Multi-branch enhanced retinal vessel segmentation network method based on stimulation guidance

Abstract

The invention provides a multi-branch enhanced retinal vessel segmentation network method based on stimulation guidance, and belongs to the technical field of medical image intelligent diagnosis. The technical problems that small blood vessels are difficult to accurately identify and boundaries are not clear in retinal blood vessel segmentation are solved. According to the technical scheme, the method comprises the following steps: S1, collecting fundus color image data to be segmented; s2, constructing a boundary enhancement module; s3, constructing a multi-scale feature aggregation module; s4, constructing a stimulation guide gating fusion module; and S5, after model training is completed, for each to-be-segmented test image. According to the method, boundary gradient enhancement, multi-scale interaction and stimulation guide gating cooperate to improve boundary continuity and detail fidelity, thin blood vessel distinguishability is improved, and robustness under a complex background and weak contrast is enhanced.

Description

Technical Field

[0001] This invention relates to the field of intelligent medical image diagnosis technology, and in particular to a stimulus-guided multi-branch enhanced retinal vessel segmentation method. Background Technology

[0002] With the rising incidence of chronic diseases such as diabetes and hypertension, fundus imaging has become a crucial tool for clinical and cognitive assessment. Subsequent vascular morphology directly reflects microcirculatory status and disease severity, playing a key role in the early detection, grading, and monitoring of diseases such as those induced by diabetes, late-stage disease, and those experiencing accelerated activation. Therefore, accurately and stably segmenting vascular structures from color fundus images, especially further refining the depiction of fine capillaries and bifurcations, has become a fundamental aspect of intelligent fundus analysis.

[0003] Existing angiogenesis methods mainly include thresholding, simulation, and morphological operators based on traditional image processing, as well as encoder-decoder structures based on deep learning. Compared with traditional methods, deep learning models perform better in complex backgrounds and multi-meaning conditions, but they still face signal challenges in practical applications: on the one hand, fundus images have problems such as signal inhomogeneity, noise interference, lesion display, and complex texture backgrounds, resulting in low signals from blood vessels and background, which easily leads to alpha responses; on the other hand, small blood vessels have weak intensity at the original resolution and are easily suppressed by textures or lost during downsampling, resulting in blurred boundaries, breaks, and insufficient locality. Although skip connections can be partially lost, their ability to model cross-dimensional information interaction is limited, making it difficult to simultaneously take into account global signal consistency and local detail deviations. To address this problem, existing research often employs strategies such as attention mechanisms, multi-scale fusion, or boundary constraints, using topological structure as a metric. However, these approaches still have limitations: relying solely on semantic fusion can easily introduce salient constraints on vessel boundaries, making it difficult to guarantee the boundary continuity of thin structures; simple multi-scale topological or additive fusion lacks topological trade-offs for different topological structures, and cross-topological and boundary information may still introduce noise; channel attention, if it only emphasizes statistics, suppresses locally sensitive small vessel signals, resulting in insufficient representation of weak structures. Therefore, a vessel segmentation method is needed that can saliently strengthen boundaries, effectively model interdisciplinary causal relationships, and easily balance global and local responses in the channel dimension, thereby improving robustness and distribution to complex backgrounds and weak, single-signal details. Summary of the Invention

[0004] The purpose of this invention is to provide a stimulation-guided multi-branch enhanced retinal vessel segmentation network method. By synergistically enhancing boundary continuity and detail fidelity through boundary gradient enhancement, multi-scale interaction, and stimulation-guided gating, it improves the distinguishability of thin vessels and enhances robustness under complex backgrounds and low contrast. It is suitable for clinical vessel segmentation and can provide support for auxiliary screening and intelligent diagnosis of diabetic retinopathy, hypertensive retinopathy, etc. It aims to improve the accuracy and generalization ability of vessel segmentation and has broad application prospects in fundus disease indication and clinical auxiliary diagnosis.

[0005] The inventive concept of this invention is as follows: This invention provides a stimulus-guided multi-branch enhanced retinal vessel segmentation network method. First, fundus color image data is acquired, and the original images are standardized and preprocessed, including size unification, intensity normalization, and necessary data augmentation, to obtain a sample set paired with pixel-level vessel annotations. The sample set is then divided into a training set and a test set. Second, a boundary attention enhancement module is constructed, which explicitly models vessel edge information through dual-scale Sobel gradients and further introduces a Sigmoid gating mechanism to adaptively modulate the edge response. Then, a multi-scale global aggregation module is designed to fuse and interact multi-scale features from different levels at a unified resolution, and channel-wise weighted summation and global feature reconstruction are achieved through scale weight modeling. Third, a stimulus-guided gating fusion module is constructed, which uses a dual-branch structure perception mechanism to model the global response and local peak features respectively, and introduces learnable gating parameters in the channel dimension to achieve dynamic balance and adaptive fusion of the two. Finally, the fused features are decoded and upsampled to restore the original image resolution, and the final segmentation result is output.

[0006] To achieve the aforementioned objectives, the present invention employs the following technical solution: a method for stimulation-guided multi-branch enhanced retinal vessel segmentation network, comprising the following steps:

[0007] S1: Collect fundus color image data to be segmented and pair it with pixel-level blood vessel annotations; perform standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and further divide the sample set into training set and test set to ensure the stability and generalization ability of model training.

[0008] S2: Construct a boundary enhancement module, which explicitly models the blood vessel edge information by combining learnable dual-scale Sobel gradients to obtain the blood vessel edge amplitude response map, and further introduces a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, thereby highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage.

[0009] S3: Construct a multi-scale feature aggregation module to fuse and interact multi-scale features from different levels at a unified resolution. Through scale weight modeling, achieve channel-wise weighting and global feature reconstruction to enhance the collaborative expression of cross-scale semantics and detailed information.

