Eye fundus blood vessel segmentation method and system based on semantic network

Abstract

The invention discloses a semantic network-based fundus blood vessel segmentation method and system in the technical field of image processing. The method comprises the steps of obtaining a fundus retina image data set; processing the fundus retina image data set to obtain a processed image data set; constructing a double-layer semantic fusion network model of enhanced convolution; training the double-layer semantic fusion network model with enhanced convolution; and inputting the processed fundus retina image data set into the trained enhanced convolution double-layer semantic fusion network model to obtain a blood vessel segmentation result. By adopting the method, sufficient blood vessel features can be extracted from the eye fundus image with a complex blood vessel structure, and more accurate blood vessels can be segmented.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and in particular relates to a method and system for segmenting fundus blood vessels based on a semantic network. Background Art

[0002] Fundus images contain a large amount of diagnostic information related to ophthalmic diseases, such as hypertension and glaucoma, which can be detected in advance by examining retinal blood vessels. Diabetic retinopathy is a disease caused by microvascular damage, which may lead to impaired vision. When using artificial intelligence or deep learning methods for automatic segmentation or detection of medical images, data annotation is required. However, manual annotation of fundus vascular images is time-consuming and labor-intensive, and requires a high level of expertise due to the low image contrast. Proliferative microvessels with smaller diameters may be missed during manual annotation. Therefore, it is necessary to develop an efficient automatic segmentation method for fundus image blood vessels.

[0003] There are multiple challenges in the task of fundus vessel segmentation. First, the vascular tree structure is complex, the width of arteries and veins is different, and the accurate segmentation of capillaries is particularly critical. Secondly, there is a class imbalance problem between vascular pixels and background pixels in fundus images, which affects the segmentation accuracy. In addition, noise interference from different acquisition devices also increases the difficulty of segmentation. Early studies mainly relied on manually annotated features for retinal vessel segmentation. These methods performed well in simple scenes, but often failed to meet their expectations on complex data sets because traditional image processing techniques cannot fully capture the complex semantics of retinal vessels.

[0004] In recent years, many traditional models for retinal vessel segmentation have been proposed. Early researchers tried to extract vascular features through wavelet transform. Soares et al. selected multi-scale two-dimensional wavelet transform to extract features within the blood vessels and used Bayesian classifier to classify each pixel into a blood vessel pixel or a background pixel. Xie et al. proposed an edge detection algorithm that is effective for retinal vessel segmentation. Ricci et al. proposed a method to treat vessel segmentation as a line detection task, which is a vessel segmentation method that uses support vector machines and line operators for classification. These line detection methods can improve the connectivity of blood vessels, but will misdetect non-vascular tissue as vascular tissue. Hoover et al. used a two-dimensional matched filter method to complement global and local features to segment fundus vessels, resulting in a true positive rate of 75%. Fraz et al. used image morphology methods to reconstruct the vascular centerline image using plane slices to obtain a segmented vascular tree. Jiang et al. proposed an adaptive local threshold method that generates target hypotheses by assuming multiple local thresholds for detecting fundus vascular images. These traditional methods for extracting vascular information have poor robustness and high complexity. They cannot distinguish between microvessels and background, resulting in failure of capillary detection. The accuracy of retinal vascular tree segmentation needs to be further improved. Deep learning methods have shown good segmentation performance in retinal vessels, which is better than traditional methods. A large number of original methods have proved this. For example, Fu et al. proposed a new method that treats retinal vascular segmentation as a boundary detection task and uses conditional random fields to obtain long-range interactions between vascular pixels. This is the first method to study contextual information. Son et al. used generative adversarial networks for retinal vascular segmentation to obtain capillaries with less error information. This is the first method to use generative adversarial networks for vascular segmentation. Oliverira et al. proposed a multi-scale network for retinal vascular segmentation, which is based on multi-scale and stationary wavelet transforms. Zheng et al. proposed a method that combines a fully convolutional network and a recursive neural network, using a recursive neural network to solve the problem of segmented blood vessels being easily broken. Chen et al. proposed a new network that realizes the fusion of low-level and high-level semantic features in the process of retinal vascular segmentation, and uses pooling layer indexes to solve the problem of saving the entire encoder feature map. Shin et al. combined graph convolutional networks and convolutional neural networks to learn the graphical connections of blood vessels. These methods can greatly improve the performance of retinal blood vessel segmentation, but there is still room for improvement in maintaining the connectivity of blood vessels. Based on the good performance of the encoder-decoder network in image segmentation, combined with a variety of improvement mechanisms such as attention mechanism, cyclic convolution mechanism and multi-scale mechanism, a number of encoder and decoder network variants have been proposed to improve segmentation performance.Zhou et al. proposed a method for image segmentation, designing a method different from the skip connection in U-Net to perform feature fusion, which is convenient for integrating multi-scale information, thereby improving the segmentation accuracy of retinal blood vessels. Li et al. proposed an IterNet model, connecting several U-Net networks together, which significantly improved the performance of fundus blood vessel segmentation. Yan et al. proposed a new segmentation loss function to balance the importance of thick blood vessels and microvessels, thereby solving the imbalance problem between capillaries and thick blood vessels in fundus images. Gu et al. proposed a context extraction method based on encoder and decoder structure, so that the spatial characteristics of retinal blood vessels are fully preserved. Jin et al. combined variable convolution with encoder and decoder structure to adapt to the shape of retinal blood vessels. Wang et al. embedded the context encoder and spatial encoder into the U-Net framework to obtain more vascular feature information. These methods can increase the continuity of blood vessel segmentation in fundus images, but the segmentation performance of microvessels affected by lesions at the edge still needs to be improved.

