Blood vessel guided and multi-instance heat map regression for fundus vessel junction detection

Abstract

The application relates to a fundus blood vessel junction point detection algorithm based on blood vessel guidance and multi-instance heat map regression, and relates to the problem of fundus blood vessel junction point detection difficulty in retinal fundus blood vessel analysis. The retinal blood vessel morphology is related to the occurrence of various eye and whole body diseases. The retinal fundus blood vessel analysis is of great significance for diagnosing eye and whole body diseases. In the retinal fundus blood vessel analysis, the fundus blood vessel junction points (branching points and intersection points) are important references for blood vessel morphology feature analysis and quantification. However, due to the complex distribution of the fundus blood vessels, the number and spatial distribution of the junction points are unknown, which leads to the difficulty in feature extraction and expression of the junction points and the insufficient utilization of context information. In order to solve the problem, the application provides the fundus blood vessel junction point detection algorithm based on blood vessel guidance and multi-instance heat map regression. Experiments show that the method has the following advantages: (1) the junction point coordinate prediction is converted into multi-instance heat map regression, so that the problem of difficult representation of the junction point features caused by sparse label annotation is solved; (2) the blood vessel attention mechanism is used to enhance the attention of the model to the blood vessel area, the multi-resolution interaction mechanism is used to enhance the perception ability of the model to the fine blood vessels, and the detection precision of the junction points is improved. The application is applied to fundus blood vessel junction point detection.

Description

Technical Field

[0001] This invention relates to an algorithm for detecting fundus vascular junctions using vascular guidance and multi-instance heatmap regression. Background Technology

[0002] The retinal vascular network is the only structure in the human circulatory system that can be directly and non-invasively observed and photographed. Retinal vascular morphology is associated with the occurrence of various ocular and systemic diseases. Retinal vascular biomarkers, such as small artery stenosis, small venule widening, vascular tortuosity, and decreased arteriovenous ratio, are considered candidate predictors of major life-threatening factors. Therefore, analyzing the retinal vascular network is of great significance for assisting in clinical disease assessment and treatment. In retinal fundus vascular analysis, fundus vascular junctions (bifurcations and intersections) are important references for the analysis and quantification of vascular morphological characteristics. For example, by measuring the degree of abnormality at bifurcations and intersections in fundus lesions, doctors can accurately assess the severity of fundus lesions. However, manual junction detection is time-consuming and subject to subjective factors; therefore, automated junction detection methods are urgently needed in clinical research.

[0003] The complexity of retinal vascular distribution and the uncertainty of the number and spatial distribution of vascular connectors make connector detection difficult. Deep learning-based methods have demonstrated superior feature extraction capabilities, outperforming traditional methods. However, challenges remain, such as difficulties in connector feature extraction and representation, and insufficient utilization of contextual information. This paper proposes a retinal vascular connector detection algorithm guided by a vessel and based on multi-instance heatmap regression. First, the retinal image and connector label map are segmented into fixed-size patches. Then, a Gaussian kernel is used to transform the labeled patches into multi-instance heatmaps. Second, a vessel-guided retinal vascular connector detection network is constructed and trained. This network enhances the model's attention to vascular regions through vessel attention mechanisms and multi-resolution interaction mechanisms, improving connector detection accuracy. Finally, multi-instance heatmap regression is used to obtain connector heatmaps, and non-maximum suppression is applied to obtain the connector detection results. This method has been widely applied in the field of retinal vascular connector detection. Summary of the Invention

[0004] The purpose of this invention is to address the difficulty in detecting fundus vascular junctions due to the complexity of fundus vascular distribution and the uncertainty of the number and spatial distribution of vascular junctions. The proposed algorithm is a fundus vascular junction detection algorithm guided by vascular guidance and multi-instance heatmap regression.

[0005] The aforementioned objectives are primarily achieved through the following technical solutions:

[0006] S1. Construct a fundus blood vessel semantic segmentation model and train it to convergence using the fundus blood vessel semantic segmentation dataset. Use this model as a pre-trained model for parameter initialization.

