A method and device for detecting thrombosis of an arteriovenous fistula

Abstract

The application discloses an arteriovenous fistula thrombus detection method and device, and relates to the technical field of image processing. The method comprises the following steps: acquiring a forearm blood vessel image to be detected; pre-processing the forearm blood vessel image to be detected; performing image enhancement on the pre-processed forearm blood vessel image based on a two-stage image enhancement method; obtaining a thrombus segmentation result of the forearm blood vessel image to be detected according to an adaptive prior shape level set evolution (APSLSE) model and the image-enhanced forearm blood vessel image; and finally performing thrombus type detection by using a machine learning or deep neural network method. The application is aimed at image analysis of shallow blood vessels in the forearm during dialysis, realizes early warning and intervention on thrombus formation of an internal fistula, and has the characteristics of intuitive data acquisition, high calculation precision and good clinical application effect.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a method and device for detecting arteriovenous fistula thrombosis. Background Technology

[0002] Currently, arteriovenous fistula thrombosis is highly likely to occur in the forearm of dialysis patients during dialysis treatment. Generally, dialysis patients need to go to the hospital regularly to have their blood drawn from their forearm using a specific device, purified, and then reinfused into their bodies. Dialysis utilizes the principle of a semi-permeable membrane to remove various harmful and excess metabolic waste products and excess electrolytes from the body through diffusion, thereby purifying the blood and correcting electrolyte and acid-base imbalances. Because of the long-term practice of drawing blood from the forearm through needles, blood vessels are physically damaged, and blood clots easily form near the puncture site, eventually leading to thrombosis. Thrombosis can easily cause vascular blockage and impaired blood flow. For patients on long-term dialysis, thrombosis also increases the difficulty and risk of dialysis. Therefore, how to detect thrombosis before it forms is currently a hot and challenging research topic in the field of hemodialysis.

[0003] In recent years, researchers have proposed their own solutions from different research perspectives regarding the detection of arteriovenous fistula thrombosis.

[0004] Traditional research on the detection and intervention of arteriovenous fistula thrombosis mainly suffers from the following problems: First, the relevant research methods are relatively complex, and the application results are not direct. For example, the Chinese patent "Method for Constructing a Predictive Model for Autologous Arteriovenous Fistula Thrombosis (CN202110440874.9)" uses a balanced iterative random forest algorithm to identify high-risk factors leading to thrombosis in mature arteriovenous fistulas and presents the established model in the form of a visualized nomogram. However, this method is difficult to use in practice. Second, although these methods can detect and intervene in thrombosis to some extent, the detection and intervention process is not intuitive, and its guidance for doctors' diagnosis and treatment is limited. For example, the Chinese patent "An Arteriovenous Fistula Thrombosis Early Warning Device and Its Control Method (CN201811424532.2)" obtains fistula thrill signals and analyzes these signals to estimate and warn of arteriovenous fistula thrombosis. The visualization effect of the data collected in the above patents is limited. Third, the reliability and accuracy of the measurement data obtained by these methods are limited. For example, the Chinese patent "Real-time Detection System, Real-time Detection Device and Blood Flow Velocity Detection Method for Arteriovenous Fistula Thrombosis (CN201710595825.6)" utilizes different placement methods and two different working modes of the ultrasonic probe to effectively eliminate the difficulty in determining the Doppler angle in practical applications of blood flow velocity measurement, solve the problem of flow velocity gradient ambiguity, and thus achieve real-time monitoring of arteriovenous fistula thrombosis and blood flow velocity. However, the above method has many sources of error during measurement, resulting in relatively limited measurement accuracy.

[0005] Therefore, how to better detect arteriovenous fistula thrombosis is a problem that needs to be solved by existing technologies. Summary of the Invention

[0006] This invention addresses the problems of existing methods for detecting arteriovenous fistula thrombosis, which are complex to use, have indirect application effects, and have limited reliability and accuracy of measurement data.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] On one hand, the present invention provides a method for detecting arteriovenous fistula thrombosis, which is implemented by an electronic device and includes:

[0009] S1. Obtain the image of the forearm blood vessels to be detected.

[0010] S2. Preprocess the forearm blood vessel image to be detected.

[0011] S3. Based on a two-stage image enhancement method, image enhancement is performed on the preprocessed forearm blood vessel image.

[0012] S4. Based on the adaptive prior shape level set evolution APSLSE model and the image-enhanced forearm blood vessel image, the thrombus type detection result of the forearm blood vessel image to be detected is obtained.

[0013] Optionally, acquiring the forearm blood vessel image to be detected in S1 includes:

[0014] Near-infrared images of superficial blood vessels in the forearm to be tested are acquired using a near-infrared image acquisition device.

[0015] Optionally, the preprocessing of the forearm blood vessel image to be detected in S2 includes:

[0016] S21. Remove the background from the forearm blood vessel image to be detected.

[0017] S22. Perform contrast stretching on the forearm blood vessel image after background removal.

[0018] S23. Perform noise removal on the forearm blood vessel image after contrast stretching.

[0019] Optionally, the two-stage image enhancement method in S3 performs image enhancement on the preprocessed forearm blood vessel image, including:

[0020] S31. Input the preprocessed forearm blood vessel image into the residual convolutional autoencoder (RCAE) model to obtain the first image output.

[0021] S32. Input the first image output into the contrast-limited adaptive histogram equalization (CLAHE) model to obtain the second image output.

[0022] S33. Calculate the combination ratio coefficient based on the preprocessed forearm vascular image.

[0023] S34. Combine the first image output and the second image output according to the combination ratio coefficient to obtain the image enhancement result of the preprocessed forearm blood vessel image.

[0024] Optionally, in S4, the thrombus type detection results of the forearm vessel image to be detected, obtained by evolving the APSLSE model based on the adaptive prior shape level set and the image-enhanced forearm vessel image, include:

[0025] S41. Based on the adaptive prior shape level set evolution APSLSE model, image segmentation is performed on the enhanced forearm blood vessel image to obtain the segmented region.

