A retinal vessel segmentation method based on full-scale single-stage decoding

Abstract

The application claims a retinal vessel segmentation method based on full-scale single-stage decoding. First, the medical image to be segmented is input into a backbone network to extract multi-level features, and the projection module is used to map them to a unified spatial resolution to construct a full-scale observation feature set. Then, the Hessian perception gated fusion module is used to extract the physical structure prior from the high-resolution shallow layer features, and the gated mechanism is used to adaptively filter the background artifacts in the deep semantic features, realizing single-stage feature fusion. Subsequently, the fused features are input into the asymmetric tubular reparameterization module, which uses multi-branch asymmetric convolution to extract tubular topological features in the training stage, and through structure reparameterization, it is equivalent to merging into a single-layer convolution for high-precision segmentation in the inference deployment stage. The application effectively reduces the decoding artifacts and structure breakage, and improves the segmentation connectivity and accuracy of the network for fine blood vessels.

Description

Technical Field

[0001] This invention belongs to the technical field of image feature extraction and segmentation methods, and in particular relates to a retinal vessel segmentation method based on full-scale single-stage decoding. Background Technology

[0002] Fully convolutional neural networks (FCNs) have become the mainstream method in image segmentation, especially in semantic segmentation tasks. Their core advantages lie in their ability to accept input images of arbitrary size, their high computational efficiency, and their ability to achieve pixel-level precision segmentation.

[0003] Since the first proposal of FCN in 2015, the academic community has conducted a large number of related studies and experiments, and a series of classic models have been derived from it, among which U-Net and its various improved U-shaped networks are the most representative. These models improve the accuracy of retinal vessel segmentation by increasing the number of network layers on the basis of FCN architecture.

[0004] However, when dealing with images of diseased retina, issues such as vessel breakpoints, missing bifurcations, and structural misjudgments remain challenges that current methods struggle to overcome effectively. Existing segmentation models perform poorly on diseased retina, primarily due to the following reasons: ① Layer-by-layer decoding mechanisms struggle to balance microvessel reconstruction with deployment efficiency. Traditional U-shaped architectures rely on layer-by-layer decoding to reconstruct segmentation maps, making it difficult to effectively capture and recover small vessels. More importantly, the traditional cascaded decoding design with skip connections introduces a large number of redundant parameters, hindering the clinical deployment of the model. ② Cross-scale fusion modules are complex and redundant, and cannot eliminate decoding artifacts. To bridge the gap between shallow details and deep semantics, existing methods often employ complex attention mechanisms and multi-level fusion strategies, resulting in parameter redundancy and an inability to eliminate artifacts. The root cause of artifacts lies in the structural defects of the decoder itself, rather than insufficient fusion strategies. ③ Isotropic feature extraction introduces significant background noise. Traditional convolutional networks primarily rely on isotropic convolutional kernels with square receptive fields for feature aggregation. However, retinal vessels possess complex morphological features characterized by high curvature, elongation, and anisotropy. When using a symmetrical square kernel to extract vascular features, a large amount of retinal background noise or lesion interference is often incorporated, and the modeling ability for long-distance topological dependence is lacking, which can easily lead to misjudgment or discontinuity in local vascular prediction.

[0005] Therefore, if a segmentation model can be designed to address the above limitations, enabling fully convolutional networks to segment retinal vessels more accurately, it will make further applications of fully convolutional networks in the diagnosis and treatment of retinal diseases possible.