[0005] Although these U-Net variants have achieved good performance, they are not sufficient to meet the challenges of fundus image segmentation. Due to the complexity of fundus retinal blood vessels and the lack of high-quality vascular imaging, the general encoder-decoder variant structure cannot stably segment all microvessels. Therefore, there is an urgent need for a fundus retinal blood vessel segmentation network method and system that can extract sufficient vascular features from fundus images with complex vascular structures and segment more accurate blood vessels. Summary of the invention

[0006] In order to solve the above-mentioned problem that the general U-shaped variation structure cannot stably segment all microvessels due to the complexity of fundus retinal blood vessels and the lack of high-quality vascular imaging, the purpose of the present invention is to provide a fundus blood vessel segmentation method and system based on a semantic network, which can extract sufficient blood vessel features from fundus images with complex vascular structures and segment more accurate blood vessels.

[0007] In order to achieve the above object, the technical solution of the present invention is as follows:

[0008] A method for segmenting fundus blood vessels based on a semantic network comprises the following steps:

[0009] S1: Obtain fundus retinal vascular dataset;

[0010] S2: performing preliminary processing on the acquired fundus retinal vascular dataset to obtain a processed fundus retinal vascular dataset;

[0011] S3: Construct a two-layer semantic fusion network model with enhanced convolution;

[0012] S4: training the enhanced convolutional double-layer semantic fusion network model to obtain a trained enhanced convolutional double-layer semantic fusion network model;

[0013] S5: Input the processed fundus retinal vascular dataset into the trained enhanced convolutional two-layer semantic fusion network model to obtain the fundus retinal vascular dataset segmentation result.

[0014] Further, the preliminary processing includes the following steps: performing channel color transformation, data normalization, contrast-limited adaptive histogram equalization and gamma correction processing on the acquired fundus retinal vascular dataset to obtain a processed fundus retinal vascular dataset.

[0015] Furthermore, a two-layer semantic fusion network model with enhanced convolution is constructed, including the following steps:

[0016] Design a surface network structure model, which simplifies basic semantic information processing and improves the efficiency of basic feature extraction; the surface network structure model includes an enhanced convolution module, a maximum pooling operation module, and a sampling process module; the sampling process module is used to restore the feature map to the original input size through a transposed convolution operation with a step size of 2, and finally use 1×1 convolution and Sigmoid function to output the fundus retinal blood vessel segmentation probability map;

[0017] A deep network structure model is designed, which is used to extract high-order semantic information vascular feature information; the deep network structure model includes a CLAFM module, an enhanced convolution module and an HLFFM module; the HLFFM module uses a maximum pooling operation to reduce the output feature map of the first three layers of the surface network encoder to a size of 32×32 to reduce computational complexity and concentrate key information; the number of channels of the feature map is adjusted to 512 through a 1×1 convolution to unify the feature expression; these processed feature maps are fused and connected in series with the output of the underlying encoder of the deep network to further optimize the feature expression.