[0026] S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image.

[0027] S43. Extract the skeleton from the binarized forearm blood vessel image to obtain the forearm blood vessel skeleton image.

[0028] S44. Extract a local image along the location of blood vessels in the forearm vascular skeleton image.

[0029] S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected.

[0030] Alternatively, based on local images and deep learning networks, the thrombus type detection results of the forearm blood vessel image to be detected can be obtained.

[0031] Optionally, the energy function in the APSLSE model based on adaptive prior shape level set evolution in S41 is shown in equation (1) below:

[0032] E(φ)=μR p (φ)+λL g (φ)+νA g (φ)+γE Shape (1)

[0033] Where μ>0; λ>0; γ>0; ν∈R; φ is the level set function; R p (φ) is the regularization term for the level set function; L g (φ) is the length term of the level set function; A g (φ) is the area term of the level set function; E ShapeThis is the shape constraint term for the level set function.

[0034] Alternatively, the area term of the level set function is shown in equation (2) below:

[0035]

[0036] Where ν1 and ν2 are coefficients, ν1, ν2 ∈ R; Ω is the domain; x ∈ Ω is a point on the domain; g is the edge indicator function; H is the Heaviside function; ε LBF It is the energy functional of the locally binary fitted LBF.

[0037] Optionally, the shape constraint term of the level set function is shown in equation (3) below:

[0038] E shape =∫ Ω (H(Φ)-H(Ψ)) 2 dx (3)

[0039] Where Ω is the domain; x∈Ω is a point in the domain; H is the Heaviside function; Φ is the evolution curve distance function; and Ψ is the distance function of the prior shape profile curve.

[0040] Optionally, the training process of the Support Vector Machine (SVM) model in S45 includes:

[0041] S451. Obtain a local image of the sample and perform enhancement processing on the local image of the sample.

[0042] S452. Extract image features from the enhanced local image of the sample to obtain multiple features of the local image of the sample; among which, multiple features include geometric features, grayscale features, gradient features and texture features.

[0043] S453. Train the Support Vector Machine (SVM) model based on multiple features to obtain a trained SVM model.

[0044] On the other hand, the present invention provides an arteriovenous fistula thrombosis detection device, which is used to implement an arteriovenous fistula thrombosis detection method, the device comprising:

[0045] The acquisition module is used to acquire images of the forearm blood vessels to be detected.

[0046] The preprocessing module is used to preprocess the forearm blood vessel images to be detected.

[0047] The image enhancement module is used to enhance the preprocessed forearm blood vessel images based on a two-stage image enhancement method.

[0048] The output module is used to obtain the thrombus type detection result of the forearm blood vessel image to be detected based on the APSLSE model evolved from the adaptive prior shape level set and the image-enhanced forearm blood vessel image.

[0049] Optionally, the acquisition module is further used for:

[0050] Near-infrared images of superficial blood vessels in the forearm to be tested are acquired using a near-infrared image acquisition device.

[0051] Optionally, the preprocessing module is further used for:

[0052] S21. Remove the background from the forearm blood vessel image to be detected.

[0053] S22. Perform contrast stretching on the forearm blood vessel image after background removal.

[0054] S23. Perform noise removal on the forearm blood vessel image after contrast stretching.

[0055] Optionally, the image enhancement module is further used for:

[0056] S31. Input the preprocessed forearm blood vessel image into the residual convolutional autoencoder (RCAE) model to obtain the first image output.

[0057] S32. Input the first image output into the contrast-limited adaptive histogram equalization (CLAHE) model to obtain the second image output.

[0058] S33. Calculate the combination ratio coefficient based on the preprocessed forearm vascular image.

[0059] S34. Combine the first image output and the second image output according to the combination ratio coefficient to obtain the image enhancement result of the preprocessed forearm blood vessel image.

[0060] Optionally, the output module is further used for:

[0061] S41. Based on the adaptive prior shape level set evolution APSLSE model, image segmentation is performed on the enhanced forearm blood vessel image to obtain the segmented region.

[0062] S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image.

[0063] S43. Extract the skeleton from the binarized forearm blood vessel image to obtain the forearm blood vessel skeleton image.

[0064] S44. Extract a local image along the location of blood vessels in the forearm vascular skeleton image.

[0065] S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected.

[0066] Alternatively, based on local images and deep learning networks, the thrombus type detection results of the forearm blood vessel image to be detected can be obtained.

[0067] Alternatively, the energy function in the APSLSE model is derived based on the adaptive prior shape level set evolution, as shown in equation (1) below:

[0068] E(φ)=μR p (φ)+λL g (φ)+νA g (φ)+γE Shape (1)

[0069] Where μ>0; λ>0; γ>0; ν∈R; φ is the level set function; R p (φ) is the regularization term for the level set function; L g (φ) is the length term of the level set function; A g (φ) is the area term of the level set function; E Shape This is the shape constraint term for the level set function.

[0070] Alternatively, the area term of the level set function is shown in equation (2) below:

[0071]

[0072] Where ν1 and ν2 are coefficients, ν1, ν2 ∈ R; Ω is the domain; x ∈ Ω is a point on the domain; g is the edge indicator function; H is the Heaviside function; ε LBF It is the energy functional of the locally binary fitted LBF.

[0073] Optionally, the shape constraint term of the level set function is shown in equation (3) below:

[0074] E shape =∫ Ω (H(Φ)-H(Ψ)) 2 dx (3)

[0075] Where Ω is the domain; x∈Ω is a point in the domain; H is the Heaviside function; Φ is the evolution curve distance function; and Ψ is the distance function of the prior shape profile curve.

[0076] Optionally, the output module is further used for:

[0077] S451. Obtain a local image of the sample and perform enhancement processing on the local image of the sample.

[0078] S452. Extract image features from the enhanced local image of the sample to obtain multiple features of the local image of the sample; among which, multiple features include geometric features, grayscale features, gradient features and texture features.