[0007] S2. The retinal fundus images and their corresponding connection point labels in the fundus vessel connectivity detection dataset are segmented into patches of fixed size. A Gaussian kernel is then used to transform the patches with binary connection point labels into multi-instance heatmap label patches. The steps are as follows:

[0008] (1) Convert the annotation information of the connection points into a binary label map, and cut the retinal fundus image and the corresponding binary label map into patches of fixed size;

[0009] (2) Use a Gaussian kernel with a fixed window size to transform the binary labels of the connection points into a multi-instance heatmap. The Gaussian kernel function is shown in formula (1):

[0010]

[0011] Where (x,y) represents the relative coordinates of the connection point, σ represents the standard deviation of the Gaussian kernel, K(x,y;σ) represents the Gaussian kernel, and e represents the natural index.

[0012] S3. Construct a vessel-guided fundus vessel connectivity detection model. This model is a multi-task learning model that includes a fundus vessel semantic segmentation network and a fundus vessel connectivity detection network. The fundus vessel semantic segmentation network and the fundus vessel connectivity detection network have the same structure. The steps are as follows:

[0013] (1) The vascular-guided fundus vascular junction detection model utilizes the vascular attention mechanism to provide prior information on vascular location for fundus vascular junction detection. The process is as follows:

[0014] First, the fundus vascular segmentation network F is used. vessel The results of fundus vessel segmentation are shown in formula (2):

[0015]

[0016] in Let represent the fundus vessel segmentation network in the t-th iteration, and x represent the original input retinal fundus image. This represents the result of fundus vessel segmentation predicted in the t-th iteration;

[0017] Then, attention weights are assigned to the fundus vessel connection point detection network using the fundus vessel segmentation results of the t-th iteration to better highlight the fundus vessels. This process is shown in Equation (3):

[0018]

[0019] Where x att Retinal fundus images indicating enhanced vascular attention. This represents the fundus vessel segmentation result predicted in the t-th iteration, where x represents the original retinal fundus image. It represents the Hadamardi (or Hadama) stack;

[0020] The retinal fundus image with enhanced vascular attention is input into the fundus vascular connectivity detection network to obtain the fundus vascular connectivity heatmap detection results, as shown in formula (4):

[0021]

[0022] in Let x represent the fundus vessel connectivity detection network in the t-th iteration. att Retinal fundus images indicating enhanced vascular attention. This represents a heatmap of fundus vessel connectivity points predicted in the t-th iteration.

[0023] (2) The retinal vascular connection point detection model guided by blood vessels utilizes a multi-resolution interaction mechanism to better preserve the local features of retinal vascular vessels. The multi-resolution interaction mechanism adds multiple convolutional layers between the skip connections of the encoder and decoder of the network and fuses adjacent features. The feature fusion process of the multi-resolution interaction mechanism is shown in Equation (5):

[0024] f i,j =[F(f i,j-1 ),D(F(f i-1,j-1 )),U(F(f i+1,j-1 (5)

[0025] Where [·] represents the feature concatenation operation, D and U represent the downsampling and upsampling operations respectively, F represents multiple cascaded convolutional layers, and f i,j This represents the characteristics at different locations in the network, where i and j represent the i-th row and j-th column of the network, respectively, and f i,j-1 f i,j Left adjacent features, f i-1,j-1 f i,j Upper-level adjacent features, f i+1,j-1 f i,j Lower-level adjacent features;

[0026] In order to reduce f i,j The number of channels, f i,j The inputs are respectively fed into a b×b convolutional layer, an a×a convolutional layer, and an a×a convolutional layer with an inflation coefficient of 2. Then, the output features are added element by element to obtain the fused features. This process is shown in formula (6):

[0027]

[0028] Where x i,j f i,jFeatures that reduce the number of channels, BN represents batch normalization, C b×b C a×a , These represent b×b convolution, a×a convolution, and a×a convolution with an inflation factor of 2, respectively.

