Fundus Vessel Image Segmentation Method and System Based on Hierarchical Diffusion Model

Abstract

The present disclosure provides a fundus vascular image segmentation method and system based on a hierarchical diffusion model, which relates to the technical field of fundus vascular segmentation, and includes obtaining an original fundus image; inputting the original fundus image into a lightweight cascaded network for feature extraction to obtain the context information of blood vessels and the position information of blood vessel details, and outputting an initial rough blood vessel segmentation image; based on the initial rough blood vessel segmentation image, using the principle of morphological dilation to extract the blood vessel skeleton and obtain a blood vessel skeleton image; taking the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions and inputting them into a diffusion model to sequentially output intermediate transitional blood vessel segmentation images at each time step; then using the original fundus image as a condition and iteratively refining the intermediate transitional blood vessel segmentation images at each time step through the denoising process of the improved diffusion model to output a final fine blood vessel segmentation image. The present disclosure improves the accuracy of blood vessel segmentation.

Description

Technical Field

[0001] The present disclosure relates to the technical field of fundus vascular segmentation, and particularly to a method and system for fundus vascular image segmentation based on a hierarchical diffusion model. Background Art

[0002] The statements in this section merely provide background technical information related to the present disclosure and do not necessarily constitute prior art.

[0003] With the emergence of computer vision, significant progress has been made in fundus retinal vascular segmentation. It is understood that retinal blood vessels can be used as a window for monitoring and evaluating various diseases, including the cardiovascular and cerebrovascular conditions of patients. Analyzing the geometric complexity of retinal arteries can serve as a valuable diagnostic biomarker for fundus and cerebrovascular diseases, and changes in the retinal vascular network can diagnose the development and progression of retinal lesions. However, the complexity and variability of the retinal vascular structure and the techniques used to acquire fundus images make retinal vascular segmentation challenging. In traditional methods, the accuracy of manual segmentation of retinal blood vessels is not high and is greatly affected by subjective factors. Therefore, developing an algorithm using computer vision technology that can reliably segment retinal blood vessels and extract vascular features from fundus images plays an important role in assisting medical staff in the diagnosis of ophthalmic and cardiovascular and cerebrovascular diseases.

[0004] In recent years, some researchers have proposed many evaluation algorithms for segmenting retinal blood vessels in fundus images, and these algorithms can be roughly divided into three categories: vessel tracking-based methods, match filtering-based methods, and machine learning-based methods.

[0005] However, the inventors have found that current vascular segmentation methods still have significant problems. Various diseases or lesions can significantly reduce image clarity, making retinal vascular segmentation challenging. Current methods usually have difficulty capturing thinner blood vessels, resulting in incomplete vascular segmentation. Summary of the Invention

[0006] To solve the above problems, the present disclosure proposes a method and system for fundus vascular image segmentation based on a hierarchical diffusion model. The vascular segmentation is regarded as a conditional generation task, and by extracting the vascular skeleton, a rough vascular segmentation image, and the original fundus image as conditions in the diffusion model, relevant information of the fundus blood vessels is obtained, thereby improving the accuracy of vascular segmentation.

[0007] According to some embodiments, the present disclosure adopts the following technical solutions:

[0008] A method for fundus vascular image segmentation based on a hierarchical diffusion model, comprising:

[0009] Obtain an original fundus image;

[0010] Input the original fundus image into a lightweight cascaded network for feature extraction to obtain the context information of blood vessels and the position information of blood vessel details, and output an initial rough blood vessel segmentation image;

[0011] Based on the initial rough blood vessel segmentation image, use the principle of morphological dilation to extract the blood vessel skeleton and obtain the blood vessel skeleton image;

[0012] Take the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions and input them into the diffusion model to sequentially output the intermediate transitional blood vessel segmentation images at each time step; then use the original fundus image as a condition and iteratively refine the intermediate transitional blood vessel segmentation images at each time step by using the denoising process of the improved diffusion model to output the final fine blood vessel segmentation image.

