Vascular Analysis Methods and Systems Based on Ultra-Wide-Angle Fundus Fluorescence Angiography Images

Abstract

This invention relates to the field of medical image classification technology, and discloses a method and system for vascular analysis based on ultra-wide-angle fundus fluorescence angiography images. The method involves acquiring ultra-wide-angle fundus fluorescence angiography images, preprocessing the fundus images to obtain a sample library, labeling fundus vessels based on the images in the sample library to construct a fundus vessel analysis training dataset, building a model, training the model using the fundus vessel analysis training dataset, and automatically extracting the vascular skeleton from the image under test using the trained model, identifying vascular nodes using a cross-point detection algorithm based on the extracted vascular skeleton, dividing the fundus image into regions, and calculating the vascular tortuosity values ​​in each region based on the vascular skeleton and vascular nodes to obtain the analysis results of fundus vessel distribution and vascular morphology. This invention can accurately observe changes in vascular tortuosity in each quadrant and at each distance, assisting clinicians in performing rapid and accurate analysis of fundus angiography images.

Description

Technical Field

[0001] This invention relates to the field of medical image classification technology, and in particular to a method and system for vascular analysis based on ultra-wide-angle fundus fluorescence angiography images. Background Technology

[0002] Ultra-widefield fluorescein angiography (UWFA), as an advanced imaging technique, provides a more comprehensive view of the peripheral retina and its vascular features. It is widely used in peripheral pathology studies in ophthalmology subspecialties such as pediatric ophthalmology, uveitis, ocular tumors, and retinal diseases, playing a crucial role in the early detection and diagnosis of eye diseases. Although ophthalmic imaging techniques such as optical coherence tomography angiography (OCTA) have made significant progress in recent years, UWFA remains irreplaceable in clinical research and diagnosis due to its wider field of view, ability to assess dynamic vascular changes, and high-contrast imaging.

[0003] However, the pathological mechanisms of many fundus diseases are not yet fully understood. With the continuous advancement of ophthalmic imaging technology, an increasing number of studies are focusing on peripheral retinal features and vascular changes in fundus lesions previously thought to affect only the posterior pole of the retina (such as epiretinal membrane, macular hole, and age-related macular degeneration). For example, studies have found that macular degeneration may have varying degrees of impact on the retina at different distances from the posterior pole or in different quadrants. Furthermore, vascular biomarkers such as vessel tortuosity are considered to play an important indicative role in the development and progression of diseases. Therefore, quantitative analysis of retinal vascular characteristics at different quadrants and distances is particularly important.

[0004] Traditional vessel extraction methods largely rely on semi-automatic techniques, such as using software like AngioTool and ImageJ. Users need to manually draw reference paths for vessels or regions of interest (ROIs), and the software then extracts the vessels based on these. While these semi-automatic methods can provide relatively good results, they suffer from drawbacks such as being time-consuming, dependent on manual operation, and limited by the user's expertise and skills, leading to poor consistency and low efficiency. Although OCTA images have high resolution, and color fundus photography has been widely used in clinical practice and has accumulated a large amount of data, many automated vessel extraction algorithms and software programs are now available, primarily targeting OCTA and color fundus photography. These programs utilize deep learning and other technologies to improve the semi-automatic analysis of traditional images, such as Airdoc.

[0005] However, automated analysis methods for ultra-wide-angle fluorescence angiography (UWFA) images remain relatively scarce. This is mainly due to the complexity of UWFA images and the special properties of fluorescence angiography images (such as color changes and dynamic images), making it difficult for existing automated analysis methods to accurately process these images. Therefore, research on UWFA images and the development of related software are still insufficient. Although some studies have performed automated segmentation and extraction analysis of fundus features (such as non-perfusion areas and neovascularization) in some UWFA images, the extraction of the vascular skeleton and the calculation of regional vascular tortuosity and related indices based on this still require further exploration.

[0006] Currently, there are no studies on the automatic calculation of VT values ​​in different regions of the fundus using ultra-wide-angle fluorescence angiography images.

