unification and intensity normalization on DWI data. Then, it automatically segments the nerve root region using an improved Attention U-Net model that incorporates a triaxial attention mechanism. Using the segmentation results as a starting point, fiber tract tracking is completed through DTI fitting and the STT algorithm, achieving three-dimensional visualization of nerve root fiber tracts and fusion with other high-resolution images. Simultaneously, the tracking coverage area after outlier removal is used as the quantitative evaluation range. Post-processing algorithms are used to determine and handle overlap between different nerve roots and the breakage of the same nerve root, accurately identifying each individual nerve root and generating FA maps. This invention overcomes the problems of time-consuming manual delineation and large individual differences in current clinical applications, enabling efficient and stable visualization of lumbosacral plexus nerve root fiber tracts and providing quantitative FA analysis for each nerve root.

[0138] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A device for analyzing lumbar-sacral plexus nerve roots based on magnetic resonance diffusion tensor imaging, characterized by, The device comprises: a segmentation module, configured to crop a region to be segmented according to the nerve root distribution characteristics after pre-processing of the diffusion-weighted image, and perform nerve root segmentation through a segmentation model; a fiber bundle tracking module, configured to track nerve root fiber bundles based on the automatic segmentation result output by the segmentation module and fitting information of the diffusion-weighted image; a visualization and fusion module, configured to convert the tracking result of the fiber bundle tracking module to a world coordinate system, perform three-dimensional visualization, and perform multi-modal fusion display with other magnetic resonance image sequences; an identification and quantitative analysis module, configured to identify nerve root overlap and fracture conditions, and generate FA value distribution curves for each separated nerve root layer by layer for quantitative analysis.

2. The lumbar-sacral plexus nerve root analysis apparatus based on magnetic resonance diffusion tensor imaging according to claim 1, characterized in that, The segmentation model adopts an Attention U-Net model based on a three-axis attention convolution module, and the specific structure is as follows: the model comprises multiple layers of encoders and corresponding decoders, each encoder comprises a three-axis attention convolution module and a down-sampling module, a three-axis attention convolution module that does not change the number of channels is used as a bottleneck layer to connect the encoder and the corresponding decoder at the deepest layer, and the features output by the last decoder are subjected to a three-dimensional convolution layer to obtain the final segmentation result.

3. The lumbar-sacral plexus nerve root analysis apparatus based on magnetic resonance diffusion tensor imaging according to claim 2, characterized in that, The three-axis attention convolution module comprises an anisotropic feature extraction branch composed of three parallel convolution layers, a lightweight attention layer for generating corresponding weights for each feature, and an output part for normalizing the weights and weightedly fusing the features.

4. The analysis method of the device for analyzing the lumbosacral plexus nerve root based on magnetic resonance diffusion tensor imaging according to claim 1, wherein, The analysis method comprises the following steps: crop a region to be segmented according to the nerve root distribution characteristics after pre-processing of the diffusion-weighted image, and perform nerve root segmentation to obtain an automatic segmentation result; fit the diffusion-weighted image, and track nerve root fiber bundles using the fitting information and the automatic segmentation result; convert the tracking result to a world coordinate system, perform three-dimensional visualization, and perform multi-modal fusion display with other magnetic resonance image sequences; identify nerve root overlap and fracture conditions, separate single nerve roots, and generate FA value distribution curves for each separated nerve root layer by layer for quantitative analysis.

5. The analysis method according to claim 4, characterized in that, Tracking nerve root fiber bundles specifically comprises: using the uniformly sampled seed points as the starting point of the tracking algorithm after direction consistency screening in the automatic segmentation result, using the STT algorithm to gradually track to obtain complete fiber bundles, and using constraint conditions to obtain consistent and physiological tracking results.

6. The analysis method according to claim 5, characterized in that, Direction consistency screening in the automatic segmentation result specifically comprises: label all connected regions in the automatic segmentation result, calculate the size of each connected region, and only keep regions with a volume greater than a preset threshold; convert the automatic segmentation result after removing the micro noise into a Boolean array to determine the initial region of the seed point generation, and the three-dimensional voxel coordinate set of the initial region is represented as: {p = (p x , p y , p z ) | segmentation(p x , p y , p z ) = 1} p x ,p y ,p z x, y, z are the coordinates of each axis respectively; Segmentation represents the segmentation result array; For each connected region, the average of the principal eigenvector direction of all voxels in the region is calculated For each candidate seed point p in the connected region, if the angle between the principal eigenvector direction v of the voxel where p locates and the average principal eigenvector direction of the connected region is greater than a preset threshold, the seed point is removed p and the average principal eigenvector direction of the connected region the angle is greater than a preset threshold, the seed point is removed, ensuring that the tracking starting direction conforms to the overall diffusion characteristics of the region, and the screening condition is:

7. The analysis method of claim 5, wherein, using the seed points retained after the direction consistency screening as effective voxels; setting the sampling density of each x, y and z axis to constructing a locally uniformly distributed sampling grid, and the coordinates of each sub-sampling point in the grid relative to the voxel origin are: d = (d x ,d y ,d z ) where d x ,d y ,d z is a positive integer, representing the number of samples along each axis within each voxel; ​ where i = 0, 1,..., d x -1; j = 0, 1,..., d y -1; k = 0, 1,..., d z -1; For each valid voxel p, sum it with all local offsets within the sampling grid to get the actual seed point coordinate: s = p + g.

8. The analysis method of claim 5, wherein, After the tracking of each independent region, further post-processing of the path is performed to remove fibers with too short length, and the tracking results are screened again for consistency in direction. The specific steps are as follows: For each fiber bundle, a main extension direction vector is fitted by principal component analysis method, and the average value of the main extension direction vectors of all fiber bundles is calculated. The whole fiber is removed if the angle between the direction and the average value is greater than a set threshold; For each fiber bundle, five consecutive points are taken as a segment, and a main extension direction vector is also fitted for the segment. The angle between the main extension direction vector and the main extension direction vector of the current fiber bundle is calculated segment by segment. If the angle is less than a preset threshold, the segment is retained. If a segment exceeds the threshold, the tracking is terminated.

9. The analysis method of claim 4, wherein, Identifying nerve root overlap and fracture, separating single nerve root specifically includes: For each independent connected region in the automatic segmentation result, fiber bundle tracking is performed, and each point on the fiber bundle is mapped back to the voxel coordinate system. The three-dimensional voxel region covered is taken as an effective mask; For the effective mask, identify the connected regions layer by layer. If there are multiple independent connected regions with an area greater than a preset value, it indicates that there is an overlap; Record the slice index of multiple connected regions as a distinguishable range. Extract the independent connected domains in the distinguishable part of the effective mask to separate the non-overlapping parts of each nerve root; Fit an average principal feature vector for each independent connected domain. For each voxel in the overlapping part, assign it to the closest connected domain according to the cosine similarity.

10. The analysis method according to claim 9, characterized in that, After separating single nerve roots, determine whether there are multiple effective masks belonging to the same nerve root, specifically including: For each effective mask, determine whether there is another effective mask connected after inflation; If there are two connected effective mask regions, further determine whether they belong to the same nerve root according to the angle between the vector of the center of mass of the two regions and the average of the average principal feature vectors of the two regions. The judgment formula is: Where c1 and c2 are the center coordinates of the two regions, v1 and v2 are the average principal feature vectors of the two regions, and θ is a preset angle threshold; If the above judgment formula is satisfied, it is considered that the two regions belong to the same nerve root, and the two regions are merged.

Images