[0013] According to some embodiments, the present disclosure adopts the following technical solutions:

[0014] A fundus blood vessel image segmentation system based on a hierarchical diffusion model, including:

[0015] A basic segmentation module that obtains the original fundus image, inputs the original fundus image into a lightweight cascaded network for feature extraction to obtain the context information of blood vessels and the position information of blood vessel details, and outputs an initial rough blood vessel segmentation image;

[0016] A blood vessel skeleton extraction module for extracting the blood vessel skeleton based on the initial rough blood vessel segmentation image by using the principle of morphological dilation to obtain the blood vessel skeleton image;

[0017] A dual-guidance module for inputting the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions into the diffusion model to sequentially output the intermediate transitional blood vessel segmentation images at each time step;

[0018] A refinement guidance segmentation module for using the original fundus image as a condition and iteratively refining the intermediate transitional blood vessel segmentation images at each time step by using the denoising process of the improved diffusion model to output the final fine blood vessel segmentation image.

[0019] Compared with the prior art, the beneficial effects of the present disclosure are:

[0020] A method for segmenting fundus blood vessels based on a hierarchical diffusion model of the present disclosure generates dual-guidance conditions through fundus image enhancement, uses a hierarchical denoising diffusion model, and obtains the information most relevant to the blood vessel position in a coarse-to-fine strategy, thereby improving the accuracy of fundus blood vessel recognition.

[0021] The present disclosure applies the generated conditions to diffusion model segmentation. Morphological processing expands thin and thick blood vessels to a unified width, enhances the attention to thin blood vessels, and provides more information about blood vessels. Information related to the blood vessel positions is learned, irrelevant information is removed, and more accurate features of blood vessel positions are obtained.

[0022] The present disclosure proposes an iterative refinement method that uses the original fundus image as a reference condition to optimize the intermediate transitional blood vessel images at each time step of the generation process, thereby further improving the accuracy of blood vessel segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] The accompanying drawings forming a part of this disclosure are used to provide a further understanding of the disclosure. The schematic embodiments and descriptions thereof of the disclosure are used to explain the disclosure and do not constitute an improper limitation of the disclosure.

[0024] Figure 1 is a schematic flowchart of the fundus blood vessel segmentation method according to an embodiment of the present disclosure;

[0025] Figure 2 is the final fine fundus blood vessel segmentation map according to an embodiment of the present disclosure;

[0026] Figure 3 is the comparison map of fundus blood vessel segmentation in the ablation experiment according to an embodiment of the present disclosure;

[0027] Figure 4 is the comparison map of the final detection results according to an embodiment of the present disclosure. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0028] The present disclosure will be further described below in conjunction with the accompanying drawings and embodiments.

[0029] It should be noted that the following detailed descriptions are all illustrative and are intended to provide further descriptions of the present disclosure. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present disclosure belongs.

[0030] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present disclosure. As used herein, unless the context clearly indicates otherwise, the singular forms are also intended to include the plural forms. In addition, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0031] Embodiment 1

[0032] In one embodiment of the present disclosure, a method for segmenting fundus blood vessel images based on a hierarchical diffusion model is provided, including the following steps:

[0033] Step 1: Obtain the original fundus image;

[0034] Step 2: Input the original fundus image into a lightweight cascaded network for feature extraction, obtain the context information of blood vessels and the position information of blood vessel details, and output an initial rough blood vessel segmentation image;

[0035] Step 3: Based on the initial rough blood vessel segmentation image, use the principle of morphological dilation to extract the blood vessel skeleton and obtain the blood vessel skeleton image;

[0036] Step 4: Take the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions and input them into the diffusion model, and sequentially output the intermediate transition blood vessel segmentation images at each time step;

[0037] Step 5: Then, use the original fundus image as a condition, and use the denoising process of the improved diffusion model to iteratively refine the intermediate transition blood vessel segmentation images at each time step, and output the final fine blood vessel segmentation image.

