Deep learning based brain high-resolution vessel wall image reconstruction system

Abstract

The application discloses a brain high-resolution blood vessel wall image reconstruction system based on deep learning, which comprises an image acquisition module, a frequency band decomposition module, a core structure reorganization module, a frequency band detail enhancement module, a frequency band synthesis optimization module and a brain blood vessel wall image reconstruction module. The application belongs to the field of image processing and specifically relates to a brain high-resolution blood vessel wall image reconstruction system based on deep learning. The scheme is characterized in that: the anatomical feature constraint filtering processing is adopted, the gray self-adaptive base function is designed, the effective interval is dynamically adjusted based on the image gray histogram, and the reliability of subsequent brain blood vessel wall image reconstruction is improved; the gradient direction consistency constraint is introduced to identify the core points, and the background interference of the non-core of the blood vessel wall image is removed; the gray self-adaptive second-order derivative weighting is adopted to realize the detail enhancement and directional noise suppression of the blood vessel wall image; the interpolation completion is performed on the missing microstructure elements, the guide factor is introduced to adjust the weight, and the stability of the brain blood vessel wall image reconstruction is improved.

Description

TECHNICAL FIELD

[0001] The application relates to the field of image processing, in particular to a brain high-resolution blood vessel wall image reconstruction system based on deep learning. BACKGROUND

[0002] The brain blood vessel wall image reconstruction system is a professional technical system for generating a high-resolution image clearly presenting the layered structure of the brain blood vessel wall and the characteristics of plaques based on MRI medical image original data through noise suppression, core structure extraction and other technologies. However, the general brain blood vessel wall image reconstruction system has the problems that noise suppression ignores the anatomical features of blood vessels, fine structures are prone to be blurred, and plaque heterogeneity is difficult to capture, thereby leading to poor reconstruction reliability; the core structure and background interference are difficult to separate, high-frequency detail enhancement and noise suppression are unbalanced, thereby leading to unstable image reconstruction accuracy. SUMMARY

[0003] In view of the above problems, the brain high-resolution blood vessel wall image reconstruction system based on deep learning is provided to overcome the defects of the prior art. The anatomical feature constraint filtering process is used to accurately suppress noise. The gray scale adaptive base function is designed, and the effective interval is dynamically adjusted based on the image gray scale histogram to dynamically match the gray scale distribution of the image, so that the gray scale characteristics of low gray scale lipid plaques, high gray scale calcified plaques and normal blood vessel walls can be more accurately decomposed. The subsequent brain blood vessel wall image reconstruction reliability is improved. The gradient direction consistency constraint is introduced to identify the core points, so that the continuity of the layered structure can be captured, and the background interference of the blood vessel wall image is removed. The gray scale adaptive second-order derivative weighting is used to realize the detail enhancement and directional noise suppression of the blood vessel wall image. The missing fine structure elements are interpolated and completed to restore the small branches and plaque textures. The guide factor is introduced to adjust the weight to adapt to the complex structure reconstruction requirement. The stability of the brain blood vessel wall image reconstruction is improved.

[0004] The technical scheme adopted by the application is as follows: The brain high-resolution blood vessel wall image reconstruction system based on deep learning provided by the application comprises an image acquisition module, a frequency band decomposition module, a core structure reorganization module, a frequency band detail enhancement module, a frequency band synthesis optimization module and a brain blood vessel wall image reconstruction module.

[0005] The image acquisition module acquires historical brain blood vessel wall images.

[0006] The frequency band decomposition module filters and denoises the historical cerebral vessel wall image by combining the blood vessel anatomical features, and defines a gray level adaptive basis function to realize frequency band decomposition.

[0007] The core structure reorganization module decomposes the low-frequency frequency band matrix, identifies the blood vessel wall core points, and reorganizes the frequency band by combining the gradient constraint.

[0008] The frequency band detail enhancement module weights the original high-frequency coefficients by the gray level adaptive basis function, and performs frequency band detail enhancement.

[0009] The frequency band synthesis optimization module interpolates and supplements the missing elements, and synthesizes the reorganized image.

[0010] The cerebral vessel wall image reconstruction module reconstructs the cerebral vessel wall image.

[0011] Further, the image acquisition module acquires historical cerebral vessel wall images, and performs gray level normalization processing, image labeling, and target vessel wall region labeling to obtain a historical cerebral vessel wall image set.