[0010] S4: Construct a stimulus-guided gating fusion module, adopt a dual-branch structure perception mechanism to model the global response and local peak features respectively, and introduce learnable gating parameters in the channel dimension to achieve dynamic balance and adaptive fusion of the two.

[0011] S5: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing to calculate and output a pixel-level blood vessel probability map with the same resolution as the original image. Then, the probability map is thresholded to remove spurious responses and smooth the boundaries, thereby obtaining the final blood vessel segmentation result of the corresponding test sample.

[0012] Furthermore, the specific steps of step S2 are as follows:

[0013] Step S2.1: Input multi-channel image pass Convolution mapping to a single-channel grayscale image Through the formula:

[0014] (1);

[0015] in express convolution, The number of input image channels, For batches, and For the height and width of the image, For indexing, For input number One channel, This represents the image after grayscale processing.

[0016] Step S2.2: At the original scale With downsampling scale Calculate separately and The Sobel gradient in the direction is calculated, and the corresponding edge gradient magnitude matrix is ​​obtained:

[0017] (2);

[0018] (3);

[0019] in and They are of direction and Gradient of direction, This is the downsampled feature map. and yes of direction and Gradient of direction, It is a numerical constant. and The obtained edge gradient magnitude matrix;

[0020] Step S2.3: Align the amplitude responses of the two scales spatially, then stitch them together, and perform pixel-level Softmax along the channel dimension to obtain adaptive fusion weights, thereby achieving adaptive fusion of edge intensity at the two scales.

[0021] (4);

[0022] (5);

[0023] in, Indicates will The result after upsampling alignment This indicates a splicing operation. This indicates the Softmax operation. and These are the pixel-level fusion weights corresponding to the two scales. For Hadama accumulation, To obtain the final fusion vector, robust normalization is used to compress the magnitude map to... interval;

[0024] Step S2.4: Apply soft thresholding to the normalized amplitude map to generate the final boundary attention map.

[0025] (6);

[0026] in for Standardized features , and These are the amplitude coefficient, bias, and threshold center, respectively. The slope is the angle; the larger the slope, the closer it is to the hard threshold. It is the Sigmoid activation function. This is the final attention map.

[0027] Furthermore, the specific steps of step S3 are as follows:

[0028] Step S3.1: Denote the multi-scale features of the decoder output at different scales as follows: ,in For each scale feature, the sequence number is used for... Convolutional projection and normalization to unify the number of channels. Based on the resolution of the last layer of the decoder, the data is upsampled to a unified spatial dimension to obtain aligned features. Global average pooling is then performed to obtain channel statistics at each scale. These statistics are then averaged and aggregated along the scale dimension, and a multilayer perceptron (MLP) is used for dimensionality adjustment to obtain channel descriptions, i.e., channel score vectors at each scale. The hidden layer dimension is... , For dimensionality reduction ratio;

[0029] Step S3.2: After stacking the scoring vectors at all scales, calculate softmax normalization along the scale dimension to obtain channel-wise attention, and then perform weighted fusion across scales:

[0030] (7);

[0031] in For the scale number, For the first Channel scoring vectors of various scales Features aligned at each scale Indicates the feature sequence index. Indicates the channel number index. and All are indexes. This is the backbone feature map after cross-scale weighted fusion;

[0032] Step S3.3: After aligning all scales, connect them along the channel dimension, and use... Convolutional layers compress channels back After normalization and activation by the GELU function, it is finally connected to the backbone channel via residual connections:

[0033] (8);

[0034] (9);

[0035] in These are the weighting coefficients. These are learnable residual coefficients. For GELU activation function, For normalization operations, For residual characteristics, Main feature map;

[0036] Step S3.4: Project the backbone and refined features onto the desired channel dimension through a fusion module consisting of convolutional layers and a channel attention mechanism:

[0037] (10);

[0038] (11);

[0039] in yes Encapsulation of convolution, normalization, and activation functions. For channel attention, For output head, The feature vector before projection. This is the final output feature.

[0040] Furthermore, the specific steps of step S4 are as follows:

[0041] Step S4.1: For the input images at various scales First use Convolution, batch normalization, and GELU activation functions are used to enhance local context representation and stabilize the distribution.

[0042] (12);

[0043] in For convolution kernel, The extracted feature map;

[0044] Step S4.2: Construct two types of complementary stimuli from a global perspective, for... Perform global average pooling and global max pooling respectively to obtain the channel-level response vector:

[0045] , (13);

[0046] in and These represent global average pooling and global max pooling, respectively. and These represent the features after average pooling and max pooling, respectively.

[0047] Step S4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require the full context.] and Inputting each data point into a dual-branch MLP generates two types of channel importance descriptions and average trend vectors. With peak significance vector :

[0048] (14);

[0049] (15);

[0050] in , , and These are the weighting coefficients. It is the ReLU activation function. and These are the average trend and peak significance, respectively, and their dimensions are mapped to... , respectively denoted as and ;

[0051] Step S4.4: To enable the model to adaptively learn the weight ratios of the average trend and peak value in each channel, the two stimulus vectors are concatenated and then... Convolution and normalization, and generating dynamic gating coefficients using Sigmoid. :

[0052] (16);

[0053] in This indicates a splicing operation. For dynamic gating coefficients;

[0054] Step S4.5: Weight the two types of candidate weights according to the dynamic gating coefficient in the channel dimension, and obtain the final attention map through Sigmoid mapping. :

[0055] (17);

[0056] in The attention map obtained from the mapping;

[0057] Apply attention and linear mapping, and introduce residual connections:

[0058] (18);

[0059] in The coefficient matrix, That is, the feature map after each scale has been gated by the channel.

