A retinal blood vessel segmentation method combining reinforced skeleton and maximum neighborhood skeleton
By constructing a retinal vessel segmentation method that combines a joint enhanced skeleton and a maximum neighborhood skeleton, the problem of discontinuity and incoherence in retinal vessel segmentation of deep learning models is solved, achieving high-precision vessel segmentation and supporting automated disease diagnosis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TIANJIN UNIV
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-05
AI Technical Summary
Existing deep learning models suffer from problems such as segmentation breaks, discontinuous fine vessels, and missed detection of terminal vessels in retinal vessel segmentation tasks, making it difficult to meet the requirements of high-precision automated segmentation.
A method for retinal vessel segmentation combining a joint enhanced skeleton and a maximum neighbor skeleton is proposed. By explicitly modeling the axial centerline and local radial scale of the vessels, the enhanced skeleton and the maximum neighbor skeleton are constructed. The difference between the two is used to construct a fine vessel constraint map, and a segmentation network guided by multiple prior supervision signals is designed.
It significantly reduces fragmentation during segmentation, improves the predictive coherence of small vessels and the overall structural integrity, provides high-precision retinal vessel segmentation results, and supports automated quantitative analysis of diseases such as diabetic retinopathy and hypertension.
Smart Images
Figure CN122156244A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of medical image processing and artificial intelligence, and particularly relates to a method for retinal vessel segmentation that combines enhanced skeleton and maximum neighborhood skeleton. Background Technology
[0002] As the only microvascular system in the human body that can be directly and non-invasively observed, the retinal vessels, with their changes in diameter, tortuosity, and vascular connectivity, are key biomarkers for diagnosing diabetic retinopathy, hypertension, and cardiovascular diseases. Therefore, accurately segmenting the retinal vascular network from fundus images is a crucial prerequisite for achieving quantitative analysis and intelligent diagnosis of these diseases.
[0003] However, retinal vascular structures have the following inherent characteristics: (1) large scale span: containing branches of varying thicknesses, from large trunks to thin terminal capillaries, with significant differences in diameter; (2) complex and delicate structure: the connectivity of the vascular network is the basis for assessing hemodynamics, and the rupture of small vessels will lead to subsequent analysis bias; (3) low imaging contrast: affected by imaging conditions, the contrast of small vessels is extremely low, making them difficult to distinguish from the background. These characteristics make high-precision automated segmentation extremely challenging.
[0004] Traditional manual annotation methods heavily rely on expert experience, resulting in limitations such as being time-consuming, labor-intensive, subjective, and lacking repeatability, making it difficult to meet the standardization requirements of large-scale clinical screening and longitudinal studies. In recent years, deep convolutional neural networks (CNNs), represented by U-Net, have made significant progress in medical image segmentation due to their superior nonlinear representation capabilities. However, in the specific task of retinal vessel segmentation, existing technologies still have the following shortcomings:
[0005] (1) Segmentation breaks and poor connectivity: Mainstream algorithms usually treat blood vessel segmentation as a simple pixel-level classification task, lacking explicit constraints on the tubular structure of blood vessels. During the training process, the model is easily dominated by the main blood vessel features with a large pixel ratio, which often leads to severe segmentation breaks in the prediction of low-contrast and structurally fragile capillaries, thus destroying the connectivity of the blood vessel network.
[0006] (2) Low accuracy in identifying fine blood vessels: Retinal blood vessels have distinct characteristics of coarseness and fineness. Existing methods usually treat all blood vessel structures as a single semantic category and do not explicitly distinguish the geometric features of coarse and fine blood vessels in the algorithm architecture. This makes it difficult for the model to balance the accuracy of coarse blood vessels with the completeness of fine blood vessel prediction, resulting in fragmented fine blood vessel segmentation.
[0007] (3) Sparsity of skeleton supervision signal: Although some studies have attempted to introduce the centerline of the blood vessel skeleton as an auxiliary supervision, it usually relies on the signal of a single pixel width extracted from the label. This supervision method has significant drawbacks: on the one hand, the signal is too sparse and gradient guidance is easily lost during feature downsampling; on the other hand, single pixel supervision lacks modeling of the direction of blood vessel extension and the range of spatial influence, which leads to low learning efficiency of the model for the axial extension trend of blood vessels and is prone to segmentation breakage.
[0008] Therefore, how to combine the geometric properties of retinal vessels to construct a joint skeleton prior and design an explicit supervision framework that can maintain the segmentation accuracy of the main vessels while reducing the segmentation of the fine vessels has become a key problem that urgently needs to be solved in the field of precision medical image analysis.