[0038] As an embodiment, the specific implementation process of a fundus blood vessel segmentation method based on a hierarchical diffusion model of the present disclosure is as follows:

[0039] Such as Figure 1 shown, first, obtain the initial rough blood vessel segmentation image and the blood vessel skeleton image to be referenced.

[0040] In this embodiment, a small lightweight cascaded network is introduced, which combines the receptive fields generated by each robust block in a cascaded manner, reduces the network complexity and eliminates the use of pooling layers. A symmetric multi-scale skip connection is adopted between the encoder and the decoder to enhance the integration of deep and shallow details, and there is a Sigmoid layer behind each output.

[0041] A further technical solution, the lightweight cascaded network specifically includes:

[0042] A basic robust block, which includes a convolutional layer, a Dropout layer, a BatchNorm layer, and a Swish activation function; a residual connection; five pairs of robust blocks, each layer contains 32 channels, with a size of 3×3 pixels; a multi-scale skip connection is added between the encoder and the symmetric decoder to better fuse deep features and shallow details.

[0043] Train the proposed lightweight cascaded network, and adopt a supervised loss function with ground truth, which is specifically expressed as:

[0044]

[0045] Among them, V represents the preliminary rough blood vessel segmentation image. pi ∈[0,1] represents the estimated probability of the model for the class with a label, and γ is an adjustable focusing parameter. Let V denote the output of the lightweight cascaded network, followed by a Sigmoid layer.

[0046] The original fundus image is input into the lightweight cascaded network, enhancing the fusion of deep and shallow details, capturing context information and location detail information, and outputting an initial rough vascular segmentation image.

[0047] Specifically, deep features usually refer to the abstract and high-level features in an image, such as the overall structure, shape, and higher-level semantic information in the image. In a deep learning model, deep features are learned through multiple layers of a deep neural network. These features are very important for identifying and understanding complex patterns and objects in an image.

[0048] Shallow features usually refer to the low-level features in an image, such as edges, textures, colors, etc. These features are usually extracted in the early processing stage of the image and have an important impact on the basic structure and appearance of the image. Shallow features are usually learned by the shallow layers of the network, and these layers are relatively fewer and closer to the input data.

[0049] Furthermore, obtaining the vascular skeleton image includes extracting the vascular skeleton based on the initial rough vascular segmentation image using the principle of morphological dilation to obtain the vascular skeleton image;

[0050] Specifically, morphological dilation is the inverse operation of the erosion operation and is similar to image dilation. The principle of morphological expansion is to find local maxima. When the white or highlighted areas of the image are expanded, the lines in the image become thicker, and the resulting operation extends beyond the highlighted areas of the original image. This is mainly used for denoising purposes.

[0051] Furthermore, processing the initial rough vascular segmentation image, the purpose of vascular skeleton extraction is to enhance the recognition ability of unrecognized areas, provide more detailed vascular information, balance the representation of thick and thin vascular pixels, so as to obtain the vascular skeleton as one of the conditions for the diffusion model. Specifically, it includes: uniformly expanding the vascular skeleton to a specific width in the obtained initial rough vascular segmentation image. The vascular skeleton extraction is specifically represented as:

[0052]

[0053] where S is the vascular skeleton, B is the convolution kernel, and V is the initial rough vascular segmentation image. Convolve the convolution kernel B with the image V, and calculate the maximum value of the pixel points in the area of the convolution kernel B and assign it to the pixel corresponding to the reference point.

[0054] Furthermore, a denoising diffusion model with fractional matching is used to generate the final fine segmentation image through dual guidance and refinement guidance.

