Retinal vessel segmentation method based on multi-dimensional space perception and frequency domain calibration

Abstract

The application relates to the technical field of medical image processing, and particularly discloses a retinal blood vessel segmentation method based on multi-dimensional space perception and frequency domain calibration, which comprises the following steps: constructing a retinal blood vessel segmentation network; the retinal blood vessel segmentation network comprises an encoder unit, a bottleneck fusion hub and a decoder unit; the encoder unit comprises multi-level feature perception blocks, each level of the feature perception block is formed by cascading a stripe convolution structure and a multi-dimensional frequency domain calibration module; the bottleneck fusion hub is located at the connection position of the encoder unit and the decoder unit; the decoding unit comprises an adaptive edge fusion block and a stripe guided attention block; based on the retinal blood vessel segmentation network, an eye fundus image to be segmented is segmented to obtain a blood vessel segmentation prediction graph. Through the synergistic effect of space form perception and frequency domain feature calibration, the capturing capacity of fine terminal blood vessels and the anti-disease noise interference capacity are significantly enhanced, and the problems of blood vessel fracture and edge blur in the segmentation result are effectively solved.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and more specifically to a method for retinal vessel segmentation based on multidimensional spatial perception and frequency domain calibration. Background Technology

[0002] Retinal vessels are the only deep microcirculatory system in the human body that can be directly and non-invasively observed. Their morphological characteristics (such as vessel diameter, branching angle, curvature, and spatial distribution) are important biomarkers for the clinical diagnosis of diabetic retinopathy, glaucoma, hypertension, and cardiovascular and cerebrovascular diseases. Therefore, precise retinal vessel segmentation has significant clinical value for achieving automated screening and early diagnosis of related diseases.

[0003] However, in real-world clinical scenarios, automated vascular segmentation faces the following significant challenges: Fundus images often suffer from uneven contrast and illumination during imaging; simultaneously, lesions such as hemorrhages, exudates, and microaneurysms on the patient's retina have color and contrast characteristics extremely similar to blood vessels, easily leading to misdiagnosis by the algorithm. The retinal vascular system comprises vessels ranging from large main trunks to extremely fine terminal vessels. Traditional convolutional neural networks (such as U-Net) easily lose structural information of small vessels after multiple downsampling operations, resulting in broken or missing terminal vessels in the segmentation results. Furthermore, there are limitations in receptive field and morphological perception. Existing deep learning methods mostly employ standard... Features are extracted using rectangular convolutional kernels. However, due to the elongated, linear geometric shape of blood vessels with varying curvature, isotropic rectangular convolutional kernels struggle to accurately capture directional information, and the local receptive field limits the model's ability to acquire long-range spatial dependencies. Furthermore, most algorithms only process intensity information in the spatial domain, neglecting the potential of image frequency domain features in separating noise and structural information, resulting in insufficient robustness against complex noise interference.

[0004] Therefore, existing segmentation methods still have significant shortcomings in handling small blood vessels, resisting lesion noise interference, and long-range context modeling. Thus, how to address these shortcomings and improve the accuracy of retinal vessel segmentation is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] In view of the above problems, the present invention proposes a retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration, so as to overcome the above problems or at least partially solve the above problems.

[0006] To achieve the above objectives, the present invention adopts the following technical solution:

[0007] A retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration includes: Construct a sample set of fundus images; A retinal vessel segmentation network is constructed. The retinal vessel segmentation network includes an encoder unit, a bottleneck fusion center, and a decoder unit. The encoder unit contains multi-level feature sensing blocks, and each level of feature sensing block is composed of a cascaded stripe convolutional structure and a multi-dimensional frequency domain calibration module. The bottleneck fusion center is located at the connection between the encoder unit and the decoder unit. The decoder unit contains an adaptive edge fusion block and a stripe-guided attention block. A retinal vessel segmentation network was trained and validated using a set of fundus image samples. The trained and validated retinal vessel segmentation network is used to segment the fundus image to be segmented, and a vessel segmentation prediction map is obtained. The process of segmenting the fundus image to be segmented includes: The directional morphological features of blood vessels in fundus images are captured by a striped convolutional structure, and the directional morphological features of blood vessels are enhanced by a multi-dimensional frequency domain calibration module to obtain frequency domain enhanced features. The bottleneck fusion hub aggregates the frequency domain enhancement features of each level in the encoder unit and performs global modeling. The gating injection mechanism is used to feed the global context back to the frequency domain enhancement features of each level to obtain the coding fusion features. In the deep nodes of the decoder unit, residual cleaning is performed on the skip path through adaptive edge fusion blocks, and the feature semantics are purified to obtain deep recovery features. After processing, the deep recovery features are input into the shallow nodes of the decoder unit as the decoding features of the shallow nodes. In the shallow nodes of the decoder unit, the decoding features of the shallow nodes are adaptively fused with the corresponding layer encoding fusion features through stripe-guided attention blocks to obtain shallow recovery features. At the end of the decoder unit, the shallow recovery features are processed to obtain the final blood vessel segmentation prediction map.

[0008] Furthermore, the process of constructing the fundus image sample set includes: The green channel component of the original fundus image is extracted, and then adaptive histogram equalization with contrast limitation and nonlinear Gamma correction are performed sequentially. An overlapping sliding window is used to cut the image, retaining image blocks that meet the blood vessel density threshold condition; Geometric transformations are performed on the selected image patches to generate an enhanced set of fundus image samples.

[0009] Furthermore, the process of capturing the directional morphological features of blood vessels in fundus images through striped convolutional structures includes: The preprocessed fundus image is subjected to a 3×3 convolution and ReLU operation to obtain the primary spatial features F. in ; Four parallel convolutional branches are used to extract the primary spatial features F. in The extremely long stripe features, medium stripe features, local perception features, and short-range perception features; Extracting channel vectors using global average pooling , channel vector Attention response values ​​for each convolutional branch are generated using a multilayer perceptron, and the weights of each convolutional branch are obtained by normalization using the Softmax function. w i The features extracted from each convolutional branch are weighted and fused using the weights of each convolutional branch, and then passed through a 1×1 convolutional layer to obtain the stripe perception features. f strip ; Using one-dimensional horizontal and vertical convolution kernels, the primary spatial features F are respectively processed. in Component aggregation is performed to capture long-range spatial dependencies in the horizontal and vertical directions, resulting in orientation-aware feature pairs. z h , z w ; Direction-aware feature pairs z h , z w The data is concatenated, and a 1×1 shared convolutional layer is used to perform channel compression and nonlinear mapping to generate the encoded spatial location vector. f temp ; Horizontal attention weights are generated through dimension splitting, convolutional mapping, and the Sigmoid activation function. a h and vertical attention weights a w Utilizing horizontal attention weights a h and vertical attention weights a w Element-wise multiplication of the input feature x yields the orientation-sensitive feature. f coord ; Extracting stripe-sensing features using global average pooling f strip With direction-sensitive features f coord The global statistics are used to generate mutual-directed attention weights through a shared multilayer perceptron and sigmoid activation function. Att strip and Att coord ; Utilizing mutual guidance attention weights Attstrip and Att coord Stripe perception features f strip With direction-sensitive features f coord Perform cross-feature modulation to obtain cross-spatial fused features. f cross ; Cross-space fusion features f cross With primary spatial features F in Perform residual connections to obtain the directional morphological features F of the blood vessels in the current layer. out .