[0009] This invention proposes a method for retinal vessel segmentation combining an enhanced skeleton and a maximum neighborhood skeleton, relating to deep learning-based medical image segmentation techniques. The invention constructs an enhanced skeleton and a maximum neighborhood skeleton by explicitly modeling the axial centerline and local radial scale of the vessels, and uses their difference to construct a fine vessel constraint map, thereby guiding the segmentation network to perform semantic decoupling and feature alignment mapping within a multi-level representation space. This invention effectively addresses the technical bottlenecks of existing deep neural networks in retinal vessel segmentation tasks, such as segmentation fragmentation, discontinuous fine vessels, and missed detections of terminal segments. Summary of the Invention
[0010] The purpose of this invention is to provide a retinal vessel segmentation method that combines enhanced skeleton and maximum neighborhood skeleton to solve the problems of segmentation breakage, discontinuity of fine vessels, and missed detection of terminal vessels in the existing deep learning models mentioned in the background art.
[0011] To achieve the above objectives, the present invention employs the following technical solution: In its first aspect, this invention proposes a method for retinal vessel segmentation that combines the enhanced skeleton and the maximum neighbor skeleton, comprising the following steps: S1. Data preparation and preprocessing: Select publicly available retinal vessel datasets and preprocess the images in the publicly available retinal vessel datasets as training sets; S2. Construct multiple skeleton prior supervision signals; Explicitly extract axial and radial geometric properties from the binary blood vessel label image to obtain a local radial scale map and a single-pixel wide axial centerline skeleton map; Based on the local radial scale map and the single-pixel wide axial centerline skeleton map, construct an enhanced skeleton map, a maximum neighborhood skeleton map, and a fine blood vessel constraint map constructed by calculating the difference between the two, as triple prior supervision signals. S3. Construct a segmentation network; the segmentation network is constructed based on an encoder-decoder architecture, including a shared encoder, a semantic decoupling module (SDM), and a master decoder; the shared encoder adopts a multi-level downsampling structure to extract multi-scale features from the input image; the semantic decoupling module is located at the end of the shared encoder and includes two independent pathways to extract the connectivity features and scale features of blood vessels, respectively, and map them to generate an enhanced skeleton prediction map and a maximum neighborhood skeleton prediction map; the master decoder gradually restores the spatial resolution through upsampling and skip connections to generate the final blood vessel segmentation prediction map; S4. Design a multi-prior joint loss function; utilize triple prior supervision signals to impose multi-dimensional constraints on the segmentation network, and design the main segmentation loss, skeleton enhancement loss, maximum neighborhood skeleton loss, and fine vessel enhancement loss as loss functions. The main segmentation loss... Using the original vascular label image Supervise the output of the master decoder Strengthening the skeleton loss use Supervision Force the model to learn the long-range connectivity of blood vessels; maximum neighborhood skeleton loss. use Supervision To achieve scale-differentiated learning; microvascular enhancement loss constrained microvascular diagram Spatially aligned with shallow features of the encoder, a parameter-aware weight allocation mechanism based on the Frobenius norm is introduced to force the network to capture fine vascular details at the feature level through weighted calculation; the joint loss function is obtained by weighted summation of the four loss components. S5. Segmentation Network Training: Within the deep learning framework, a network model is written using the Python programming language, and image data and three corresponding prior signals are input into the network. The segmentation network is trained using backpropagation with a joint loss function, and the weight parameters with the best performance are saved. S6. Retinal vessel segmentation: The trained segmentation network is used to process the fundus image to obtain a vessel probability map. The vessel probability map is then subjected to threshold binarization to obtain the vessel segmentation result.
[0012] Through the above technical solution, this invention can achieve high-precision segmentation of retinal vessels based on multiple skeleton prior guidance. The generated segmentation results show a significant reduction in fine vessel breaks, high integrity of the vascular network, and accurate reconstruction of vessel thickness variations, extension directions, and connection relationships.
[0013] This solution not only provides high-precision segmentation results for automated screening of fundus diseases, but can also be extended to image segmentation tasks of other organs with typical tubular geometric features, such as the coronary arteries and airways. By reducing reliance on manual high-precision annotation, this invention improves the reliability of automated diagnostic systems and has broad clinical application value.
[0014] Preferably, the data preparation and preprocessing in S1 are as follows: We used the publicly available retinal vascular datasets CHASE_DB1, DRIVE, and STARE, and divided them into training and testing sets. We used a computer to read retinal fundus image data and performed augmentation processing on the training set data to improve the model's generalization ability.
[0015] Preferably, in step S2, multiple skeleton prior supervision signals are constructed as follows: Based on binary vascular tagging The distance from each pixel in the blood vessel region to the nearest boundary is calculated using Euclidean distance transform (EDT), and a local radial scale map is extracted. The Zhang-Suen algorithm was used to create binary label images of blood vessels. Refine the drawing and extract the centerline skeleton. Centerline skeleton diagram Introducing maximum trust distance For each skeleton point, a trust neighborhood is constructed to build a reinforced skeleton graph. Based on local radial scale map Introducing a scale-distinguishing threshold For each skeleton point, construct a neighborhood range and build a maximum neighborhood skeleton graph. Extract the difference between the enhanced skeleton and the maximum neighbor skeleton to construct a fine blood vessel constraint map. .