[0018] Furthermore, the CLAFM module includes a dual attention mechanism, each branch includes a spatial attention module and a channel module; the first attention module is used to strengthen the correlation of important channels in the feature map; the second attention module enhances the expression of key features by adjusting the channel weights; the feature fusion strategy adopts fine concatenation and upsampling operations; the feature maps extracted by encoders at different levels are first upsampled and 1×1 convolved, and then concatenated.

[0019] Furthermore, the method further comprises an evaluation step S6: using accuracy, sensitivity, F1 score and area under the receiver operating characteristic curve as evaluation indicators to evaluate the fundus retinal blood vessel segmentation result.

[0020] Furthermore, before constructing the enhanced convolutional two-layer semantic fusion network model, an enhanced convolutional two-layer semantic fusion network is first constructed.

[0021] Furthermore, in the two-layer semantic fusion network model of enhanced convolution, the specific steps of constructing enhanced convolution include: replacing the 3×3 convolution layer in the convolution in the deep learning network with a 1×1+3×3+1×1 convolution layer, adding FRN normalization and Mish activation function for pre-activation before convolution, and adding a Dropout layer after the 3×3 convolution layer.

[0022] Furthermore, the specific steps of constructing a two-layer semantic fusion network include: weighted addition of the vascular probability map output by the surface network and the vascular probability map output by the deep network to generate a final prediction probability map.

[0023] Further, a fundus retinal blood vessel segmentation network system includes an acquisition module, a processing module, a construction module, a training module, a segmentation module and an evaluation module connected in sequence;

[0024] The acquisition module is used to obtain the fundus retinal blood vessel data set;

[0025] The processing module is used to perform preliminary processing on the fundus retinal blood vessel data set acquired by the acquisition module to obtain a processed fundus retinal blood vessel data set;

[0026] The building module is used to construct a two-layer semantic fusion network model with enhanced convolution;

[0027] The training module is used to train the enhanced convolutional double-layer semantic fusion network model to obtain the trained enhanced convolutional double-layer semantic fusion network model;

[0028] The segmentation module is used to input the retinal image data set processed by the processing module into the trained enhanced convolutional double-layer semantic fusion network model to obtain the fundus retinal blood vessel segmentation result;

[0029] The evaluation module is used to evaluate the fundus retinal blood vessel segmentation result using accuracy (ACC), sensitivity (SE), F1 score (F1-Score) and area under the receiver operating characteristic curve (ROC) (AUC) as evaluation indicators.

[0030] The above scheme has the following beneficial effects:

[0031] 1. This scheme constructs a two-layer semantic fusion network model of enhanced convolution to extract microvessels in fundus retinal images. Compared with the existing related technologies, this scheme constructs enhanced convolution blocks to optimize computational efficiency and accurately extract vascular information, so as to better extract vascular features; uses CLAFM through deep networks to optimize the coding layer structure, fully considers the differences between different input channels in extracting vascular information, and enhances the recognition ability of fundus blood vessels; HLFFM more effectively retains the semantic information of the features, and fully maintains the spatial details of the features, thereby significantly improving the accuracy and efficiency of retinal blood vessel segmentation.

[0032] 2. This solution uses FRN and Mish in the convolutional block in the RCP module together. FRN eliminates the dependence on batch size and improves the stability of the model. The Mish activation mechanism is introduced to provide an unbounded activation environment to avoid activation saturation and keep the gradient flow smooth, thereby significantly improving the accuracy and efficiency of retinal vascular image segmentation as a whole, solving the problems of too many layers, slow convergence and insufficient feature extraction in traditional methods. Residual connections are introduced in the Res-PDC module to avoid the degradation problem that often affects model prediction as the number of layers increases.