[0007] Therefore, there is an urgent need for vascular analysis methods and systems based on ultra-wide-angle fundus fluorescence angiography images, which can accurately observe changes in vascular curvature in each quadrant and at each distance, and assist clinicians in performing rapid and accurate analysis of fundus angiography images. Summary of the Invention

[0008] To address the aforementioned technical problems, this invention provides a vascular analysis method and system based on ultra-wide-angle fundus fluorescence angiography images, which can accurately observe changes in vascular curvature in each quadrant and at each distance, assisting clinicians in performing rapid and accurate analysis of fundus angiography images.

[0009] This invention provides a method for vascular analysis based on ultra-wide-angle fundus fluorescence angiography images, comprising the following steps:

[0010] S1. Obtain ultra-wide-angle fundus fluorescence angiography images, and preprocess the ultra-wide-angle fundus fluorescence angiography images to obtain a sample library;

[0011] S2. Based on the images in the sample library, annotate the retinal blood vessels and construct a training dataset for retinal blood vessel analysis;

[0012] S3. Construct the DSP-Unet model. Train the DSP-Unet model using the fundus vascular analysis training dataset. Automatically extract the vascular skeleton from the ultra-wide-angle fundus fluorescence angiography image under test using the trained DSP-Unet model.

[0013] S4. Based on the extracted vascular skeleton, the cross-point detection algorithm is used to identify vascular nodes;

[0014] S5. Divide the ultra-wide-angle fundus fluorescence angiography image into regions, calculate the tortuosity value of the blood vessels in each region based on the vascular skeleton and vascular nodes, and obtain the analysis results of fundus blood vessel distribution and vascular morphology.

[0015] Furthermore, in S1, preprocessing includes: artifact removal, contrast enhancement, grayscale correction, segmentation, and data cleaning.

[0016] Furthermore, in S3, constructing the DSP-Unet model includes:

[0017] Based on the Unet model, the convolutional modules of the Unet model are replaced with depthwise separable residual convolutional modules. A multi-scale feature space pyramid module is embedded and fused between the last depthwise separable residual convolutional module and the first decoder to construct the DSP-Unet model.

[0018] Furthermore, the loss function of the DSP-Unet model is a joint loss function, expressed as follows:

[0019] ;

[0020] Among them, L B Let L represent the joint loss function, λ represent the weighting coefficients, and L represent the weighting coefficients. Ce L represents the cross-entropy loss function. Dice This represents the Dice coefficient loss function.

[0021] Furthermore, in S4, based on the extracted vascular skeleton, the cross-point detection algorithm is used to identify vascular nodes, including:

[0022] S41. Define the pixels on the center line of the vascular skeleton as skeleton points, and determine whether the current pixel is a skeleton point; if yes, continue to the next step, otherwise move to the next pixel for judgment.

[0023] S42. Determine whether there are more than or equal to a preset number of skeleton points among the 8 adjacent pixels of the skeleton point; if so, determine that the skeleton point is a blood vessel node; otherwise, determine that it is not a blood vessel node.

[0024] Furthermore, in S5, the tortuosity value of blood vessels in each region is calculated using the following formula:

[0025] ;

[0026] Wherein, VT represents the tortuosity of a certain segment of blood vessel, L represents the actual length of the centerline of the segment of blood vessel, and D represents the straight-line distance between the start and end points of the segment of blood vessel.

[0027] Here, the blood vessel between two adjacent blood vessel nodes is defined as a blood vessel segment.

[0028] Furthermore, in S5, the extended ETDRS ring partitioning method is used to divide the ultra-wide-angle fundus fluorescence angiography image into regions.

[0029] The present invention also provides a vascular analysis system based on ultra-wide-angle fundus fluorescence angiography images, for performing the vascular analysis method based on ultra-wide-angle fundus fluorescence angiography images described in any of the above claims, the system comprising the following modules:

[0030] The data acquisition module is used to acquire ultra-wide-angle fundus fluorescence angiography images and preprocess the ultra-wide-angle fundus fluorescence angiography images to obtain a sample library;

[0031] The dataset construction module is used to annotate fundus blood vessels based on images in the sample library and build a fundus blood vessel analysis training dataset.