[0016] Preferably, the extraction of the local radial scale map The details are as follows: Based on binary vascular tag map For any pixel position Local radial scale map extracted by Euclidean distance transformation :
[0017] in, Represents the set of pixels in the blood vessel region. Represents the set of background pixels; Indicates the position of the corresponding pixel Distance to the nearest boundary.
[0018] Preferably, the extraction of the centerline skeleton diagram The details are as follows: The Zhang-Suen algorithm was used to process the binary blood vessel labeling image. Skeletonization is performed by deleting boundary pixels through multiple rounds of parallel iterations, resulting in a centerline skeleton map that maintains the connectivity of the original region and the single-pixel width of key structural points. .
[0019] Preferably, the construction of the reinforced skeleton diagram The details are as follows: For any skeleton point Its trusted neighborhood is defined as:
[0020]
[0021] The enhanced skeleton graph is obtained by taking the union of the trusted neighborhoods of all skeleton points. :
[0022] in, Centerline skeleton diagram The set of centerline points with a width of one pixel, composed of non-zero pixels.
[0023] Preferably, the construction of the maximum neighborhood skeleton graph The details are as follows: Based on local radial scale map Introducing a scale-distinguishing threshold Calculate the trust distance for each skeleton point:
[0024] For capillaries, the supervised region is equivalent to their actual vascular morphology; for coarse vessels, the supervised region is normalized to a region with a width of [missing information]. A strip-shaped area of a pixel; Based on trust distance Constructing the neighborhood range:
[0025]
[0026] Obtain the union of the neighborhood ranges of all skeleton points to obtain an initial supervised region. ;Will With binary vascular labeling Taking the intersection, we obtain the final maximum neighborhood skeleton graph:
[0027] in, This represents the maximum neighborhood skeleton graph.
[0028] Preferably, the segmentation network is constructed in S3 as follows: The shared encoder adopts a standard U-Net design; the main decoder adopts a standard U-Net decoding structure. The semantic decoupling module includes two structurally symmetrical but parameter-independent pathways, used to extract axial connectivity features and radial scale features of blood vessels from deep semantic information, respectively; each pathway contains two layers. Convolutional units, followed by a layer The convolutional unit compresses the feature dimension to a preset embedding dimension; each convolutional unit contains convolution, batch normalization, and ReLU activation functions; the input to the semantic decoupling module is the deepest feature map extracted by the shared encoder. Output feature maps that encode the axial connectivity and radial scale information of blood vessels. ; The two feature maps are fed into lightweight auxiliary prediction heads, each consisting of multiple convolutional-BN-ReLU layers. These layers progressively upsample the features and map them to a single-channel probability map with the same resolution as the input image, resulting in the enhanced skeleton prediction map. And the maximum neighborhood skeleton prediction map .
[0029] Preferably, the shared encoder and master decoder are as follows: The shared encoder contains five levels of downsampling operations, each consisting of a double convolutional block, which is composed of two consecutive... The system consists of convolutional layers, each followed by batch normalization and the ReLU activation function; the first... The level encoder output feature map is ; The main decoder gradually recovers the spatial resolution through bilinear interpolation and incorporates encoder features through skip connections. Channel concatenation is performed with the upsampled features of the corresponding level; Level decoding features are Finally, the shallowest feature map of the main decoder is obtained. ,pass Convolutional layers map features to blood vessel segmentation prediction maps. .
[0030] Preferably, the main segmentation loss in S4 is as follows: Main segmentation loss Using the original binary vascular label image Blood vessel segmentation prediction map output by the supervised master decoder Specifically: Using binary cross-entropy With Dice coefficient loss The combination of is denoted as For the prediction graph and real label images , The calculation is as follows:
[0031] in:
[0032]
[0033] in, N Total number of pixels For the first The predicted probability value of each pixel. For the first The actual label value of each pixel. This is the smoothing constant.
[0034] Preferably, the enhanced skeleton loss and the maximum neighborhood skeleton loss in S4 are specifically as follows: Strengthening the skeleton loss Using the same combined loss function, the enhanced skeleton prediction map output by the semantic decoupling module is supervised. :
[0035] Maximum Neighborhood Skeleton Loss Using the same combined loss function, the maximum neighborhood skeleton prediction map output by the supervised semantic decoupling module is compared. :
[0036] in, Represents binary cross-entropy With Dice coefficient loss The combination of .