[0033] 3. This scheme uses the maximum pooling operation through the HLFFM module to reduce the output feature map of the first three layers of the surface network encoder to a size of 32×32 to reduce computational complexity and concentrate key information; adjust the number of channels of the feature map to 512 through 1×1 convolution to unify the feature expression; fuse these processed feature maps and connect them in series with the output of the bottom encoder of the deep network to further optimize the feature expression. The CLAFM module includes a dual attention mechanism, each branch includes a spatial attention module and a channel module; the first attention module is used to strengthen the correlation of important channels in the feature map; the second attention module enhances the expression of key features by adjusting the channel weights; the feature fusion strategy uses fine concatenation and upsampling operations; the feature maps extracted by encoders at different levels are first upsampled and 1×1 convolved, and then concatenated.

[0034] In summary, this scheme can better extract the key features in the fundus retinal image, thereby improving the accuracy and robustness of microvascular feature extraction. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 This is the convolution (NPCD) graph of the present invention.

[0036] Figure 2 This is the enhanced convolution (RCP) graph of the present invention.

[0037] Figure 3 This is a diagram of the hierarchical feature fusion method (HLFFM) of the present invention.

[0038] Figure 4 This is a diagram of the cross-stage dual attention mechanism (CLAFM) of the present invention.

[0039] Figure 5 This is the network framework diagram of DSFN-Net of the present invention.

[0040] Figure 6 This is a visualization effect diagram of the segmentation of the present invention.

[0041] Figure 7 The present invention is a flowchart of a network method for segmenting fundus retinal blood vessels.

[0042] Figure 8 The present invention is a block diagram of a fundus retinal blood vessel segmentation network system. DETAILED DESCRIPTION

[0043] The following is further described in detail through specific implementation methods:

[0044] Example

[0045] Basically as attached Figure 1-8 As shown:

[0046] A fundus retinal blood vessel segmentation network method comprises the following steps:

[0047] S1: Obtain the fundus retinal vascular dataset. The fundus retinal vascular dataset includes DRIVE, CHASE_DB1, and STARE. The specific information is shown in Table 1.

[0048] Table 1 Detailed information of three fundus retinal vascular datasets: DRIVE, CHASE_DB1 and STARE

[0049] Dataset name Total number of pictures Number of training / testing Image resolution DRIVE 40 20 / 20 565×584 CHASE_DB1 28 20 / 8 999×960 STARE 20 19 / 1(Leave one out) 700×605

[0050] S2: Preliminary processing is performed on the acquired fundus retinal vascular dataset to obtain a processed fundus retinal vascular dataset; the specific preliminary processing method is to perform channel color transformation, data normalization, limited contrast adaptive histogram equalization and gamma correction processing on the acquired fundus retinal vascular dataset to obtain a processed fundus retinal vascular dataset.

[0051] Before S3, in order to better extract vascular features, the present invention designs enhanced convolution and double-layer networks. The deep network combines surface and deep feature extraction to optimize feature extraction and computational efficiency. The surface network simplifies the processing of basic semantic information and improves the efficiency of basic feature extraction; the deep network focuses on high-order semantic information, processes the same retinal image data through independent parameters, and ensures the accurate extraction of high-order features. In the convolution layer part, a residual convolution block (RCP) composed of NormalizationPreactivation Convolution Drop-out (NPCD) is used. The enhanced convolution strategy, namely NPCD convolution, is innovatively introduced. Through Filter Re-sponse Normalization (FRN) and Mish activation mechanism, as well as Dropout=0.2 strategy, the performance of the convolution structure is optimized. NPCD convolution eliminates the dependence on batch size through the convolution sequence of 1×1, 3×3 and 1×1 again, and cooperates with FRN to improve the stability of the model, as shown in formula (1). The Mish activation mechanism provides a boundless activation environment to avoid activation saturation and keep the gradient flow smooth, as shown in formula (2). These improvements significantly improve the accuracy and efficiency of retinal vascular image segmentation as a whole, and solve the problems of too many levels, slow convergence and insufficient feature extraction in traditional methods.

[0052]

[0053] The filter response normalization method is performed in the (H, W) dimension, f is an N-dimensional (H, W) feature vector, δ is usually set to 1e-6, λ is a learnable parameter (the parameter is a real number R), i∈N.