[0055] The generation process of the denoising diffusion model is specifically expressed as:

[0056]

[0057] Among them, β t ∈ [0, 1] is the fixed variance. Therefore, the noise target x from x0 t The distribution is expressed as:

[0058]

[0059]

[0060] Among them, X0 represents the original clean image, and x t represents the image with noise at time t; Specifically, DDPM uses a deep neural network (usually UNet) to predict x by taking the mean and covariance of x t-1 as the input. The generation process is represented by a parametric Gaussian transformation: t Among them, μ is derived using Bayes' theorem

[0061]

[0062] (x θ , t): t To perform the learning of the denoising process, first generate a sample x by adding Gaussian noise θ to x0

[0063]

[0064] ~q(x t |x0), and then train the model ∈ t (x θ , t) to predict the added noise: t In this embodiment, the vascular skeleton and the denoised initial rough vascular image obtained are connected with x

[0065]

[0066] as the input for future predictions. Therefore, this is the improved parametric Gaussian transition for conditional generation: t Among them, μ

[0067]

[0068] Among them, μ θrepresents the posterior conditional mean, and the conditional denoising method learns the noise prediction of the initial rough vascular image with additional vascular skeletons and denoising, denoted as

[0069]

[0070] A further technical solution, the training step of using the obtained vascular skeleton image and rough vascular segmentation image as conditions in the diffusion model to obtain the intermediate transitional vascular segmentation image at each time step of the diffusion model includes: using the obtained initial rough fundus vascular segmentation image and vascular skeleton image as guiding conditions, taking the intermediate transitional vascular segmentation result as the output, modifying the UNet network structure in the diffusion model to achieve posterior prediction, and training a bi-conditional guided diffusion network; fine-tuning the training model by randomly changing the image with gray pixels.

[0071] The modified UNet model network of the diffusion model specifically includes:

[0072] An encoder composed of a series of residual and downsampling convolutional layers; a decoder composed of a series of residual layers and corresponding upsampling convolutional layers; connecting these components through skip connections, and the skip connections connect the intermediate layers of the same spatial size.

[0073] In the above, the training model is fine-tuned by randomly changing the image with gray pixels, using as the unconditional representation. To connect the input image with two additional conditions (i.e., the initial rough vascular segmentation image and the vascular skeleton), the input channels of the model are extended to seven. For 30% of each condition during the sampling process, the ratio between the preliminary rough vascular segmentation image and the vascular skeleton is controlled by the following linear combination:

[0074]

[0075] Furthermore, taking the obtained intermediate transitional vascular segmentation result as the input and the original fundus image as the condition, the denoising process of the diffusion model is used to iteratively refine each intermediate result to obtain a more accurate vascular segmentation result.

[0076] The obtained vascular segmentation images still show differences from reality. Therefore, it is necessary to provide control over the accuracy of the output with respect to the input, which requires realism control. In addition to the information from the 2D classifier guiding the vascular skeleton and the initial rough vascular segmentation image, the proposed method involves iteratively refining the latent variables, applying iterative latent variable refinement to refine each intermediate transition in the downsampled reference image during the generation process. The proposed realism control allows for an additional trade-off between the consistency of the provided vascular skeleton / initial rough fundus vascular segmentation image and the distance to the target data distribution (i.e., the real image). Using the original downsampled fundus image, each intermediate transition in the image generation process is enhanced. Let LP represent a linear low-pass filtering process that downsamples to a modified size N before returning to the original scale. The actual adjustment during the conditional generation process at each time step t can be expressed as follows:

[0077]

[0078]

[0079] where I is the original fundus retinal image. Figure 1 The iterative refinement process using the original fundus image as a condition is shown. In the diffusion model, the intermediate transition results at each time step are refined under the condition of the original image. Since there are a total of T steps and each step is obtained based on the result of the previous step, it is called iterative refinement.

