Blood vessel segmentation method based on perceptual tissue rule

Abstract

The invention discloses a blood vessel segmentation method based on perceptual tissue rules, and belongs to the field of image processing. Comprising the following steps: establishing perceptual organization rules; establishing a neural network based on a perceptual organization rule; establishing a semi-supervised learning method based on a covariant rate; setting a 3D image mode; and defining a loss function and setting an experiment. The method is inspired by a human visual system and comprises two key parts: a neural network based on a perceptual organization rule and a semi-supervised learning method based on a covariant rate. And by calculating perceptual tissue rules such as similarity, proximity and continuity, expression of vascular coherence structure features is enhanced, so that the segmentation effect is improved. The blood vessel segmentation method disclosed by the invention shows excellent diagnostic ability in assisting doctors in clinical diagnosis and treatment, and can be used for quickly analyzing and processing the condition of a patient.

Description

Technical Field

[0001] The present invention belongs to the technical field of medical image processing, and in particular relates to a blood vessel segmentation method based on perceptual tissue rules and its application. Background Art

[0002] The vascular structure in living organisms plays a vital role in maintaining physiological functions, ensuring that oxygen and nutrients are properly delivered to tissues and organs. The geometry and topology of blood vessels vary significantly and are complex, and these features are often associated with different pathological conditions.

[0003] With advances in imaging technology, it is now possible to more accurately observe and analyze the structure and function of the vascular system through two-dimensional (2D) or three-dimensional (3D) images. This enhanced capability helps to gain a deeper understanding of the progression of diseases and the effects of therapeutic interventions. For example, optical coherence tomography angiography (OCTA) is a non-invasive imaging technique that vividly depicts the structure of the vascular network in the eye by scanning changes in blood flow in the retinal vascular area. OCTA can clearly display capillaries and microvessels, making it valuable in detecting various ophthalmic diseases such as glaucoma, diabetic retinopathy, and choroidal neovascularization. In addition, studies have shown that changes in microvascular morphology revealed by OCTA are associated with various neurodegenerative diseases including Alzheimer's disease and mild cognitive impairment.

[0004] Similarly, magnetic resonance angiography (MRA) provides detailed imaging of blood vessels in the brain and other parts of the body, enabling non-invasive examination of the vascular system. MRA uses magnetic resonance imaging (MRI) to produce high-resolution, three-dimensional images that help visualize large and small vessels. This imaging technique is essential for diagnosing and monitoring cerebrovascular diseases such as aneurysms, arteriovenous malformations, and stroke. It also plays an important role in assessing the progression of atherosclerosis and the effects of therapeutic interventions. Accurately visualizing and measuring cerebral blood flow and vascular morphology is essential to understanding the pathophysiology of various neurological diseases and developing targeted treatment strategies. Studies have shown that MRA can reveal vascular changes associated with diseases such as multiple sclerosis and Parkinson's disease, thereby enhancing the understanding of these diseases and improving patient care.

[0005] Therefore, automatic detection and quantitative analysis of vascular structures in medical images are of great value for clinical research, diagnosis and treatment. At present, many applications of deep learning in the field of medical image segmentation have made significant progress. Deep learning models, such as convolutional neural networks (CNNs), can effectively identify and segment different tissues in medical images. These models usually use a large amount of training data to learn the mapping from input images to output labels, thereby achieving high-quality segmentation results.

[0006] However, the task of vessel segmentation still faces challenges. First, manual labeling requires highly trained professionals, consuming a lot of time and effort, making it difficult to implement on large-scale, highly dynamic, and high-resolution data. In addition, the presence of noise and the impact of disease on image clarity and tissue integrity further exacerbate the complexity of vessel segmentation. In contrast, human doctors use their rich prior knowledge to accurately identify and label retinal vessels when faced with discontinuous or blurred images. Therefore, the present invention aims to develop a method that draws on human experience and can reliably and accurately segment vessels in 2D and 3D medical images with a small number of labeled samples. Summary of the invention

[0007] The purpose of the present invention is to provide a blood vessel segmentation method based on perceptual organization rules; in order to solve the problem of limited medical image annotation data, the present invention constructs a CF-SSL method. The method redefines visual rules and applies them to semi-supervised learning tasks, making small samples in low-dimensional manifolds denser. This reduces overfitting during model training and improves the generalization ability of the model. The blood vessel segmentation method based on perceptual organization rules proposed in the present invention has excellent diagnostic capabilities, so that it can assist doctors in their clinical work and quickly diagnose and treat patients' conditions.