[0060] Furthermore, this invention also provides a stimulus-guided multi-branch enhanced retinal vessel segmentation network system, which, when applied using the method described in this invention, includes the following steps:

[0061] The acquisition module is configured to perform the following process: segmenting fundus color image data and pairing it with pixel-level blood vessel annotations; performing standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and dividing the sample set into a training set and a test set;

[0062] The boundary enhancement module is configured to perform the following process: explicitly modeling the blood vessel edge information by combining learnable dual-scale Sobel gradients to obtain the blood vessel edge amplitude response map, and further introducing a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage.

[0063] The multi-scale feature aggregation module is configured to perform the following process: fuse and interact multi-scale features from different levels at a uniform resolution, realize channel-wise weighted and global feature reconstruction through scale weight modeling, and enhance the collaborative expression capability of cross-scale semantics and detailed information.

[0064] The stimulus-guided gating fusion module is configured to perform the following process: a dual-branch structure perception mechanism is used to model the global response and local peak features respectively, and learnable gating parameters are introduced in the channel dimension to achieve dynamic balance and adaptive fusion of the two.

[0065] The vessel segmentation module is configured to perform the following process: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing, calculates and outputs a pixel-level vessel probability map with the same resolution as the original image, and then thresholds the probability map to remove spurious responses and smooth the boundaries, thereby obtaining the final vessel segmentation result for the corresponding test sample.

[0066] Meanwhile, the present invention proposes an electronic device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the computer program is executed, it implements the steps of the method described in the present invention.

[0067] Furthermore, the present invention proposes a computer-readable storage medium having a computer program stored thereon, the computer program being configured to implement the steps of the method described in the present invention when invoked by a processor.

[0068] Finally, the present invention provides a computer program product comprising a computer program / instructions that, when executed by a processor, implement the steps of the method described in the present invention.

[0069] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0070] 1. This invention utilizes a multi-stent feature aggregation module to fuse multi-layer features at a unified resolution and perform stent weight modeling. This achieves channel-by-channel weighting and global feature reconstruction, enhancing the dynamic balance between global response and local hotspots in span-range vascular channels. This strengthens channel expansion and improves the issues of fragile small vessels and noise interference. This solution can be used for fine vascular segmentation in public datasets and clinical scenarios, providing more reliable fundamental technical support for auxiliary channels and intelligent diagnosis of fundus diseases.

[0071] 2. Enhanced ability to express vascular boundaries and details: This invention introduces a vascular enhancement module, which performs salient modeling of vascular edges through dual-scale Sobel microscopy, and combines gating to enhance the response of small and weak vascular boundaries, effectively improving the continuity and defect details of vascular boundaries.

[0072] 3. Enhance cross-dimensional signal and global feature reconstruction: This invention employs a multi-dimensional signal aggregation module to fuse different signal systems at a unified resolution, and achieves progressive channel signal and global feature reconstruction through dimensional weight modeling, thereby enhancing the consistency between multi-dimensional signals and texture information and improving the problems of fine signal branches and insufficient topology.

[0073] 4. Achieving a balance between global response and local hotspot gain: This invention proposes a stimulus-guided gating fusion module that uses a bi-branch stimulus mechanism to characterize global trends and local hotspot features, and uses learnable gating parameters to dynamically enhance the expression of features between channels and improve the biasing ability for splitting, bifurcation and complex background regions.

[0074] 5. Improve robustness and generalization performance under complex logic conditions: Through the optimization of boundary reinforcement, multidisciplinary aggregation and stimulus-guided gating, this invention is stronger in cases of signal inhomogeneity, noise interference, complex background texture and lesion display, reduces alpha response and false negative risk, and improves generalization ability under different datasets and clinical scenarios.

[0075] 6. Balancing accuracy and efficiency to promote practical deployment and application: This invention adopts a strategy enhancement strategy to improve accuracy and stability without significantly increasing the inference burden. It outputs pixel-level vascular maps for subsequent quantitative analysis and clinical interpretation, providing reliable technical support for the auxiliary causal relationship and intelligent diagnosis of diseases such as diabetes-driven diseases. Attached Figure Description

[0076] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used to explain the invention but do not constitute a limitation thereof.

[0077] Figure 1 This is a schematic diagram of the overall framework of the stimulus-guided multi-branch enhanced retinal vessel segmentation network of the present invention.

[0078] Figure 2 This invention relates to a neural network structure for a stimulus-guided multi-branch enhanced retinal vessel segmentation network.

[0079] Figure 3 This invention relates to a boundary enhancement module structure for a stimulus-guided multi-branch enhanced retinal vessel segmentation network.

[0080] Figure 4 This invention relates to a multi-scale feature aggregation module structure based on a stimulus-guided multi-branch enhanced retinal vessel segmentation network.

[0081] Figure 5 This is a structural diagram of the stimulation-guided gating fusion module based on a stimulation-guided multi-branch enhanced retinal vessel segmentation network of the present invention.

[0082] Figure 6 This image shows a comparison of the segmentation performance of the stimulus-guided multi-branch enhanced retinal vessel segmentation network of this invention with mainstream models. Detailed Implementation

[0083] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. Of course, the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0084] Example 1: See Figures 1 to 6 The present invention provides a technical solution for a stimulus-guided multi-branch enhanced retinal vessel segmentation network to select a pair of images from a dataset. For example, the process from raw data images to the final predicted result includes the following steps:

[0085] S1: Collect fundus color image data to be segmented and pair it with pixel-level blood vessel annotations; perform standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and further divide the sample set into training set and test set to ensure the stability and generalization ability of model training.