[0037] Preferably, the loss of microvascular enhancement in S4 is specifically as follows: Let the feature maps output from the first three layers of the shared encoder be respectively , , Firstly, through a lightweight projection head The features from each layer are mapped to a unified single-channel space to obtain the feature embedding. :
[0038] Each projection head By one Convolutional layers are implemented, followed by batch normalization and ReLU activation; this yields the embedded features. , , ; Accordingly, the fine blood vessel constraint diagram Downsampling is performed using bilinear interpolation to match the spatial dimensions of each layer's features:
[0039] Let the first The parameter set of each encoder module is as follows Its adaptive weights The calculation is as follows:
[0040] in, Denotes the Frobenius norm. To prevent division by zero for small constants; the weights satisfy the normalization condition. ; Ultimately, microvascular enhancement loss The weighted sum of the losses at each layer:
[0041] in, Represents binary cross-entropy With Dice coefficient loss The combination of .
[0042] In a second aspect, the present invention proposes a retinal vessel segmentation system that combines a reinforced framework with a maximum neighborhood framework, including a segmentation network; The segmentation network includes a shared encoder, a semantic decoupling module, and a main decoder.
[0043] Compared with the prior art, the beneficial effects of the present invention are: The method in this invention extracts axial and radial geometric features from binary blood vessel labels, constructs a reinforced skeleton and a maximum neighborhood skeleton, and uses the difference between the two to construct a fine blood vessel constraint map, thereby establishing explicit prior constraints within a deep learning framework.
[0044] By combining multiple prior signals, the segmentation network's ability to perceive low-contrast fine blood vessels is enhanced. This significantly reduces segmentation interruptions while maintaining the segmentation accuracy of main blood vessels, improving the predictive coherence and overall structural integrity of small blood vessels. Consequently, it provides reliable, high-precision segmentation results for automated quantitative analysis of diseases such as diabetic retinopathy and hypertension. Attached Figure Description
[0045] Figure 1This is a structural block diagram of the retinal vascular segmentation network that combines the enhanced framework and the maximum neighbor framework in this invention. Figure 2 The following are the segmentation results of this invention on different datasets ( Figure 2 (a) shows the segmentation results of the CHASE_DB1 dataset. Figure 2 (b) shows the segmentation results of the DRIVE dataset. Figure 2 (c) is a diagram showing the segmentation results of the STARE dataset. Detailed Implementation
[0046] 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, and 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.
[0047] Example 1: A method for retinal vessel segmentation combining enhanced scaffold and maximum neighbor scaffold includes the following steps: Step 1: Data preparation and preprocessing.
[0048] This invention uses the publicly available retinal vascular datasets STARE, DRIVE, and CHASE_DB1, all of which are equipped with manually annotated medical gold standard images.
[0049] The CHASE_DB1 dataset contains 28 retinal images, collected from the eyes of 14 child subjects, with a resolution of 999×960 pixels per image. Since this dataset does not pre-divide into training and testing sets, this invention employs a random partitioning method, selecting 22 images for model training and the remaining 6 for performance testing.
[0050] The DRIVE dataset contains 40 fundus images, including 7 from diseased eyes and 33 from healthy subjects, with a resolution of 768×584 pixels per image. This invention follows the experimental protocol of this dataset: 20 images are used for model training, and the remaining 20 are used for performance testing.
[0051] The STARE dataset contains 20 color fundus images, each with a resolution of 700×605 pixels. Fifteen images were randomly selected for model training, and the remaining five were used for performance testing.
[0052] During training, data augmentation processes such as random flipping, random rotation, and random cropping are performed on the training set images to improve the model's generalization ability.
[0053] Step 2: Construct multiple skeleton prior supervision signals.
[0054] For each binary blood vessel label image in the training set The present invention first performs two preprocessing operations to obtain the axial and radial geometric properties of blood vessels.
[0055] (1) Extracting the local radial scale map : Given a binary blood vessel label map Define the set of foreground (vascular region) pixels. and background pixel set For any pixel position Local radial scale map obtained through Euclidean distance transformation (EDT) :
[0056] In practical calculations, a two-pass scanning algorithm is used to implement the Euclidean distance transformation, with a time complexity of O(n log n). The result It reflects the distance from the point to the nearest boundary: the larger the value, the closer the point is to the axial center of the blood vessel and the larger the diameter; the smaller the value, the closer the point is to the edge of the blood vessel or the smaller the diameter.
[0057] (2) Extract the centerline skeleton diagram : To capture the axial connectivity structure of vascular networks, this invention employs the Zhang-Suen fast parallel thinning algorithm. The algorithm performs skeletonization by deleting boundary pixels through multiple rounds of parallel iterations, resulting in a single-pixel wide skeleton that preserves the connectivity of the original region and key structural points (such as endpoints and bifurcation points). .set up For the skeletonization operator, the centerline skeleton diagram is... Defined as:
[0058] in, The non-zero pixels in the matrix constitute the centerline point set with a width of one pixel. This set characterizes the axial trunks and connections of the vascular network.
[0059] based on and The present invention further constructs three skeleton prior maps rich in semantic information as multi-dimensional supervision signals.