[0032] The model building module is used to build the DSP-Unet model. The DSP-Unet model is trained using the fundus vascular analysis training dataset. The trained DSP-Unet model is used to automatically extract the vascular skeleton of the ultra-wide-angle fundus fluorescence angiography image to be tested.

[0033] The vascular node recognition module is used to identify vascular nodes based on the extracted vascular skeleton using a cross-point detection algorithm.

[0034] The analysis module is used to divide the ultra-wide-angle fundus fluorescence angiography image into regions. Based on the vascular skeleton and vascular nodes, it calculates the vascular tortuosity value in each region and obtains the analysis results of fundus vascular distribution and vascular morphology.

[0035] The embodiments of the present invention have the following technical effects:

[0036] This invention performs vascular analysis based on ultra-wide-angle fundus fluorescence angiography images, effectively avoiding the limitations of traditional fundus imaging techniques, such as limited field of view and difficulty in comprehensively observing the peripheral retina. It enables more comprehensive capture of fundus vascular information. Targeted improvements are made to the Unet deep learning model, combining a depthwise separable residual convolution module and a multi-scale feature space pyramid module, effectively reducing computational load and enhancing feature representation capabilities. Compared to traditional image analysis methods, it exhibits higher accuracy and stronger stability. Furthermore, this invention optimizes the model's learning ability for different fundus vascular features through a joint loss function, resulting in good generalization ability across various fundus images. The vascular analysis results of this invention can provide important evidence for the early diagnosis and research of ophthalmic diseases, and are of great significance for improving the accuracy and efficiency of clinical ophthalmic diagnosis. Attached Figure Description

[0037] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0038] Figure 1 This is a flowchart of a vascular analysis method based on ultra-wide-angle fundus fluorescence angiography images provided in an embodiment of the present invention;

[0039] Figure 2 This is a schematic diagram of the architecture of the DSP-Unet model provided in an embodiment of the present invention;

[0040] Figure 3 This is a schematic diagram of the depth-separable residual convolution module in the DSP-Unet model structure provided in this embodiment of the invention;

[0041] Figure 4 This is a schematic diagram of the structure of the DSP-Unet model structure that integrates the multi-scale feature space pyramid module provided in the embodiment of the present invention;

[0042] Figure 5 This is a schematic diagram of identifying vascular nodes based on a vascular skeleton provided in an embodiment of the present invention;

[0043] Figure 6 This is a schematic diagram of region division of ultra-wide-angle fundus fluorescence angiography images provided in an embodiment of the present invention. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0045] This invention proposes a vascular analysis method based on ultra-wide-angle fundus fluorescence angiography images. Figure 1 This is a flowchart of the vascular analysis method based on ultra-wide-angle fundus fluorescence angiography images provided in this embodiment of the invention. See also... Figure 1 Specifically, it includes:

[0046] S1. Obtain ultra-wide-angle fundus fluorescence angiography images, and preprocess the ultra-wide-angle fundus fluorescence angiography images to obtain a sample library.

[0047] The preprocessing includes: artifact removal, contrast enhancement, grayscale correction, segmentation, and data cleaning. Artifact removal, contrast enhancement, and grayscale correction are used to eliminate the effects of residual contrast agent artifacts and uneven illumination.

[0048] Fundus images from typical patients were selected for acquisition to ensure the diversity and representativeness of ultra-wide-angle fundus fluorescein angiography images. Before image acquisition, the parameters of the imaging equipment (such as exposure time, aperture size, gain, etc.) were set appropriately to obtain high-quality image data.

[0049] After data acquisition, image quality is checked, including sharpness, contrast, and noise levels, to ensure there are no blurry, overexposed, or underexposed images. Noise reduction processing is performed on the fundus images to remove noise and improve image sharpness.

[0050] Simultaneously, image contrast enhancement is performed to improve the accuracy of subsequent processing. Color correction is applied to eliminate color inconsistencies caused by different lighting conditions, ensuring color consistency between images. The preprocessed images are cropped into 512×512 pixel sample images. High-quality images with uniform distribution are selected from a large number of cropped sample images to construct a sample library, ensuring a balanced distribution of blood vessels and fundus background, and avoiding highly imbalanced samples.