[0086] S2: Construct a boundary enhancement module, which explicitly models the blood vessel edge information by combining learnable dual-scale Sobel gradients to obtain the blood vessel edge amplitude response map, and further introduces a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, thereby highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage.

[0087] S3: Construct a multi-scale feature aggregation module to fuse and interact multi-scale features from different levels at a unified resolution. Through scale weight modeling, achieve channel-wise weighting and global feature reconstruction to enhance the collaborative expression of cross-scale semantics and detailed information.

[0088] S4: Construct a stimulus-guided gating fusion module, adopt a dual-branch structure perception mechanism to model the global response and local peak features respectively, and introduce learnable gating parameters in the channel dimension to achieve dynamic balance and adaptive fusion of the two.

[0089] S5: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing to calculate and output a pixel-level blood vessel probability map with the same resolution as the original image. Then, the probability map is thresholded to remove spurious responses and smooth the boundaries, thereby obtaining the final blood vessel segmentation result of the corresponding test sample.

[0090] Specifically, the steps of step S2 are as follows:

[0091] Step S2.1: First, to enable the model to learn more directly the channel combination that is most discriminative for edges, input the multi-channel image. pass Convolution mapping to a single-channel grayscale image Through the formula:

[0092] (1)

[0093] in express convolution, The number of input image channels, It is 16. and Here, the height and width of the image are both 64. For indexing, For input number One channel, among which for , and They are respectively:

[0094]

[0095]

[0096] Step S2.2: Then, at the original scale With downsampling scale Calculate separately and The Sobel gradient in the direction is calculated, and the corresponding edge gradient magnitude matrix is ​​obtained:

[0097] (2)

[0098] (3)

[0099] in and They are of direction and Gradient of direction, This is the downsampled feature map. and yes of direction and Gradient of direction, This is a numerical constant, here taken as 10. -12 , and The resulting edge gradient magnitude matrices are as follows:

[0100]

[0101]

[0102] Step S2.3: Subsequently, the amplitude responses of the two scales are aligned in the spatial dimension and then stitched together. Pixel-level Softmax is then performed along the channel dimension to obtain adaptive fusion weights, thereby achieving adaptive fusion of edge intensity at the two scales.

[0103] (4)

[0104] (5)

[0105] in, Indicates will The result after upsampling alignment This indicates a splicing operation. This indicates the Softmax operation. and These are the pixel-level fusion weights corresponding to the two scales. For Hadama accumulation, This is the final fused vector. Meanwhile, to ensure most pixels fall within a good contrast range, we use robust normalization to compress the magnitude map to the [0,1] interval, and minimize the impact of extreme values. and They are respectively:

[0106]

[0107]

[0108] Step S2.4: Finally, apply soft thresholding to the normalized amplitude map to generate the final boundary attention map.

[0109] (6)

[0110] in for Standardized features =0.80、 =0.20 and =0.18 represents the amplitude coefficient, bias, and threshold center. =6.0 represents the slope; the larger the slope, the closer it is to the hard threshold. It is the Sigmoid activation function. This is the final attention map, with the following values:

[0111]

[0112] Specifically, the steps of step S3 are as follows:

[0113] Step S3.1: First, denote the multi-scale features output by the decoder at different scales as follows: , , and Perform each scale feature Convolutional projection and normalization are performed to unify the number of channels to 64. Based on the resolution of the last layer of the decoder, the channel is upsampled to a unified spatial dimension to obtain aligned features. Global average pooling is then performed to obtain channel statistics at each scale. Finally, average aggregation is performed on the scale dimension and multilayer perceptron (MLP) is used to perform dimensionality increase and decrease operations to obtain channel descriptions, i.e., channel scoring vectors at each scale, where the dimension of the hidden layer is 8.

[0114] Step S3.2: Next, after stacking the scoring vectors at all scales, softmax normalization is calculated in the scale dimension to obtain channel-wise attention, and then weighted fusion is performed across scales:

[0115] (7)

[0116] in For the scale number, For the first Channel scoring vectors of various scales Features aligned at each scale Indicates the feature sequence index. Indicates the channel number index. and All are indexes. The backbone feature map is obtained after cross-scale weighted fusion. The Softmax calculation result is as follows:

[0117]

[0118] Step S3.3: Then, after aligning all scales, connect them along the channel dimension, compress the channels back to 64 using a 1×1 convolutional layer, normalize them, and pass them through the GELU activation function. Finally, connect them to the backbone channels through residual connections.

[0119] (8)

[0120] (9)

[0121] in These are the weighting coefficients. =0.1 is the learnable residual coefficient. For GELU activation function, For normalization operations, For residual characteristics, As the backbone feature map, here R and They are respectively:

[0122]

[0123]

[0124] Step S3.4: Finally, the backbone and refined features are projected onto the desired channel dimension through a fusion module consisting of convolutional layers and channel attention mechanisms.

[0125] (10)

[0126] (11)

[0127] in yes Encapsulation of convolution, normalization, and activation functions. For channel attention, For output head, The feature vector before projection. The final output features are as follows:

[0128]

[0129] Specifically, the steps of step S4 are as follows:

[0130] Step S4.1: For the input images at various scales First use Convolution, batch normalization, and GELU activation functions are used to enhance local context representation and stabilize the distribution.

[0131] (12)

[0132] in For convolution kernel, The extracted feature map is as follows:

[0133]

[0134] Step S4.2: Next, construct two types of complementary stimuli from a global perspective, and apply them to... Perform global average pooling and global max pooling respectively to obtain the channel-level response vector:

[0135] , (13)

[0136] in and These represent global average pooling and global max pooling, respectively. and These represent the features after average pooling and max pooling, respectively.