[0012] Further, the frequency band decomposition module specifically includes:

[0013] Noise suppression; suppress noise by weighted average of similar neighborhood pixels, and introduce blood vessel wall anatomical structure features to dynamically adjust neighborhood weights; train a lightweight U-Net model, input the noise-suppressed historical cerebral vessel wall image, output the binary mask of the target blood vessel wall, and only keep the pixels within the mask to obtain the extracted and retained cerebral vessel wall image ;

[0014] Customized decomposition; design a gray level adaptive basis function, dynamically adjust the effective interval based on the image gray level histogram; based on the 3rd derivative of the gray level adaptive basis function, decompose into low-frequency band coefficients and high-frequency band coefficients.

[0015] Further, the core structure reorganization module first singular value decomposes the low-frequency frequency band matrix; identifies the blood vessel wall core points, defines the blood vessel wall core points, and introduces the gradient direction consistency constraint; and reorganizes the frequency band from the identified sparse structure vector.

[0016] Further, the frequency band detail enhancement module weights the original high-frequency coefficients by the 2nd derivative of the gray level adaptive basis function.

[0017] Further, the frequency band synthesis optimization module interpolates the missing elements of the identified sparse structure vector; synthesizes the low-frequency coefficients of the sparse structure reorganization and the enhanced high-frequency coefficients to obtain the final reorganized image; evaluates the reorganization effect, adjusts the parameters by using the particle swarm algorithm, and dynamically adjusts the blood vessel curvature and plaque area as the guide factor.

[0018] Further, the cerebral vascular wall image reconstruction module acquires the cerebral vascular wall image in real time, after gray scale normalization processing, sequentially input to the frequency band decomposition module, the core structure reorganization module, the frequency band detail enhancement module and the frequency band synthesis optimization module, and the final reorganization image obtained by the frequency band synthesis optimization module is taken as the cerebral vascular wall image reconstruction result.

[0019] The beneficial effects achieved by the above-mentioned scheme are as follows:

[0020] (1) For the problem that the general cerebral vascular wall image reconstruction system ignores the vascular anatomical features in noise suppression, the fine structure is easy to blur, and it is difficult to capture the plaque heterogeneity, thereby leading to poor reliability of the reconstruction, the scheme realizes accurate noise suppression through anatomical feature constraint filtering processing; the gray scale adaptive basis function is designed, and the effective interval is dynamically adjusted based on the image gray scale histogram, which dynamically matches the gray scale distribution of the image, and can more accurately decompose the low gray scale lipid plaque, high gray scale calcified plaque and gray scale features of normal vascular wall; thereby improving the reliability of subsequent cerebral vascular wall image reconstruction.

[0021] (2) For the problem that the general cerebral vascular wall image reconstruction system has difficulty in separating the core structure and background interference, the high-frequency detail enhancement and noise suppression are unbalanced, thereby leading to unstable image reconstruction accuracy, the scheme introduces gradient direction consistency constraint for core point discrimination, ensures the continuity of capturing layered structure, and eliminates the background interference of the vascular wall image; based on the gray scale adaptive second derivative weighting, the detail enhancement and noise directional suppression of the vascular wall image are realized; the missing fine structure elements are interpolated and completed, and the small branches and plaque texture are restored; the guide factor is introduced to adjust the weight, and the complex structure reconstruction demand is adapted; thereby improving the stability of the cerebral vascular wall image reconstruction. BRIEF DESCRIPTION OF DRAWINGS

[0022] Figure 1 The module schematic diagram of the cerebral high-resolution vascular wall image reconstruction system based on deep learning provided by the present application is provided.

[0023] The accompanying drawings are used to provide a further understanding of the present application, and constitute a part of the specification, together with the embodiments of the present application, to explain the present application, and do not constitute a limitation on the present application. DETAILED DESCRIPTION

[0024] The technical solutions in the embodiments of the present application will be described clearly and completely below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative labor are within the scope of protection of the present application.

[0025] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inner", "outer" and the like indicate the orientation or positional relationship based on the orientation or positional relationship shown in the drawings, and are only for the convenience of describing the present application and simplifying the description, and do not indicate or imply that the devices or elements referred to must have a particular orientation, be constructed and operated in a particular orientation, and therefore cannot be understood as a limitation on the present application.

[0026] Embodiment one, refer to Figure 1 The present application provides a brain high-resolution blood vessel wall image reconstruction system based on deep learning, which comprises an image acquisition module, a frequency band decomposition module, a core structure reorganization module, a frequency band detail enhancement module, a frequency band synthesis optimization module and a brain blood vessel wall image reconstruction module.

[0027] The image acquisition module acquires historical brain blood vessel wall images; and sends data to the frequency band decomposition module.

[0028] The frequency band decomposition module filters and denoises the historical brain blood vessel wall images by combining the blood vessel anatomical features, defines a gray scale adaptive basis function, realizes frequency band decomposition, and sends data to the core structure reorganization module.