[0029] S4. Using a pre-trained fundus vessel semantic segmentation model, initialize the parameters of both the fundus vessel semantic segmentation network and the fundus vessel connection point detection network of the vessel-guided fundus vessel connection point detection model. Train the vessel-guided fundus vessel connection point detection model until convergence using the fundus vessel connection point detection dataset and the fundus vessel semantic segmentation dataset. The steps are as follows:

[0030] (1) The parameters of the fundus vascular semantic segmentation network and the fundus vascular connection point detection network of the vascular-guided fundus vascular connection point detection model were initialized using the pre-trained fundus vascular semantic segmentation model.

[0031] (2) The fundus vessel semantic segmentation dataset and the fundus vessel connectivity detection dataset were used to train the vessel-guided fundus vessel connectivity detection model. The loss function of the vessel-guided fundus vessel connectivity detection model includes vessel segmentation loss, heatmap regression loss and feature matching loss:

[0032] The loss function for fundus vessel segmentation is binary cross-entropy loss, and the vessel segmentation loss is shown in Equation (7):

[0033]

[0034] Where L vessel This represents the blood vessel segmentation loss, where N represents the total number of pixels. This represents the true value of the blood vessel at pixel n. This represents the predicted blood vessel value at pixel n;

[0035] The regression analysis of fundus vascular junction heatmaps used Dice loss as the loss function, as shown in formula (8):

[0036]

[0037] Where L heahmap Indicates the heatmap regression loss, y keypoint This represents the true value of the heatmap of the connection points. This represents the predicted value of the heatmap of the connection point, and |.| represents the modulo operation;

[0038] Feature matching loss is used to align the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network. The loss function is shown in Equation (9):

[0039] L feature =|fvessel -f keypoint |1 (9) where L feature Let f represent the feature matching loss, |·|1 represent the L1 loss, and f vessel with f keypoint These represent the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network, respectively.

[0040] The loss of the vessel-guided fundus vessel connectivity detection model is a weighted sum of vessel segmentation loss, heatmap regression loss, and feature matching loss, as shown in the following formula:

[0041] L model =αL vessel +βL heatmap +γL feature (10)

[0042] Where L model L represents the total loss of the vascular-guided fundus vascular junction detection model. vessel Indicates vessel segmentation loss, L heahmap L represents the heatmap regression loss. feature represents the feature matching loss, and α, β, and γ represent the loss balancing weights, respectively;

[0043] (3) The fundus vascular connection point detection model was trained to convergence using the fundus vascular segmentation dataset and the fundus vascular connection point detection dataset.

[0044] S5. Input the retinal fundus image into the vascular-guided fundus vascular connectivity detection model to obtain the connectivity heatmap regression results. Then, obtain the fundus vascular connectivity detection results through non-maximum suppression. The steps are as follows:

[0045] (1) Input the retinal fundus image into the fundus vessel connection point detection model guided by blood vessels to obtain the fundus vessel segmentation result and the fundus vessel connection point heatmap;

[0046] (2) Non-maximum suppression is applied to the connection point heatmap using an m×m window to filter out the connection point edges and duplicate predictions, thus obtaining the detection results of fundus vascular connection points.