[0137] Step S4.3: Subsequently, and Inputting each data point into a dual-branch MLP generates two types of channel importance descriptions and average trend vectors. With peak significance vector :

[0138] (14);

[0139] (15);

[0140] in , , and These are the weighting coefficients. It is the ReLU activation function. and These are the average trend and peak significance, specifically:

[0141]

[0142]

[0143] Map its dimensions to , respectively denoted as and ;

[0144] Step S4.4: To enable the model to adaptively learn the weight ratios of the average trend and peak value in each channel, the two stimulus vectors are concatenated and then... Convolution and normalization, and generating dynamic gating coefficients using Sigmoid. :

[0145] (16);

[0146] in This indicates a splicing operation. The dynamic gating coefficient is as follows:

[0147]

[0148] Step S4.5: Subsequently, the two types of weight candidates are weighted and fused according to the dynamic gating coefficient at the channel dimension, and the final attention map is obtained through Sigmoid mapping. :

[0149] (17)

[0150] in The attention map obtained from the mapping;

[0151] Finally, attention and linear mapping are applied, and residual connections are introduced:

[0152] (18)

[0153] in The coefficient matrix, That is, the feature map after channel gating at each scale, and its matrix is:

[0154]

[0155] Example 2

[0156] Referring to Example 1, in this example, we use the parameters and results obtained in Example 1, and introduce various existing traditional methods as comparison objects to conduct further experimental comparisons and analyses. By conducting unified testing and systematic evaluation of the performance of each model on multiple evaluation indicators, the results show that the present invention performs better overall on all indicators, and can achieve a more comprehensive and shared advantage in segmentation, thereby verifying the superiority and advancement of the present invention over traditional methods.

[0157] 1. Traditional model

[0158] The following comparison model was selected in this embodiment:

[0159] 1) U-Net Model: U-Net is a classic encoder-decoder segmentation network. It extracts high-level semantics through downsampling, restores spatial resolution through upsampling, and uses skip connections to fuse shallow layers with semantics, thus gaining widespread application in medical image segmentation. Its advantages include a simple structure and easy training, but it is relatively patterned for small sample tasks. Due to the loss of boundary details caused by multiple downsampling operations, and the limited modeling of long-range dependencies by dense receptive fields, it suffers from problems such as minor interruptions in detail updates and boundary adhesion when dealing with thin, weakly sparse, or complex lesion-sparse tasks.

[0160] 2) TransU-Net Model: TransU-Net introduces a Transformer encoder into the U-Net framework, combining local sparse features with global self-focus modeling, enabling the model to simultaneously capture local texture and long-range dependency information. Its advantage is enhanced global understanding, which is beneficial for cross-regional vascular deformability modeling and structural consistency representation. However, the computational complexity of the Transformer and its significant inherent problems, coupled with the lack of specific constraints for boundary / fine-grained structures, may still result in insufficient edge enhancement of thin vessels, and it relies more heavily on engineering tuning in high-resolution fundus images.

[0161] 3) SGAT-Net Model: The SGAT-Net (Stimulus-guided Adaptive Transformer) network introduces the probability of "stimulus" through sparse vascular guidance. It typically highlights vascular responses in complex contexts through joint modeling of local and global features and adaptive receptive field adjustment. Its advantages include consideration of both global and local details, particularly improving performance in small vessels and weak signal regions. However, when dealing with stimulus / gating hyperparameters and feature fusion path probabilities, training stability and generalization performance become more dependent on data distribution and enhancement strategies. Furthermore, it may still experience false positives and false negatives in cases of minor changes or complex focal disease types.

[0162] 4) TDCAU-Net Model: TDCAU-Net integrates Transformer and Dilated Convolutional attention modules on the U-Net body. It expands the receptive field through holes and combines channel and spatial attention to enhance key information, thereby enhancing vascular expression while preserving its integrity. Its advantage lies in its strong representation ability for both main and microvessels and its ability to alleviate the problem of multi-signal structures. However, this type of structure is typically used for module target and cardiac calculations, and parameters such as hole rate and attention intensity need to be adjusted. In noisy scenes with weak or non-detailed boundaries, the lack of significant detail boundary shaping may also result in insufficient focus.

[0163] 5) TUnet-LBF Model: TUnet-LBF typically combines the multi-scale modeling capabilities of Transformer-Unet with the idea of ​​energy function / local binarization fitting, using additional structural constraints to guide vascular structures to be smoother and closer to the topological morphology. Its advantage lies in maintaining the continuity of vascular formation and small structures more effectively, and reducing interruptions and voids with certain programming. However, the presence of energy terms or additional terms often introduces extra parameters and complexity, requiring more coordinated weight balancing for different datasets; furthermore, when lesion interference is strong, excessive introduction may expand the false detection area.

[0164] 6) HiDiffSeg Model: HiDiffSeg is based on the diffusion concept for segmentation modeling. It typically employs a coarse-to-fine generation approach / denoised inference path, allowing the network to gradually refine local boundaries and detailed textures after global structural modeling. Its advantages include good structural consistency, stronger morphological expression potential, and the ability to handle complex, fine blood vessels entangled with background textures. However, diffusion models usually involve many inference steps, resulting in higher computational costs, and are also sensitive to the scale and stability of training data. For clinical real-time or lightweight configurations, further acceleration and simplification strategies are needed.

[0165] 7) HRD-Net Model: HRD-Net emphasizes high-resolution feature preservation and adaptive learning capabilities. It typically maintains the spatial representation through hotspots or high-definition backbones, and then enhances signal representation using feature enhancement / global aggregation modules to reduce localization bias caused by downsampling. It emphasizes more accurate boundary localization, higher detail fidelity, and greater consistency for thin vessels and fragmented branches. However, the focus on hotspots tends to lead to greater memory and computational burden, and the multi-branch fusion strategy is complex to design. Furthermore, if its details lack a gating noise function, it may still produce pseudo-neural responses when noise textures are strong or lesions are obvious.