[0008] Specifically, the method is inspired by the human visual system and includes two parts: a neural network based on perceptual organization rules (PoUNet) and a semi-supervised learning method based on covariance rate (CF-SSL). PoUNet uses the characteristics of human visual perception to enhance the expression of the coherent structural features of blood vessels by calculating perceptual organization rules such as similarity, proximity and continuity, thereby improving the segmentation performance. In order to achieve the above purpose, the technical solution of the present invention is: a blood vessel segmentation method based on perceptual organization rules, the method comprising the following steps:

[0009] Step 1: Establish a neural network structure based on perceptual organization rules;

[0010] The neural network based on perceptual organization rules is an encoder-decoder architecture of U-Net, and the encoder-decoder architecture includes a shallow part and a deep part, and the corresponding shallow part and deep part in the encoder-decoder are jump-connected; the shallow part is a convolution operation to extract local low-level features in the image, and the jump connection of the shallow part transfers the extracted local features; the deep part is a perceptual organization encoder-perceptual rule decoder, and the jump connection of the deep part transfers the perceptual rule coefficient matrix;

[0011] The method of the perceptual organization encoder is: calculate the similarity, proximity, and continuity between the image blocks of the input image, and then calculate the coefficient matrix matrix using the following formula: c ;

[0012] matrix c =Similarity+Proximity·Continuity

[0013] Coefficient matrix matrix c The output of the perceptual organization encoder is obtained after weighting with the current feature;

[0014] The method of the perceptual rule decoder is to weight the current feature with the coefficient matrix transmitted through the jump connection and then perform random inactivation to obtain the output of the perceptual rule decoder;

[0015] Step 2: Use a semi-supervised learning method to learn the neural network structure established in step 1;

[0016] Step 3: Use the learned neural network structure to perform blood vessel segmentation on the image to be detected.

[0017] Furthermore, the semi-supervised learning method in step 2 is a semi-supervised learning method based on covariance rate;

[0018] The semi-supervised learning method based on covariance rate includes fully supervised learning and unsupervised learning. The fully supervised learning is to use labeled data to pre-train the neural network structure based on perceptual organization rules established in step 1 to obtain a pre-trained model; the unsupervised learning is to use the pre-trained model to predict unlabeled data to obtain unsupervised prediction values; background drift processing is performed on the unlabeled data, and the data is labeled with the unsupervised prediction values ​​corresponding to the unlabeled data, and then the data is used to train the pre-trained model to obtain the final optimized model.

[0019] Furthermore, the loss function of the fully supervised learning in step 2 is for:

[0020]

[0021] in, represents the weighted binary cross entropy loss function, represents the Dice coefficient loss function, y represents the true value, represents the predicted probability, w0 and w1 are the weights assigned to the background and foreground, respectively.

[0022] Furthermore, the loss function in the unsupervised learning is

[0023]

[0024] in, represents the embedding vector consistency loss function, represents the prospect prediction consistency loss function, and represents the quantized posterior probability distribution of real samples and artificial samples, respectively, defined as a one-hot vector; n represents the number of encoding neurons; D(·) is a discriminant function that returns 0 when the two input one-hot vectors are the same and 1 when they are different; y t and f Represent the model's predicted probability of the prospect in real samples and artificial samples, respectively.

[0025] Furthermore, since artificial samples inevitably contain noise components, it is very difficult to constrain the consistency of the embedding vectors of artificial samples and real samples in continuous space, so the vector quantization method is introduced to measure the consistency of the embedding vector in discrete space; first, labeled samples are used to construct a codebook in fully supervised learning. The codebook randomly initializes K D-dimensional vectors to record different vascular patterns, including straight vessels, curved vessels, and bifurcated vessels in different orientations; for an embedding vector Z output by the encoder, the nearest neighbor search is used to map it to one of the K vectors in the codebook to generate a discrete representation vector z q ; Finally, the decoder is used to decode z q Decode to get the predicted probability map; overall training loss function as follows:

[0026]

[0027] in, represents the segmentation loss function, represents the constraint loss function that prevents the embedding space from expanding infinitely, z represents the embedding vector output by the encoder, sg represents the gradient truncation operation, F(·) represents the decoder, and y represents the true value. Represents the predicted probability, β is set as a constant; after the model for z q After optimization iterations, a discrete codebook with feature representation capability is obtained.