[0079] S453. Train the Support Vector Machine (SVM) model based on multiple features to obtain a trained SVM model.

[0080] On one hand, an electronic device is provided, comprising a processor and a memory, wherein the memory stores at least one instruction, which is loaded and executed by the processor to implement the above-described method for detecting arteriovenous fistula thrombosis.

[0081] On the one hand, a computer-readable storage medium is provided, wherein at least one instruction is stored in the storage medium, the at least one instruction being loaded and executed by a processor to implement the above-described method for detecting arteriovenous fistula thrombosis.

[0082] The beneficial effects of the technical solutions provided in the embodiments of the present invention include at least the following:

[0083] The above scheme demonstrates that the designed method can acquire shallow blood vessel images relatively clearly and intuitively, which is beneficial for the design and application of various image processing and pattern recognition algorithms.

[0084] Secondly, the designed method can simultaneously detect and label / intervene arteriovenous fistula thrombosis, and has the advantages of being easy to use, having reliable information processing, and being highly automated, which can significantly improve patients' satisfaction with the diagnosis and treatment results.

[0085] Furthermore, the designed method comprehensively employs the active contour / level set method with geometric constraints for superficial vessel segmentation, which has the advantages of reliable calculation results and high accuracy.

[0086] Finally, the designed method exhibits a high level of intelligence, comprehensively employing level set methods and machine learning techniques to achieve accurate calculations for blood vessel segmentation and identification. The use of deep learning networks for the classification and identification of arteriovenous fistula thrombosis demonstrates advantages such as high detection accuracy and computational robustness. Experiments show that the designed method achieves high computational efficiency (over 90%) for both genders. Attached Figure Description

[0087] 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.

[0088] Figure 1 This is a schematic diagram of the arteriovenous fistula thrombosis detection method provided in the embodiments of the present invention;

[0089] Figure 2 This is a flowchart of the two-stage image enhancement process provided in the embodiments of the present invention;

[0090] Figure 3 This is a structural diagram of the RCAE model provided in the embodiments of the present invention;

[0091] Figure 4 This is an example diagram of the preprocessed image and its mask provided in an embodiment of the present invention;

[0092] Figure 5 This is a near-infrared vascular image and its typical region diagram provided in the embodiments of the present invention;

[0093] Figure 6 This is an image processing flowchart provided in an embodiment of the present invention;

[0094] Figure 7 This is a block diagram of the arteriovenous fistula thrombosis detection device provided in an embodiment of the present invention;

[0095] Figure 8 This is a schematic diagram of the structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0096] To make the technical problems, technical solutions and advantages of the present invention clearer, a detailed description will be given below in conjunction with the accompanying drawings and specific embodiments.

[0097] like Figure 1 As shown, this embodiment of the invention provides a method for detecting arteriovenous fistula thrombosis, which can be implemented by electronic equipment. Figure 1 The flowchart shown illustrates a method for detecting arteriovenous fistula thrombosis. This method's processing flow may include the following steps:

[0098] S1. Obtain the image of the forearm blood vessels to be detected.

[0099] Optionally, acquiring the forearm blood vessel image to be detected in S1 includes:

[0100] Near-infrared images of superficial blood vessels in the forearm to be tested are acquired using a near-infrared image acquisition device.

[0101] In one feasible implementation, the near-infrared image acquisition device may include a near-infrared camera, a near-infrared light source, a camera and light source support and slide rail, a marking / intervention robotic arm, a robotic arm slide rail, and a marking / intervention robotic arm end effector.

[0102] Furthermore, the near-infrared image acquisition device can employ an 850nm wavelength light source and a response camera; the light source and camera are mounted on bracket 1 to irradiate and photograph the forearm; opposite bracket 1 is bracket 2, which is equipped with a transmission robotic arm. Based on the thrombus detection results from bracket 1, the arm can locate the thrombus position and deliver the end effector of the robotic arm to the thrombus location for early intervention using methods such as heat therapy, electrotherapy, or special topical medications. Brackets 1 and 2 are mounted on a slide rail and can move along the same side of the arm.

[0103] Furthermore, to address the problem of detecting arteriovenous fistula thrombosis, this invention designs a device that integrates a near-infrared camera, a near-infrared light source, and a robotic arm, and employs image processing and pattern recognition techniques to detect arteriovenous fistula thrombosis. It uses near-infrared imaging in typical spectral bands for superficial vessel detection and image segmentation. Based on this, machine learning technology is used for thrombus classification and identification, offering advantages such as intuitive data acquisition, high computational accuracy, and good clinical application results.

[0104] S2. Preprocess the forearm blood vessel image to be detected.

[0105] Optionally, the preprocessing of the forearm blood vessel image to be detected in S2 includes:

[0106] S21. Remove the background from the forearm blood vessel image to be detected.

[0107] In one feasible implementation, to facilitate subsequent processing, the present invention first standardizes the size of the near-infrared vascular images before removing the background. Their size can be set to 1024×768 pixels. In the forearm near-infrared vascular image, the grayscale ranges of the forearm and background regions differ significantly. Background removal of the near-infrared vascular image is necessary to reduce background interference and make the algorithm more adaptive. The forearm region is significantly brighter than the background because it reflects more near-infrared light. Therefore, the arm portion can be preserved while the background is easily removed. The expression for background removal is shown in equation (1) below:

[0108]

[0109] Where I(x,y) and N(x,y) are the gray values ​​at position (x,y) of the original image and the image after background removal, respectively; T H It is the threshold for background removal.

[0110] S22. Perform contrast stretching on the forearm blood vessel image after background removal.

[0111] In one feasible implementation, the near-infrared vascular image has a wide grayscale distribution range due to different aperture settings of the near-infrared camera or different thicknesses of fat and muscle in the patient's forearm, directly increasing the difficulty of subsequent arm vascular morphology analysis. Contrast stretching can increase the difference between the foreground and background, expand the dynamic range of grayscale, and make it easier to distinguish blood vessels in the image. The equation for contrast stretching is defined as follows (2):

[0112]

[0113] Where I(x,y) and N(x,y) are the gray values ​​at point (x,y) of the original image and the image after contrast stretching, respectively; I min and I max These are the minimum and maximum grayscale values ​​of the original image, respectively; MAX I It is the maximum value of the image pixel brightness. For grayscale images, if each pixel is represented by 8 bits of data, then MAX is... I =2 8 -1 = 255.