[0054] Considering the differences in the extraction of vascular information by the encoding layer under different input channels, the encoding layer structure is optimized to enhance the recognition ability of fundus blood vessels. The present invention proposes a hierarchical feature fusion method (HLFFM). The core of the HLFFM method is to effectively combine features from different depths by optimizing the feature capture and fusion process. Specifically, the output feature map of the first three layers of network encoders is first reduced to a size of 32×32 using the maximum pooling operation to reduce the computational complexity and concentrate key information. Then, the number of channels of the feature map is adjusted to 512 by 1×1 convolution to unify the feature expression. Subsequently, these processed feature maps are fused and connected in series with the output of the underlying encoder to further optimize the feature expression. This method not only effectively retains the semantic information of the features, but also fully maintains the spatial details of the features, thereby significantly improving the accuracy and efficiency of retinal blood vessel segmentation. The innovation of this method lies in its fine control of the change in the number of initial input channels and the optimized design in the feature fusion process. Compared with the traditional direct feature splicing method, it can better balance the richness of information and computational efficiency and effectively solve the limitations of existing methods.

[0055] In order to effectively utilize the vascular information extracted from each encoding layer, the span fusion technology is studied. This paper proposes a cross-level feature fusion method (CLAFM), whose main innovations and advantages are reflected in the following aspects: First, CLAFM adopts a dual attention mechanism, combining CBAM (Convolutional Block Attention Module) and SE (Squeeze-and-Excitation) modules. These two attention mechanisms work in parallel in the network. CBAM strengthens the correlation of important channels in the feature map through the channel attention mechanism, while SE enhances the expression of key features through channel weight adjustment. By using these two modules in parallel, CLAFM can comprehensively and dynamically capture the spatial and channel information of the feature map, thereby effectively improving the extraction accuracy and segmentation performance of vascular features. Secondly, CLAFM adopts fine concatenation and upsampling operations in the feature fusion strategy. Specifically, it first upsamples and 1×1 convolutions the feature maps extracted by encoders at different levels, and then performs concatenation operations. This strategy not only avoids the information redundancy and computational complexity problems caused by traditional direct splicing, but also preserves the uniqueness of each level of features, providing richer feature inputs for subsequent attention mechanisms. Finally, CLAFM is optimized for deep networks, avoiding unnecessary computational overhead for shallow networks, thereby ensuring the performance advantage of the model when running efficiently. This specific optimization strategy for deep features enables CLAFM to more accurately extract key features when processing complex retinal vascular structures, thereby improving the accuracy of vascular segmentation.

[0056] Specifically, according to the U-Net design concept, the present invention designs a double-layer semantic fusion network model with enhanced convolution; its structure is shown in the attached figure. Figure 5 As shown: DSFM-Net (deep learning network of double-layer semantic fusion network with enhanced convolution) includes two parts: surface network and deep network. The deep network is designed to focus on high-order semantic information and process the same retinal image data through independent parameters to ensure the accurate extraction of high-order features. It also includes three modules, including RCP module, CLAFM module and HLFFM module. The surface network simplifies the basic semantic information processing, improves the efficiency of basic feature extraction, and can achieve end-to-end learning. In DSFM-Net, the surface network consists of three modules, including RCP module, maximum pooling operation module and sampling process operation. The sampling process includes a transposed convolution operation with a step size of 2 to restore the feature map to the original input size. Finally, 1×1 convolution and Sigmoid function are used to output the probability map of fundus retinal blood vessel segmentation.

[0057] S4: training the enhanced convolutional double-layer semantic fusion network model to obtain a trained enhanced convolutional double-layer semantic fusion network model;

[0058] In the training phase, the learning rate is set to 0.0001, the Adam algorithm is selected as the optimization algorithm, the learning rate decays to 0 after 600 epochs, and the binary cross entropy is used as the loss function.

[0059] Then, the DSFM-Net model was trained based on PyTorch as the backend on an experimental platform with 24G memory and NVIDIA RTX3090 GPU.