[0047] Invention Effects

[0048] This invention provides a fundus vascular connection point detection algorithm guided by blood vessels and based on multi-instance heatmap regression. The method first segments the retinal fundus image and connection point label map into fixed-size patches, then uses a Gaussian kernel to transform the labeled patches into multi-instance heatmaps, increasing learnable information through this heuristic strategy. Secondly, a blood vessel-guided fundus vascular connection point detection network is constructed and trained. This network mainly consists of a blood vessel attention mechanism and a multi-resolution interaction mechanism. The blood vessel attention mechanism enhances the model's focus on vascular regions, providing effective attention allocation for connection point detection. The multi-resolution interaction mechanism better preserves image detail information, enhancing the model's ability to perceive minute blood vessels. Finally, a connection point heatmap is obtained using multi-instance heatmap regression, and the connection point detection results are obtained through non-maximum suppression. Experiments show that the advantages of this method are as follows: (1) It transforms connection point coordinate prediction into multi-instance heatmap regression, increasing the learnable information of the model through a heuristic strategy, thus solving the problem of difficulty in representing connection point features due to sparse label annotations. (2) The vascular-guided fundus vascular connection point detection network model uses the vascular attention mechanism to enhance the model’s attention to the vascular region and the multi-resolution interaction mechanism to enhance the model’s perception of fine blood vessels, thereby improving the accuracy of connection point detection. Attached Figure Description

[0049] Figure 1 This is a flowchart of the fundus vascular connection point detection algorithm using vascular guidance and multi-instance thermal regression in an example of the present invention.

[0050] Figure 2 This is a flowchart illustrating the generation of multi-instance heatmaps from connection point label annotations in an example of the present invention.

[0051] Figure 3 This is a structural diagram of the retinal vessel connection point detection model guided by blood vessels in an example of the present invention;

[0052] Figure 4 This is a structural diagram of the fundus blood vessel connection point detection network and the fundus blood vessel segmentation network in an example of the present invention;

[0053] Figure 5 This is a flowchart of the vascular attention mechanism in an example of the present invention;

[0054] Figure 6 This is a diagram illustrating the feature fusion process of the multi-resolution interaction mechanism in an example of the present invention. Specific implementation methods Specific implementation method one:

[0056] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0057] like Figure 1 The retinal vascular connectivity detection algorithm based on vascular guidance and multi-instance heatmap regression presented in this paper includes the following steps:

[0058] S1. Train the retinal fundus vessel semantic segmentation model FR-UNet using the DRIVE dataset. FR-UNet is a full-resolution network for vessel segmentation. After the model converges, save its model parameters and use them as a pre-trained model for parameter initialization.

[0059] S2. The retinal fundus images and their corresponding connection point labels in the fundus vessel connectivity detection dataset DRIVE are cropped into patches of 48×48 size. Then, a Gaussian kernel window with a window size of 7×7 is used to transform the patch of binary label map of the connection points into a multi-instance heatmap label patch. The fundus vessel connectivity detection dataset DRIVE is a connectivity annotation provided by researchers based on the initial fundus vessel semantic segmentation dataset DRIVE. The two have the same retinal fundus images.

[0060] S3. Construct a vascular-guided fundus vascular junction detection model. This model is a multi-task learning model that includes a fundus vascular semantic segmentation network and a fundus vascular junction detection network. The fundus vascular semantic segmentation network and the fundus vascular junction detection network have the same network structure as FR-UNet.

[0061] S4. Using the pre-trained fundus vessel semantic segmentation model FR-UNet, initialize the parameters of the fundus vessel semantic segmentation network and the fundus vessel connector network of the fundus vessel guided fundus vessel connector model, respectively. Train the fundus vessel guided fundus vessel connector detection model until convergence using the fundus vessel connector detection dataset DRIVE and the fundus vessel semantic segmentation dataset DRIVE.

[0062] S5. Input the retinal fundus image to be detected into the vascular-guided fundus vascular junction model to obtain the junction heatmap regression results. Use nonmaximum suppression with a window size of 7×7 to obtain the fundus vascular junction detection results.

[0063] The following is a detailed description of examples of the present invention:

[0064] The specific implementation of this invention is as follows:

[0065] S1, such as Figure 1 As shown, the FR-UNet retinal fundus vessel semantic segmentation model is trained using the DRIVE dataset. FR-UNet is a full-resolution network for vessel segmentation. After the model converges, its model parameters are saved and used as a pre-trained model for parameter initialization.