[0051] The segmentation process divides the ultra-wide-angle fundus fluorescence angiography image into several image blocks. The image blocks are then cleaned by removing those with no blood vessels or only a few blood vessels, as well as those with unclear images due to occlusion or other factors. Only high-quality image blocks are retained to form a sample library for subsequent training.

[0052] S2. Based on the images in the sample library, annotate the retinal blood vessels and construct a training dataset for retinal blood vessel analysis.

[0053] In some embodiments, 90% of the image data in the sample library is used as training samples and 10% as test samples. The sample images are labeled according to a binary classification task, namely ocular vessels and fundus background, to construct a training dataset for fundus vessel analysis.

[0054] S3. Construct the DSP-Unet model. Train the DSP-Unet model using the fundus vascular analysis training dataset. Automatically extract the vascular skeleton of the ultra-wide-angle fundus fluorescence angiography image under test using the trained DSP-Unet model.

[0055] In some embodiments, Figure 2 This is a schematic diagram of the architecture of the DSP-Unet model provided in an embodiment of the present invention. Figure 3This is a schematic diagram of the depth-separable residual convolution module in the DSP-Unet model structure provided in this embodiment of the invention. Figure 4 This is a schematic diagram of the structure of the DSP-Unet model structure fused with the multi-scale feature space pyramid module provided in the embodiment of the present invention. See also... Figures 2-4 The construction of the DSP-Unet model includes:

[0056] Based on the Unet model, the convolutional modules of the Unet model are replaced with depthwise separable residual convolutional modules. A multi-scale feature space pyramid module is embedded and fused between the last depthwise separable residual convolutional module and the first decoder to construct the DSP-Unet model.

[0057] The depthwise separable residual convolution module includes multiple residual convolutional layers to downsample the features obtained from the input feature extraction module. In the encoding stage, this module replaces the traditional convolution process, significantly reducing computational cost while increasing the number of channels. At the end of the encoding stage, a multi-scale feature space pyramid module (ASPP module in the figure) is added, consisting of multiple dilated convolutional layers and a global average pooling layer. This effectively changes the model's receptive field and enhances feature representation capabilities by fusing features from different scales. In the decoding stage, the skip connection structure of the original model is retained. Skip connection layers concatenate deep and shallow features, optimizing the model's ability to learn different ground cover features.

[0058] The DSP-Unet model employs depthwise separable convolution. This technique decomposes the traditional standard convolution operation into two independent steps: depthwise convolution and pointwise convolution. In the depthwise convolution stage, each input channel is independently processed by a convolutional kernel, reducing computational cost and preserving channel independence. Subsequently, pointwise convolution uses a 1×1 convolutional kernel to fuse the feature maps after depthwise convolution, adjusting the number of feature channels to generate a combined output. This decomposition reduces the model's computational complexity and improves its efficiency and performance.

[0059] The DSP-Unet model introduces a Fusion Multi-Scale Feature Space Pyramid (ASPP) module, which, with the help of convolutional kernels with different dilation rates, can efficiently capture contextual information at different scales in an image. Through this fusion of multi-scale information, the model's receptive field is significantly expanded, enabling a more comprehensive understanding of complex structures and details in the image.

[0060] The input to the model consists of a batch of 1×512×512 images and a batch of 512×512 labels. The label value is 1 or 0, representing that there are two classes (blood vessels or background). Here, batch is the number of samples used for training in each batch, that is, the number of pixels in the image patch.

[0061] The specific process includes:

[0062] a. Feature extraction is performed on the image. After data fusion, the image passes through a convolutional layer and a pooling layer to obtain a 128x128x64 feature map.

[0063] b. Use four depthwise separable residual convolutional module layers to further extract features from the feature map obtained in the previous step, and finally obtain a deep feature map of 16×16×2048.

[0064] c. The 16×16×2048 feature map obtained in the second step is processed using the multi-scale feature space pyramid module to further increase the receptive field of the model. Furthermore, a global pooling layer is used to fuse features from different scales, which can effectively enhance the feature representation capability.

[0065] d. Upsample the obtained feature map using 4 upsampling layers, and connect the deep and shallow features using a skip connection module to restore the size to 512×512×2 (2 is the number of classes).