[0006] A search revealed that patent application CN115049682A proposes a retinal vessel segmentation method based on multi-scale dense networks. This method constructs a multi-scale dense network model based on traditional encoder-decoder structures for feature extraction and vessel segmentation. While this method improves segmentation accuracy to some extent, this traditional multi-scale cascaded decoding architecture still has significant technical defects and shortcomings: ① Layer-by-layer cascaded decoding can lead to microvascular topological breaks and artifacts: Traditional architectures rely on layer-by-layer upsampling mechanisms to reconstruct the segmentation map. This multi-level upsampling and error accumulation easily cause structural deformation and breakage of terminal small blood vessels, resulting in decoding artifacts that are difficult to eliminate. ② Complex dense fusion modules have high redundancy, which is not conducive to lightweight clinical deployment: To bridge the semantic gap between shallow and deep layers, the patent often piles up dense cross-scale feature fusion operations. This not only fails to improve the structural defects of the decoder itself, but also introduces a large number of redundant parameters and large computational overhead, making it difficult to deploy on clinical portable devices with limited computing resources. ③ Isotropic convolution lacks the ability to model the topology of tubular structures: Traditional networks mainly rely on isotropic square convolution kernels, which lack the ability to perceive the long-distance topological continuity of highly curved, slender and anisotropic morphology of retinal blood vessels, and are prone to introducing retinal background noise interference such as lesions or uneven illumination.

[0007] To address the shortcomings and deficiencies of the existing technologies, this invention proposes a retinal vessel segmentation method based on full-scale single-stage decoding. The specific solutions are as follows:

[0008] 1. Breaking the cascaded decoding paradigm, this invention proposes a full-scale single-stage decoding strategy: Addressing the topological breaks and parameter redundancy issues caused by layer-by-layer decoding, this invention uses a lightweight projection module to align all deep and shallow layer encoded features to the target spatial resolution in one step, performing global single-stage direct fusion. This avoids error amplification and structural deformation caused by multi-level upsampling, significantly reducing the number of network parameters while ensuring the complete connectivity of micro-networks.

[0009] 2. Introducing a Hessian-sensing gated fusion mechanism to filter background artifacts: To address the problem of noise and artifacts introduced by dense fusion, this invention directly mines zero-parameter continuous vascular edge hessian features from shallow high-resolution features.

[0010] Physical prior. By utilizing this physical prior in conjunction with an adaptive gating mechanism, background clutter and lesion artifacts in deep semantic features can be accurately stripped and filtered, reducing the false positive rate in low-contrast environments.

[0011] 3. Design of an asymmetric tubular reparameterization module, balancing tubular perception and zero-cost accelerated inference: Addressing the limitations and deployment challenges of isotropic convolution modeling, this invention utilizes a multi-branch asymmetric structure to explicitly learn the anisotropic topological features of slender tubular blood vessels during the training phase. In the testing and deployment phase, structural reparameterization technology is used to equivalently merge these features into a single standard 7×7 convolution, achieving high-precision segmentation without increasing any additional inference computation overhead, perfectly meeting the needs of lightweight and efficient clinical deployment. Summary of the Invention

[0012] This invention aims to solve the problems of the prior art. It breaks away from the traditional hierarchical cascade decoding paradigm, improving the segmentation connectivity of fine blood vessels while reducing the number of model parameters.

[0013] A method for retinal vessel segmentation based on full-scale single-stage decoding is proposed. The technical solution of this invention is as follows:

[0014] A retinal vessel segmentation method based on full-scale single-stage decoding includes the following steps:

[0015] Step 1: Obtain the fundus color image to be segmented, and perform preprocessing operations such as grayscale transformation, standardization, contrast-limited histogram equalization and gamma correction. Input the preprocessed image into the backbone feature extraction network for downsampling to obtain multi-level retinal vascular coding features.