[0010] Furthermore, the process of enhancing the directional morphological features of blood vessels through a multi-dimensional frequency domain calibration module includes: Directional morphological features F out As the input feature of the multi-dimensional frequency domain calibration module, the input feature F out Perform a Fast Fourier Transform to convert it to frequency domain features; By constructing three independent Gaussian frequency band masks corresponding to the low-frequency, mid-frequency, and high-frequency regions respectively, and introducing learnable channel-wise weight parameters combined with the softplus activation function, an adaptive filter for each frequency band is obtained. The adaptive filters of the three frequency bands are linearly combined to form the overall frequency domain filter. F total ; The frequency domain features are filtered based on the overall frequency domain filter, and the filtered features are mapped back to the spatial domain features through inverse Fourier transform. Spatial domain features and input features F out Perform residual connections to obtain preliminary frequency domain enhancement features. ; Employing a hybrid attention mechanism to enhance initial frequency domain features Feature selection and enhancement are performed to obtain the final frequency domain enhanced features. .

[0011] Furthermore, a hybrid attention mechanism is employed to enhance the frequency domain output features. The process of feature selection and enhancement includes: On the channel, preliminary frequency domain enhancement features are aggregated in parallel using global average pooling and global max pooling. The spatial information is used to generate two different spatial context feature vectors; these two spatial context feature vectors are then fed into a shared multilayer perceptron to extract the dependencies between channels, and finally a channel weight map is generated using a sigmoid activation function. M c ; Using channel weight graph M c Preliminary frequency domain enhancement features Perform channel calibration to obtain channel refinement features. ; Channel refinement features Average pooling and max pooling are applied along the channel dimension to generate two two-dimensional spatial information feature vectors. The two spatial information feature vectors are concatenated and fused through a 7×7 convolutional layer. A spatial weight map M is generated by applying a sigmoid activation function. s ; Using spatial attention map Ms and channel refinement features Pixel-by-pixel multiplication yields the final frequency domain enhancement feature f. att .

[0012] Furthermore, the process of aggregating the frequency domain enhancement features of each level in the encoder unit through the bottleneck fusion hub and performing global modeling, and then using a gated injection mechanism to feed the global context back to the frequency domain enhancement features of each level includes: The frequency domain enhancement features of each level of the encoder unit are used as the coding features F of each level. i ; The encoding features F at different scales at each level of the encoder unit are used to encode the features. i Project and pool to a uniform 12×12 fixed size, and obtain the full-scale converged feature S2 by pixel-level summation; The full-scale converged feature S2 is flattened along the spatial dimension, transforming it into a two-dimensional sequence form suitable for Transformer processing. The sequence z is input into the global context transformer for modeling; After performing long-range modeling, the output feature sequence is reconstructed in terms of spatial dimensions, and it is remapped into a two-dimensional tensor that conforms to the original spatial topology. Z global ; By using 1×1 convolution operations in conjunction with batch normalization, the two-dimensional tensor Z is transformed. global The number of channels is mapped from the embedding dimension C back to the original physical channel depth C of the corresponding encoder unit level. i To obtain the global features after channel alignment at the corresponding level. ; Bilinear interpolation algorithm is used to perform global feature matching after channel alignment. Perform spatial domain resampling to obtain spatially aligned enhanced global features at the corresponding level. ; Will enhance global features Input a gated branch consisting of a 3×3 convolutional layer and a sigmoid activation function, and generate a spatially gated weight map. Gate i Spatial gating weight map Gate i Used to identify prominent vascular areas from a global perspective; Introducing learnable residual injection coefficients g i Adaptive injection calibration is performed on the coding features at each level of the encoder unit to obtain the coding fusion features. The calibration formula is:

[0013] Where i represents the i-th level of the encoder unit.

[0014] Furthermore, the process of obtaining deep recovery features through adaptive edge fusion blocks includes: Residual path processing is performed on the deep coding fusion features of the encoder unit. The encoded features after residual path processing are mapped to a unified dimensional space and concatenated with the decoded features of the corresponding level to obtain feature F. res ; Utilizing the bottleneck structure of an autoencoder to analyze feature F res Perform spatial domain purification to generate a spatial gating graph. M gate ; Using spatial gating diagrams M gate For feature F res Adaptive reweighting is performed to obtain refined features. ; For feature F res Perform max pooling and average pooling, and calculate the difference between the two pooling results to obtain the display edge map. E edge ; Purification characteristics through splicing With display edge map E edge This yields the final deep recovery features. .

[0015] Furthermore, the generation process of shallow restoration features includes: Deep recovery features After processing by the input convolutional module, a 2x upsampling operation is performed to obtain the decoding features of the shallow nodes of the decoder unit. ; The encoded fusion features of the i-th layer of the encoder unit are concatenated with the encoded features of the corresponding layer. Then, residual path processing is performed on the concatenated features. Finally, the encoded features after residual path processing are combined with the decoded features of the shallow nodes. Mapping to a unified dimensional space and concatenating the results yields the features. F sum ; Using bidirectional stripe convolution to pair features F sum Feature extraction is performed to obtain the directional morphological response of blood vessels, and the features are obtained after batch normalization and ReLU activation function. f strip ; Using 1×1 convolutional layers for features f strip A non-linear mapping along the channel dimension is performed, and a morphology-aware weight map is generated by combining it with the Sigmoid activation function. W In the shape-perceived weight map W In the image, the pixel value corresponding to the blood vessel area is close to 1, while the pixel value of the background area is close to 0. Using the generated shape-aware weight map W Asynchronous gating fusion is performed on the decoding features of shallow nodes and the encoding features of the corresponding layers to obtain shallow recovery features.

[0016] Furthermore, when training the retinal vessel segmentation network, the binary cross-entropy loss and Dice loss are weighted and summed to form the total loss function.

[0017] Furthermore, at the end of the decoder unit, a lightweight processing module consisting of multiple convolutional layers is set up to perform nonlinear transformation and channel compression on the shallow recovery features to obtain the final blood vessel segmentation prediction map.

[0018] As can be seen from the above technical solution, compared with the prior art, the present invention has the following beneficial effects: (1) To improve the geometric accuracy of morphological perception, this invention constructs a striped convolutional structure and uses a long strip receptive field instead of a single square receptive field for feature extraction, which significantly enhances the ability to perceive the direction of small blood vessels and solves the problem of blurred edges.

[0019] (2) To address the interference resistance of complex backgrounds, this invention proposes a frequency domain calibration module, which uses a learnable frequency domain mask to separate noise and vascular signals, significantly amplifying the difference between vascular and background lesion elements, and improving the continuity of vascular edge segmentation.

[0020] (3) To address the overall synergy of long-range contexts, this invention constructs a bottleneck fusion center, which captures cross-scale vascular spatial distribution features by performing global modeling on a 12×12 resolution feature map. Furthermore, through a bidirectional gating fusion mechanism, it effectively repairs large vessel ruptures and peripheral defects, ensuring the integrity of the segmentation results.