[0114] S23. Perform noise removal on the forearm blood vessel image after contrast stretching.

[0115] In one feasible implementation, to suppress noise in near-infrared vascular images of the forearm, this invention uses median filtering to remove it. The median filter is a nonlinear filter that can solve the problem of blurred details and has a good effect on removing pulse interference and image noise. Its idea is to arrange the local pixels surrounding the pixel according to their grayscale intensity levels and take the median grayscale value as the pixel's grayscale value. The denoising process of median filtering can be expressed as the following equation (3):

[0116] N(x,y)=Median{I(xk,yl),(k,l)∈w} (3)

[0117] Where I(x,y) and N(x,y) are the gray values ​​of the original image and the denoised image, respectively; w is the template window, and the shape of w can be a straight line, square, circle, rhombus, etc.

[0118] S3. Based on a two-stage image enhancement method, image enhancement is performed on the preprocessed forearm blood vessel image.

[0119] Optionally, the two-stage image enhancement method in S3 performs image enhancement on the preprocessed forearm blood vessel image, including:

[0120] S31. Input the preprocessed forearm blood vessel image into the residual convolutional autoencoder (RCAE) model to obtain the first image output.

[0121] In one feasible implementation, for the first step of convolutional neural network enhancement, consider using an RCAE (Residual Convolutional Auto-Encoders) architecture to train the network on the image. This allows for targeted enhancement of blood vessel shape.

[0122] S32. Input the first image output into the CLAHE (Contrast Limited Adaptive Histogram Equalization) model to obtain the second image output.

[0123] S33. Calculate the combination ratio coefficient α based on the preprocessed forearm vascular image.

[0124] S34. Combine the first image output and the second image output according to the combination ratio coefficient to obtain the image enhancement result of the preprocessed forearm blood vessel image.

[0125] In one feasible implementation, the flowchart of the two-stage image enhancement process is as follows: Figure 2 As shown, image enhancement can effectively solve problems such as low contrast, blurred vascular lines, and loss of vascular details in near-infrared vascular images. This invention employs a two-stage enhancement method that combines a deep convolutional neural network autoencoder with traditional enhancement methods. This method combines the advantages of convolutional neural networks and traditional image enhancement methods, effectively enhancing vascular structures while suppressing non-vascular structures.

[0126] Furthermore, the linear combination method is shown in equation (4) below:

[0127] I=αI1+(1-α)I2 (4)

[0128] The combination ratio α is determined by the information entropy of the preprocessed forearm vascular image, and its definition is shown in equation (5):

[0129]

[0130] Information entropy reflects the amount of information in an image. When the information entropy is low, a larger coefficient α is chosen to achieve better image enhancement. The calculation method for coefficient α is shown in equation (6) below:

[0131]

[0132] Where I is the result image of the two-stage image enhancement; I1 and I2 are the result images of the second and first enhancement steps, respectively; p i It represents the proportion of pixels with grayscale value i in the image; based on experimental observations, M is selected. H1 and M H2.

[0133] Furthermore, the structure of the RCAE model is as follows: Figure 3 As shown, the RCAE consists of encoder and decoder blocks connected by residuals. The encoder uses an input vascular image, and the decoder attempts to reconstruct an enhanced version of it. In this invention, the input image is grayscale and has a fixed size of 1024×768 pixels; therefore, the encoder of the RCAE consists of 3 blocks, each of which includes convolution, activation, and normalization. These blocks are sequentially connected to each other via pooling operations, which reduces the spatial dimension of the effective input at each stage. In this invention, max pooling is performed on a 2×2 window of the encoder. At the decoder end, there are 3 consecutive decoding blocks: each consists of a convolution and a deconvolution filter, as well as activation and normalization layers. This invention reserves a stride of 2 for the deconvolution in the decoder. The network includes residual transfer for each pair of blocks in the encoder and decoder. This allows the differential components to learn the differences between the input and the target across layers. ReLU (Rectified Linear Unit) is chosen as the activation operator in both the encoder and decoder; and a batch normalization layer is interspersed in each block to reduce dependence on the training dataset and improve the network's generalization ability.

[0134] Furthermore, the size of the convolutional kernel is a crucial factor in determining the effective receptive field of the input and subsequent layers of the deep network. Since the width or thickness of blood vessel shapes varies within a given representation, the optimal kernel size may not be easily determined. Relatively large filters are better able to learn the spatial relationships between distant pixels in an image; smaller filters focus on encoding features within local patches of the input. In this invention, the model's convolutional kernels are gradually reduced from 9×9 to 3×3 in the encoder, while a gradual increase in kernel size is observed in the decoder layer. The number of channels within the network is also significantly increased to 64 to introduce the complexity required for the task. For training the RCAE, this invention uses an Adam optimizer with a learning rate of 0.001 for computation. Initially, the weights are initialized with random values ​​centered at 0 and Gaussian normalized to a bias of 0.05.

[0135] Figure 4 The original image and its mask are shown, along with an example of a reference image. (a) is the preprocessed image; (b) is the mask of (a); and (c) is the reference image. The synthesized enhanced near-infrared vessel image is treated as the reference output during training. This invention treats the enhanced image (reference image) as a linear combination of the actual input image and binary masks of vessel labels. Therefore, this invention allows for manual annotation of vessel locations in the form of binary masks for the training dataset. If I pre It is the input near-infrared vascular image, I maskIf it is a binary mask with a description of the vascular structure, then reference image I can be obtained. ref As shown in equation (7):

[0136] I ref =βI pre +(1-β)I mask (7)

[0137] Where β refers to the fixed weight parameter; I pre I mask and I ref The same size, I mask The colors need to be reversed in advance.