[0080] Experimental process

[0081] In this embodiment, experiments on retinal vascular segmentation are conducted on three publicly available datasets, DRIVE, STARE, and CHASE_DB1, to evaluate the performance of the method proposed in this disclosure. And comparisons are made with other state-of-the-art methods on these datasets. Among them, the DRIVE dataset comes from a diabetic retinopathy screening project in the Netherlands. A total of 40 images are randomly selected, of which 33 show no symptoms of diabetic retinopathy and 7 show mild early symptoms. Each image is taken in JPEG format with a resolution of 584×565 pixels. The STARE dataset consists of 20 color fundus images, 10 of which contain lesions and 10 do not. The fundus image resolution is 605×700, and it is also widely regarded as one of the most widely used fundus image repositories. The CHASE_DB1 dataset uses 28 retinal images captured by a portable Nidek NM-200-D fundus camera with a field of view (FOV) of 30°. The resolution of each image is 999×960, and specific techniques are used to determine the binary mask and the segmentation ground truth.

[0082] In this embodiment, the original sizes of these three datasets are different. Therefore, the image sizes in these datasets are adjusted and uniformly resized to 512×512 pixels. In addition, data augmentation techniques are adopted, including flipping and histogram stretching of all pictures in the three datasets, etc., expanding to three times the size of the training set of the original dataset, generating 80, 48, and 80 images respectively. Among them, 208 are training samples and 32 are test samples, and both the training set and the test set contain accurate vascular segmentation results of known fundus images.

[0083] As metrics, accuracy (Acc), sensitivity (Se), specificity (Sp), F1-score (F1), and area under the curve (AUC) are used as measures. All experiments were conducted on a PC equipped with an i5 quad-core 2.59GHz CPU, 8GB RAM, and an NVIDIA RTX A30 GPU. The segmentation experiment results on the DRIVE, STARE, and CHASE_DB1 datasets are presented in Figure 2 and compared with the results of other state-of-the-art segmentation algorithms, as described in Table 1.

[0084] Table 1 Comparison data of vascular segmentation

[0085]

[0086] Compared with other algorithms, the proposed scheme in this embodiment can achieve more accurate segmentation results.

[0087] In this embodiment, in order to verify the influence of different factors on the model performance, extensive ablation experiments were conducted on the DRIVE, STARE, and CHASE_DB1 datasets. Specifically, the effects of the dual-guidance module and the refinement-guided segmentation module on vascular segmentation were studied. The ablation study segmentation data of these two parts are shown as Figure 3 shown and in Table 2 below.

[0088] Table 2 Comparison data of ablation experiments

[0089]

[0090] To evaluate the accuracy of the proposed model with randomly selected image samples, cross-validation experiments were conducted. Specifically, cross-validation was performed on the DRIVE dataset and the STARE dataset without fine-tuning, and the corresponding results are shown in Table 3. Compared with the CS-Net model, the model proposed in this embodiment shows higher accuracy and effectiveness in all metrics.

[0091] Table 3 Results of the cross-validation experiments described in the embodiment

[0092]

[0093] Example 2

[0094] This embodiment provides a fundus vascular image segmentation system based on a hierarchical diffusion model, including:

[0095] A basic segmentation module that obtains the original fundus image, inputs the original fundus image into a lightweight cascaded network for feature extraction, obtains the context information of blood vessels and the position information of blood vessel details, and outputs an initial rough blood vessel segmentation image;

[0096] A blood vessel skeleton extraction module that is used to extract the blood vessel skeleton based on the initial rough blood vessel segmentation image by using the principle of morphological dilation and obtains a blood vessel skeleton image;

[0097] A dual-guidance module that is used to input the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guidance conditions into the diffusion model, and sequentially outputs the intermediate transitional blood vessel segmentation images at each time step;

[0098] A refinement guidance segmentation module that is used to use the original fundus image as a condition and iteratively refine the intermediate transitional blood vessel segmentation images at each time step by using the denoising process of the improved diffusion model to output the final fine blood vessel segmentation image.