[0060] (1) Strengthening the skeleton diagram Build: To alleviate the severity of single-pixel skeleton supervision, a maximum trust distance is introduced. A trust neighborhood is constructed for each skeleton point. That is, starting from the skeleton point, in... The region within a pixel distance is considered the enhanced supervision region for that skeleton point. For any skeleton point Its trusted neighborhood is defined as:
[0061] in Take the union of the trusted neighborhoods of all skeleton points to obtain the enhanced skeleton graph:
[0062] The enhanced skeleton expands the monitoring area from a single-pixel "line" to a fault-tolerant "band" for explicitly guiding the connectivity of blood vessels.
[0063] (2) Maximum neighborhood skeleton graph Build: To more precisely characterize the radial features of retinal vessels with clearly defined thicknesses, this invention further proposes a maximum neighborhood skeleton map. Its goal is to achieve scale-aware, fine-grained supervision. A scale-discrimination threshold is introduced. Combined with local radial scale diagram Calculate the trust distance for each skeleton point:
[0064] For capillaries ( The monitored area is equivalent to its actual vascular morphology to prevent the monitoring signal from overflowing into the background and introducing noise; for thick vessels ( The supervised area was normalized to a width of approximately The strip-shaped region of the pixel retains the guidance of its center continuity while avoiding the loss of edge accuracy due to an excessively wide supervision area.
[0065] Based on trust distance Constructing the neighborhood range:
[0066] Take the union of the neighborhood ranges of all skeleton points to obtain an initial supervised region. To ensure that the monitoring signal is strictly located inside a real blood vessel, the initial region... With vascular tags Taking the intersection, we obtain the final maximum neighborhood skeleton graph:
[0067] (3) Microvascular constraint diagram Build: By taking the difference between the reinforced framework and the maximum neighborhood framework, we obtain the microvascular constraint map:
[0068] This region corresponds to the halo-like area around the ends and small branches of capillaries. It is the most anatomically fragile capillary region and the most easily overlooked by the model during training.
[0069] Step 3: Construct a segmentation network.
[0070] like Figure 1 The diagram shows the retinal vessel segmentation structure proposed in this invention. This network is built on the U-Net encoder-decoder architecture, including a shared encoder module, a semantic decoupling module (SDM), and a main decoder module.
[0071] (1) Shared encoder: The shared encoder is responsible for extracting multi-scale features with rich semantics from the input image. This encoder adopts a standard U-Net design and includes five levels of downsampling operations, each consisting of a double convolutional block, denoted as . A double convolutional block consists of two consecutive... It consists of convolutional layers, each followed by batch normalization and ReLU activation functions to enhance nonlinear representation capabilities.
[0072] Given an input retinal fundus image Let the first The feature map output by the level encoder is ,in , For the input image The encoding process is as follows: First, the input image is processed through an initial double convolutional block to obtain the first-level features:
[0073] For subsequent levels ( Each level first passes through a step with a size of 2. The max pooling layer performs spatial downsampling, followed by a double convolutional block for feature transformation.
[0074] in, express Max pooling operation. For the first The feature map output is halved in size at each level, and the number of channels... Increasing step by step, specifically as follows: Deepest feature map Located at the bottleneck layer, it has the largest receptive field and carries the richest global semantic information, which will serve as the input to the semantic decoupling module.
[0075] (2) Main decoder: The main decoder employs a standard U-Net decoding structure, gradually restoring spatial resolution through bilinear interpolation. To compensate for information loss during downsampling, skip connections are used to transfer encoder features. Channel concatenation is performed with the upsampled features of the corresponding level. Let... Indicates the first Level decoding features, among which The upsampling process decreases from deeper to shallower layers. It can be represented as:
[0076] in, . This indicates a bilinear interpolation upsampling operation. This represents a feature concatenation operation along the channel dimension. This represents a convolutional processing block, which contains a A convolutional layer, followed by batch normalization and the ReLU activation function.
[0077] Finally, the shallowest feature map of the decoder is obtained. ,pass Convolutional layers map features into single-channel prediction maps. :
[0078] in, This represents the Sigmoid activation function. This indicates a convolution operation with a kernel size of 1×1, used to map the number of channels in the feature map to the target dimension. Represents each pixel The probability that it belongs to a blood vessel.
[0079] (3) Semantic Decoupling Module (SDM): The semantic decoupling module is located at the end of the encoder, and its input is the bottleneck features extracted by the shared encoder. This module contains two structurally symmetrical but parameter-independent pathways, used to extract axial connectivity features and radial scale features of blood vessels from deep semantic information, respectively. Each pathway contains two layers. Convolutional units are designed to increase the receptive field and follow a layer of... Convolutional units compress the feature dimension to a preset embedding dimension. Each convolutional unit contains convolution, batch normalization (BN), and ReLU activation functions.