[0138] S4. Based on the adaptive prior shape level set evolution APSLSE model and the image-enhanced forearm blood vessel image, the thrombus type detection result of the forearm blood vessel image to be detected is obtained.

[0139] Optionally, in S4, the thrombus type detection results of the forearm vessel image to be detected, obtained by evolving the APSLSE model based on the adaptive prior shape level set and the image-enhanced forearm vessel image, include:

[0140] S41. Based on the adaptive prior shape level set evolution APSLSE model, image segmentation is performed on the enhanced forearm blood vessel image to obtain the segmented region.

[0141] In one feasible implementation, this invention proposes an APSLSE (Adaptive Prior ShapeLevel Set Evolution) method for vessel segmentation. Compared to earlier methods, it successfully addresses the sensitivity issue to the initial contour, adapts to the direction of level set evolution, and better suits vessel segmentation applications.

[0142] Since the initial contour has a significant impact on most segmentation models, choosing different initial contours can lead to completely different segmentation results. This invention effectively reduces the impact of the coarse segmentation result on the final segmentation result by using the coarse segmentation result as the initial contour for iteration. Figure 5 Examples of near-infrared images and their typical regions are given. APSLSE designed an energy function for the evolution curve, defined as shown in equation (8):

[0143] E(φ)=μR p (φ)+λL g (φ)+νA g (φ)+γE Shape (8)

[0144] Where μ>0; λ>0; γ>0; ν∈R is the energy functional Rp (φ), L g (φ), A g (φ) and E Shape The coefficients; φ is the level set function, which is the energy term E(φ), R p (φ), L g (φ), A g (φ), E Shape The variable; R p (φ) is the regularization term for the level set function; L g (φ) is the length term of the level set function; A g (φ) is the area term of the level set function; E Shape This is the shape constraint term for the level set function.

[0145] Furthermore, the regularization term of the level set function is shown in equation (9) below:

[0146]

[0147] in, is the gradient operator; p is the potential (or energy density) function; the preferred potential function of the distance regularization term is the double-well potential, expressed as equation (10):

[0148]

[0149] Similar to DRLSE, this invention introduces a distance regularization term into APSLSE, which adds an internal function to the energy function in the traditional active contour model to correct the deviation between the level set function and the signed distance function. Therefore, the purpose of the level set function does not need to be periodically reinitialized during evolution.

[0150] Furthermore, in order to regularize the zero-order profile of φ, the present invention also uses the length of the zero-order curve (surface) of φ, given by the following equation (11):

[0151]

[0152] Where δ is the Dirac trigonometric function; g is the edge indicator function. In practice, the function L... g The Dirac trigonometric function δ in the equation is given by the following smooth function δ ε (x) approximates, as defined in many level set methods, as in equation (12):

[0153]

[0154] Where ε is an empirically set parameter. The edge indicator function g is defined as follows (13):

[0155]

[0156] Among them, G σ It is a Gaussian kernel with a standard deviation of σ. Convolution is used to smooth images to reduce noise. The value of the function g at object boundaries is typically smaller than its value at other locations.

[0157] Optionally, for many level set models where curve evolution is unidirectional, the model is combined with the LBF (Local Binary Fitting) method to adaptively change the direction of motion of the level set model. The area term of the level set function is shown in equation (14) below:

[0158]

[0159] Where ν1 and ν2 are coefficients, and ν1, ν2 ∈ R; Ω is the domain; x ∈ Ω is a point on the domain; g is the edge indicator function; H is the Heaviside function, which is smoothed by the following equation (15) H ε (x) approximates; ε LBF It is the energy functional of the locally binary fitted LBF.

[0160]

[0161] Clearly, the area term is a variable. When it is locally positive, it undergoes contraction; conversely, when it is locally negative, it undergoes expansion. Therefore, regardless of whether the initial contour intersects with the target, the model of this invention can perform appropriate movements based on image information.

[0162] Furthermore, to overcome the sensitivity of most level set methods to the initial contour, this invention proposes a shape constraint based on the Hessian matrix. Since the vascular structure in an image is linear, a multi-scale vascular enhancement filter based on the eigenvalues ​​of the Hessian matrix can effectively utilize this characteristic to extract blood vessels. It constructs a vascular similarity function V(x) using the relationship between the eigenvalues ​​of the Hessian matrix and the linear target to detect linear targets, and uses multi-scale fusion to handle cases with different blood vessel sizes. After multi-scale filtering, the vascular similarity function at vascular targets has a large V(x) value (V(x)∈[0,1]), while the similarity function values ​​at other non-vascular targets are small or 0. The vascular contour curve can then be defined as C... shape ={x|V(x)=T}, where the threshold T∈[0,1].

[0163] In this invention, the signed distance function proposed by Rousson et al. is used to represent the shape of blood vessels, as shown in equation (16):

[0164]

[0165] Where d(x,C) shape ) is the pixel x and the curve C shape The Euclidean distance between them; the shape of blood vessels in an image can be represented by a signed distance function; R C It is in the image plane Ω by C shape The defined region. Therefore, the initialization level set function can be set to (x, C). shape As shown in equation (17):

[0166] Φ(0,x,C)=Ψ(x,C shape (17)

[0167] Where Φ is the evolution curve distance function, i.e., the level set representation of the target edge curve; Ψ is the distance function of the prior shape contour curve. The prior shape level set segmentation model uses the curve shape energy function to penalize curves that deviate from the prior shape.

[0168] Optionally, the shape constraint term of the level set function is shown in equation (18) below:

[0169] E shape =∫ Ω (H(Φ)-H(Ψ)) 2 dx (18)

[0170] Where Ω is the domain; x∈Ω is a point in the domain; H is the Heaviside function; Φ is the evolution curve distance function; and Ψ is the distance function of the prior shape profile curve.

[0171] S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image.

[0172] S43. Extract the skeleton from the binarized forearm blood vessel image to obtain the forearm blood vessel skeleton image.