[0066] S2, such as Figure 2 As shown, retinal fundus images and their corresponding connection point labels from the DRIVE fundus vessel connectivity detection dataset are cropped into 48×48 patches. Then, a 7×7 Gaussian kernel window is used to transform the patch of binary label images of the connectivity points into a multi-instance heatmap label patch. The DRIVE fundus vessel connectivity detection dataset is a connectivity annotation dataset provided by researchers based on the initial DRIVE fundus vessel semantic segmentation dataset. Both datasets use the same retinal fundus images. The steps are as follows:

[0067] (1) Read the coordinate information of the connection point and create a binary label map of the same size as the corresponding retinal fundus image. Assign the value of 255 to the position of the connection point and 0 to the other positions. Cut the retinal fundus image and the corresponding binary label map into patches of size 48×48.

[0068] (2) Use a Gaussian kernel with a window size of 7×7 to convert the binary labels of the connection points into a multi-instance heatmap. The Gaussian kernel function used is shown in formula (1):

[0069]

[0070] Where (x,y) represents the relative coordinates of the connection point, σ represents the standard deviation of the Gaussian kernel, K(x,y;σ) represents the Gaussian kernel, and e represents the natural index.

[0071] S3. Construct a vessel-guided fundus vessel connectivity detection model. This model is a multi-task learning model that includes a fundus vessel semantic segmentation network and a fundus vessel connectivity detection network. The model structure is as follows: Figure 3 As shown, the fundus vessel semantic segmentation network and the fundus vessel connection point detection network have the same network structure as FR-UNet, as shown in the figure. Figure 4 As shown, the steps are as follows:

[0072] (1) As Figure 5 As shown, the vascular-guided fundus vascular connectivity detection model utilizes the vascular attention mechanism to provide prior information on vascular location for fundus vascular connectivity detection. The process is as follows:

[0073] First, the fundus vascular segmentation network F is used. vesselThe results of fundus vessel segmentation are shown in formula (2):

[0074]

[0075] in Let represent the fundus vessel segmentation network in the t-th iteration, and x represent the original input retinal fundus image. Let represent the results of fundus vessel segmentation predicted in the t-th iteration;

[0076] Then, attention weights are assigned to the fundus vessel connection point detection network using the fundus vessel segmentation results of the t-th iteration to better highlight the fundus vessels. This process is shown in Equation (3):

[0077]

[0078] Where x att Retinal fundus images indicating enhanced vascular attention. This represents the fundus vessel segmentation result predicted in the t-th iteration, where x represents the original retinal fundus image. It represents the Hadamardi (or Hadama) stack;

[0079] The retinal fundus image with enhanced vascular attention is input into the fundus vascular connectivity detection network to obtain the fundus vascular connectivity heatmap detection results, as shown in formula (4):

[0080]

[0081] in Let x represent the fundus vessel connectivity detection network in the t-th iteration. att Retinal fundus images indicating enhanced vascular attention. This represents a heatmap of fundus vessel connectivity points predicted in the t-th iteration.

[0082] (2) The retinal vascular connection point detection model guided by blood vessels utilizes a multi-resolution interaction mechanism to better preserve the local features of retinal vascular vessels. The multi-resolution interaction mechanism adds multiple convolutional layers between the skip connections of the encoder and decoder of the network and fuses adjacent features. The feature fusion process of the multi-resolution interaction mechanism is shown in Equation (5):

[0083] f i,j =[F(f i,j-1 ),D(F(f i-1,j-1 )),U(F(f i+1,j-1 (5) where [·] represents the feature concatenation operation, D and U represent the downsampling and upsampling operations respectively, F represents two 3×3 convolutional layers, and f i ,jThis represents the characteristics at different locations in the network, where i and j represent the i-th row and j-th column of the network, respectively, and f i,j-1 f i,j Left adjacent features, f i-1,j-1 f i,j Upper-level adjacent features, f i+1,j-1 f i,j Lower-level adjacent features;

[0084] like Figure 5 As shown, in order to reduce f i,j The number of channels, f i,j The inputs are respectively fed into a 1×1 convolutional layer, a 3×3 convolutional layer, and a 3×3 convolutional layer with an inflation coefficient of 2. Then, the output features are added element by element to obtain the fused features. This process is shown in formula (6):

[0085]

[0086] Where x i,j f i,j Features that reduce the number of channels, BN represents batch normalization, C 1×1 C 3×3 , These represent 1×1 convolution, 3×3 convolution, and 3×3 convolution with an expansion factor of 2, respectively.