[0066] e. Loss Function: For each pixel of the 512×512×2 feature map and the 512×512 label, a joint loss function is calculated. Considering the class imbalance problem, the DSP-Unet model uses a joint loss function combining the cross-entropy loss function and the Dice coefficient loss function. By combining the advantages of multiple loss functions, the problem of sample imbalance can be better solved, while effectively preventing overfitting. The expression of the joint loss function is as follows:

[0067] ;

[0068] Among them, L B Let L represent the joint loss function, λ represent the weighting coefficients, and L represent the weighting coefficients. Ce L represents the cross-entropy loss function. Dice This represents the Dice coefficient loss function.

[0069] The cross-entropy loss function measures the difference between the probability distribution predicted by the model and the probability distribution of the true labels. The smaller the value, the more accurate the model prediction. The specific description is as follows:

[0070] ;

[0071] Where N represents the number of samples (i.e., the number of pixels in an image patch), and i represents the i-th sample (i.e., the i-th pixel). Let y represent the true label of the i-th sample. i Let y represent the predicted label of the i-th sample. i =1 indicates that the sample is a positive class, i.e., blood vessels, y i =0 indicates that the sample is a negative class, i.e., the background.

[0072] The Dice loss function, also known as the F1 loss function, is a variant based on the Dice coefficient and is commonly used in image segmentation tasks. The Dice coefficient is a statistical measure of the similarity between two samples. The Dice loss function is robust to class imbalance because it focuses on the overlap between the predicted and ground truth regions, rather than simply calculating the absolute number of classes. Furthermore, for segmenting small objects or fine structures, the Dice loss function can provide better performance than traditional loss functions because it depends on the overlap of regions, not the number of pixels. The Dice coefficient is described as follows:

[0073] ;

[0074] The definition of is:

[0075] ;

[0076] Where |A∩B| represents the size of the intersection of sets A and B, i.e., the number of correctly predicted samples.

[0077] |A| represents the size of set A, which is the total number of positive samples predicted by the model.

[0078] |B| represents the size of set B, which is the actual total number of positive samples.

[0079] S4. Based on the extracted vascular skeleton, the cross-point detection algorithm is used to identify vascular nodes.

[0080] After automatically extracting the vascular skeleton using a trained model, the extracted vascular skeleton is analyzed using an intersection detection algorithm to automatically identify and locate intersections in the vascular network, thereby achieving accurate detection of vascular nodes.

[0081] In some embodiments, Figure 5 This is a schematic diagram of identifying vascular nodes based on the vascular skeleton provided in an embodiment of the present invention. See also... Figure 5 S4 specifically includes the following sub-steps:

[0082] S41. Define the pixels on the center line of the vascular skeleton as skeleton points, and determine whether the current pixel is a skeleton point.

[0083] If yes, proceed to the next step; otherwise, move to the next pixel for judgment.

[0084] S42. Determine whether there are more than or equal to a preset number of skeleton points among the 8 adjacent pixels of the skeleton point.

[0085] If so, the skeleton point is determined to be a blood vessel node; otherwise, it is determined not to be a blood vessel node.

[0086] The preset number can be 3, meaning that at least 3 of the 8 adjacent pixels of the current skeleton point are also skeleton points. It can also be adjusted adaptively according to the actual situation, because intersections usually connect multiple blood vessel branches, so there will be multiple skeleton points in their neighborhood.

[0087] S5. Divide the ultra-wide-angle fundus fluorescence angiography image into regions, calculate the tortuosity value of the blood vessels in each region based on the vascular skeleton and vascular nodes, and obtain the analysis results of fundus blood vessel distribution and vascular morphology.

[0088] In some embodiments, Figure 6 This is a schematic diagram illustrating the region division of an ultra-wide-angle fundus fluorescein angiography image provided in an embodiment of the present invention. See also... Figure 6 This invention employs an extended ETDRS ring partitioning method to divide ultra-wide-angle fundus fluorescein angiography images into regions. The ETDRS (Early Treatment Diabetic Retinopathy Study) ring is a standardized partitioning method for assessing diabetic retinopathy. It divides the retinal region using concentric rings (typically in diameter ranges of 1°, 3°, and 6°) centered on the fovea and into four quadrants: superior, inferior, nasal, and temporal. Traditional ETDRS partitioning is limited by the imaging range (approximately 30°-50° field of view), while ultra-wide-angle fundus fluorescein angiography can cover up to 200° of field of view. Based on this, this invention extends the diameter of the ETDRS ring to 21 mm (almost covering the entire retina), dividing the fundus image into 21 standardized regions. This partitioning method retains the core logic of ETDRS (concentric circles + quadrant division) while adapting to the anatomical characteristics of ultra-wide-angle imaging, providing a quantitative standard for locating peripheral retinal features.