[0173] S44. Extract a local image along the location of blood vessels in the forearm vascular skeleton image.

[0174] In one feasible implementation, a local image is cropped with points on the skeleton as the center, and an SVM model is called to perform thrombosis detection.

[0175] S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected.

[0176] In one feasible implementation, the skeleton is traversed at a certain interval to complete the thrombus detection of blood vessels in the entire image.

[0177] Furthermore, if no thrombus is found in any local image, then the image is considered to be free of thrombus. If a thrombus is found in a local image, it is marked with a red box to indicate that a thrombus exists at the current location.

[0178] Optionally, the training process of the SVM (Support Vector Machine) model in S45 includes:

[0179] S451. Obtain a local image of the sample and perform enhancement processing on the local image of the sample.

[0180] S452. Extract image features from the enhanced local image of the sample to obtain multiple features of the local image of the sample; among which, multiple features include geometric features, grayscale features, gradient features and texture features.

[0181] In one feasible implementation, the method for calculating the geometric feature (roundness) is as shown in the following equation (19):

[0182]

[0183] Where A is the area of ​​the region and P is the perimeter of the region. Calculate the roundness of each connected region after segmentation; if it contains multiple regions, take the value closest to 1.

[0184] Gray-scale features include gray-scale mean, gray-scale variance, equivalent number of views, and entropy.

[0185] The method for calculating the average grayscale value is shown in equation (20) below:

[0186]

[0187] Where f(x,y) represents the gray value of the pixel (x,y) corresponding to image f; M and N are the width and height of the image, respectively.

[0188] The calculation method for grayscale variance is shown in equation (21) below:

[0189] D var (f)=∑ y ∑ x |f(x,y)-μ| 2 (twenty one)

[0190] Where μ is the average gray value of the entire image.

[0191] The method for calculating the equivalent number of views is shown in equation (22) below:

[0192]

[0193] The entropy is calculated as shown in equation (23):

[0194]

[0195] Where, p i L represents the probability of a pixel with gray value i appearing in the image, where L is the total number of gray levels (usually 256).

[0196] Gradient features include average gradient, Brenner gradient, energy gradient, and Vollath function.

[0197] The average gradient is calculated as shown in equation (24):

[0198]

[0199] in, and These represent the gradients in the horizontal and vertical directions, respectively.

[0200] The Brenner gradient is calculated as shown in equation (25):

[0201] D Brenner (f)=∑ y ∑ x |f(x+2,y)-f(x,y)| 2 (25)

[0202] The energy gradient is calculated as shown in equation (26):

[0203] D ene (f)=∑ y ∑ x (|f(x+1,y)-f(x,y)| 2 +|f(x,y+1)-f(x,y)| 2 (26)

[0204] The Vollath gradient is calculated as shown in equation (27):

[0205] D Vollath (f)=∑ y ∑ x f(x,y)*f(x+1,y)-M*N*μ 2 (27)

[0206] Texture features (gray-level co-occurrence matrix) include:

[0207] The energy calculation method is shown in equation (28) below:

[0208] En(f)=∑ y ∑ x f(x,y) 2 (28)

[0209] The method for calculating contrast is shown in equation (29) below:

[0210]

[0211] The correlation calculation method is shown in equation (30) below:

[0212]

[0213] The entropy is calculated as shown in equation (31):

[0214] Ent(f)=-∑ y ∑ x f(x,y)log(f(x,y)) (31)

[0215] Where, μ x =∑ y ∑ x xf(x,y); μ y =∑ y ∑ x yf(x,y);

[0216]

[0217] S453. Train the Support Vector Machine (SVM) model based on multiple features to obtain a trained SVM model.

[0218] In one feasible implementation, Support Vector Machine (SVM) is considered one of the best-performing algorithms in supervised learning, exhibiting excellent stability and accuracy when learning highly complex nonlinear equations. Its advantage lies in its ability to find good classification models in exceptionally complex high-dimensional feature spaces. In this application, due to the limited number of samples, the traditional machine learning algorithm SVM, which requires a relatively small number of training samples, is chosen.

[0219] Optionally, the thrombus type detection result of the forearm blood vessel image to be detected can be obtained based on the local image and the deep learning network.

[0220] In one feasible implementation, early thrombosis identification can be achieved using either a support vector machine or a deep learning network. Using a deep learning network eliminates the need for manual feature design, as the network automatically calculates abstract features based on the labeled results.

[0221] Furthermore, the computational flowchart proposed in this invention is as follows: Figure 6As shown, the image of the forearm vessels was obtained from a near-infrared image acquisition device, and then image preprocessing was performed first. Preprocessing included background removal, contrast stretching, and noise removal. Image enhancement can effectively solve problems such as low contrast, blurred vessel lines, and loss of vessel details in near-infrared vessel images. This invention employs a two-stage enhancement method, combining a deep convolutional neural network autoencoder with traditional enhancement methods. This method combines the advantages of convolutional neural networks and traditional image enhancement methods, effectively enhancing vascular structures and suppressing non-vascular structures. Subsequently, vessel shape segmentation and extraction are performed. In this invention, an APSLSE is proposed to achieve vessel segmentation. Compared with earlier methods, it can successfully solve the sensitivity problem to the initial contour, adapt to the direction of level set evolution, and better suit vessel segmentation applications.

[0222] In this embodiment of the invention, the designed method can acquire shallow blood vessel images more clearly and intuitively, which is beneficial to the design and application of various image processing and pattern recognition algorithms.

[0223] Secondly, the designed method can simultaneously detect and label / intervene arteriovenous fistula thrombosis, and has the advantages of being easy to use, having reliable information processing, and being highly automated, which can significantly improve patients' satisfaction with the diagnosis and treatment results.

[0224] Furthermore, the designed method comprehensively employs the active contour / level set method with geometric constraints for superficial vessel segmentation, which has the advantages of reliable calculation results and high accuracy.