[0087] S4. Using the pre-trained fundus vessel semantic segmentation model FR-UNet, initialize the parameters of both the fundus vessel semantic segmentation network and the fundus vessel connector network of the vessel-guided fundus vessel connector model. Train the vessel-guided fundus vessel connector detection model to convergence using the fundus vessel connector detection dataset DRIVE and the fundus vessel segmentation dataset DRIVE. The steps are as follows:

[0088] (1) The parameters of the fundus vascular semantic segmentation network and the fundus vascular connection point detection network of the vascular-guided fundus vascular connection point detection model were initialized using the pre-trained fundus vascular semantic segmentation model FR-UNet.

[0089] (2) The fundus vessel semantic segmentation dataset DRIVE and the fundus vessel connectivity detection dataset DRIVE were used to train the vessel-guided fundus vessel connectivity detection model. The loss function of the vessel-guided fundus vessel connectivity detection model includes vessel segmentation loss, heatmap regression loss and feature matching loss:

[0090] The loss function for fundus vessel segmentation is binary cross-entropy loss, and the vessel segmentation loss is shown in Equation (7):

[0091]

[0092] Where Lvessel This represents the blood vessel segmentation loss, where N represents the total number of pixels. This represents the true value of the blood vessel at pixel n. This represents the predicted blood vessel value at pixel n;

[0093] The regression analysis of fundus vascular junction heatmaps used Dice loss as the loss function, as shown in formula (8):

[0094]

[0095] Where L heahmap Indicates the heatmap regression loss, y keypoint This represents the true value of the heatmap of the connection points. This represents the predicted value of the heatmap of the connection point, and |.| represents the modulo operation;

[0096] Feature matching loss is used to align the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network. The loss function is shown in Equation (9):

[0097] L feature =|f vessel -f keypoint |1 (9) Where L feature Let f represent the feature matching loss, |·|1 represent the L1 loss, and f vessel with f keypoint These represent the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network, respectively. Figure 4 The first row of feature maps shown;

[0098] The loss of the vessel-guided fundus vessel connectivity detection model is a weighted sum of vessel segmentation loss, heatmap regression loss, and feature matching loss, as shown in the following formula:

[0099] L model =αL vessel +βL heatmap +γL feature (10)

[0100] Where L model L represents the total loss of the vascular-guided fundus vascular junction detection model. vessel Indicates vessel segmentation loss, L heahmap L represents the heatmap regression loss. feature Let α represent the feature matching loss, and β and γ represent the loss balancing weights, where α = 1, β = 5, and γ = 1.

[0101] (3) The fundus vascular connection point detection model was trained to convergence using the DRIVE fundus vascular connection point detection dataset and the DRIVE fundus vascular segmentation dataset.

[0102] S5. Input the retinal fundus image to be detected into the vascular-guided fundus vascular connectivity model to obtain the connectivity heatmap regression results. Then, obtain the fundus vascular connectivity detection results through nonmaximum suppression with a window size of 7×7. The steps are as follows:

[0103] (1) Input the retinal fundus image into the fundus vessel connection point detection model guided by blood vessels to obtain the fundus vessel segmentation result and the fundus vessel connection point heatmap;

[0104] (2) Non-maximum suppression was applied to the connection point heatmap using a 7×7 window to filter out the connection point edges and duplicate predictions, thus obtaining the detection results of fundus vascular connection points.