[0166] 1. Comparison Indicators

[0167] This embodiment constructs a multi-dimensional evaluation system for retinal vessel segmentation tasks, comprehensively assessing model performance from multiple perspectives, including pixel-level classification accuracy, positive class detection capability, negative class exclusion capability, and overall discriminative ability. Specifically, this study uses accuracy (ACC), sensitivity (SE), specificity (SP), F1 score (F1), and area under the curve (AUC) as model performance evaluation metrics. The models are compared with mainstream models on three public datasets: DRIVE, CHASEDB1, and STARE. Through comprehensive analysis of these metrics, the actual performance of different models in small vessel detection, background staining, structural defects, and overall robustness can be more accurately reflected, thus providing a correlation for comparing the merits of different methods.

[0168] ACC (Accuracy): Represents the proportion of pixels correctly predicted by the model out of the total number of pixels, indicating the overall accuracy of segmentation. This metric is insightful and easy to understand, suitable for evaluating the overall classification performance of the model. However, in blood vessel segmentation tasks, background pixels often far outnumber blood vessel pixels, resulting in a significant imbalance in class distribution. In such cases, relying solely on ACC may overlook the model's ability to miss small blood vessels, leading to an overestimation of the model's capabilities.

[0169] Sensitivity (SE): This represents the proportion of pixels that are actually blood vessels, correctly identified as blood vessels by the model. It is used to quickly assess the model's ability to detect positive blood vessel classes. This metric is crucial for scenarios like disease where "it's better to detect more than less" and reflects the model's ability to capture small blood vessels and low-risk areas. However, SE only focuses on the model's positive vessel detection, and the model also introduces false blood vessel responses (false detections). Therefore, SE may exist while the actual segmentation quality is not ideal. Thus, it is usually evaluated in conjunction with metrics such as SP and F1.

[0170] SP (Specificity): Represents the proportion of pixels that are correctly identified as background by the model out of all pixels that are actually background. It is used to evaluate the model's ability to repair the background of a region. This metric can reflect the model's ability to handle delay noise, lesion interference, and false detection of non-vascular structures. However, SP mainly focuses on negative class performance. If the model over-classifies (causally identifies), it may lead to a high SP and a low SE, resulting in missed detection of small blood vessels. Therefore, it needs to be analyzed in conjunction with SE to demonstrate the balance.

[0171] The F1 score (F1-score) is the adjusted average of ACC and SE, used to comprehensively evaluate the model's ability to balance low false positives and low false negatives. F1 is sensitive to class imbalance and can more realistically reflect the overall quality of positive class recognition in vessel segmentation tasks, making it particularly suitable for evaluating the segmentation performance of small vessels and bifurcation regions. However, F1 depends on the threshold selection; changes in the threshold will affect the precision and precision-recall values. Therefore, when comparing different methods, a consistent threshold or a unified post-processing strategy should be adopted.

[0172] AUC (Area Under ROC Curve): The ROC curve depicts the relationship between the TPR (SE) and FPR of a model under different thresholds. The area under the AUC curve simplifies the model's overall discriminative ability and threshold-independent stability. A UC closer to 1 indicates a stronger ability to distinguish blood vessels from the background, comprehensively reflecting the model's overall performance under different decision thresholds. However, in pixel-level highly imbalanced datasets, AUC may be relatively optimistic and may not fully reflect the true segmentation quality of small blood vessel regions when used alone. Therefore, it needs to be analyzed in conjunction with other metrics such as F1 and SE.

[0173] In summary, each index describes the model performance from different perspectives and has its own limitations. This embodiment adopts a comprehensive evaluation method of ACC, SE, SP, F1 and AUC to compare and analyze the segmentation effect of the present invention with that of existing methods, so as to explore the advantages of the present invention in terms of blood vessel detection capability, background occlusion capability and overall robustness in a more in-depth way.

[0174] 2. The comparison results are shown in Table 1 below.

[0175] Table 1

[0176]

[0177] As can be seen from Table 1, this invention has been comprehensively evaluated on three mainstream public datasets for retinal vessel segmentation: DRIVE, CHASEDB1, and STARE, and has achieved best or highly competitive performance on several key metrics.

[0178] On ACC, this invention achieved the highest values ​​on all three datasets: DRIVE (0.9773), CHASEDB1 (0.9816), and STARE (0.9885). Compared to U-Net, ACC improved by 0.0212, 0.0237, and 0.0304 respectively, and achieved the highest value among all comparison methods on STARE, demonstrating the stability improvement of this invention in overall pixel-level classification accuracy.

[0179] On SE, this invention achieves 0.8318 and 0.8471 on DRIVE and CHASEDB1, respectively. On STARE, the SE of this invention is 0.8398, slightly lower than HiDiffSeg's 0.8433, but still significantly higher than U-Net's 0.7943, indicating that this invention has a stronger overall detection capability for positive blood vessel pixels, especially small blood vessels and weakly repaired areas.

[0180] On SP, this invention achieved the highest values ​​of 0.9875 and 0.9834 on DRIVE and STARE respectively, indicating stronger background region restoration capabilities and fewer false positives. On CHASEDB1, the invention's SP was 0.9825, slightly lower than HiDiffSeg's 0.9837, but still in the top tier, demonstrating a competitive advantage in background texture and noise suppression.