[0029] The core structure reorganization module decomposes the low-frequency frequency band matrix, identifies the blood vessel wall core points and reorganizes the frequency band in combination with the gradient constraint; and sends data to the frequency band detail enhancement module.

[0030] The frequency band detail enhancement module weights the original high-frequency coefficients by the gray scale adaptive basis function, performs frequency band detail enhancement, and sends data to the frequency band synthesis optimization module.

[0031] The frequency band synthesis optimization module inserts and supplements missing elements, synthesizes and reorganizes the image of the frequency band, and sends data to the brain blood vessel wall image reconstruction module.

[0032] The brain blood vessel wall image reconstruction module reconstructs the brain blood vessel wall image.

[0033] Embodiment two, refer to Figure 1 Based on the above embodiment, the image acquisition module acquires historical brain blood vessel wall images, acquires high-resolution brain blood vessel wall images of the target blood vessels in the brain by 3.0T and above MRI equipment, covers the complete running area of the blood vessel wall, and the spatial resolution is not less than 0.5mmx0.5mmx1.0mm; and performs gray scale normalization processing, adopts adaptive z-score normalization; performs image labeling, labels the target blood vessel wall area, and labels the blood vessel wall boundary, lumen and tube wall boundary; obtains a set of historical brain blood vessel wall images; and the image labeling is only used for verifying the performance of the module.

[0034] Embodiment three, see Figure 1 , which is based on the above embodiment, the frequency band decomposition module specifically comprises:

[0035] Noise suppression; make use of the characteristics of strong spatial correlation and small gray fluctuation of blood vessel wall pixels in historical brain blood vessel wall images, suppress noise by weighted average of similar neighborhood pixels, and introduce blood vessel wall anatomical structure features to dynamically adjust neighborhood weight and enhance the specificity of noise suppression in complex structure area, expressed as: ; ; wherein, is the curvature of the blood vessel, the tubular structure of the image is first enhanced by Frangi filtering, and then the blood vessel center axis is extracted by morphological thinning; the local segment of the axis is fitted by B-spline curve, and the curvature is calculated according to the first and second derivatives of the curve; is the thickness of the blood vessel wall, based on the denoised image, the intima and adventitia are located by Canny edge detection combined with the blood vessel mask, the distance between the intima and adventitia boundary on the same radial line is calculated along the direction perpendicular to the blood vessel axis, and the wall thickness at the corresponding position is obtained; is the filter denoising; x is a pixel point, y is any pixel in the neighborhood of x h is the adjustment parameter, the value range is [5, 40]; is the normalized constant, ; is the gray value; is the weight coefficient of the anatomical structure constraint term, the value range is [0.3, 0.8]; is the similarity distance of blood vessel anatomical structure; is the weight coefficient of the wall thickness difference in the anatomical distance, the value range is [0.6, 1.2]; train the lightweight U-Net model, input the historical brain blood vessel wall image after noise suppression, and output the binary mask M of the target blood vessel wall, only keep the pixels in the mask, expressed as: ; is the brain blood vessel wall image reserved after extraction;

[0036] Customized decomposition; for the tubular layered structure of the blood vessel wall, a gray self-adaptive basis function is designed to ensure that the wavelet matches the gray distribution of the blood vessel wall, the effective interval is dynamically adjusted based on the image gray histogram, and the decomposition ability for heterogeneous plaques is enhanced, expressed as: ; wherein, is the gray self-adaptive basis function; k is the gray value of the image pixel; is the 10% quantile of the gray histogram, corresponding to the low gray lipid plaque; is the 90% quantile of the gray histogram, corresponding to the high gray calcified plaque; is the interval midpoint; based on the 3rd derivative of the gray self-adaptive basis function, Decomposed into low-frequency coefficients and high-frequency coefficients (low-frequency class 1, high-frequency class 3), represented as: ; ; ;in, It is a combination function of the third derivative of the gray-level adaptive basis functions, used to capture drastic gray-level changes in blood vessel wall images; It is the first derivative term; It is a combination of third-order derivative terms, capturing drastic changes in grayscale.

[0037] By performing the above operations, this solution addresses the problems of conventional cerebral vascular wall image reconstruction systems, such as noise suppression ignoring vascular anatomical features, blurring of fine structures, difficulty in capturing plaque heterogeneity, and consequently poor reconstruction reliability. It achieves precise noise suppression through anatomical feature-constrained filtering; designs an adaptive gray-level basis function and dynamically adjusts the effective range based on the image's gray-level histogram to dynamically match the image's gray-level distribution, enabling more accurate decomposition of low-gray-level lipid plaques, high-gray-level calcified plaques, and normal vascular wall gray-level features; thereby improving the reliability of subsequent cerebral vascular wall image reconstruction.