[0225] Finally, the designed method exhibits a high level of intelligence, comprehensively employing level set methods and machine learning techniques to achieve accurate calculation of blood vessel segmentation. The use of deep learning networks for the classification and identification of arteriovenous fistula thrombosis demonstrates advantages such as high detection accuracy and computational robustness. Experiments show that the designed method achieves high computational efficiency (over 90%) for both genders.

[0226] like Figure 7 As shown, this embodiment of the invention provides an arteriovenous fistula thrombosis detection device 700. This device 700 is used to implement a method for detecting arteriovenous fistula thrombosis. The device 700 includes:

[0227] The acquisition module 710 is used to acquire images of the forearm blood vessels to be detected.

[0228] The preprocessing module 720 is used to preprocess the forearm blood vessel images to be detected.

[0229] Image enhancement module 730 is used to enhance the preprocessed forearm blood vessel image based on a two-stage image enhancement method.

[0230] The output module 740 is used to obtain the thrombus type detection result of the forearm blood vessel image to be detected based on the APSLSE model evolved from the adaptive prior shape level set and the image-enhanced forearm blood vessel image.

[0231] Optionally, the acquisition module 710 is further used for:

[0232] Near-infrared images of superficial blood vessels in the forearm to be tested are acquired using a near-infrared image acquisition device.

[0233] Optionally, the preprocessing module 720 is further used for:

[0234] S21. Remove the background from the forearm blood vessel image to be detected.

[0235] S22. Perform contrast stretching on the forearm blood vessel image after background removal.

[0236] S23. Perform noise removal on the forearm blood vessel image after contrast stretching.

[0237] Optionally, the image enhancement module 730 is further used for:

[0238] S31. Input the preprocessed forearm blood vessel image into the residual convolutional autoencoder (RCAE) model to obtain the first image output.

[0239] S32. Input the first image output into the contrast-limited adaptive histogram equalization (CLAHE) model to obtain the second image output.

[0240] S33. Calculate the combination ratio coefficient based on the preprocessed forearm vascular image.

[0241] S34. Combine the first image output and the second image output according to the combination ratio coefficient to obtain the image enhancement result of the preprocessed forearm blood vessel image.

[0242] Optionally, the output module 740 is further used for:

[0243] S41. Based on the adaptive prior shape level set evolution APSLSE model, image segmentation is performed on the enhanced forearm blood vessel image to obtain the segmented region.

[0244] S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image.

[0245] S43. Extract the skeleton from the binarized forearm blood vessel image to obtain the forearm blood vessel skeleton image.

[0246] S44. Extract a local image along the location of blood vessels in the forearm vascular skeleton image.

[0247] S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected.

[0248] Alternatively, based on local images and deep learning networks, the thrombus type detection results of the forearm blood vessel image to be detected can be obtained.

[0249] Alternatively, the energy function in the APSLSE model is derived based on the adaptive prior shape level set evolution, as shown in equation (1) below:

[0250] E(φ)=μR p (φ)+λL g (φ)+νA g (φ)+γE Shape (1)

[0251] Where μ>0; λ>0; γ>0; ν∈R; φ is the level set function; R p (φ) is the regularization term for the level set function; L g (φ) is the length term of the level set function; A g (φ) is the area term of the level set function; E Shape This is the shape constraint term for the level set function.

[0252] Alternatively, the area term of the level set function is shown in equation (2) below:

[0253]

[0254] Where ν1 and ν2 are coefficients, ν1, ν2 ∈ R; Ω is the domain; x ∈ Ω is a point on the domain; g is the edge indicator function; H is the Heaviside function; ε LBF It is the energy functional of the locally binary fitted LBF.

[0255] Optionally, the shape constraint term of the level set function is shown in equation (3) below:

[0256] E shape =∫ Ω (H(Φ)-H(Ψ)) 2 dx (3)

[0257] Where Ω is the domain; x∈Ω is a point in the domain; H is the Heaviside function; Φ is the evolution curve distance function; and Ψ is the distance function of the prior shape profile curve.

[0258] Optionally, the output module is further used for:

[0259] S451. Obtain a local image of the sample and perform enhancement processing on the local image of the sample.

[0260] S452. Extract image features from the enhanced local image of the sample to obtain multiple features of the local image of the sample; among which, multiple features include geometric features, grayscale features, gradient features and texture features.

[0261] S453. Train the Support Vector Machine (SVM) model based on multiple features to obtain a trained SVM model.

[0262] In this embodiment of the invention, the designed method can acquire shallow blood vessel images more clearly and intuitively, which is beneficial to the design and application of various image processing and pattern recognition algorithms.

[0263] Secondly, the designed method can simultaneously detect and label / intervene arteriovenous fistula thrombosis, and has the advantages of being easy to use, having reliable information processing, and being highly automated, which can significantly improve patients' satisfaction with the diagnosis and treatment results.

[0264] Furthermore, the designed method comprehensively employs the active contour / level set method with geometric constraints for superficial vessel segmentation, which has the advantages of reliable calculation results and high accuracy.

[0265] Finally, the designed method exhibits a high level of intelligence, comprehensively employing level set methods and machine learning techniques to achieve accurate calculation of blood vessel segmentation. The use of deep learning networks for the classification and identification of arteriovenous fistula thrombosis demonstrates advantages such as high detection accuracy and computational robustness. Experiments show that the designed method achieves high computational efficiency (over 90%) for both genders.

[0266] Figure 8 This is a schematic diagram of the structure of an electronic device 800 provided in an embodiment of the present invention. The electronic device 800 can vary considerably due to differences in configuration or performance, and may include one or more central processing units (CPUs) 801 and one or more memories 802. The memory 802 stores at least one instruction, which is loaded and executed by the processor 801 to implement the following method for detecting arteriovenous fistula thrombosis:

[0267] S1. Obtain the image of the forearm blood vessels to be detected.

[0268] S2. Preprocess the forearm blood vessel image to be detected.

[0269] S3. Based on a two-stage image enhancement method, image enhancement is performed on the preprocessed forearm blood vessel image.