[0089] Vascular tortuosity (VT) is a quantitative indicator that measures the degree of curvature of a blood vessel's centerline. In some embodiments, the formula for calculating the VT value of a segment of the blood vessel's centerline C (defined as the segment between two adjacent vascular nodes) is as follows:

[0090] ;

[0091] Wherein, VT represents the tortuosity of a certain segment of blood vessel, L represents the actual length of the centerline of the blood vessel segment, that is, the arc length along the centerline of the blood vessel, and D represents the straight-line distance between the start and end points of the blood vessel segment.

[0092] For the vascular centerline C, discretize it as follows: Points P1, P2, ..., P n The actual length L of the blood vessel centerline can be calculated using the following formula:

[0093] ;

[0094] Among them, (x) j y j ) and (x j+1 y j+1 ( ) are points P j and P j+1 The coordinates are given by n, where n represents the number of discrete points and j represents the j-th discrete point.

[0095] The straight-line distance D of the vascular segment is the distance between the starting point P1 and the ending point P of the vascular segment. n Euclidean distance between them:

[0096] ;

[0097] Where (x1, y1) and (x n y n Points P1 and P2 are respectively. n The coordinates.

[0098] By calculating and analyzing VT values ​​and extracting vascular skeletons, the distribution characteristics of retinal blood vessels can be accurately obtained, providing an important basis for the diagnosis and research of ophthalmic diseases.

[0099] This invention performs vascular analysis based on ultra-wide-angle fundus fluorescence angiography images, effectively avoiding the limitations of traditional fundus imaging techniques, such as limited field of view and difficulty in comprehensively observing the peripheral retina. It enables more comprehensive capture of fundus vascular information. Targeted improvements are made to the Unet deep learning model, combining a depthwise separable residual convolution module and a multi-scale feature space pyramid module, effectively reducing computational load and enhancing feature representation capabilities. Compared to traditional image analysis methods, it exhibits higher accuracy and stronger stability. Furthermore, this invention optimizes the model's learning ability for different fundus vascular features through a joint loss function, resulting in good generalization ability across various fundus images. The vascular analysis results of this invention can provide important evidence for the early diagnosis and research of ophthalmic diseases, and are of great significance for improving the accuracy and efficiency of clinical ophthalmic diagnosis.

[0100] This invention also provides a vascular analysis system based on ultra-wide-angle fundus fluorescence angiography images. This system is used to execute the vascular analysis method based on ultra-wide-angle fundus fluorescence angiography images described in the above embodiments. The system includes the following modules:

[0101] The data acquisition module is used to acquire ultra-wide-angle fundus fluorescence angiography images and preprocess the ultra-wide-angle fundus fluorescence angiography images to obtain a sample library;

[0102] The dataset construction module is used to annotate fundus blood vessels based on images in the sample library and build a fundus blood vessel analysis training dataset.

[0103] The model building module is used to build the DSP-Unet model. The DSP-Unet model is trained using the fundus vascular analysis training dataset. The trained DSP-Unet model is used to automatically extract the vascular skeleton of the ultra-wide-angle fundus fluorescence angiography image to be tested.

[0104] The vascular node recognition module is used to identify vascular nodes based on the extracted vascular skeleton using a cross-point detection algorithm.

[0105] The analysis module is used to divide the ultra-wide-angle fundus fluorescence angiography image into regions. Based on the vascular skeleton and vascular nodes, it calculates the vascular tortuosity value in each region and obtains the analysis results of fundus vascular distribution and vascular morphology.

[0106] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the technical solutions of the embodiments of the present invention.