[0028] Furthermore, in the unsupervised learning process, the encoder of the pre-trained model is used to encode the unlabeled data, and the embedded vector is output after encoding. The embedded vector and the vascular pattern closest to the codebook are calculated to annotate the unlabeled data, and the encoding corresponding to the closest vascular pattern is used as the posterior probability distribution. The embedded vector consistency loss function is used. To optimize the pre-trained model.

[0029] Furthermore, the calculation method of the similarity in step 1 is:

[0030]

[0031] Among them, q and k represent the feature vectors of all pixels in the query queue and key respectively, i and j represent the index of the pixel position of the feature map, and F S represents the softmax function;

[0032] The calculation method of the close rate is:

[0033]

[0034] Among them, c i,j Represents the position coordinates of all pixels, σ is a learnable parameter;

[0035] The calculation method of continuity rate is:

[0036] The composite convolution kernel obtained by multiplying the Gabor convolution kernel and the learnable convolution kernel is used to extract the directional response features of different image blocks. i,j ; Use cosine similarity to calculate the directional difference of directional response features of different image blocks. The directional difference represents the continuity rate. The calculation method is:

[0037]

[0038] Among them, i and j represent the index of the feature map pixel position.

[0039] Furthermore, when the perceptual organization rule blood vessel segmentation method is used for 2D images, the convolution kernel in the neural network structure established in step 1 is a 2D convolution kernel, and the Gabor convolution kernel for calculating the continuity rate Continuity is a 2D Gabor function; when used for 3D images, the convolution kernel in the neural network structure established in step 1 is a 3D convolution kernel, and the Gabor convolution kernel for calculating the continuity rate Continuity is a 3D Gabor function.

[0040] Compared with the prior art, the present invention has the following beneficial effects:

[0041] 1) This paper integrates the theory of human visual system into the deep learning model and proposes a new vascular segmentation model based on perceptual organization rules. The model designs and integrates calculation rules such as similarity, proximity and continuity, which enhances the coherence expression of vascular structure and has better segmentation performance than existing methods.

[0042] 2) This paper proposes a semi-supervised learning method based on covariance rate. This method uses feature quantization to construct a vascular pattern code table, uses background drift technology to generate a large number of artificial samples, and optimizes the model using the coding consistency of discrete space, so that the model has better robustness and generalization in few-sample learning tasks.

[0043] 3) The method of the present invention has universal applicability, can be applied to two-dimensional and three-dimensional image modes, and has high clinical application potential. BRIEF DESCRIPTION OF THE DRAWINGS

[0044] Figure 1 This is a structural diagram of the perceptual organization rule-based neural network (PoUNet) in the present invention.

[0045] Figure 2 2 is a structural diagram of the perceptual organization encoder / decoder module in the present invention.

[0046] Figure 3 Schematic diagram of the covariance rate-based semi-supervised learning method (CF-SSL) in the present invention. DETAILED DESCRIPTION

[0047] In practical applications, medical images are often affected by poor imaging quality and noise distribution, resulting in discontinuities and blurs in the tissue structure of the image, making it difficult for the algorithm to classify it as a whole. However, humans can easily identify and separate objects in complex scenes because of the role of perceptual organization rules. Perceptual organization rules refer to some principles in the human perceptual system that can help perceive and understand information in the environment more effectively.

[0048] The present invention provides a blood vessel segmentation method and application based on perceptual tissue rules, comprising the following steps:

[0049] Step S1: Establishment of perceptual organization rules;

[0050] The present invention designs calculation modules of similarity, proximity and continuity for the complex curved structures of blood vessels and nerve fibers, and integrates them into the encoder, thereby enhancing the model's ability to group structural features. Through a non-local method based on a self-attention mechanism, the present invention effectively constructs an expression model of these perceptual organization rules, helps learn long-distance dependencies in image features, and thus improves the segmentation effect of complex structures in medical images. In order to surpass the inductive bias of traditional convolutional neural networks, the present invention adopts a non-local method based on a self-attention mechanism to construct an expression model of similarity, proximity and continuity. This design helps to learn long-distance dependencies in image features, thereby better expressing and utilizing perceptual organization rules and improving the segmentation effect of complex structures in medical images.