[0016] Step 2: Construct a full-scale single-stage decoding network. This network mainly includes a lightweight projection module, a Hessian sensing gated fusion module, and an asymmetric tubular reparameterization module. Through the lightweight projection module, the encoded features from all levels extracted in Step 1 are uniformly mapped to the same number of channels and spatial resolution. Scale alignment is then performed using bilinear interpolation to obtain a full-scale observation feature set in a unified spatial domain. ;

[0017] Step 3: Introduce the Hessian perception-gated fusion module to extract the tubular physical structure prior of retinal vessels from the high-resolution shallow features of the full-scale observation feature set. Then, use a prior-gated network to filter background artifacts and perform adaptive residual enhancement on the low-resolution deep semantic features, achieving single-stage feature decoding and fusion to obtain the fused features. ;

[0018] Step 4: Merge features The input is fed into the asymmetric tubular reparameterization module, where the topological continuity of slender retinal vessels is explicitly learned during the training phase using asymmetric convolutions with anisotropic multi-branching. During the testing and deployment phase, the multi-branching is equivalently converted to a single-layer standard. Convolution enables lossless and accelerated inference;

[0019] Step 5: Set the joint loss function BCE-Dice Loss and iteratively train the network model until the network converges to obtain a trained retinal vessel segmentation model;

[0020] Step 6: Input the fundus image to be tested into the trained retinal vessel segmentation model and output the highly connected pixel-level segmentation results of the vessels.

[0021] Furthermore, the lightweight projection module in step 2 specifically comprises:

[0022] The first Layer coding features Passing through in sequence Convolutional layers, hidden layers Convolutional layers, batch normalization layers, and ReLU activation functions unify channel dimension to a unified feature channel dimension. Subsequently, it is synchronously mapped to the target image resolution through bilinear interpolation. The full-scale observation feature set is obtained. ,in , The total number of levels output by the backbone network; this full-scale mapping process is expressed by the following formula:

[0023]

[0024]

[0025] In the above formula, This represents the input feature map, and BN represents batch normalization. Represents the ReLU activation function. express Convolutional layer express Convolutional layer This indicates a bilinear interpolation operation, where H and W represent the interpolated target height and width, respectively.

[0026] Furthermore, the Hessian perception gating fusion module in step 3 specifically includes the following steps:

[0027] Full-scale observation feature set Divided into shallow feature subsets according to the original resolution with deep feature subset , respectively using Convolution is performed to compress the channel dimension and then processed using the ReLU activation function and batch normalization, which are then aggregated to obtain a shallow structure representation. With deep semantic representation Regularized features are obtained by applying two-dimensional spatial Dropout to the deep representation. The formula is expressed as:

[0028]

[0029]

[0030] in, This indicates the concatenation of channel dimensions. and express Convolution kernel;

[0031] Construct a zero-parameter Hessian prior extractor to extract multi-channel shallow features. Compressed into a single-channel structural field by arithmetic mean Using a fixed Laplace discrete difference operator As the physical core and Perform convolution operations and obtain continuously differentiable Hessian structure priors at the retinal vascular margin by mapping through a temperature-calibrated sigmoid function. :

[0032]

[0033]

[0034] In the above formula, This is the sensitivity adjustment factor, with a value of 10. ;

[0035] Constructing a priori gating networks ,Will , as well as Channel concatenation is used as input, and a gating weight map at the target boundary is adaptively calculated through multi-layer convolution. The obtained gating weight map is then used to perform background artifact filtering and adaptive residual enhancement on deep semantic features to obtain cleaned features. Finally, the shallow details and the purified deep features are directly spliced ​​and fused in a single stage to obtain the fused features. :

[0036]

[0037]

[0038]

[0039] In the above formula, This represents the standard Sigmoid activation function. Represents the Hadamard product. and They represent and The convolution operation.

[0040] Furthermore, the asymmetric tubular reparameterization module in step 4 specifically comprises:

[0041] During the training phase, a system based on standards is constructed. Convolutional branches Horizontal anisotropy Convolutional branches Vertical anisotropy Convolutional branches and the direct branch of the identity mapping The four-branch parallel asymmetric structure consists of branches that do not contain bias terms. Its Conv-BN cascaded forward propagation process and multi-directional topological feature superposition output are used to form the structure. Represented as:

[0042]

[0043]