[0038] Example 4, see Figure 1 This embodiment is based on the above embodiment. The core structure reconstruction module is a frequency band that contains a vascular core structure but still retains brain tissue background. Therefore, singular value decomposition is first performed on the low-frequency band matrix, which is represented as: ;in, It is a low-frequency band matrix; It is a left-valued vector matrix, corresponding to the horizontal features of the blood vessel wall; It is a diagonal singular value matrix; It is a right heterovalued vector matrix, corresponding to the vertical features of the blood vessel wall; the subscripts H, W, and r are the row and column numbers, and T is the transpose; for the identification of the core points of the blood vessel wall, the core points of the blood vessel wall are defined as gray-level abrupt change points or start and end points in the heterovalued vector, corresponding to the plaque edge and the blood vessel endpoint, respectively, and gradient direction consistency constraints are introduced to accurately capture the continuous features of the layered structure. The identification formula is expressed as: ; ; by the identified sparse vectors , The recombined frequency band is represented as: ;in, It is the core output after screening; It is the value of the outlier vector at position i; It is the value at position i-1; It is the gradient direction angle; It is the reference gradient direction; is the allowable deviation range of the gradient direction; m is the vector endpoint.

[0039] Example 5, see Figure 1 This embodiment is based on the above embodiment. The frequency band detail enhancement module contains high-frequency details such as patch edges and the intima-media boundary, but these details are easily affected by MRI noise. Therefore, the second derivative of the gray-level adaptive basis function is used to weight the original high-frequency coefficients to achieve detail enhancement + noise suppression, as shown below: ; ;in, It is the high-frequency coefficient that is enhanced and suppresses noise; It is the second derivative of the gray-scale adaptive basis function; It is the original high-frequency coefficient; to avoid noise amplification and ensure that the enhanced details are all valid information.

[0040] Example 6, see Figure 1 This embodiment is based on the above embodiment. The frequency band synthesis optimization module is based on the fact that the outlier vector of the blood vessel wall has a locally linear distribution, therefore... , The missing elements are interpolated, and the missing elements correspond to the fine structures of the blood vessel wall, represented as follows: ; low-frequency coefficients of sparse structure recombination With enhanced high-frequency coefficients The final reconstructed image is obtained by synthesis. , represented as: ;in, These are the missing element values ​​obtained after interpolation, corresponding to the positions in the outlier vector of the blood vessel wall that are identified as non-core points but belong to fine structures; and These are the core points on either side of the missing element; It is the core point corresponding to the outlier vector value; It is an inverse 3D binary tree decomposition; the reconstruction effect is evaluated using vessel wall thickness error and plaque edge matching degree, which are respectively the difference in vessel wall thickness between the reconstructed image and the pathological gold standard, and the overlap rate of plaque edges between the reconstructed image and the gold standard; an error threshold and a matching degree threshold are set. If the vessel wall thickness error after reconstruction is higher than the error threshold or the plaque edge matching degree is lower than the matching degree threshold, the parameters are adjusted and reconstructed. The particle swarm optimization algorithm is used to adjust the parameters. For the inertial weight of the particle swarm, the vessel curvature and plaque area S are used as guiding factors for dynamic adjustment. When the curvature is large and the plaque area is large, the weight is reduced to enhance the local search capability, which is expressed as: ;in, These are the inertial weights used in the particle swarm optimization algorithm; and These are the maximum and minimum weights, respectively; q is the vascular curvature, and S is the plaque area, which needs to be normalized.

[0041] By performing the above operations, this solution addresses the problems of difficulty in separating core structures from background interference and imbalance between high-frequency detail enhancement and noise suppression in general cerebral blood vessel wall image reconstruction systems, which leads to unstable image reconstruction accuracy. This is achieved by introducing gradient direction consistency constraints to identify core points, ensuring the continuity of captured layered structures and eliminating non-core background interference in the blood vessel wall image; using gray-level adaptive second derivative weighting to achieve detail enhancement and targeted noise suppression in the blood vessel wall image; interpolating and completing missing fine structural elements to restore fine branches and plaque textures; and introducing a guiding factor to adjust weights to adapt to the needs of complex structure reconstruction, thereby improving the stability of cerebral blood vessel wall image reconstruction.