[0051] Step S11: Definition of similarity.

[0052] The similarity measures the similarity of local features such as color, brightness, and texture between image blocks. The query queue q and key k represent the pixel set of the current feature map. The similarity of the feature vector is calculated by the inner product, and then converted into a probability distribution through the softmax function, which represents the contribution of each input element to the output element (i.e., the coefficient matrix). Finally, the coefficient matrix is ​​weighted for the current eigenvalue, so that the pixel set with a higher degree of similarity will be activated and sent to the subsequent feature extraction step. The calculation formula for the similarity is as follows:

[0053]

[0054] Where q and k represent the feature vectors of all pixels in the query queue and key, respectively, i and j represent the indices of the pixel positions in the feature map, and F S Represents the softmax function.

[0055] Step S12: Definition of the approach rate.

[0056] The closeness rate measures the physical distance between image blocks and is a strong perceptual organization clue. The present invention measures the distance between image blocks with Euclidean distance and uses Gaussian function to simulate the probability of whether image blocks belong to the same group. The calculation formula of the closeness rate is as follows:

[0057]

[0058] where c i,j represents the position coordinates of all pixels, and σ is a learnable parameter.

[0059] Step S13: Definition of continuity rate.

[0060] The continuity rate measures the difference in directional response between image blocks. The present invention constructs a directional perception module to capture the texture and edge direction information of blood vessels or nerve fibers. Gabor convolution can adapt to features of different sizes and directions by adjusting the scale and direction of the filter, and is similar to some functions of human vision, and can effectively simulate the perception of the human visual system. The two-dimensional Gabor function in the present invention adopts the following definition:

[0061]

[0062] Where x′=x cosθ+y sinθ, y′=-x sinθ+y cosθ. x and y represent the position coordinates of the image pixels, λ represents the wavelength, θ represents the directional parameter, γ represents the spatial aspect ratio, ψ represents the phase offset, and σ represents the standard deviation of the Gaussian factor. The present invention instantiates the imaginary part of the formula, and the convolution kernel size is configured as 7×7 with 16 different directional parameters θ (θ=2π / i, i=1,2,...16), which is used to build a Gabor filter library for the direction perception module. The direction perception module uses a composite convolution kernel obtained by multiplying a Gabor convolution kernel with a learnable convolution kernel to extract the directional response features of the image. The directional response features of different image blocks will use cosine similarity to measure the directional differences, calculated as follows:

[0063]

[0064] Among them i,j represents the directional response of all pixels, and i and j represent the index of the pixel position of the feature map.

[0065] Step S14: Combination of perceptual organization rules.

[0066] This study believes that proximity is a stronger perceptual organization cue. Even if the elements are oriented in the same direction, the distance will still directly affect the grouping of elements. Therefore, the perceptual rule coefficient matrix matrix c The construction of adopts the form of proximity weighting and continuity rate, and uses position information to constrain the continuity of elements. The difference is that the similarity rate does not use the proximity rate to weight it. This is because similarity carries the self-attention relationship of diverse features. Retaining its long-distance perception ability can narrow the difference between classes and reduce the number of category centers, which is more conducive to the segmentation task of this study. In summary, the matrix c The calculation of is as follows:

[0067] matrix c =Similarity+Proximity·Continuity.

[0068] The matrix c The output feature of the perceptual organization encoder is obtained by weighting the current feature. The perceptual organization rules established in the encoder also act on the decoding of the feature, matrix c The features are fed into the perceptual organization decoder through the skip connection layer and weighted. The advantage of such decoding is that the high-level semantic features can be reconstructed in the direction of the same group of elements through the rules recorded by the encoder, which is conducive to building a more continuous and stable curve structure.

[0069] Step S2: Establishment of a neural network (PoUNet) based on perceptual organization rules.