[0044] in Represents the fused feature map. For the trainable parameters and runtime statistics of the corresponding branch batch normalization layer, This represents a block of combined operations that combine a specific type of convolution operation with batch normalization. This represents the weight tensor of the convolutional layer. In batch normalization formulas, a very small constant (usually set to a very small value, such as...) is introduced to prevent the denominator from being zero and to ensure the stability of numerical calculations. );

[0045] During the testing and deployment phase, the structure reparameterization technique was used to fully integrate the batch normalized layer parameters of each branch into the convolutional kernel weights and biases, thereby obtaining an equivalent kernel for a single branch. and equivalent bias For identity mappings, direct connection branches are used. The Dirac tensor is then subjected to an equivalent transformation; subsequently, zero-filling operations are performed. Align the equivalent feature matrices of all branches to a uniform form. Spatial dimensions are determined, and the aligned weights and biases are linearly superimposed and merged into a single standard. Convolutional layers:

[0046]

[0047]

[0048]

[0049]

[0050]

[0051] Among them, let subscript These represent different branches of convolution and identity mapping, respectively. and These represent the equivalent convolutional kernel weights and biases of each branch after fusing the parameters of the corresponding batch normalized layer; This indicates a boundary zero-padding operation (its subscripts represent the number of zeros padded in the upper, lower, left, and right spatial dimensions, respectively). This represents the transformed convolution kernel after the equivalent weights of each branch have been zero-padded, aligned, and standardized to a uniform size; and These represent the globally equivalent single convolution kernel and bias obtained by merging the multi-branch structures, respectively.

[0052] The final deployment output is represented as:

[0053]

[0054] This enables retinal vascular anisotropic topological connectivity perception with zero additional computational overhead during deployment inference.

[0055] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, when the processor executes the program, it implements a retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of the claims.

[0056] A non-transitory computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of the claims.

[0057] The advantages and beneficial effects of this invention are as follows:

[0058] This invention breaks away from the hierarchical decoding paradigm, effectively mitigating decoding artifacts. It abandons the traditional layer-by-layer cascaded upsampling reconstruction chain, instead employing a full-scale, single-stage decoding strategy to align all encoded features to a target unified spatial resolution in one step for direct global fusion. This design avoids the error amplification and structural deformation caused by multi-level upsampling in conventional decoders, mitigating decoding artifact generation and ensuring the complete topological connectivity of micro-vessels.

[0059] The zero-parameter Hessian prior extractor adaptively guides and significantly suppresses background noise. This invention's innovative Hessian-sensing gated fusion module cleverly utilizes the discrete Hessian physical difference operator to directly extract high-frequency, continuous vessel edge prior anchor points from shallow, high-resolution features. Combined with a highly dynamic gated adaptive mechanism, it can accurately strip away and filter background clutter caused by upsampling interpolation or fundus lesions, thereby reducing the false positive noise rate of capillary vessels in extremely low-contrast environments.

[0060] Lightweight and accelerated inference, ideal for lightweight clinical deployment. The asymmetric tubular reparameterization module designed in this invention uses an anisotropic asymmetric multi-branch structure during training to enhance orientation sensitivity to tubular boundaries, while in actual clinical testing and deployment, it is linearly fused into a single convolutional layer through reparameterization. When combined with a lightweight backbone network (such as MobileViT-xxs), the model requires only a minimal number of parameters (1.02M) and a computational overhead of 1.03G FLOPs to achieve competitive segmentation accuracy, possessing practical value for large-scale deployment and rapid inference on portable wireless ophthalmic clinical devices. Attached Figure Description

[0061] Figure 1 This is a flowchart of a preferred embodiment of the present invention for retinal vessel segmentation based on full-scale single-stage decoding.

[0062] Figure 2 This is a diagram of the overall structure of a full-scale, single-stage decoding network.

[0063] Figure 3 This is a structural diagram of the Haisen perception gating fusion module.

[0064] Figure 4 This is a diagram of the overall structure of the asymmetric tubular reparameterization module. Detailed Implementation

[0065] The technical solutions of the embodiments of the present invention will be clearly and thoroughly described below with reference to the accompanying drawings. The described embodiments are merely some embodiments of the present invention.