[0042] Example 7, see Figure 1 This embodiment is based on the above embodiment. The cerebral blood vessel wall image reconstruction module acquires cerebral blood vessel wall images in real time. After grayscale normalization processing, the images are sequentially input to the frequency band decomposition module, the core structure reconstruction module, the frequency band detail enhancement module, and the frequency band synthesis optimization module. The final reconstructed image obtained by the frequency band synthesis optimization module is used as the cerebral blood vessel wall image reconstruction result.

[0043] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0044] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. A deep learning-based high-resolution brain blood vessel wall image reconstruction system, characterized in that: The system includes an image acquisition module, a frequency band decomposition module, a core structure reconstruction module, a frequency band detail enhancement module, a frequency band synthesis optimization module, and a cerebral blood vessel wall image reconstruction module. The image acquisition module acquires historical images of cerebral blood vessel walls; The frequency band decomposition module performs frequency band decomposition on historical cerebral blood vessel wall images by combining vascular anatomical features with filtering and denoising, and by defining a gray-level adaptive basis function. The core structure reconstruction module decomposes the low-frequency band matrix, identifies the core points of the blood vessel wall, and reconstructs the frequency bands in combination with gradient constraints. The frequency band detail enhancement module enhances frequency band details by weighting the original high-frequency coefficients using a gray-scale adaptive basis function. The frequency band synthesis and optimization module interpolates and supplements missing elements, and synthesizes and reconstructs images from the frequency bands; The cerebral blood vessel wall image reconstruction module reconstructs images of the cerebral blood vessel wall; The frequency band decomposition module specifically includes: Noise suppression: Noise is suppressed by weighted averaging of similar neighboring pixels, and anatomical features of the blood vessel wall are incorporated to dynamically adjust the neighborhood weights; a lightweight U-Net model is trained, inputting a noise-suppressed historical image of the cerebral blood vessel wall, and outputting a binary mask of the target blood vessel wall, retaining only the pixels within the mask to obtain the extracted and preserved image of the cerebral blood vessel wall. ; Customized decomposition; design of gray-level adaptive basis functions to ensure that the wavelet matches the gray-level distribution of the blood vessel wall, expressed as: ;in, It is a gray-level adaptive basis function; k is the gray value of the image pixel; It is the 10th percentile of the grayscale histogram, corresponding to low grayscale lipid plaques; It is the 90th percentile of the grayscale histogram, corresponding to high grayscale calcified patches; It is the midpoint of the interval; based on the third derivative of the gray-scale adaptive basis function, Decomposed into low-frequency coefficients and high-frequency coefficients, expressed as: ; ; ;in, It is a combination function of the third derivative of the gray-level adaptive basis functions, used to capture drastic gray-level changes in blood vessel wall images; It is the first derivative term; It is a combination of third-order derivative terms, capturing drastic changes in grayscale.

2. The brain high-resolution blood vessel wall image reconstruction system based on deep learning according to claim 1, characterized in that: The core structure reorganization module first performs singular value decomposition on the low-frequency band matrix; identifies the core points of the blood vessel wall, defines the core points of the blood vessel wall, and introduces gradient direction consistency constraints; and reorganizes the frequency bands from the identified sparse vectors.

3. The brain high-resolution blood vessel wall image reconstruction system based on deep learning according to claim 2, characterized in that: The frequency band detail enhancement module uses the second derivative of the gray-scale adaptive basis function to weight the original high-frequency coefficients.

4. The brain high-resolution blood vessel wall image reconstruction system based on deep learning according to claim 3, characterized in that: The frequency band synthesis optimization module interpolates the missing elements of the identified sparse vector; it synthesizes the low-frequency coefficients of the sparse reconstruction with the enhanced high-frequency coefficients to obtain the final reconstructed image; it evaluates the reconstruction effect and uses a particle swarm optimization algorithm to adjust the parameters. For the inertial weight of the particle swarm, the vascular curvature and plaque area are used as guiding factors for dynamic adjustment.

5. The deep learning-based high-resolution brain blood vessel wall image reconstruction system according to claim 4, characterized in that: The image acquisition module acquires historical images of cerebral blood vessel walls; performs grayscale normalization processing, performs image annotation, and annotates the target blood vessel wall region to obtain a set of historical cerebral blood vessel wall images.

6. The brain high-resolution blood vessel wall image reconstruction system based on deep learning according to claim 5, characterized in that: The cerebral blood vessel wall image reconstruction module acquires images of the cerebral blood vessel wall in real time. After grayscale normalization, the images are sequentially input into the frequency band decomposition module, the core structure reconstruction module, the frequency band detail enhancement module, and the frequency band synthesis optimization module. The final reconstructed image obtained by the frequency band synthesis optimization module is used as the cerebral blood vessel wall image reconstruction result.

Images