[0270] S4. Based on the adaptive prior shape level set evolution APSLSE model and the image-enhanced forearm blood vessel image, the thrombus type detection result of the forearm blood vessel image to be detected is obtained.

[0271] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions that can be executed by a processor in a terminal to complete the arteriovenous fistula thrombosis detection method described above. For example, the computer-readable storage medium may be a ROM, random access memory (RAM), CD-ROM, magnetic tape, floppy disk, or optical data storage device, etc.

[0272] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0273] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method for detecting arteriovenous fistula thrombosis, characterized in that, The method includes: S1. Obtain an image of the forearm blood vessels to be detected; S2. Preprocess the forearm blood vessel image to be detected; S3. Based on a two-stage image enhancement method, image enhancement is performed on the preprocessed forearm blood vessel image; S4. Based on the adaptive prior shape level set evolution APSLSE model and the image-enhanced forearm blood vessel image, the thrombus type detection result of the forearm blood vessel image to be detected is obtained. The thrombus type detection results of the forearm blood vessel image to be detected in S4, based on the adaptive prior shape level set evolution APSLSE model and the image enhancement of the forearm blood vessel image, include: S41. Image segmentation of forearm blood vessel images after image enhancement based on the adaptive prior shape level set evolution APSLSE model. S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image. S43. Extract the skeleton from the binarized forearm blood vessel image to obtain a forearm blood vessel skeleton image. S44. Extract a local image along the blood vessel locations in the forearm vascular skeleton image; S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected; Alternatively, based on the local image and the deep learning network, the thrombus type detection result of the forearm blood vessel image to be detected can be obtained; The energy function in the APSLSE model based on adaptive prior shape level set evolution in S41 is shown in equation (1) below: （1） in, >0; >0; >0; R; It is a level set function; For the level set function, it is the regularization term; The length term of the level set function; This is the area term of the level set function; For the shape constraint term of the level set function; The area term of the level set function is shown in equation (2) below: （2） in, and For coefficients, , R; For domain; Let be a point on the domain; For edge indicator functions; For the Heaviside function; The energy functional of the locally binary fitted LBF; The shape constraint term of the horizontal set function is shown in equation (3) below: （3） in, For domain; Let be a point on the domain; For the Heaviside function; The distance function of the evolution curve; It is the distance function of the prior shape profile curve.

2. The method according to claim 1, characterized in that, The acquisition of the forearm blood vessel image to be detected in S1 includes: Near-infrared images of superficial blood vessels in the forearm to be tested are acquired using a near-infrared image acquisition device.

3. The method according to claim 1, characterized in that, The preprocessing of the forearm blood vessel image to be detected in S2 includes: S21. Remove the background from the forearm blood vessel image to be detected; S22. Perform contrast stretching on the forearm blood vessel image after background removal; S23. Perform noise removal on the forearm blood vessel image after contrast stretching.

4. The method according to claim 1, characterized in that, The two-stage image enhancement method in S3, which enhances the preprocessed forearm blood vessel image, includes: S31. Input the preprocessed forearm blood vessel image into the residual convolutional autoencoder (RCAE) model to obtain the first image output. S32. Input the first image output into the contrast-limited adaptive histogram equalization (CLAHE) model to obtain the second image output; S33. Calculate the combination ratio coefficient based on the preprocessed forearm vascular image; S34. Combine the first image output and the second image output according to the combination ratio coefficient to obtain the image enhancement result of the preprocessed forearm blood vessel image.

5. The method according to claim 1, characterized in that, The training process of the Support Vector Machine (SVM) model in S45 includes: S451. Obtain a local image of the sample and perform enhancement processing on the local image of the sample; S452. Extract image features from the enhanced local image of the sample to obtain multiple features of the local image of the sample; wherein, the multiple features include geometric features, grayscale features, gradient features and texture features; S453. Train the Support Vector Machine (SVM) model based on the multiple features to obtain a trained SVM model.

6. A device for detecting arteriovenous fistula thrombosis, characterized in that, The device includes: The acquisition module is used to acquire images of the forearm blood vessels to be detected. The preprocessing module is used to preprocess the forearm blood vessel image to be detected; The image enhancement module is used to enhance the preprocessed forearm blood vessel image based on a two-stage image enhancement method. The output module is used to obtain the thrombus type detection result of the forearm blood vessel image to be detected based on the APSLSE model evolved from the adaptive prior shape level set and the image-enhanced forearm blood vessel image. The step of obtaining the thrombus type detection result of the forearm vessel image to be detected based on the adaptive prior shape level set evolution APSLSE model and the image enhancement of the forearm vessel image includes: S41. Image segmentation of forearm blood vessel images after image enhancement based on the adaptive prior shape level set evolution APSLSE model. S42. Perform binarization processing on the segmented forearm blood vessel image to obtain a binarized forearm blood vessel image. S43. Extract the skeleton from the binarized forearm blood vessel image to obtain a forearm blood vessel skeleton image. S44. Extract a local image along the blood vessel locations in the forearm vascular skeleton image; S45. Based on the local image and the trained support vector machine (SVM) model, obtain the thrombus type detection result of the forearm blood vessel image to be detected; Alternatively, based on the local image and the deep learning network, the thrombus type detection result of the forearm blood vessel image to be detected can be obtained; The energy function in the APSLSE model based on adaptive prior shape level set evolution in S41 is shown in equation (1) below: （1） in, >0; >0; >0; R; It is a level set function; For the level set function, it is the regularization term; The length term of the level set function; This is the area term of the level set function; For the shape constraint term of the level set function; The area term of the level set function is shown in equation (2) below: （2） in, and For coefficients, , R; For domain; Let be a point on the domain; For edge indicator functions; For the Heaviside function; The energy functional of the locally binary fitted LBF; The shape constraint term of the horizontal set function is shown in equation (3) below: （3） in, For domain; Let be a point on the domain; For the Heaviside function; The distance function of the evolution curve; It is the distance function of the prior shape profile curve.

Images