[0070] The PoUNet model proposed in the present invention is based on the encoder-decoder architecture of U-Net and uses jump connections to restore fine edge features in the shallow layer. In the shallow part, the model continues to use traditional convolution operations to extract local low-level features in the image, such as color, brightness and direction information; in the deep part, the original structure is replaced by the perceptual organization encoder-decoder to better integrate local features and generate representations that conform to the perceptual organization rules. This design can not only effectively reduce the computational cost, but is also particularly suitable for training smaller-scale image data sets. By adopting maximum pooling as a downsampling strategy and bilinear interpolation as an upsampling method, the present invention further reduces the number of parameters of the model, thereby alleviating the overfitting problem and improving the training efficiency of the model in a limited data environment.

[0071] Step S3: Establishment of covariance rate based semi-supervised learning method (CF-SSL).

[0072] Medical applications based on artificial intelligence usually rely on a large number of manual annotations, but it is extremely difficult to obtain a large number of detailed medical expert annotations. The present invention proposes a new semi-supervised learning method (CF-SSL) for image segmentation tasks based on the covariance rate in perceptual organization. Covariance rate refers to the fact that when certain elements in a scene move or change simultaneously or in correlation, these elements are more easily identified as a whole rather than independent parts. The characteristic of covariance rate is that it can effectively integrate complex scenes and dynamic events into a meaningful whole. The present invention deals with static image segmentation tasks. In order to simulate the movement of the foreground, a background drift method is specially designed to simulate the generation of diverse artificial samples. The present invention is based on the manifold hypothesis, that is, within a local neighborhood, the data should have a similar representation. Under this assumption, artificial samples generated by background drift of unlabeled data use consistency loss to bring artificial samples closer to real samples on a low-dimensional manifold, thereby optimizing model performance.

[0073] Aiming at the problem of difficulty in manual labeling, the present invention proposes a new semi-supervised learning method (CF-SSL) for image segmentation based on the covariance rate in perceptual organization. The covariance rate can identify related elements as a whole and effectively integrate complex scenes. The present invention designs a background drift method to simulate and generate diverse artificial samples, and through the manifold assumption and consistency loss, makes the unlabeled data closer to the real samples on the low-dimensional manifold, thereby optimizing the model performance.

[0074] Step S31: Definition of covariance rate.

[0075] Covariance refers to the fact that elements that move or change at the same time are more likely to be perceived as a whole rather than as independent individuals. According to the concept of relative motion, the movement of the background will not affect the perception of the object. In order to apply the principle of common fate to the static image segmentation task, the present invention designs a background drift method. The method simulates the movement of the foreground by moving the background pixels, thereby generating a variety of artificial samples. The background drift module keeps the predicted foreground area position in the original image unchanged, and performs a random overall displacement of the background pixels in four directions of up, down, left and right. At this time, the image will have unfilled pixels caused by the movement of the background. This method uses a random sampling method with replacement to fill the image and finally obtain an artificial sample. The sampled pixels come from the background pixel group around the foreground of the real sample. According to the rule of covariance, the foreground pixels before and after the movement should be regarded as a whole, so the model's potential representation of the artificial sample after encoding and the prediction of the foreground should be consistent with the real sample.

[0076] Step S32: Definition of feature quantification method.

[0077] In unsupervised training, the consistency constraints imposed on the model by the present invention include two aspects: the consistency of the latent representation after encoding and the consistency of the predicted probability after decoding. However, approximating the embedding vectors of artificial samples and real samples in a continuous space is challenging and easily leads to the collapse of model optimization. To solve this problem, the present invention introduces a feature quantization method to convert the encoded features into a more abstract representation. The present invention constructs a codebook in a discrete space for encoding different vascular patterns, and measures the consistency of the embedded vector by evaluating whether two sets of features encode the same vascular pattern. This method provides greater flexibility for the latent variable space and better supports unsupervised training.

[0078] In the specific implementation, the present invention defines a codebook Where K is the number of pattern categories and D is the dimension of the latent space vector e. In the fully supervised training process, labeled samples are used to construct the codebook. For an embedding vector Z, a nearest neighbor search is used to map it to one of the K vectors in the codebook to generate a discrete representation vector z q Finally, q Decode and get the predicted probability map. The overall training loss function is as follows:

[0079]

[0080] Where sg represents the gradient clipping operation. F(·) represents the decoder. β is set to 0.25.