[0080] Define connectivity feature paths and geometric scale feature pathways The decoupling process can be represented as:
[0081] in, These represent feature maps that encode axial connectivity and radial scale information of blood vessels, respectively.
[0082] The two feature maps mentioned above are input into lightweight auxiliary prediction heads. Each auxiliary prediction head consists of multiple convolutional-BN-ReLU layers, which progressively upsample the features and map them into a single-channel probability map with the same resolution as the input image.
[0083] in, These represent the enhanced skeleton prediction map and the maximum neighborhood skeleton prediction map, respectively. and These represent the auxiliary prediction heads corresponding to connectivity features and geometric scale features, respectively.
[0084] Step 4: Design a joint loss function with multiple priors.
[0085] To ensure the model's segmentation accuracy of retinal vessels, this invention designs corresponding loss functions for skeleton prior signals of different dimensions. The total loss function consists of four weighted components: main segmentation loss, enhanced skeleton loss, maximum neighborhood skeleton loss, and fine vessel enhancement loss, forming a joint optimization objective.
[0086] (1) Principal Segmentation Loss : The main segmentation loss uses a vessel label map To supervise the target segmentation map output by the main decoder module This invention employs binary cross-entropy. With Dice coefficient loss The combination of is denoted as For the prediction graph and real label images , The calculation is as follows:
[0087] in:
[0088]
[0089] in, N Total number of pixels For the first The predicted probability value of each pixel. For the first The actual label value of each pixel. This is the smoothing constant.
[0090] (2) Loss of reinforcement skeleton : The enhanced skeleton loss uses the same combined loss function, and the enhanced skeleton prediction map output by the supervised semantic decoupling module is used. :
[0091] (3) Maximum neighborhood skeleton loss : The maximum neighborhood skeleton loss uses the same combined loss function, and the maximum neighborhood skeleton prediction map is output by the supervised semantic decoupling module. :
[0092] (4) Loss of microvascular enhancement : To effectively incorporate prior information about fine blood vessel regions into the network's feature learning, this invention designs a multi-scale feature alignment strategy. This strategy imposes constraints on the shallow layers of the encoder, ensuring that the network focuses on the detailed features of fine blood vessels from an early stage.
[0093] Let the feature maps output from the first three layers of the encoder be respectively , , First, through a lightweight projection head. The features from each layer are mapped to a unified single-channel space to obtain the feature embedding. :
[0094] Each projection head By one The process involves convolutional layers followed by batch normalization and ReLU activation to obtain the embedded features. , , .
[0095] Accordingly, the fine blood vessel constraint diagram Downsampling is performed using bilinear interpolation to match the spatial dimensions of each layer's features:
[0096] Network layers of different depths have varying abilities to sense small blood vessels. Superficial layer characteristics. Containing rich detail and location information, it is crucial for capturing the texture of fine blood vessels; deep features It contains high-level semantics, but its localization of fine blood vessels is relatively coarse. To adaptively balance the intensity of the supervision signals in each layer, this invention introduces a parameter-aware weight allocation mechanism based on the Frobenius norm. Let the... The parameter set of each encoder module is as follows Its adaptive weights The calculation is as follows:
[0097] in, Denotes the Frobenius norm. To prevent division by zero for small constants. The weights satisfy the normalization condition. .
[0098] Ultimately, the enhancement and loss of microvessels The weighted sum of the losses at each layer:
[0099] This loss function encourages the network to be effective in representing features at different scales. The indicated capillary region produces a high response, thereby enabling the capture of fine blood vessels at the feature level.
[0100] (5) Total loss function: The total loss function of this invention is obtained by weighted summation of the above four loss components:
[0101] The balance coefficient is set as follows: , , .
[0102] Step 5: Model Training and Validation This invention is implemented using Python within the PyTorch deep learning framework. The computer was configured with an i7-7700K CPU at 4.20GHz, a 1080ti graphics card, and Ubuntu 20.04 operating system. The effectiveness of the proposed vessel segmentation method was validated on the CHASE_DB1, DRIVE, and STARE datasets.
[0103] For training configuration, all input images were uniformly scaled to 512×512 pixels. The Adam optimizer was used to optimize the network, with an initial learning rate set to 0.0085 and a weight decay coefficient set to [value missing]. The batch size for training all datasets was set to 2, and the total number of training epochs for the model was set to 200.
[0104] In terms of skeleton prior construction, a method based on maximum trust distance is adopted. Build a reinforced skeleton Scale-based threshold Construct the maximum neighborhood skeleton and based on and Constructing a fine vascular constraint map During training, the image data from the training set and its corresponding multiple skeleton prior signals are used as inputs to the neural network. The network output is then fed into various loss functions for joint optimization, and the optimal weight parameters are saved after training via backpropagation.