[0181] In terms of F1, this invention achieves the highest values ​​on CHASEDB1 and STARRE, reaching 0.8463 and 0.8392 respectively, indicating a better trade-off between SE and SP. On DRIVE, the invention's F1 is 0.8467, slightly lower than HiDiffSeg's 0.8479, but the difference is small. Overall, its performance still significantly supports U-Net (0.8162) and most Transformer / UNet variants, demonstrating a balanced stability and robustness.

[0182] In terms of AUC, this invention achieved the highest AUC on all three datasets: DRIVE (0.9896), CHASEDB1 (0.9869), and STARE (0.9889). Particularly on STARE, the AUC improved by 0.1063 compared to U-Net's 0.8826, indicating that this invention has stronger background detection capabilities for CVD and more stable generalization under different threshold conditions.

[0183] Example 3: This example proposes a stimulus-guided multi-branch enhanced retinal vessel segmentation network system, applying the steps of the method described in this invention. The system includes:

[0184] The acquisition module is configured to perform the following process: segmenting fundus color image data and pairing it with pixel-level blood vessel annotations; performing standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and dividing the sample set into a training set and a test set;

[0185] The boundary enhancement module is configured to perform the following process: explicitly modeling the blood vessel edge information by combining learnable dual-scale Sobel gradients to obtain the blood vessel edge amplitude response map, and further introducing a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage.

[0186] The multi-scale feature aggregation module is configured to perform the following process: fuse and interact multi-scale features from different levels at a uniform resolution, realize channel-wise weighted and global feature reconstruction through scale weight modeling, and enhance the collaborative expression capability of cross-scale semantics and detailed information.

[0187] The stimulus-guided gating fusion module is configured to perform the following process: a dual-branch structure perception mechanism is used to model the global response and local peak features respectively, and learnable gating parameters are introduced in the channel dimension to achieve dynamic balance and adaptive fusion of the two.

[0188] The vessel segmentation module is configured to perform the following process: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing, calculates and outputs a pixel-level vessel probability map with the same resolution as the original image, and then thresholds the probability map to remove spurious responses and smooth the boundaries, thereby obtaining the final vessel segmentation result for the corresponding test sample.

[0189] Example 4: This example proposes an electronic system, including: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, the instructions being executed by the at least one processor to enable the at least one processor to perform the method steps of the present invention.

[0190] Example 5: This example proposes a computer-readable storage medium storing a computer program thereon. When the computer program is executed by a processor, it implements the steps of the method described in this invention, which will not be repeated here.

[0191] Example 6: This example proposes a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, they implement the steps of the method described in this invention, which will not be repeated here.

[0192] It should be noted that the processing flow of embodiments 3-6 corresponds to the specific steps of the method provided in embodiment 1 of the present invention, and has the corresponding functional modules and beneficial effects of the method. Technical details not described in detail in this embodiment can be found in the method provided in embodiment 1 of the present invention.

[0193] The program code used to implement the methods of this application may be written in any combination of one or more programming languages. This program code may be provided to a processor or controller of a general-purpose computer, special-purpose computer, or other programmable data processing device, such that when executed by the processor or controller, the functions / operations specified in the flowcharts and / or block diagrams are implemented. The program code may be executed entirely on a machine, partially on a machine, as a standalone software package partially on a machine and partially on a remote machine, or entirely on a remote machine or server.

[0194] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A stimulus-guided method for segmenting multi-branch enhanced retinal vessels, characterized in that, Includes the following steps: Step S1: Collect fundus color image data to be segmented and pair it with pixel-level blood vessel annotations; perform standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and further divide the sample set into training set and test set; Step S2: Construct a boundary enhancement module. By combining learnable dual-scale Sobel gradients to explicitly model blood vessel edge information, obtain the blood vessel edge amplitude response map, and introduce a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, thereby highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage. Step S3: Construct a multi-scale feature aggregation module to fuse and interact multi-scale features from different levels at a unified resolution. Through scale weight modeling, achieve channel-wise weighting and global feature reconstruction to enhance the collaborative expression capability of cross-scale semantics and detailed information. Step S4: Construct a stimulus-guided gating fusion module, adopt a dual-branch structure perception mechanism to model the global response and local peak features respectively, and introduce learnable gating parameters in the channel dimension to achieve dynamic balance and adaptive fusion of the two. Step S5: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing to calculate and output a pixel-level blood vessel probability map with the same resolution as the original image. Then, the probability map is thresholded to remove spurious responses and smooth the boundaries, thereby obtaining the final blood vessel segmentation result of the corresponding test sample.

2. The stimulus-guided multi-branch enhanced retinal vessel segmentation network according to claim 1, characterized in that, Step S2 includes the following steps: Step S2.1: Input multi-channel image pass Convolution mapping to a single-channel grayscale image Through the formula: （1）； in express convolution, The number of input image channels, For batches, and For the height and width of the image, For indexing, For input number One channel, This represents the image after grayscale processing. Step S2.2: At the original scale With downsampling scale Calculate separately and The Sobel gradient in the direction is calculated, and the corresponding edge gradient magnitude matrix is ​​obtained: （2）； （3）； in and They are of direction and Gradient of direction, This is the downsampled feature map. and yes of direction and Gradient of direction, It is a numerical constant. and The obtained edge gradient magnitude matrix; Step S2.3: Align the amplitude responses of the two scales spatially, then stitch them together, and perform pixel-level Softmax along the channel dimension to obtain adaptive fusion weights, thereby achieving adaptive fusion of edge intensity at the two scales. （4）； （5）； in, Indicates will The result after upsampling alignment This indicates a splicing operation. This indicates the Softmax operation. and These are the pixel-level fusion weights corresponding to the two scales. For Hadama accumulation, To obtain the final fusion vector, robust normalization is used to compress the magnitude map to... interval; Step S2.4: Apply soft thresholding to the normalized amplitude map to generate the final boundary attention map. （6）； in for Standardized features , and These are the amplitude coefficient, bias, and threshold center, respectively. The slope is the angle; the larger the slope, the closer it is to the hard threshold. It is the Sigmoid activation function. This is the final attention map.