[0099] In a further technical solution, in the basic segmentation module, the original fundus image is input into the lightweight cascaded network, enhancing the fusion of deep and shallow details, capturing context information and position detail information, and taking the initial rough fundus blood vessel segmentation image as the output.

[0100] In a further technical solution, in the blood vessel skeleton extraction module, the obtained initial rough fundus blood vessel segmentation image is used as the input, and the principle of morphological dilation is used to balance the representation of thick blood vessel pixels and thin blood vessel pixels, enhancing the recognition ability of unrecognized regions, providing more detailed blood vessel information, and taking the blood vessel skeleton map as the output.

[0101] In a further technical solution, the training steps of the dual-guidance module include: using the rough fundus blood vessel segmentation map obtained through the basic segmentation module and the blood vessel skeleton image obtained through the blood vessel skeleton extraction module as guidance conditions, taking the intermediate transitional blood vessel segmentation result as the output, modifying the UNet network structure in the diffusion model to achieve posterior prediction, and training the dual-condition guidance diffusion network; fine-tuning the training model by randomly changing the image with gray pixels.

[0102] In a further technical solution, in the refinement guidance segmentation module, in each time step of the diffusion process, the intermediate transitional blood vessel segmentation result obtained through the dual-guidance module is used as the input, the original fundus image is used as the condition, and the denoising process of the improved diffusion model is used to iteratively refine each intermediate result to obtain a more accurate blood vessel segmentation result.

[0103] The present disclosure is described with reference to the flowcharts and / or block diagrams of methods, apparatuses (systems), and computer program products according to embodiments of the present disclosure. It should be understood that each flow and / or block in the flowchart and / or block diagram, and the combination of flows and / or blocks in the flowchart and / or block diagram, can be implemented by computer program instructions. These computer program instructions can be provided to the processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing devices to generate a machine, such that the instructions executed by the processor of the computer or other programmable data processing devices generate means for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0104] These computer program instructions can also be loaded onto a computer or other programmable data processing device, such that a series of operation steps are executed on the computer or other programmable device to generate a computer-implemented process, so that the instructions executed on the computer or other programmable device provide steps for implementing the functions specified in one Figure 1 flow or multiple flows and / or blocks Figure 1 or multiple blocks.

[0105] Although the specific embodiments of the present disclosure have been described above in conjunction with the accompanying drawings, they are not limitations on the protection scope of the present disclosure. Those skilled in the art should understand that, based on the technical solutions of the present disclosure, various modifications or variations that can be made by those skilled in the art without creative efforts are still within the protection scope of the present disclosure.

Claims

1. A method for segmenting fundus vascular images based on a hierarchical diffusion model, characterized in that Including: Obtain the original fundus image; Input the original fundus image into a lightweight cascaded network for feature extraction to obtain the context information of blood vessels and the position information of blood vessel details, and output an initial rough blood vessel segmentation image; Based on the initial rough blood vessel segmentation image, use the principle of morphological dilation to extract the blood vessel skeleton and obtain a blood vessel skeleton image; Take the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions and input them into a diffusion model to sequentially output intermediate transition blood vessel segmentation images at each time step; then use the original fundus image as a condition and use the denoising process of the improved diffusion model to iteratively refine the intermediate transition blood vessel segmentation images at each time step to output the final fine blood vessel segmentation image; Among them, using the denoising process of the improved diffusion model to iteratively refine the intermediate transition blood vessel segmentation images at each time step to output the final fine blood vessel segmentation image includes: Based on the intermediate transition blood vessel segmentation image at each time step, iteratively refine the latent variables, apply iterative latent variable refinement to refine each intermediate transition blood vessel segmentation image in the downsampled reference image during the generation process, and use the original fundus image to enhance each intermediate transition during the image generation process to output the final fine blood vessel segmentation image.

2. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein, The lightweight cascaded network combines the receptive fields generated by each robust block in a cascaded manner, adopts symmetric multi-scale skip connections between the encoder and the decoder to enhance the integration of deep and shallow details, and there is a Sigmoid layer behind each output. Each robust block specifically includes: a convolutional layer, a Dropout layer, a BatchNorm layer, and a Swish activation function.

3. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein, The lightweight cascaded network adopts a loss function supervised with ground truth, specifically expressed as: Among them, V represents the initial rough blood vessel segmentation image; represents the estimated probability of the model for the class with the label, and is an adjustable focusing parameter; let V be the output of the lightweight cascade network, followed by a Sigmoid layer.

4. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein, Based on the initial rough blood vessel segmentation image, use the principle of morphological dilation to extract the blood vessel skeleton and obtain a blood vessel skeleton image, and then uniformly expand the blood vessel skeleton to a specific width. The extraction method is: Where S is the blood vessel skeleton, B is the convolution kernel, and V is the initial rough blood vessel segmentation image; convolve the convolution kernel B with the image V, and calculate the maximum value of the pixel points in the area of the convolution kernel B and assign it to the pixel corresponding to the reference point.

5. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein, The generation process of the diffusion model is specifically expressed as: wherein, is the fixed variance; thus, the noise target from is represented as: Among them, ; the diffusion model uses a UNet deep neural network, by taking 's mean and covariance as inputs to predict representing the image obtained at the current time t.

6. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein, Input the obtained initial rough blood vessel segmentation image and blood vessel skeleton image as guiding conditions into the diffusion model, and connect them with as the input of the diffusion model, which is an improved parametric Gaussian transition for conditional generation: Among them, represents the posterior conditional mean, and the conditional denoising method learns the noise prediction of the initial rough vessel image with additional vessel skeletons and denoising.

7. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 1, wherein Using the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guiding conditions, control the ratio between the initial rough blood vessel segmentation image and the blood vessel skeleton through linear combination, take the intermediate transition blood vessel segmentation image as the output, modify the UNet network structure in the diffusion model to achieve subsequent iterative refinement, and train a bi-conditional guided diffusion network; fine-tune the diffusion model by randomly changing the image with gray pixels.

8. The method for segmenting fundus vascular images based on a hierarchical diffusion model according to claim 7, wherein The modified UNet model network of the diffusion model specifically includes: An encoder composed of a series of residual and downsampling convolutional layers; a decoder composed of a series of residual layers and corresponding upsampling convolutional layers; connect the intermediate layers of the same spatial size through skip connections.

9. A fundus vascular image segmentation system based on a hierarchical diffusion model, characterized in that, Including: The basic segmentation module obtains the original fundus image, inputs the original fundus image into a lightweight cascaded network for feature extraction, obtains the context information of blood vessels and the location information of blood vessel details, and outputs an initial rough blood vessel segmentation image; The blood vessel skeleton extraction module is used to extract the blood vessel skeleton based on the initial rough blood vessel segmentation image by using the principle of morphological dilation to obtain a blood vessel skeleton image; The dual-guidance module is used to input the obtained initial rough blood vessel segmentation image and the blood vessel skeleton image as guidance conditions into the diffusion model, and sequentially output the intermediate transitional blood vessel segmentation images at each time step; The refined guidance segmentation module is used to use the original fundus image as a condition, and iteratively refine the intermediate transitional blood vessel segmentation images at each time step by using the denoising process of the improved diffusion model to output the final fine blood vessel segmentation image; Among them, using the denoising process of the improved diffusion model to iteratively refine the intermediate transitional blood vessel segmentation images at each time step and output the final fine blood vessel segmentation image includes: Based on the intermediate transitional blood vessel segmentation images at each time step, iteratively refine the latent variables, apply iterative latent variable refinement to refine each intermediate transitional blood vessel segmentation image in the downsampled reference image during the generation process, and use the original fundus image to enhance each intermediate transition during the image generation process to output the final fine blood vessel segmentation image.

Images