[0105] This invention may have other embodiments. Without departing from the spirit and essence of this invention, those skilled in the art can make various corresponding changes and modifications according to this invention, but these corresponding changes and modifications should all fall within the scope of this invention.

Claims

1. A fundus vascular connectivity detection algorithm guided by blood vessels and based on multi-instance heatmap regression, characterized in that, It includes the following steps: S1. Construct a fundus blood vessel semantic segmentation model and train it to convergence using the fundus blood vessel semantic segmentation dataset. Use it as a pre-trained model for parameter initialization. S2. Cut the retinal fundus images and their corresponding connection point labels of the fundus vascular junction detection dataset into patches of fixed size, and use a Gaussian kernel to transform the patch of binary connection point labels into a multi-instance heatmap label patch. S3. Construct a vascular-guided fundus vascular junction detection model. This model is a multi-task learning model that includes a fundus vascular semantic segmentation network and a fundus vascular junction detection network. The fundus vascular semantic segmentation network and the fundus vascular junction detection network have the same structure. S31. The vascular-guided fundus vascular junction detection model utilizes the vascular attention mechanism to provide prior information on vascular location for fundus vascular junction detection. The process is as follows: First, the network of blood vessels in the fundus is segmented. The results of fundus vessel segmentation are shown in formula (1): (1) wherein denotes the fundus blood vessel segmentation network of the t-th iteration, x denotes the input original retinal fundus image, denotes the fundus blood vessel segmentation result predicted by the t-th iteration Then, attention weights are assigned to the fundus vessel connection point detection network using the fundus vessel segmentation results of the t-th iteration to better highlight the fundus vessels. This process is shown in Equation (2): (2) wherein a retinal fundus image representing blood vessel attention enhancement, a fundus blood vessel segmentation result of the t-th iteration prediction, x represents an original retinal fundus image, represents a Hadamard product; The retinal fundus image with enhanced vascular attention is input into the fundus vascular connectivity detection network to obtain the fundus vascular connectivity heatmap detection results, as shown in formula (3): (3) wherein denotes the fundus blood vessel junction point detection network of the tth iteration, denotes the retinal fundus image with enhanced blood vessel attention, denotes the predicted fundus blood vessel junction heat map of the tth iteration; S32: The vascular-guided fundus vascular connection point detection model utilizes a multi-resolution interaction mechanism to better preserve the local features of retinal and fundus vessels. The multi-resolution interaction mechanism adds multiple convolutional layers between the skip connections of the encoder and decoder of the network and fuses adjacent features. The feature fusion process of the multi-resolution interaction mechanism is shown in Equation (4): (4) in This indicates a feature concatenation operation. and These represent downsampling and upsampling operations, respectively. This represents multiple cascaded convolutional layers. This represents the characteristics at different locations in the network, where i and j represent the i-th row and j-th column of the network, respectively. express Left adjacent features, express Upper-level adjacent features, express Lower-level adjacent features; In order to reduce The number of channels, The inputs are respectively fed into a b×b convolutional layer, an a×a convolutional layer, and an a×a convolutional layer with an inflation coefficient of 2. Then, the output features are added element by element to obtain the fused features. This process is shown in formula (5): (5) in express Features that reduce the number of channels, BN represents batch normalization. , , These represent b×b convolution, a×a convolution, and a×a convolution with an inflation factor of 2, respectively. S4. Using the pre-trained fundus blood vessel semantic segmentation model, initialize the parameters of the fundus blood vessel semantic segmentation network and the fundus blood vessel connection point detection network of the blood vessel-guided fundus blood vessel connection point detection model, respectively. Train the blood vessel-guided fundus blood vessel connection point detection model until convergence using the fundus blood vessel connection point detection dataset and the fundus blood vessel semantic segmentation dataset. S5. Input the retinal fundus image into the vascular-guided fundus vascular junction detection model to obtain the junction heatmap regression results, and obtain the fundus vascular junction detection results through nonmaximum suppression.