3. The stimulus-guided multi-branch enhanced retinal vessel segmentation network according to claim 1, characterized in that, Step S3 includes the following steps: Step S3.1: Denote the multi-scale features of the decoder output at different scales as follows: ,in For each scale feature, the sequence number is used for... Convolutional projection and normalization to unify the number of channels. Based on the resolution of the last layer of the decoder, the data is upsampled to a unified spatial dimension to obtain aligned features. Global average pooling is then performed to obtain channel statistics at each scale. These statistics are then averaged and aggregated along the scale dimension, and a multilayer perceptron (MLP) is used for dimensionality adjustment to obtain channel descriptions, i.e., channel score vectors at each scale. The hidden layer dimension is... , For dimensionality reduction ratio; Step S3.2: After stacking the scoring vectors at all scales, calculate softmax normalization along the scale dimension to obtain channel-wise attention, and then perform weighted fusion across scales: （7）； in For the scale number, For the first Channel scoring vectors of various scales Features aligned at each scale Indicates the feature sequence index. Indicates the channel number index. and All are indexes. This is the backbone feature map after cross-scale weighted fusion; Step S3.3: After aligning all scales, connect them along the channel dimension, and use... Convolutional layers compress channels back After normalization and activation by the GELU function, it is finally connected to the backbone channel via residual connections: （8）； （9）； in These are the weighting coefficients. These are learnable residual coefficients. For GELU activation function, For normalization operations, For residual characteristics, Main feature map; Step S3.4: Project the backbone and refined features onto the desired channel dimension through a fusion module consisting of convolutional layers and a channel attention mechanism: （10）； （11）； in yes Encapsulation of convolution, normalization, and activation functions. For channel attention, For output head, The feature vector before projection. This is the final output feature.

4. The stimulus-guided multi-branch enhanced retinal vessel segmentation network according to claim 1, characterized in that, Step S4 includes the following steps: Step S4.1: For the input images at various scales First use Convolution, batch normalization, and GELU activation functions are used to enhance local context representation and stabilize the distribution. （12）； in For convolution kernel, The extracted feature map; Step S4.2: Construct two types of complementary stimuli from a global perspective, for... Perform global average pooling and global max pooling respectively to obtain the channel-level response vector: ， （13）； in and These represent global average pooling and global max pooling, respectively. and These represent the features after average pooling and max pooling, respectively. Step S4.3: [The text appears to be incomplete and contains several grammatical errors. A more accurate translation would require and Inputting each data point into a dual-branch MLP generates two types of channel importance descriptions and average trend vectors. With peak significance vector : （14）； （15）； in , , and These are the weighting coefficients. It is the ReLU activation function. and These are the average trend and peak significance, respectively, and their dimensions are mapped to... , respectively denoted as and ; Step S4.4: To enable the model to adaptively learn the weight ratios of the average trend and peak value in each channel, the two stimulus vectors are concatenated and then... Convolution and normalization, and generating dynamic gating coefficients using Sigmoid. : （16）； in This indicates a splicing operation. For dynamic gating coefficients; Step S4.5: Weight the two types of candidate weights according to the dynamic gating coefficient in the channel dimension, and obtain the final attention map through Sigmoid mapping. : （17）； in The attention map obtained from the mapping; Apply attention and linear mapping, and introduce residual connections: （18）； in The coefficient matrix, That is, the feature map after each scale has been gated by the channel.

5. A stimulus-guided multi-branch enhanced retinal vessel segmentation network system, characterized in that, The system comprising the steps of applying the method according to any one of claims 1 to 4, wherein the system includes: The acquisition module is configured to perform the following process: segmenting fundus color image data and pairing it with pixel-level blood vessel annotations; performing standardized preprocessing operations on the original images, including size unification, intensity normalization and necessary data augmentation, to obtain a standardized sample set, and dividing the sample set into a training set and a test set; The boundary enhancement module is configured to perform the following process: explicitly modeling the blood vessel edge information by combining learnable dual-scale Sobel gradients to obtain the blood vessel edge amplitude response map, and further introducing a parameterized Sigmoid gating mechanism to adaptively modulate the edge response, highlighting small blood vessels and low-contrast boundary structures in the feature extraction stage. The multi-scale feature aggregation module is configured to perform the following process: fuse and interact multi-scale features from different levels at a uniform resolution, realize channel-wise weighted and global feature reconstruction through scale weight modeling, and enhance the collaborative expression capability of cross-scale semantics and detailed information. The stimulus-guided gating fusion module is configured to perform the following process: a dual-branch structure perception mechanism is used to model the global response and local peak features respectively, and learnable gating parameters are introduced in the channel dimension to achieve dynamic balance and adaptive fusion of the two. The vessel segmentation module is configured to perform the following process: After the model training is completed, for each test image to be segmented, the network sequentially performs boundary enhancement, multi-scale global aggregation and stimulus-guided gating fusion cascade processing, calculates and outputs a pixel-level vessel probability map with the same resolution as the original image, and then thresholds the probability map to remove spurious responses and smooth the boundaries, thereby obtaining the final vessel segmentation result for the corresponding test sample.

6. 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 computer program is executed, it implements the steps of the method as described in any one of claims 1 to 4.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, The computer program is configured to implement the steps of the method according to any one of claims 1 to 4 when invoked by a processor.

8. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the method according to any one of claims 1 to 4.

Images