[0081] Step S4: 3D image mode setting.

[0082] The method designed by the present invention supports 2D image and 3D image modes. The 3D image mode mainly differs in three aspects:

[0083] 1) The 2D Gabor function of the continuity rate in step S1 is transformed into a 3D Gabor function, which is calculated as follows:

[0084]

[0085] The present invention instantiates the imaginary part of the formula, the convolution kernel size is configured as 7×7, and 16 different spatial parameters are used to initialize the 3D Gabor function, and the calculation is as follows:

[0086]

[0087] 2) The PoUNet model established in step S2 is transformed from a 2D convolution operation module to a 3D convolution operation module;

[0088] 3) The background drift method in step S31 is transformed from random movement in 4 directions to random movement in 8 directions.

[0089] Step S5: Definition of loss function and experimental setup.

[0090] The semi-supervised learning method constructed by the present invention includes two parts: a fully supervised learning loss function and an unsupervised learning loss function. In order to cope with the sparse foreground of vascular images, the fully supervised learning loss function is composed of a weighted binary cross entropy loss function and a Dice coefficient loss function, and is calculated as follows:

[0091]

[0092] Where y represents the true value, represents the predicted probability. w0 and w1 are the weights assigned to the background and foreground, respectively. In this embodiment, w0 is set to 0.1 and w1 is set to 0.9.

[0093] The unsupervised learning loss mainly constrains the consistency of the model's perception of artificial samples and real samples. According to the rule of covariance rate, the foreground pixels before and after the motion should be regarded as a whole. The present invention calculates the consistency loss function for the encoded potential representation and the prediction of the foreground, which is calculated as follows:

[0094]

[0095] in and represents the quantized posterior probability distribution of real samples and artificial samples, respectively, defined as a one-hot vector. n represents the number of encoding neurons. D(·) is a discriminant function that returns 0 when two input one-hot vectors are the same and 1 when they are different.t and f Represent the model's predicted probability of the prospect in real samples and artificial samples, respectively.

[0096] In semi-supervised learning training, the semi-supervised learning loss combines the fully supervised learning loss with the unsupervised consistency loss during training. The total training loss function is calculated as follows:

[0097]

[0098] Where E a represents the total number of iterations, E c Indicates the current iteration. As the training progresses, the weight of the consistency constraint α gradually decreases. During training, the adaptive moment estimation gradient descent algorithm is used to optimize the loss function. The initial learning rate is set to 0.001, and the coefficient is multiplied by 0.9 after each iteration of training. The model is iterated for a total of 200 times.

[0099] The above contents are detailed descriptions of the embodiments of the present invention in combination with the accompanying drawings, but it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, deformation, substitution, combination or simplification made according to the principle of the technical solution of the present invention are all regarded as equivalent replacement methods and should be regarded as belonging to the protection scope of the present invention.

Claims

1. A blood vessel segmentation method based on perceptual tissue rules, the method comprising the following steps: Step 1: Establish a neural network structure based on perceptual organization rules; The neural network based on perceptual organization rules is an encoder-decoder architecture of U-Net, the encoder-decoder architecture includes a shallow part and a deep part, and the corresponding shallow part and deep part in the encoder-decoder are jump-connected; the shallow part is a convolution operation to extract local low-level features in the image, and the jump connection of the shallow part transmits the extracted local features; The deep part is the perceptual organization encoder-perceptual rule decoder, and the skip connection in the deep part transmits the perceptual rule coefficient matrix; The method of the perceptual organization encoder is: calculate the similarity, proximity, and continuity between the image blocks of the input image, and then calculate the coefficient matrix matrix using the following formula: c ; matrix c =Similarity+Proximity·Continuity Coefficient matrix matrix c The output of the perceptual organization encoder is obtained after weighting with the current feature; The method of the perceptual rule decoder is to weight the current feature with the coefficient matrix transmitted through the jump connection and then perform random inactivation to obtain the output of the perceptual rule decoder; Step 2: Use a semi-supervised learning method to learn the neural network structure established in step 1; Step 3: Use the learned neural network structure to perform blood vessel segmentation on the image to be detected.