[0066] The technical solution of the present invention to solve the above-mentioned technical problems is:

[0067] The retinal vessel segmentation method based on full-scale single-stage decoding provided in this embodiment has the following core setup and operation process in a computer environment:

[0068] Step 1: Data Preprocessing and Data Augmentation. First, preprocess the fundus dataset (such as DRIVE, STARE, or CHASE_DB1 datasets) by segmenting the fundus color images into... Small image patches are used to achieve data augmentation and expand the training samples. The image patches are divided into training, validation, and test sets in an 8:1:1 ratio.

[0069] Step 2: Construct a full-scale, single-stage decoding and segmentation network, mainly including a feature encoding stage, a unified encoding scale stage in the encoding layer, and a single-stage decoding stage. For example... Figure 1 and Figure 2 As shown, the preprocessed training image blocks are fed into a preset backbone feature extraction network (such as EfficientNetV2 or lightweight MobileViT_v2). The backbone network uses cascaded convolution and multi-layer downsampling to gradually reduce the spatial size and multiply the number of channels, eventually leading to an L-layer feature matrix layer with different semantic abstraction levels and resolutions.

[0070] Step 3: Standardize the coding scale at the coding layer. (Combined with...) Figure 2 In Stage 2 of the algorithm, each layer of multi-resolution features undergoes dimensionality reduction transformation through its corresponding lightweight projection operator. The first order feature is by Convolution compresses and aligns its diverse channels to the exact same baseline dimension, then crosses a layer with batch normalization. Convolution is used for multi-scale feature extraction and normalization. After obtaining feature maps of the same channel, bilinear interpolation is used to perform spatial upsampling and scaling, restoring the target to its initial input size and generating a feature observation set at the same resolution. This lays the foundation for a structurally consistent spatial field for subsequent single-stage parallel decoding.

[0071] Step 4: Introduce the Hessian sensing gating fusion module. For example... Figure 3 As shown, the full-scale observation set is divided into a shallow subset and a deep semantic subset. Through... Convolutions output shallow detail features and deep semantic features separately. A random 2D Dropout with a retention rate of 0.8 is applied to the shallow features to prevent overfitting. A zero-parameter Hessian prior extractor is used to calculate the arithmetic mean between channels of the deep features to obtain a single-channel continuous structure field. A sliding window operation is performed on the structure field using a fixed Laplacian physical template to capture local second-order gradient changes. The absolute value is then used to generate a soft-edge structure prior matrix using a Sigmoid mapping. Subsequently, a prior gating module integrating two convolutional layers is constructed. The shallow features, regularized deep features, and structure field prior are stacked and concatenated along the channel axis as joint input to generate a gating weight coefficient map, completing high-frequency cleansing and selective residual enhancement of the deep feature matrix. Finally, the shallow high-frequency details and cleaned deep semantic features are concatenated by channel and aggregated through a fusion convolution layer to form a global single-stage decoding feature that combines semantic depth and micro-edge connectivity.

[0072] Step 5: Introduce the asymmetric tubular reparameterization module. For example... Figure 4 As shown, during the iterative training phase of the network model, the decoded features are imported into the asymmetric tubular reparameterization module. This module simultaneously initiates four unbiased mapping branches during training: standard... Convolutional layer, anisotropic horizontal strip Convolutional layer, anisotropic vertical strip The convolutional layers and an identity mapping branch that directly passes forward features are used. The feature maps from the four branches undergo element-wise matrix accumulation at the output, and the entire dataset is processed by the ReLU activation function to obtain the output of the vascular topology sensing. During the training convergence and testing phase, a reparameterization algorithm is used to perform lossless network structure fusion. The Conv-BN cascaded structures in each branch are equivalently merged into a single equivalent weight and corresponding equivalent independent bias. A zero-padding operator aligns the convolutional kernels of each branch to a uniform spatial size. Finally, the four-way aligned weight matrix and biases are directly linearly superimposed and combined, equivalently fusing into a single standard weight matrix. A single convolutional layer enables lightweight inference deployment.