[0105] During the inference phase, the test image is input into the trained network model using the optimal weight parameters saved during the training phase to obtain the probability map output by the main segmentation head. Binarization is performed with a threshold of 0.5 to obtain the final binary segmentation result.
[0106] Experimental results: This invention was experimentally validated on three publicly available retinal vessel segmentation datasets: CHASE_DB1, DRIVE, and STARE. To comprehensively evaluate the performance of the proposed method, the region overlap indexes Dice and IoU, as well as the fidelity index clDice for measuring structural connectivity, were used. The quantitative evaluation results of the segmentation performance of the proposed method and the baseline model U-Net on different datasets are shown in Table 1.
[0107] Table 1. Quantitative evaluation results of the segmentation performance of the method of the present invention and U-Net on different datasets.
[0108] As shown in Table 1, the experimental data demonstrate that the method of this invention achieves superior segmentation performance compared to U-Net on three publicly available datasets. Specifically, it improves the performance by an average of 1.47 and 2.25 percentage points on the region overlap index Dice and IoU, respectively, and by an average of 1.04 percentage points on the structural connectivity fidelity index clDice.
[0109] Figure 2 These are the segmentation results of this invention on different datasets. Figure 2 (a) shows the segmentation results of the CHASE_DB1 dataset. Figure 2 (b) shows the segmentation results of the DRIVE dataset. Figure 2(c) Segmentation results for the STARE dataset. It can be seen that the present invention can accurately restore the thickness variations, extension direction, and connectivity of blood vessels in fundus images from different datasets.
[0110] The above results demonstrate that by introducing multiple prior signals composed of an enhanced skeleton, a maximum neighborhood skeleton, and a fine vessel constraint map, this invention can effectively solve the core technical bottleneck of retinal vessel segmentation, which is characterized by the fragility of small vessels and poor predictive coherence. It significantly improves the structural integrity of the vessel prediction results and provides high-precision segmentation results to support the automated screening of fundus diseases.
[0111] The above description is only for the purpose of helping to understand the method and core essence of the present invention, but the scope of protection of the present invention is not limited thereto. For those skilled in the art, any equivalent substitutions or modifications made to the technical solution and inventive concept disclosed in the present invention within the scope of the technology disclosed in the present invention should be covered within the scope of protection of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.
Claims
1. A method for retinal vessel segmentation combining enhanced framework and maximum neighbor framework, characterized in that, Includes the following steps: S1. Data Preparation and Preprocessing: Select a publicly available retinal vessel dataset for preprocessing as the training set; S2. Construct multiple skeleton prior supervision signals; Explicitly extract axial and radial geometric properties from the binary blood vessel label image to obtain a local radial scale map and a single-pixel wide axial centerline skeleton map; Based on the local radial scale map and the single-pixel wide axial centerline skeleton map, construct an enhanced skeleton map, a maximum neighborhood skeleton map, and a fine blood vessel constraint map constructed by calculating the difference between the two, as triple prior supervision signals. S3. Construct a segmentation network; the segmentation network is constructed based on an encoder-decoder architecture, including a shared encoder, a semantic decoupling module, and a main decoder; the shared encoder extracts multi-scale features from the input image; the semantic decoupling module is located at the end of the shared encoder and includes two independent pathways, which extract the connectivity features and scale features of blood vessels respectively, and map them to generate an enhanced skeleton prediction map and a maximum neighborhood skeleton prediction map; the main decoder generates the final blood vessel segmentation prediction map. S4. Design a joint loss function with multiple priors: Use triple prior supervision signals to constrain the segmentation network in multiple dimensions, and design the main segmentation loss, the enhanced skeleton loss, the maximum neighborhood skeleton loss, and the micro-vessel enhancement loss as loss functions. The joint loss function is obtained by weighted summation of the four loss components. S5. Training the segmentation network: Backpropagation training of the segmentation network is performed using the joint loss function; S6. Retinal vessel segmentation: The trained segmentation network is used to process the fundus image to obtain a vessel probability map. The vessel probability map is then subjected to threshold binarization to obtain the vessel segmentation result.
2. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 1, characterized in that, In S2, multiple skeleton prior supervision signals are constructed as follows: Based on binary vascular tagging The distance from each pixel in the blood vessel region to the nearest boundary is calculated using Euclidean distance transformation, and a local radial scale map is extracted. The Zhang-Suen algorithm was used to create binary label images of blood vessels. Refine the drawing and extract the centerline skeleton. ; Based on centerline skeleton diagram Introducing maximum trust distance For each skeleton point, a trust neighborhood is constructed to build a reinforced skeleton graph. Based on local radial scale map Introducing a scale-distinguishing threshold For each skeleton point, construct a neighborhood range and build a maximum neighborhood skeleton graph. Extract the difference between the enhanced skeleton and the maximum neighbor skeleton to construct a fine blood vessel constraint map. .
3. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 2, characterized in that, The construction of a reinforced skeleton diagram The details are as follows: For any skeleton point Its trusted neighborhood is defined as: Obtain the union of the trusted neighborhoods of all skeleton points to obtain the enhanced skeleton graph. : in, Centerline skeleton diagram The set of centerline points with a width of one pixel, composed of non-zero pixels.
4. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 2, characterized in that, The construction of the maximum neighborhood skeleton graph The details are as follows: Based on local radial scale map Introducing a scale-distinguishing threshold Calculate the trust distance for each skeleton point: For capillaries, the supervised region is equivalent to their actual vascular morphology; for coarse vessels, the supervised region is normalized to a region with a width of [missing information]. A strip-shaped area of a pixel; Based on trust distance Constructing the neighborhood range: Obtain the union of the neighborhood ranges of all skeleton points to obtain an initial supervised region. ;Will With binary vascular labeling Taking the intersection, we obtain the final maximum neighborhood skeleton graph: in, This represents the maximum neighborhood skeleton graph.
5. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 1, characterized in that, The segmentation network is constructed in S3 as follows: The shared encoder adopts a standard U-Net design; the main decoder adopts a standard U-Net decoding structure. The semantic decoupling module includes two structurally symmetrical but parameter-independent pathways, which are used to extract the axial connectivity features and radial scale features of blood vessels from deep semantic information, respectively. Each pathway contains two layers. Convolutional units, followed by a layer The convolutional unit compresses the feature dimension to a preset embedding dimension; each convolutional unit contains convolution, batch normalization, and ReLU activation functions; the input to the semantic decoupling module is the deepest feature map extracted by the shared encoder. Output feature maps that encode the axial connectivity and radial scale information of blood vessels. ; The two feature maps are input into lightweight auxiliary prediction heads, each consisting of multiple convolutional-BN-ReLU layers. These layers progressively upsample the features and map them to a single-channel probability map with the same resolution as the input image, resulting in the enhanced skeleton prediction map. And the maximum neighborhood skeleton prediction map .
6. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 5, characterized in that, The shared encoder and master decoder are as follows: The shared encoder contains five levels of downsampling operations, each consisting of a double convolutional block, which is composed of two consecutive... The system consists of convolutional layers, each followed by batch normalization and the ReLU activation function; the first... The level encoder output feature map is ; The main decoder gradually recovers the spatial resolution through bilinear interpolation and incorporates encoder features through skip connections. Channel concatenation is performed with the upsampled features of the corresponding level; Level decoding features are Finally, the shallowest feature map of the main decoder is obtained. ,pass Convolutional layers map features to blood vessel segmentation prediction maps. .
7. The retinal vessel segmentation method based on the combined enhanced framework and the maximum neighbor framework according to claim 2, characterized in that, The main segmentation loss in S4 is as follows: Main segmentation loss Using the original binary vascular label image Blood vessel segmentation prediction map output by the supervised master decoder Specifically: Using binary cross-entropy With Dice coefficient loss The combination is denoted as For the prediction graph and real label images , The calculation is as follows: in: in, N Total number of pixels For the first The predicted probability value of each pixel. For the first The actual label value of each pixel. This is the smoothing constant.
8. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 7, characterized in that, The enhanced skeleton loss and maximum neighborhood skeleton loss in S4 are specifically as follows: Strengthening the skeleton loss Enhanced skeleton prediction graph used to supervise the output of the semantic decoupling module : Maximum Neighborhood Skeleton Loss The maximum neighborhood skeleton prediction graph used to supervise the output of the semantic decoupling module. : in, Represents binary cross-entropy With Dice coefficient loss The combination of .
9. The retinal vessel segmentation method based on the combined enhanced skeleton and the maximum neighbor skeleton according to claim 2, characterized in that, The loss of microvascular enhancement in S4 is specifically as follows: Let the feature maps output from the first three layers of the shared encoder be respectively , , Firstly, through a lightweight projection head The features from each layer are mapped to a unified single-channel space to obtain the feature embedding. : Each projection head By one Convolutional layers are used, followed by batch normalization and ReLU activation to obtain embedded features. , , ; Accordingly, the fine blood vessel constraint diagram Downsampling is performed using bilinear interpolation to match the spatial dimensions of each layer's features: Let the first The parameter set of each encoder module is as follows Its adaptive weights The calculation is as follows: in, Describing the Frobenius norm, To prevent division by zero for small constants; the weights satisfy the normalization condition. ; Ultimately, microvascular enhancement loss The weighted sum of the losses at each layer: in, Represents binary cross-entropy With Dice coefficient loss The combination of .
10. A retinal vessel segmentation system for combining a reinforced skeleton and a maximum neighbor skeleton obtained by the method described in any one of claims 1-9, characterized in that, Including segmentation networks; The segmentation network includes a shared encoder, a semantic decoupling module, and a main decoder.