2. A blood vessel segmentation method based on perceptual organization rules as claimed in claim 1, characterized in that: The semi-supervised learning method in step 2 is a semi-supervised learning method based on covariance rate; The semi-supervised learning method based on covariance rate includes fully supervised learning and unsupervised learning. The fully supervised learning is to use labeled data to pre-train the neural network structure based on perceptual organization rules established in step 1 to obtain a pre-trained model; the unsupervised learning is to use the pre-trained model to predict unlabeled data to obtain an unsupervised prediction value; The unlabeled data is processed for background drift, and the unsupervised prediction value corresponding to the unlabeled data is used to label the data. Then, the data is used to train the pre-trained model to obtain the final optimized model.

3. The blood vessel segmentation method based on perceptual organization rules as claimed in claim 1, characterized in that: The calculation method of similarity in step 1 is: Among them, q and k represent the feature vectors of all pixels in the query queue and key respectively, i and j represent the index of the pixel position of the feature map, and F S represents the softmax function; The calculation method of the close rate is: Among them, c i,j Represents the position coordinates of all pixels, σ is a learnable parameter; The calculation method of continuity rate is: The composite convolution kernel obtained by multiplying the Gabor convolution kernel and the learnable convolution kernel is used to extract the directional response features of different image blocks. i,j ; Use cosine similarity to calculate the directional difference of directional response features of different image blocks. The directional difference represents the continuity rate. The calculation method is: Among them, i and j represent the index of the feature map pixel position.

4. The blood vessel segmentation method based on perceptual organization rules as claimed in claim 2, characterized in that: The loss function of the fully supervised learning in step 2 is for: in, represents the weighted binary cross entropy loss function, represents the Dice coefficient loss function, y represents the true value, represents the predicted probability, w0 and w1 are the weights assigned to the background and foreground, respectively.

5. The blood vessel segmentation method based on perceptual organization rules as claimed in claim 2, characterized in that: The loss function in the unsupervised learning is in, represents the embedding vector consistency loss function, represents the prospect prediction consistency loss function, and represents the quantized posterior probability distribution of real samples and artificial samples, respectively, defined as a one-hot vector; n represents the number of encoding neurons; D(·) is a discriminant function that returns 0 when the two input one-hot vectors are the same and 1 when they are different; y t and f Represent the model's predicted probability of the prospect in real samples and artificial samples, respectively.

6. The blood vessel segmentation method based on perceptual organization rules as claimed in claim 3, characterized in that: When the perceptual organization rule blood vessel segmentation method is used for 2D images, the convolution kernel in the neural network structure established in step 1 is a 2D convolution kernel, and the Gabor convolution kernel for calculating the continuity rate Continuity is a 2D Gabor function; when used for 3D images, the convolution kernel in the neural network structure established in step 1 is a 3D convolution kernel, and the Gabor convolution kernel for calculating the continuity rate Continuity is a 3D Gabor function.

7. The blood vessel segmentation method based on perceptual tissue rules as claimed in claim 4, characterized in that: First, a codebook is constructed using labeled samples in fully supervised learning. The codebook randomly initializes K D-dimensional vectors to record different vascular patterns, including straight vessels, curved vessels, and bifurcated vessels in different orientations. For an embedding vector Z output by the encoder, the nearest neighbor search is used to map it to one of the K vectors in the codebook to generate a discrete representation vector z q ; Finally, the decoder is used to decode z q Decode to get the predicted probability map; overall training loss function as follows: in, represents the segmentation loss function, represents the constraint loss function that prevents the embedding space from expanding infinitely, z represents the embedding vector output by the encoder, sg represents the gradient truncation operation, F(·) represents the decoder, and y represents the true value. Represents the predicted probability, β is set as a constant; after the model for z q After optimization iterations, a discrete codebook with feature representation capability is obtained.

8. The blood vessel segmentation method based on perceptual tissue rules as claimed in claim 7, characterized in that: In the unsupervised learning process, the encoder of the pre-trained model is used to encode the unlabeled data, and the embedded vector is output after encoding. The embedded vector and the vascular pattern closest to the codebook are calculated to annotate the unlabeled data. The encoding corresponding to the closest vascular pattern is used as the posterior probability distribution, and the embedded vector consistency loss function is used. To optimize the pre-trained model.

Images