[0073] Step 6: Model Training and Testing. In the training optimization phase, this invention uses a joint loss function to simultaneously optimize the parameters of the above modules: the standard binary cross-entropy loss function and the Dice similarity loss function, which measures the similarity of overlapping regions, are combined with equal weights. The specific calculation expression of the Dice loss function is as follows: ,in To prevent the minimum value of the denominator being zero After the network model training converges, the test set images are fed into the segmentation model as input, and then processed through equivalent... After inference convolution, connect one layer The convolutional layer outputs the final pixel-level binary hard-label prediction map of retinal blood vessels, completing the image segmentation.

[0074] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of the claims.

[0075] A non-transitory computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of the claims.

[0076] A computer program product includes a computer program that, when executed by a processor, implements the retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of the claims.

[0077] The systems, devices, modules, or units described in the above embodiments can be implemented by computer chips or entities, or by products with certain functions.

[0078] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0079] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0080] The above embodiments should be understood as illustrative only and not as limiting the scope of protection of the present invention. After reading the description of the present invention, those skilled in the art can make various alterations or modifications to the present invention, and these equivalent changes and modifications also fall within the scope defined by the claims of the present invention.

Claims

1. A method for retinal vessel segmentation based on full-scale single-stage decoding, characterized in that, Includes the following steps: Step 1: Obtain the fundus color image to be segmented, and perform preprocessing operations such as grayscale transformation, standardization, contrast-limited histogram equalization and gamma correction. Input the preprocessed image into the backbone feature extraction network for downsampling to obtain multi-level retinal vascular coding features. Step 2: Construct a full-scale single-stage decoding network. This network mainly includes a lightweight projection module, a Hessian sensing gated fusion module, and an asymmetric tubular reparameterization module. Through the lightweight projection module, the encoded features from all levels extracted in Step 1 are uniformly mapped to the same number of channels and spatial resolution. Scale alignment is then performed using bilinear interpolation to obtain a full-scale observation feature set in a unified spatial domain. ; Step 3: Introduce the Hessian perception-gated fusion module to extract the tubular physical structure prior of retinal vessels from the high-resolution shallow features of the full-scale observation feature set. Then, use a prior-gated network to filter background artifacts and perform adaptive residual enhancement on the low-resolution deep semantic features, achieving single-stage feature decoding and fusion to obtain the fused features. ; Step 4: Merge features The input is fed into the asymmetric tubular reparameterization module, where the topological continuity of slender retinal vessels is explicitly learned during the training phase using asymmetric convolutions with anisotropic multi-branching. During the testing and deployment phase, the multi-branching is equivalently converted to a single-layer standard. Convolution enables lossless and accelerated inference; Step 5: Set the joint loss function BCE-Dice Loss and iteratively train the network model until the network converges to obtain a trained retinal vessel segmentation model; Step 6: Input the fundus image to be tested into the trained retinal vessel segmentation model and output the highly connected pixel-level segmentation results of the vessels.

2. The retinal vessel segmentation method based on full-scale single-stage decoding according to claim 1, characterized in that, The lightweight projection module in step 2 is specifically: The first Layer coding features Passing through in sequence Convolutional layers, hidden layers Convolutional layers, batch normalization layers, and ReLU activation functions unify channel dimension to a unified feature channel dimension. Subsequently, it is synchronously mapped to the target image resolution through bilinear interpolation. The full-scale observation feature set is obtained. ,in , The total number of levels output by the backbone network; this full-scale mapping process is expressed by the following formula: In the above formula, This represents the input feature map, and BN represents batch normalization. Represents the ReLU activation function. express Convolutional layer express Convolutional layer This indicates a bilinear interpolation operation, where H and W represent the interpolated target height and width, respectively.