[0021] (4) To address the purity and precision of feature recovery, this invention constructs an asynchronous recovery link consisting of an adaptive edge fusion block and a stripe-guided attention block. It utilizes the compression and reconstruction mechanism of an autoencoder to eliminate deep semantic redundancy and enhance explicit edges. Combined with the stripe perception gating mechanism, it accurately locks the superficial vascular terminal, achieving high-precision extraction of microvessels and effective suppression of lesion artifacts. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0023] Figure 1 This is a flowchart of the retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration provided in the embodiments of the present invention; Figure 2 This is an overall architecture diagram of the retinal vessel segmentation network provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the striped convolution structure provided in an embodiment of the present invention; Figure 4 This is a schematic diagram of the multi-dimensional frequency domain calibration module provided in an embodiment of the present invention; Figure 5 This is a schematic diagram of the bottleneck fusion hub provided in an embodiment of the present invention; Figure 6 This is a schematic diagram of the adaptive edge blending block and stripe-guided attention block provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the fundus image feature visualization process provided in an embodiment of the present invention; Figure 8 This is a schematic diagram of intelligent slicing and data augmentation of the experimental dataset provided in this embodiment of the invention; Figure 9 This is a comparative visualization of the CHASE_DB1 dataset provided in this embodiment of the invention. Figure 1 ; Figure 10 This is a comparative visualization of the DRIVE dataset provided in this embodiment of the invention. Figure 2 . Detailed Implementation

[0024] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] like Figure 1 As shown, this embodiment of the invention discloses a retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration, including: Construct a sample set of fundus images; A retinal vessel segmentation network is constructed. The retinal vessel segmentation network includes an encoder unit, a bottleneck fusion center, and a decoder unit. The encoder unit contains multi-level feature sensing blocks, and each level of feature sensing block is composed of a cascaded stripe convolutional structure and a multi-dimensional frequency domain calibration module. The bottleneck fusion center is located at the connection between the encoder unit and the decoder unit. The decoder unit contains an adaptive edge fusion block and a stripe-guided attention block. A retinal vessel segmentation network was trained and validated using a set of fundus image samples. The trained and validated retinal vessel segmentation network is used to segment the fundus image to be segmented, and a vessel segmentation prediction map is obtained. The process of segmenting the fundus image to be segmented includes: The directional morphological features of blood vessels in fundus images are captured by a striped convolutional structure, and the directional morphological features of blood vessels are enhanced by a multi-dimensional frequency domain calibration module to obtain frequency domain enhanced features. The bottleneck fusion hub aggregates the frequency domain enhancement features of each level in the encoder unit and performs global modeling. The gating injection mechanism is used to feed the global context back to the frequency domain enhancement features of each level to obtain the coding fusion features. In the deep nodes of the decoder unit, residual cleaning is performed on the skip path through adaptive edge fusion blocks, and the feature semantics are purified to obtain deep recovery features. The obtained deep recovery features are then fed into the shallow nodes of the decoder unit after convolution and upsampling to obtain the decoding features of the shallow nodes of the decoder unit. In the shallow nodes of the decoder unit, the decoding features of the shallow nodes are adaptively fused with the encoding fusion features of the corresponding layer through stripe-guided attention blocks to obtain shallow recovery features. At the end of the decoder unit, the shallow recovery features are processed to obtain the final blood vessel segmentation prediction map.

[0026] In one specific embodiment, the process of constructing a fundus image sample set includes: 1) Image feature visualization: specifically as follows Figure 7 As shown, the original fundus color image is first acquired, and the green channel component, which is most sensitive to vascular features, is extracted from the original fundus image. To eliminate uneven illumination and enhance the contrast between blood vessels and the background, contrast-limited adaptive histogram equalization (CLAHE) and nonlinear gamma correction are performed sequentially to enhance the edges of blood vessels.

[0027] 2) Intelligent slicing and density screening: Specifically, as shown in... Figure 8 As shown, overlapping sliding windows are used to cut the image. To ensure the quality of the training set, only images that meet the blood vessel density threshold are retained. Threshold Conditional image patch; blood vessel density threshold Threshold The calculation formula is:

[0028] Where PatchSize is the image patch size and Ratio is the preset density ratio coefficient.

[0029] 3) Perform geometric transformations on the selected image blocks, including original, rotated (90 degrees, 180 degrees, 270 degrees) and horizontally flipped images, to generate an enhanced fundus image sample set.

[0030] The enhanced fundus image sample set is divided into a training sample set and an independent test set. The training sample set is randomly divided into a training subset and a validation subset according to a preset ratio. During training, the parameters of the segmentation network are updated using the training subset, and the performance is periodically evaluated and hyperparameters are filtered using the validation subset until the segmentation network converges or reaches a preset number of iterations. Finally, the segmentation network weights with the best performance during the validation process are extracted and applied to the independent test set to output the final retinal vessel segmentation prediction result.

[0031] In this embodiment, the publicly available and standard retinal vessel segmentation datasets DRIVE and CHASEDB1 were selected, and the datasets were partitioned and augmented. For example, the DRIVE dataset contains images with a resolution of 565×584, and is divided into 20 training sets and 20 test sets. The CHASEDB1 dataset contains images with a resolution of 999×960, and is divided into 20 training sets and 8 test sets. During training, the training portions of both datasets are divided into training and validation subsets in an 8:2 ratio. Finally, eight-fold symmetric augmentation (including rotations of 90°, 180°, and 270°, and horizontal flipping) is performed on the selected image patches to form the final training image set.

[0032] Next, we will explain in detail the process of segmenting fundus images using the retinal vessel segmentation network.

[0033] like Figure 2 As shown, the retinal vessel segmentation network includes an encoder unit, a bottleneck fusion center, and a decoder unit. The encoder unit contains multi-level feature sensing blocks, each of which is composed of a strip-coordinated selective convolutional structure (SCS) and a frequency-enhanced attention module (FEA Module). The bottleneck fusion center is located at the connection between the encoder unit and the decoder unit and mainly utilizes a full-scale convergence module (GCT-GIM). The decoder unit includes an adaptive edge fusion block (R_CAE Block, ResPath-based CleanAdaptive Edge) and a strip-guided attention block (R-SGA, ResPath-based Strip Guided Attention).

[0034] 1. Encoder Unit: The encoder unit consists of three levels of feature sensing blocks, each level being a cascaded integrated stripe convolution structure and a multi-dimensional frequency domain calibration module.

[0035] 1) such as Figure 3 As shown, the process of capturing the directional morphological features of blood vessels in fundus images using a striped convolutional structure includes: ① The fringe convolution structure of the encoder unit adopts a two-stage feature extraction strategy. The preprocessed fundus image X is subjected to a 3×3 convolution and ReLU operation to obtain the primary spatial features F. in :

[0036] in, This represents the ReLU activation function. This represents a two-dimensional batch normalization layer.

[0037] ② Subsequently, primary spatial features The mathematical implementation process for shape perception in the SCS core module is as follows: Dynamic weight allocation in Selective Kernel Fusion (SK-Fusion): Four parallel convolutional branches (K×1, 1×K, 3×3, and 1×1) are used to extract input features respectively. Extremely long stripe features, medium stripe features, local perception features, and short-range perception features f long , f mid ,f local , f short Among them, the striped convolution branch uses elongated receptive fields to capture the directional morphology of blood vessels, and the kernel size of the striped convolution... K The value range is 5 to 15. Specifically, this invention uses 11×1 and 1×11 strip convolutions to extract large-scale long strip contextual information, obtaining features. f long 7×1 and 1×7 strip convolutions are used to capture medium-range strip features, resulting in feature... f mid 3×3 convolution is used to capture local neighborhood details and obtain features. f local A 1×1 convolution is used to integrate cross-channel information, preserving fine point features to obtain the feature set. f short .

[0038] ③ The features extracted from the four convolutional branches are weighted and fused to obtain the aggregated feature S1:

[0039] The obtained aggregated features S1 are used to extract channel vectors through global average pooling (GAP) operation. Then will Attention response values ​​for each branch and each channel are generated using a multilayer perceptron (MLP). Then, the channel-wise branch weights are obtained by normalizing each channel along the branch dimension using a softmax function. w i Finally, the weights of each branch are calculated. w i With the corresponding original feature f i Weighted fusion is performed to obtain stripe perception features. f strip The specific calculation process is as follows:

[0040]

[0041] in, , Z Channel statistics are used to describe the importance of each channel; i Indicates the first Attention weights for each convolutional branch; Indicates the use of generating the first A fully connected mapping function for the weights of each branch; This represents the fused features of the output, i.e., stripe perception features.