2. The fundus vascular connectivity detection algorithm based on vascular guidance and multi-instance thermal regression as described in claim 1, characterized in that, Step S2 involves cutting the retinal fundus images and their corresponding connection point labels from the fundus vessel connectivity detection dataset into fixed-size patches, and then using a Gaussian kernel to transform the patch of binary connection point labels into a multi-instance heatmap label patch. This includes the following steps: S21. Convert the annotation information of the connection points into a binary label map, and cut the retinal fundus image and the corresponding binary label map into patches of fixed size; S22. Use a Gaussian kernel with a fixed window size to transform the binary labels of the connection points into a multi-instance heatmap. The Gaussian kernel function is shown in formula (6): (6) in Indicates the relative coordinates with respect to the connection point. The standard deviation of the Gaussian kernel is represented. Represents the Gaussian kernel. This represents the natural index.

3. The fundus vascular connectivity detection algorithm based on vascular guidance and multi-instance thermal regression as described in claim 1, characterized in that, Step S4 uses a pre-trained fundus vessel semantic segmentation model to initialize the parameters of both the fundus vessel semantic segmentation network and the fundus vessel connection point detection network of the vessel-guided fundus vessel connection point detection model. Training the vessel-guided fundus vessel connection point detection model to convergence using the fundus vessel connection point detection dataset and the fundus vessel semantic segmentation dataset includes the following steps: S41. Use the pre-trained fundus blood vessel semantic segmentation model to initialize the parameters of the fundus blood vessel semantic segmentation network and the fundus blood vessel connection point detection network of the blood vessel-guided fundus blood vessel connection point detection model, respectively. S42. The fundus vessel semantic segmentation dataset and the fundus vessel connectivity detection dataset are used to train the vessel-guided fundus vessel connectivity detection model. The loss function of the vessel-guided fundus vessel connectivity detection model includes vessel segmentation loss, heatmap regression loss, and feature matching loss. The loss function for fundus vessel segmentation is binary cross-entropy loss, and the vessel segmentation loss is shown in Equation (7): (7) in This represents the blood vessel segmentation loss, where N represents the total number of pixels. This represents the true value of the blood vessel at pixel n. This represents the predicted blood vessel value at pixel n; The regression analysis of fundus vascular junction heatmaps used Dice loss as the loss function, as shown in formula (8): (8) in Indicates the heatmap regression loss. This represents the true value of the heatmap of the connection points. Represents the predicted value of the heatmap of the connection point, and | . | represents the modulo operation; Feature matching loss is used to align the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network. The loss function is shown in Equation (9): (9) in Represents feature matching loss, Indicates L1 loss, and These represent the intermediate layer features of the fundus vessel segmentation network and the fundus vessel connection point detection network, respectively. The loss of the vessel-guided fundus vessel connectivity detection model is a weighted sum of vessel segmentation loss, heatmap regression loss, and feature matching loss, as shown in the following formula: (10) in This represents the total loss of the vascular-guided fundus vascular junction detection model. Indicates loss of blood vessel segmentation. Indicates the heatmap regression loss. Represents feature matching loss, , , These represent the loss balance weights; S43. Train the vascular-guided vascular connection point detection model to convergence using the retinal vascular connection point detection dataset and the retinal vascular segmentation dataset.

4. The fundus vascular connectivity detection algorithm based on vascular guidance and multi-instance thermal regression as described in claim 1, characterized in that, Step S5 inputs the retinal fundus image into the vascular-guided fundus vessel connectivity detection model to obtain the connectivity heatmap regression results. The fundus vessel connectivity detection results are obtained through nonmaximum suppression, which includes the following steps: S51. Input the retinal fundus image into the fundus vessel connection point detection model guided by blood vessels to obtain the fundus vessel segmentation result and the fundus vessel connection point heatmap. S52. Use an m×m window to perform nonmaximum suppression on the connection point heatmap to filter out connection point edges and duplicate predictions, and obtain the detection results of fundus vascular connection points.

Images