3. The retinal vessel segmentation method based on full-scale single-stage decoding according to claim 2, characterized in that, The Hessen perception gating fusion module in step 3 specifically includes the following steps: Full-scale observation feature set Divided into shallow feature subsets according to the original resolution with deep feature subset , respectively using Convolution is performed to compress the channel dimension and then processed using the ReLU activation function and batch normalization, which are then aggregated to obtain a shallow structure representation. With deep semantic representation Regularized features are obtained by applying two-dimensional spatial Dropout to the deep representation. The formula is expressed as: in, This indicates the concatenation of channel dimensions. and express Convolution kernel; Construct a zero-parameter Hessian prior extractor to extract multi-channel shallow features. Compressed into a single-channel structural field by arithmetic mean Using a fixed Laplace discrete difference operator As the physical core and Perform convolution operations and obtain continuously differentiable Hessian structure priors at the retinal vascular margin by mapping through a temperature-calibrated sigmoid function. : In the above formula, This is the sensitivity adjustment factor, with a value of 10. ; Constructing a priori gating networks ,Will , as well as Channel concatenation is used as input, and a gating weight map at the target boundary is adaptively calculated through multi-layer convolution. The obtained gating weight map is then used to perform background artifact filtering and adaptive residual enhancement on deep semantic features to obtain cleaned features. Finally, the shallow details and the purified deep features are directly spliced ​​and fused in a single stage to obtain the fused features. : In the above formula, This represents the standard Sigmoid activation function. Represents the Hadamard product. and They represent and The convolution operation.

4. The retinal vessel segmentation method based on full-scale single-stage decoding according to claim 1, characterized in that, The asymmetric tubular reparameterization module in step 4 specifically refers to: During the training phase, a system based on standards is constructed. Convolutional branches Horizontal anisotropy Convolutional branches Vertical anisotropy Convolutional branches and the direct branch of the identity mapping The four-branch parallel asymmetric structure consists of branches that do not contain bias terms. Its Conv-BN cascaded forward propagation process and multi-directional topological feature superposition output are used to form the structure. Represented as: in Represents the fused feature map. For the trainable parameters and runtime statistics of the corresponding branch batch normalization layer, This represents a block of combined operations that combine a specific type of convolution operation with batch normalization. This represents the weight tensor of the convolutional layer. In batch normalization formulas, a very small constant (usually set to a very small value, such as...) is introduced to prevent the denominator from being zero and to ensure the stability of numerical calculations. ); During the testing and deployment phase, the structure reparameterization technique was used to fully integrate the batch normalized layer parameters of each branch into the convolutional kernel weights and biases, thereby obtaining an equivalent kernel for a single branch. and equivalent bias For identity mappings, direct connection branches are used. The Dirac tensor is then subjected to an equivalent transformation; subsequently, zero-filling operations are performed. Align the equivalent feature matrices of all branches to a uniform form. Spatial dimensions are determined, and the aligned weights and biases are linearly superimposed and merged into a single standard. Convolutional layers: Among them, let subscript These represent different branches of convolution and identity mapping, respectively. and These represent the equivalent convolutional kernel weights and biases of each branch after fusing the parameters of the corresponding batch normalized layer; This indicates a boundary zero-padding operation (its subscripts represent the number of zeros padded in the upper, lower, left, and right spatial dimensions, respectively). This represents the transformed convolution kernel after the equivalent weights of each branch have been zero-padded, aligned, and standardized to a uniform size; and These represent the globally equivalent single convolution kernel and bias obtained by merging the multi-branch structures, respectively; The final deployment output is represented as: This enables retinal vascular anisotropic topological connectivity perception with zero additional computational overhead during deployment inference.

5. An electronic device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements a retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of claims 1 to 4.

6. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements a retinal vessel segmentation method based on full-scale single-stage decoding as described in any one of claims 1 to 4.

Images