[0042] ④ Next, coordinate-aware branches are constructed synchronously through coordinate localization encoding to extract orientation-sensitive features. f coord : First, in order to extract input features Figure X Sensitive information about height (H) and width (W) is utilized by a one-dimensional horizontal convolution kernel. H ,1) with vertical convolution kernel (1, W ), employing directional average pooling (Mean) to perform operations in the H and W directions on the input feature F in Performing Mean operations along the H and W dimensions preserves spatial information in the height and width dimensions, capturing long-range spatial dependencies in the horizontal and vertical directions, and obtaining orientation-aware feature pairs. z h , z w Then, the orientation-aware features are paired. z h , z w The data is concatenated, and a 1×1 shared convolutional layer is used to perform channel compression and nonlinear mapping to generate the encoded spatial location vector. f temp Finally, horizontal attention weights are generated through dimensionality splitting, convolutional mapping, and the Sigmoid activation function. a h and vertical attention weights a w Utilizing horizontal attention weights a h and vertical attention weights a w Element-wise multiplication of the input feature x yields the orientation-sensitive feature. f coord This enables precise anchoring of blood vessel locations. The specific calculation process is as follows:

[0043]

[0044]

[0045]

[0046]

[0047] Where [ , ] represents splicing operations along spatial dimensions; BN Indicates batch normalization; Represents a nonlinear activation function; Represents element-wise multiplication; represents the Sigmoid activation function; Split represents the splitting operation along the spatial dimension; x represents the original input feature.

[0048] ⑤ Cross-spatial interaction: Utilizing the cross-spatial learning module to perform f coord and f strip To achieve deep synergy between stripe perception features and spatial localization features, this invention introduces a cross-spatial interactive learning mechanism at the tail of the SCS structure. This mechanism enhances the joint modeling of slender blood vessels and their spatial locations through mutual guidance between features. The specific calculation process is as follows: First, stripe-sensing features are extracted using global average pooling. f strip With direction-sensitive features f coord The global statistics are used to generate mutual-directed attention weights through a shared multilayer perceptron and sigmoid activation function. Att strip and Att coord ; Subsequently, mutual guidance attention weights were utilized. Att strip and Att coord Stripe perception features f strip With direction-sensitive features f coord Perform cross-feature modulation to obtain cross-spatial fused features. f cross The specific calculation process is as follows:

[0049]

[0050]

[0051] Finally, the cross-space fusion feature f cross With input features F in Perform residual connections to obtain the directional morphological features F of the blood vessels in the current layer. out The specific calculation process is as follows:

[0052] in, It carries the a priori linear morphology of blood vessels; and It carries the prior knowledge of the spatial location of blood vessels; This represents the Sigmoid activation function; `.` indicates element-wise multiplication; `Shortcut(.)` indicates residual concatenation.

[0053] 2) such as Figure 4 As shown, the process of enhancing the directional morphological features of blood vessels through a multi-dimensional frequency domain calibration module includes: ①The directional morphological features F out As the input feature of the multi-dimensional frequency domain calibration module, the input feature F out Perform a Fast Fourier Transform to convert it to frequency domain features; construct three independent Gaussian frequency band masks through filter branches. M l , M m , M h These correspond to the low-frequency, mid-frequency, and high-frequency regions, respectively; learnable channel-wise weight parameters are introduced. The adaptive filter for each frequency band is obtained by combining the softplus activation function. The adaptive filters of the three frequency bands are linearly combined to form the overall frequency domain filter h. total The frequency domain features are filtered using an overall frequency domain filter, and the filtered features are then mapped back to the spatial domain features via inverse Fourier transform. ; and combine it with the input feature F out Perform residual connections to obtain preliminary frequency domain enhancement features. The entire process is represented as follows:

[0054]

[0055]

[0056]

[0057]

[0058]

[0059] in, Indicates the input feature map, ( u, v ( ) represents the frequency coordinates; Indicates Fast Fourier Transform; Represents the inverse Fourier transform; resscale represents a learnable channel-wise scaling factor; Clamp() restricts the scaling factor to a specific value. Within the range; This represents the output characteristics after initial enhancement in the frequency domain.

[0060] Through the aforementioned frequency domain filtering mechanism, this invention can adaptively enhance the information expression capability of different frequency components, thereby improving the network's ability to perceive fine-grained structural features.

[0061] ② Employing a hybrid attention mechanism to enhance initial frequency domain features Feature selection and enhancement are performed to obtain the final frequency domain enhanced features. This process is divided into two sequential stages: channel attention mechanism and spatial attention mechanism, which further highlight important semantics and suppress noise. The specific process includes: On the channel, preliminary frequency domain enhancement features are aggregated in parallel using global average pooling and global max pooling. The spatial information is used to generate two different spatial context feature vectors; these two spatial context feature vectors are then fed into a shared multilayer perceptron to extract the dependencies between channels, and finally a channel weight map is generated using a sigmoid activation function. M c Using channel weight graph M c Preliminary frequency domain enhancement features Perform channel calibration to obtain channel refinement features. ; refine channel features Average pooling and max pooling are applied along the channel dimension to generate two two-dimensional spatial information feature vectors. The two spatial information feature vectors are concatenated and fused through a 7×7 convolutional layer. A spatial weight map M is generated by applying a sigmoid activation function. s Using spatial attention map Ms and channel refinement features Pixel-by-pixel multiplication yields the final frequency domain enhancement feature F. att The entire process can be represented by the following formula:

[0062]

[0063]

[0064]

[0065] Where AvgPool represents average pooling; MaxPool represents global pooling; Mean represents a multilayer perceptron; channel Indicates channel average pooling; Max channel Indicates channel max pooling; [ ; ] represents the Sigmoid activation function; [ ; ] represents the concatenation operation along the channel dimension; f 7×7This represents a convolution operation with a kernel size of 7×7; .

[0066] Through the above-mentioned secondary calibration process, the segmentation network can focus on significant vascular structures from a complex background and suppress the feature responses of non-vascular regions.

[0067] The encoder unit uses a three-level cascaded structure to progressively compress spatial resolution and increase channel dimensions to capture multi-scale features ranging from primary texture to high-level semantics. F i Indicates the encoder unit number i The output characteristics of the layer, and its data evolution process, satisfy the following cascading logic:

[0068] In this evolution formula, This represents spatial perception operations that integrate the SCS structure. This represents frequency domain purification and attention calibration operations. Pool This represents a max-pooling operator with a step size of 2. This logic ensures that the network can obtain pure vascular features after frequency domain calibration at each scale, providing a high-quality skip connection source for the asynchronous recovery of the subsequent decoder.

[0069] 2. The bottleneck fusion hub aims to break through the local limitations of the receptive field of traditional convolution. By using the Global Context Transformer-Gated Injection Module (GCT-GIM), the multi-scale features {F1, F2, F3} obtained by the three-layer encoders obtained by the SCS and FEA modules are projected and converged into a unified low-resolution space to perform global topology modeling. Then, the global context is adaptively fed back to the frequency domain enhancement features of each layer using a gated injection mechanism to repair the phenomenon of blood vessel rupture.

[0070] In this embodiment, the input channel is set to C=64, and the initial resolution is 96×96. Therefore, the dimensions of the input encoder features F1, F2, and F3 are 96×96×64, 48×48×128, and 24×24×256, respectively. Figure 5 As shown, the specific process includes: ① The frequency domain enhancement features of each level of the encoder unit are used as the coding features F of each level. i .

[0071] ②Use 1×1 convolutional layers to encode the features F at different scales of each level of the encoder unit. iProjecting to a uniform embedding dimension of 96, and using adaptive average pooling to uniformly scale the features to a fixed 12×12 resolution, the full-scale converged feature S2 is obtained by pixel-level summation; the specific calculation formula is as follows:

[0072]

[0073] Where BN stands for batch normalization; ReLU represents the activation function; Conv 1×1 This represents a 1×1 convolution operation used for smoothing features; , , ; Indicates adaptive average pooling; features after pooling .