Claims

1. A method of blood vessel analysis based on ultra-wide angle fundus fluorescein angiography images, characterized in that, The method comprises the following steps: S1, obtaining an ultra-wide-angle fundus fluorescence angiography image, and preprocessing the ultra-wide-angle fundus fluorescence angiography image to obtain a sample library; S2, labeling fundus blood vessels based on images in the sample library, and constructing a fundus blood vessel analysis training data set; S3, constructing a DSP-Unet model, training the DSP-Unet model by using the fundus blood vessel analysis training data set, and automatically extracting a blood vessel skeleton of a to-be-tested ultra-wide-angle fundus fluorescence angiography image by using the trained DSP-Unet model; specifically comprising: based on the Unet model, replacing a convolution module of the Unet model with a deep separable residual convolution module, inlaying a multi-scale feature space pyramid module between a last deep separable residual convolution module and a first decoder, and constructing a DSP-Unet model; S4, identifying a blood vessel node by using a cross-point detection algorithm according to the extracted blood vessel skeleton; specifically comprising: S41, defining a pixel point on a center line of the blood vessel skeleton as a skeleton point, and judging whether the current pixel point is a skeleton point; if yes, the next step is continued to be executed, otherwise, the next pixel point is moved to be judged; S42, judging whether there are more than or equal to a preset number of skeleton points in adjacent eight pixels of the skeleton point; if yes, the skeleton point is determined to be a blood vessel node, otherwise, it is determined to be not a blood vessel node; S5, dividing a region of the ultra-wide-angle fundus fluorescence angiography image, calculating a blood vessel tortuosity value in each region according to the blood vessel skeleton and the blood vessel node, and obtaining an analysis result of fundus blood vessel distribution and blood vessel morphology.

2. The method of claim 1, wherein the method is based on ultra-widefield fundus fluorescein angiography images. In S1, the preprocessing comprises: artifact removal processing, contrast enhancement, grayscale correction, segmentation processing and data cleaning.

3. The method of claim 1, wherein the method is based on ultra-widefield fundus fluorescein angiography images. The loss function of the DSP-Unet model is a joint loss function, and the expression is as follows: ； wherein, L B represents a joint loss function, λ represents a weight coefficient, L Ce represents a cross-entropy loss function, L Dice represents a Dice coefficient loss function.

4. The method of claim 1, wherein the method is based on ultra-widefield fundus fluorescein angiography images. In S5, the blood vessel tortuosity value in each region is calculated, and the calculation formula is as follows: ； wherein VT represents the blood vessel tortuosity of a certain section of blood vessel, L represents the actual length of the blood vessel center line of the section of blood vessel, and D represents the straight line distance between the starting point and the ending point of the section of blood vessel. In S5, an extended ETDRS ring partition method is used to divide the region of the ultra-wide-angle fundus fluorescence angiography image.

5. The method of claim 1, wherein the method is based on ultra-widefield fundus fluorescein angiography images. The system comprises the following modules:

6. A system for blood vessel analysis based on ultra-widefield fundus fluorescein angiography images for implementing the method for blood vessel analysis based on ultra-widefield fundus fluorescein angiography images according to any one of the preceding claims 1 to 5, characterized in that, A data acquisition module is configured to obtain an ultra-wide-angle fundus fluorescence angiography image, and preprocess the ultra-wide-angle fundus fluorescence angiography image to obtain a sample library; A data set construction module is configured to label fundus blood vessels based on images in the sample library, and construct a fundus blood vessel analysis training data set; A model construction module is configured to construct a DSP-Unet model, train the DSP-Unet model by using the fundus blood vessel analysis training data set, and automatically extract a blood vessel skeleton of a to-be-tested ultra-wide-angle fundus fluorescence angiography image by using the trained DSP-Unet model; A blood vessel node identification module is configured to identify a blood vessel node by using a cross-point detection algorithm according to the extracted blood vessel skeleton; ​ An analysis module is configured to divide the ultra-wide-angle fundus fluorescence angiography image into regions, calculate a blood vessel tortuosity value in each region according to the blood vessel skeleton and the blood vessel nodes, and obtain an analysis result of fundus blood vessel distribution and blood vessel morphology.

Images