[0074] ③ Flatten the full-scale convergence feature S2 along the spatial dimension and convert it into a two-dimensional sequence form suitable for Transformer processing. The sequence z is input into the Global ContextTransformer for modeling. The Global ContextTransformer adopts a standard encoder architecture, which consists of a layer normalization (LN) layer, a multi-head self-attention (MSA) layer, and a linear mapping-based feedforward network (FFN) layer connected in series. It aims to capture the global dependencies and topological connectivity of vascular structures in the image.

[0075] After performing long-range modeling, the output feature sequence is reconstructed in terms of spatial dimensions, and it is remapped into a two-dimensional tensor that conforms to the original spatial topology. Z global The calculation formula for this process is as follows:

[0076]

[0077] in, z The input feature sequence is represented by LayerNorm, which represents the layer normalization operation; MSA(.) represents the multi-head self-attention mechanism; z' represents the output sequence of the self-attention layer; FFN(.) represents a feedforward network containing two fully connected layers and a nonlinear activation function; Reshape(.) represents the spatial transformation operation that restores a one-dimensional sequence to a two-dimensional tensor. Z global This represents the feature map after global modeling.

[0078] ③ To achieve precise alignment between low-resolution global prior information and local detail features at each level in dimensional space, a 1×1 convolution operation combined with batch normalization is used to transform the two-dimensional tensor Z...global The number of channels is mapped from the embedding dimension C back to the original physical channel depth C of the corresponding encoder unit level. i This process ensures that the global features are semantically aligned with the local features, resulting in channel-aligned global features at the corresponding level. The specific calculation formula is as follows:

[0079] Among them, Conv 1×1 This represents a convolution operation with a kernel size of 1×1; This represents the global feature after channel alignment in the i-th layer.

[0080] ④ After completing the channel dimension calibration, a bilinear interpolation algorithm is used to align the global features after channel alignment. Spatial domain resampling is performed, stretching the geometric resolution from an extremely low 12×12 pixel scale to the original resolution corresponding to each level. This process ensures that the higher-level guidance weights can be accurately applied to the blood vessel edge pixels at different scales, resulting in spatially aligned enhanced global features at the corresponding levels. The specific calculation formula is as follows:

[0081] Where Upsample(.) represents the bilinear upsampling operation. This represents the enhanced global features after spatial alignment.

[0082] ⑤ Applying spatial weights to local features enhances global features. The input consists of a gated branch containing a 3×3 convolutional layer and a sigmoid activation function. The spatial gated weight map is generated by smoothing out artifacts caused by interpolation through convolution operations. Gate i Spatial gating weight map Gate i Used to identify significant vascular regions from a global perspective; introduces learnable residual injection coefficients. g i Adaptive injection calibration is performed on the coding features at each level of the encoder unit to obtain the coding fusion features. The specific calculation formula is as follows:

[0083]

[0084] Where i represents the i-th level of the encoder unit, Represents the Sigmoid activation function; Conv 3×3 This represents a 3×3 convolutional layer used for smoothing features; F represents element-wise multiplication; i This represents the i-th layer encoded fusion feature of the input, i.e. Figure 5 F1, F2, F3; g i This represents a learnable scaling factor with an initial value of 0, used to dynamically control the injection strength of global information; This represents the encoded fusion feature of the final output, i.e. Figure 5 In , , .

[0085] This step uses global information to weight and calibrate local responses, effectively suppressing background interference in non-vascular regions while preserving the high-frequency edge details of the original features.

[0086] 3. In the decoder stage, an asynchronous recovery strategy is employed, utilizing the R-CAE and R-SGA modules to specifically address deep semantic gaps and shallow edge noise, respectively. (Reference) Figure 6 The specific implementation details are as follows: 1) At deep nodes, the R-CAE module is used for residual cleaning and autoencoder purification to obtain deep recovery features. The specific process includes: ① In deeper layers of the decoder unit (e.g., 12×12 to 24×24 scales), to bridge the semantic gap between the codec and decoder, this embodiment utilizes the R-CAE module to perform residual cleanup on skip connection paths. First, the i-th layer encoded fusion features obtained from GCT-GIM are... The data is fed into a convolutional block to obtain the refined features. This module consists of two-layer convolution, batch normalization, and ReLU activation, designed to enhance feature representation and suppress noise. The refined features are then processed... Performing a 2x upsampling convolution operation yields the decoding features of shallow nodes in the decoder unit. .

[0087] Subsequently, the encoded fusion features obtained from the i-th layer encoder unit are... and the corresponding level of coding features Feature concatenation is performed, and then the aggregated features are passed through a residual path (ResBlock) of length L=2. The features processed by the residual path are then combined with the decoded features. The features are uniformly mapped to the same dimensional space and multi-scale feature integration is performed through channel concatenation to achieve cross-level information complementarity and optimization, resulting in feature F. res The specific calculation formula is as follows:

[0088]

[0089]

[0090] ② Utilize the bottleneck structure of the autoencoder to analyze feature F res Spatial domain purification is performed on this structure by compressing the channel dimension by a factor of 1 / 4, followed by a dimension-up projection layer. Conv up Reconstruct and generate spatial gating maps M gate The specific calculation formula is as follows:

[0091] The final residual path output is F res = Batch normalization (BN); ReLU represents the ReLU activation function; Dropout represents regularization; ConvTranspose2d() represents the upsampling convolution operation. Represents the Sigmoid activation function; Conv 3×3 This represents a 3×3 area used for smoothing features; F represents the coding fusion feature of the i-th layer of the encoder unit; enci This represents the decoding features of the corresponding decoder unit; Concat() represents the feature concatenation operation, Conv down Represents a dimension-reduced projection layer with a 1 / 4 compression ratio; Conv up This indicates a higher-dimensional projection layer.

[0092] ③ Utilizing spatial gating diagrams M gate The feature F obtained from the residual path res Adaptive reweighting is performed to obtain refined features. To compensate for blurred boundaries in deep features, the module performs morphological gradient calculations to generate explicit edge maps. E edge , for features Perform max pooling (to capture dilation features) and average pooling (to capture erosion features), and calculate the difference between the two pooling results to obtain the edge map. E edge ; Purification features through splicing With display edge map E edge This yields the final deep recovery features. The entire calculation process is represented as follows:

[0093]

[0094]

[0095] Among them, MaxPool and AvgPool represent the max pooling and average pooling operations with a kernel size of 3×3, respectively; [ , ] indicates element-wise multiplication; [ , ] indicates concatenation operation.

[0096] 2) For shallow nodes, stripe morphology sensing fusion is performed through the R-SGA module. In this embodiment, the R-SGA module is used to accurately capture the terminal of small blood vessels at the shallow nodes of the decoder. The specific process includes: ① First, the deep recovery features obtained from R-CAE are... The data is fed into the convolutional module (Conv block), which consists of two layers of convolution, batch normalization, and ReLU activation, designed to enhance feature representation and suppress noise. The refined features are then processed... Perform a 2x upsampling operation to obtain the decoding features of the shallow nodes of the decoder. .

[0097] The i-th layer encoded fusion feature obtained from GCT-GIM and the corresponding level's coding features F i Feature concatenation is performed, and then the aggregated features are passed through a residual path of length L=3 to reduce information loss of encoder features during transmission. The encoded features processed by the residual path are then combined with the decoded features F of the corresponding level. deci Mapping to a unified dimensional space and concatenating the results yields the features. F sum :

[0098]

[0099]

[0100] in, This represents the coding fusion feature of the i-th layer of the encoder unit; This indicates the decoding features of the corresponding layer.

[0101] ② Use bidirectional striped convolution (1×K, K×1) to analyze the features. F sum Feature extraction is performed to obtain the directional morphological response of blood vessels, and the features are obtained after batch normalization and ReLU activation function. f strip ;feature f strip It can significantly enhance the response intensity of slender blood vessel targets, while suppressing dotted background noise through its elongated receptive field.

[0102] To transform morphological features into gating weights that can be used for feature fusion, a 1×1 convolutional layer is used to gating the features. f strip A non-linear mapping along the channel dimension is performed, and a morphology-aware weight map is generated by combining it with the Sigmoid activation function. W In the shape-perceived weight map W In the image, the pixel values ​​corresponding to blood vessel regions are close to 1, while those of background regions are close to 0. This weight will serve as the control signal for subsequent asynchronous gating fusion, guiding the segmentation network to perform complementary information integration between the encoding and decoding paths. The entire calculation process is represented as follows:

[0103]

[0104] in, Indicates the activation function; Indicates hierarchical normalization 1; , This represents convolution operations with kernel sizes of K×1 and 1×K; This represents a convolution operation with a kernel size of 1×1; Indicates the Sigmal activation function ③ Utilizing the generated shape-aware weight map W Asynchronous gating fusion is performed on the shallow encoded fusion features and decoded features to obtain shallow recovered features. This mechanism uses (1- W Suppress background artifacts in the encoding path, while utilizing W Enhance the vascular details in the decoding path to obtain superficial recovery features. The specific calculation formula is as follows:

[0105] in, This represents the coding feature of the i-th layer of the encoder unit. This represents the decoding feature of the i-th layer of the decoder unit.

[0106] 3) Finally, restore the shallow features. A lightweight processing module consisting of multiple convolutional layers is used to further perform nonlinear transformation and channel compression on the obtained shallow recovery features to obtain the final blood vessel segmentation prediction map.

[0107] In one embodiment, when training the retinal vessel segmentation network, a hybrid loss function is constructed to balance pixel-level classification accuracy with the topological similarity of vessel morphology. During the testing phase, multi-dimensional quantitative metrics are used to comprehensively evaluate the segmentation performance. Specific implementation details are as follows: To address the severe class imbalance problem (i.e., background pixels far outnumber vessel pixels) in retinal vessel segmentation, this invention employs a weighted combination of binary cross-entropy loss (BCE Loss) and Dice loss. The binary cross-entropy loss measures the distribution difference between the predicted value and the ground truth label on a pixel-by-pixel basis. The Dice loss aims to maximize the overlap between the predicted region and the actual vessel region, providing strong constraints on preserving the morphology of small vessels. Finally, the two losses are aggregated using weighted coefficients. The specific calculation formula is as follows:

[0108] Where N is the total number of pixels. y i For the first i The true label (0 or 1) of each pixel. p i To predict the probability that a pixel belongs to a blood vessel for the segmentation network.

[0109] in, It is a smoothing factor used to prevent the denominator from being zero.

[0110] The total loss function is:

[0111] In this embodiment, the following settings are provided: = 0.5. This hybrid loss mechanism ensures that the segmentation network converges quickly in the early stages of training and allows for precise fine-tuning of slender blood vessel branches in the later stages.

[0112] During the testing phase, an evaluation mechanism was established, including the use of accuracy (Acc), sensitivity (Sen), specificity (Spe), F1 score (F1), and area under the ROC curve (AUC) to assess retinal vessel segmentation. The relevant parameters include the following:

[0113]

[0114]

[0115]

[0116]

[0117] In this context, TP represents the number of pixels that correctly predict blood vessels, TN represents the number of pixels that correctly predict the background, FP represents the number of pixels that incorrectly predict the background as blood vessels, and FN represents the number of pixels that incorrectly predict blood vessels as background. The Dice coefficient is used to measure the similarity between the predicted image and the label. Sensitivity represents the degree of segmentation between blood vessels and the background, reflecting the algorithm's ability to segment small blood vessels. Higher sensitivity means that more small blood vessels have been segmented. Specificity represents the ability to identify background elements.

[0118] Overall, to address the issues of blurred vessel segmentation edges and the easy loss of small vessels, this invention utilizes a multi-branch, elongated receptive field (including horizontal 1×K and vertical K×1 convolutional branches) to specifically extract the directional contour information of vessels, effectively simulating the characteristics of stripe-sensitive cells in biological vision. Simultaneously, a frequency domain filter is introduced in the encoding stage, utilizing a three-band Gaussian mask (low-frequency mask). Intermediate frequency mask High-frequency mask The residual calibration of the intra-layer features is performed, where the mask standard deviation is dynamically adjusted by learnable weight parameters, which amplifies the difference between blood vessels and background without increasing the training burden of the model. To address the issues of limited receptive fields and numerous vessel breaks due to interference from lesion backgrounds, this invention utilizes a full-scale feature aggregation module to achieve global modeling on a 12×12 resolution feature map, and leverages long-range contextual information as prior guidance for convolutional feature extraction. By constructing an asynchronous recovery link and employing residual cleanup and gated attention mechanisms for targeted repair of deep semantics and shallow edges, lesion artifacts are effectively removed and broken vessels are repaired.

[0119] Next, the method of this invention was experimentally verified. The experimental environment was as follows: the experiment was conducted on an NVIDIA 3090 GPU with 24GB of memory, using Python 3.8.19 and CUDA 12.4. The method of this invention was trained and tested on a PyTorch platform of version 2.4.0.

[0120] Parameter settings: The batch size is set to 64, and the segmentation network will undergo 40 rounds of training to learn vascular features. This invention utilizes the Adam optimizer to update the network parameters, setting the initial learning rate to 1e-5. Batch Normalization and ReLU are applied sequentially after each convolutional layer to accelerate the convergence of the segmentation network and address the gradient vanishing problem. Finally, the probability map obtained from the segmentation network is converted to binary form using a threshold of 0.5.

[0121] During training, data augmentation techniques were applied, such as flipping, rotating, and adding Gaussian noise.

[0122] To verify the effectiveness of the proposed method, it was compared with several representative segmentation networks, including U-Net, LadderNet, AttU-Net, AAU-Net, DEF-Net, DPF-Net, EDAE-Net, MSMA-Net, and LMAF-Net. Experiments were conducted on two widely used public datasets (DRIVE and CHASE_DB1), and the segmentation performance of each method was systematically evaluated using five evaluation metrics (AUC, F1, ACC, SEN, and SPE). Tables 1 and 2 present the performance evaluation results of each method on the DRIVE and CHASE_DB1 datasets, respectively.

[0123] Table 1 Experimental Results of the DRIVE Dataset

[0124] Table 2 Experimental Results of CHASE_DB1 Dataset

[0125] As shown in Table 1, on the DRIVE dataset, the method of this invention achieved the best performance in AUC (0.9886), F1 (0.8440), ACC (0.9734), SEN (0.8636), and SPE (0.9836), ranking first in all metrics. The Sen metric improved by approximately 3.19% compared to the suboptimal model LMAF-Net, fully demonstrating the high sensitivity of this structure in capturing extremely fine blood vessel features. Table 2 shows that on the CHASE_DB1 dataset, the method of this invention performed best in AUC (0.9901), ACC (0.9765), and SEN (0.8548), ranking first in all three metrics; while ranking fifth in F1 (0.8229) and fourth in SPE (0.9849). Notably, the method of this invention maintained its leading position in AUC, which measures the overall performance of the model, and showed a significant leap in SEN, reflecting sensitivity, compared to the comparison model. In summary, a higher SEN index means the model can identify more small blood vessels and boundary pixels. This result fully demonstrates the superior ability of this invention in microvascular structure extraction. The experimental results show that the method of this invention outperforms existing competing models in overall performance, especially in AUC and SEN indices. The significant improvement in AUC and SEN indicates that this model has a clear advantage in reducing missed blood vessels (especially small blood vessels in complex backgrounds), thereby effectively improving the reliability of medical image analysis in clinical diagnosis.

[0126] Figure 7This is a schematic diagram of the fundus image preprocessing process of the present invention. It shows that the green channel with the highest contrast is extracted from the original color image, and the contrast-limited adaptive histogram equalization technique is applied to visualize the blood vessel edge response. Then, the image brightness distribution is nonlinearly reconstructed by combining a Gamma correction with a coefficient of 1.2, thereby achieving in-depth visualization of microvascular features.

[0127] Figure 8 This is a schematic diagram of the experimental dataset slicing and data augmentation of the present invention. It shows that a 96×96 pixel sliding window is used to simultaneously segment the preprocessed image and the corresponding label, and eight symmetric geometric augmentations, including clockwise rotation, horizontal flip, vertical flip and compound transformation, are performed on the generated sample blocks. By constructing pixel-level synchronized augmentation training pairs, the aim is to improve the segmentation network's ability to capture anisotropic blood vessel morphology.

[0128] Figure 9 and Figure 10 The qualitative segmentation results of the segmentation network proposed in this invention and several mainstream comparison algorithms are shown on the DRIVE and CHASE_DB1 datasets. From left to right, the figures show the original fundus image, the preprocessed image, a magnified view of the region of interest (ROI), the gold standard label (GroundTruth), and the prediction masks of each comparison model. The comparison of the magnified ROI regions reveals that traditional models such as U-Net and Att-UNet generally suffer from severe missed detections or pixel disconnections when processing extremely fine vascular terminals, and are prone to noise interference in complex backgrounds. In contrast, the model proposed in this invention exhibits superior vascular topology preservation capabilities. Its extracted vascular texture is smooth with clear edges, accurately reconstructing the anatomical structure of small branches, and showing no obvious adhesions or artifacts at vascular intersections. Visual comparison results further confirm that, due to the introduction of multi-dimensional spatial morphology perception and full-scale frequency domain calibration mechanisms, the model of this invention has stronger robustness in capturing anisotropic vascular orientations. Visually, the prediction results of this invention are more accurate, effectively solving the technical challenge of vascular rupture in low-contrast environments and significantly improving the structural integrity of the retinal vascular network.

[0129] To verify the effectiveness of the proposed SCS, FEA, GCT-GIM, and R_CAE / SGA modules, a systematic ablation experiment was conducted on the CHASE_DB1 dataset. U-Net was used as the baseline segmentation network to evaluate the contribution of each module to the performance improvement. The relevant quantitative results are summarized in Table 3.

[0130] Table 3 Summary of Quantitative Results of Ablation Experiments

[0131] The effectiveness of the SCS module: The visualization results of the ablation experiments clearly demonstrate the effectiveness of integrating the Spatial Coordinate Sensing Module (SCS) into the baseline model. Compared with the original baseline, the introduction of the SCS module significantly enhanced the model's ability to capture anisotropic vessel orientations, especially at small vessel branches where detection was more sensitive. The examination of the quantitative information shown in Table 3 validated this improvement; after integrating the SCS module, the SEN index increased significantly from 0.7861 to 0.8264. The increase in the SEN index indicates a significant reduction in missed detections, validating the crucial role of the SCS module in enhancing spatial morphological feature extraction.

[0132] Effectiveness of the FEA Module: To evaluate the role of the Frequency Domain Calibration Module (FEA), this invention integrates it into the segmentation network containing SCS. The introduction of FEA significantly improves the clarity of vessel edges and effectively suppresses background redundancy noise. The quantitative results in Table 3 further show that after integrating the FEA module, the F1 score, which measures segmentation accuracy, improves from 0.8021 to 0.8108, and the AUC score recovers to 0.9885. This fully demonstrates the effectiveness of FEA in frequency domain component calibration and enhancing the separability of early features.

[0133] Effectiveness of the GCT-GIM module: Integrating the full-scale feature convergence module (GCT-GIM) demonstrates the positive impact of global topology modeling on segmentation continuity. Particularly in complex intersection regions and major blood vessel trunks, the connectivity of the segmentation results is significantly enhanced, effectively addressing the issue of vessel rupture caused by limited local receptive fields. The SEN index further increased from 0.8218 to 0.8423, and the F1 score increased to 0.8193, validating the core value of GCT-GIM in capturing long-range spatial dependencies and improving structural integrity.

[0134] Synergistic Effect of R-CAE / SGA Modules: Ultimately, integrating the asynchronous recovery modules (R-CAE / R-SGA) into the complete architecture resulted in optimal performance for core metrics such as SEN, ACC, F1, and AUC, reaching 0.8548, 0.9765, 0.8229, and 0.9901, respectively. The R-CAE / SGA modules, through self-encoding and purification of deep features and morphological-guided fusion of superficial details, demonstrated exceptional detail recovery capabilities in the depiction of vascular terminals. Notably, while the SPE metric slightly decreased from 0.9895 to 0.9849 compared to the baseline, this reflects the necessary trade-off between SEN and SPE. Overall, the quantitative metrics clearly demonstrate that each component in this invention framework makes a practical contribution, and through the synergistic effect of these modules, the segmentation network achieves optimal retinal vessel segmentation results.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration, characterized in that, include: Construct a sample set of fundus images; A retinal vessel segmentation network is constructed. The retinal vessel segmentation network includes an encoder unit, a bottleneck fusion center, and a decoder unit. The encoder unit contains multi-level feature sensing blocks, and each level of feature sensing block is composed of a cascaded stripe convolutional structure and a multi-dimensional frequency domain calibration module. The bottleneck fusion center is located at the connection between the encoder unit and the decoder unit. The decoder unit contains an adaptive edge fusion block and a stripe-guided attention block. A retinal vessel segmentation network was trained and validated using a set of fundus image samples. The trained and validated retinal vessel segmentation network is used to segment the fundus image to be segmented, and a vessel segmentation prediction map is obtained. The process of segmenting the fundus image to be segmented includes: The directional morphological features of blood vessels in fundus images are captured by a striped convolutional structure, and the directional morphological features of blood vessels are enhanced by a multi-dimensional frequency domain calibration module to obtain frequency domain enhanced features. The bottleneck fusion hub aggregates the frequency domain enhancement features of each level in the encoder unit and performs global modeling. The gating injection mechanism is used to feed the global context back to the frequency domain enhancement features of each level to obtain the coding fusion features. In the deep nodes of the decoder unit, residual cleaning is performed on the skip path through adaptive edge fusion blocks, and the feature semantics are purified to obtain deep recovery features. After processing, the deep recovery features are input into the shallow nodes of the decoder unit as the decoding features of the shallow nodes. In the shallow nodes of the decoder unit, the decoding features of the shallow nodes are adaptively fused with the corresponding layer encoding fusion features through stripe-guided attention blocks to obtain shallow recovery features. At the end of the decoder unit, the shallow recovery features are processed to obtain the final blood vessel segmentation prediction map.

2. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The process of constructing the fundus image sample set includes: The green channel component of the original fundus image is extracted, and then adaptive histogram equalization with contrast limitation and nonlinear Gamma correction are performed sequentially. An overlapping sliding window is used to cut the image, retaining image blocks that meet the blood vessel density threshold condition; Geometric transformations are performed on the selected image patches to generate an enhanced set of fundus image samples.

3. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The process of capturing the directional and morphological features of blood vessels in fundus images using striped convolutional structures includes: The preprocessed fundus image is subjected to a 3×3 convolution and ReLU operation to obtain the primary spatial features F. in ; Four parallel convolutional branches are used to extract the primary spatial features F. in The extremely long stripe features, medium stripe features, local perception features, and short-range perception features; Extracting channel vectors using global average pooling , channel vector Attention response values ​​for each convolutional branch are generated using a multilayer perceptron, and the weights of each convolutional branch are obtained by normalization using the Softmax function. w i The features extracted from each convolutional branch are weighted and fused using the weights of each convolutional branch, and then passed through a 1×1 convolutional layer to obtain the stripe perception features. f strip ; Using one-dimensional horizontal and vertical convolution kernels, the primary spatial features F are respectively processed. in Component aggregation is performed to capture long-range spatial dependencies in the horizontal and vertical directions, resulting in orientation-aware feature pairs. z h , z w ; Direction-aware feature pairs z h , z w The data is concatenated, and a 1×1 shared convolutional layer is used to perform channel compression and nonlinear mapping to generate the encoded spatial location vector. f temp ; Horizontal attention weights are generated through dimension splitting, convolutional mapping, and the Sigmoid activation function. a h and vertical attention weights a w Utilizing horizontal attention weights a h and vertical attention weights a w Element-wise multiplication of the input feature x yields the orientation-sensitive feature. f coord ; Extracting stripe-sensing features using global average pooling f strip With direction-sensitive features f coord The global statistics are used to generate mutual-directed attention weights through a shared multilayer perceptron and sigmoid activation function. Att strip and Att coord ; Utilizing mutual guidance attention weights Att strip and Att coord Stripe perception features f strip With direction-sensitive features f coord Perform cross-feature modulation to obtain cross-spatial fused features. f cross ; Cross-space fusion features f cross With primary spatial features F in Perform residual connections to obtain the directional morphological features F of the blood vessels in the current layer. out .

4. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The process of enhancing the directional morphological features of blood vessels using a multi-dimensional frequency domain calibration module includes: Directional morphological features F out As the input feature of the multi-dimensional frequency domain calibration module, the input feature F out Perform a Fast Fourier Transform to convert it to frequency domain features; By constructing three independent Gaussian frequency band masks corresponding to the low-frequency, mid-frequency, and high-frequency regions respectively, and introducing learnable channel-wise weight parameters combined with the softplus activation function, an adaptive filter for each frequency band is obtained. The adaptive filters of the three frequency bands are linearly combined to form the overall frequency domain filter. F total ; The frequency domain features are filtered based on the overall frequency domain filter, and the filtered features are mapped back to the spatial domain features through inverse Fourier transform. Spatial domain features and input features F out Perform residual connections to obtain preliminary frequency domain enhancement features. ; Employing a hybrid attention mechanism to enhance initial frequency domain features Feature selection and enhancement are performed to obtain the final frequency domain enhanced features. .

5. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 4, characterized in that, Employing a hybrid attention mechanism to enhance the output features in the frequency domain The process of feature selection and enhancement includes: On the channel, preliminary frequency domain enhancement features are aggregated in parallel using global average pooling and global max pooling. The spatial information is used to generate two different spatial context feature vectors; these two spatial context feature vectors are then fed into a shared multilayer perceptron to extract the dependencies between channels, and finally a channel weight map is generated using a sigmoid activation function. M c ; Using channel weight graph M c Preliminary frequency domain enhancement features Perform channel calibration to obtain channel refinement features. ; Channel refinement features Average pooling and max pooling are applied along the channel dimension to generate two two-dimensional spatial information feature vectors. The two spatial information feature vectors are concatenated and fused through a 7×7 convolutional layer. A spatial weight map M is generated by applying a sigmoid activation function. s ; Using spatial attention map Ms and channel refinement features Pixel-by-pixel multiplication yields the final frequency domain enhancement feature f. att .

6. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The process of pooling the frequency domain enhancement features of each level in the encoder unit through the bottleneck fusion hub and performing global modeling, and then feeding the global context back to the frequency domain enhancement features of each level using a gated injection mechanism, includes: The frequency domain enhancement features of each level of the encoder unit are used as the coding features F of each level. i ; The encoding features F at different scales at each level of the encoder unit are used to encode the features. i Project and pool to a uniform 12×12 fixed size, and obtain the full-scale converged feature S2 by pixel-level summation; The full-scale converged feature S2 is flattened along the spatial dimension, transforming it into a two-dimensional sequence form suitable for Transformer processing. The sequence z is input into the global context transformer for modeling; After performing long-range modeling, the output feature sequence is reconstructed in terms of spatial dimensions, and it is remapped into a two-dimensional tensor that conforms to the original spatial topology. Z global ; By using 1×1 convolution operations in conjunction with batch normalization, the two-dimensional tensor Z is transformed. global The number of channels is mapped from the embedding dimension C back to the original physical channel depth C of the corresponding encoder unit level. i To obtain the global features after channel alignment at the corresponding level. ; Bilinear interpolation algorithm is used to perform global feature matching after channel alignment. Perform spatial domain resampling to obtain spatially aligned enhanced global features at the corresponding level. ; Will enhance global features Input a gated branch consisting of a 3×3 convolutional layer and a sigmoid activation function, and generate a spatially gated weight map. Gate i Spatial gating weight map Gate i Used to identify prominent vascular areas from a global perspective; Introducing learnable residual injection coefficients g i Adaptive injection calibration is performed on the coding features at each level of the encoder unit to obtain the coding fusion features. The calibration formula is: Where i represents the i-th level of the encoder unit.

7. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The process of obtaining deep recovery features through adaptive edge fusion blocks includes: Residual path processing is performed on the deep coding fusion features of the encoder unit. The encoded features after residual path processing are mapped to a unified dimensional space and concatenated with the decoded features of the corresponding level to obtain feature F. res ; Utilizing the bottleneck structure of an autoencoder to analyze feature F res Perform spatial domain purification to generate a spatial gating graph. M gate ; Using spatial gating diagrams M gate For feature F res Adaptive reweighting is performed to obtain refined features. ; For feature F res Perform max pooling and average pooling, and calculate the difference between the two pooling results to obtain the display edge map. E edge ; Purification characteristics through splicing With display edge map E edge This yields the final deep recovery features. .

8. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, The generation process of shallow restoration features includes: Deep recovery features After processing by the input convolutional module, a 2x upsampling operation is performed to obtain the decoding features of the shallow nodes of the decoder unit. ; The encoded fusion features of the i-th layer of the encoder unit are concatenated with the encoded features of the corresponding layer, and then residual path processing is performed on the concatenated features; the encoded features after residual path processing are then combined with the decoded features of the shallow nodes. Mapping to a unified dimensional space and concatenating the results yields the features. F sum ; Using bidirectional stripe convolution to pair features F sum Feature extraction is performed to obtain the directional morphological response of blood vessels, and the features are obtained after batch normalization and ReLU activation function. f strip ; Using 1×1 convolutional layers for features f strip A non-linear mapping along the channel dimension is performed, and a morphology-aware weight map is generated by combining it with the Sigmoid activation function. W In the shape-perceived weight map W In the image, the pixel value corresponding to the blood vessel area is close to 1, while the pixel value of the background area is close to 0. Using the generated shape-aware weight map W Decoding features of shallow nodes and the coding features of the corresponding layer Asynchronous gating fusion is performed to obtain shallow recovery features.

9. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, When training the retinal vessel segmentation network, the binary cross-entropy loss and Dice loss are weighted and summed to form the total loss function.

10. The retinal vessel segmentation method based on multidimensional spatial perception and frequency domain calibration as described in claim 1, characterized in that, At the end of the decoder unit, a lightweight processing module consisting of multiple convolutional layers is set up to perform nonlinear transformation and channel compression on the shallow recovery features to obtain the final blood vessel segmentation prediction map.

Images