[0060] S5: inputting the processed fundus retinal blood vessel dataset into the trained enhanced convolutional double-layer semantic fusion network model to obtain the fundus retinal blood vessel dataset segmentation result;

[0061] S6: Accuracy (ACC), sensitivity (SE), F1-score (F1-Score) and area under the receiver operating characteristic curve (ROC) (AUC) are used as evaluation indicators to evaluate the fundus retinal blood vessel segmentation results. The calculation formulas of each evaluation indicator are:

[0062]

[0063] Among them, TP is a true positive, indicating that it is predicted as a blood vessel point pixel, and the expert standard is also a blood vessel point pixel; TN is a true negative, indicating that it is predicted as a background point pixel, and the expert standard is also a background point pixel; FP is a false positive, indicating that it is predicted as a blood vessel point pixel, and the expert standard is a background point pixel; FN is a false negative, indicating that it is predicted as a background point pixel, and the expert standard is a blood vessel point pixel.

[0064] The ROC curve describes the increase and decrease of the true positive rate and the false positive rate. The area under the ROC curve can be used to evaluate the performance of the fundus retinal vessel segmentation network. The closer the AUC value is to 1, the better the network performance is.

[0065] Experimental verification

[0066] In order to verify the effectiveness of the model proposed in this paper, an ablation experiment was carried out on the DRIVE dataset, mainly conducting experimental analysis from two aspects: qualitative and quantitative.

[0067] Qualitative analysis:

[0068] Figure 6The figure shows the fundus retinal vessel segmentation effect in the ablation experiment. (a) represents the original RGB image, (b) represents the expert standard segmentation image, (c) S1 represents no enhanced convolution, (d) S2 represents no multi-level semantic fusion module, (e) S3 represents no hierarchical feature fusion method, (f) S4 represents no cross-stage fusion attention mechanism, and (g) S5 represents the proposed DSFN-Net method. It can be clearly seen from the highlighted parts that in terms of the segmentation of fine blood vessels and vascular coherence, the algorithm proposed in the present invention is better than the fundus retinal vessel segmentation effect of the U-Net benchmark model.

[0069] Quantitative analysis:

[0070] Table 2 records the results of various evaluation indicators in the ablation experiment. By comparing the values ​​of various evaluation indicators, it can be seen that the model proposed in the present invention achieves higher results than the baseline model on the DRIVE dataset.

[0071] Table 2 Experimental results based on DRIVE dataset

[0072] Method ACC SE SP F1 AUC S1 97.00 82.05 98.46 82.66 98.78 S2 97.07 81.49 98.59 82.92 98.82 S3 97.08 82.00 98.55 83.02 98.81 S4 97.09 79.41 98.81 82.62 98.83 S5 97.11 82.21 98.56 83.21 98.83

[0073] In order to verify the superiority of the model proposed in the present invention, we compared the experimental results of the model of the present invention with the experimental results of the existing fundus retinal blood vessel segmentation method. Tables 3, 4 and 5 show the evaluation index values ​​of the present invention and the existing literature methods on the DRIVE, CHASE_DB1 and STARE datasets, respectively. By comparing the values ​​of various evaluation indicators, it can be seen that the performance of the method of the present invention is better, and relatively balanced values ​​have been achieved in the five evaluation indicators of accuracy, sensitivity, sensitivity, F1 and AUC.

[0074] Table 3 Comparative experimental results based on DRIVE dataset

[0075] Method ACC SE SP F1 AUC MPC-EM 95.74 80.83 97.96 - 98.22 DUNet 95.66 79.63 98.00 - 98.02 SID2Net 95.20 - - 81.63 97.54 NFN+ 95.82 79.96 98.13 - 98.30 HAnet 95.81 79.91 98.13 82.93 98.23 MMDC-Net 96.07 80.74 97.55 - 96.13 BCU-Net 96.67 81.42 98.08 80.96 97.91 RVS-FDSC 96.92 82.21 98.78 82.71 98.56 FSE-Net 96.27 81.50 98.34 98.39 DSFM-Net 97.11 82.21 98.56 83.21 98.83

[0076] Table 4 Comparative experimental results based on the CHASE_DB1 dataset

[0077] Method ACC SE SP F1 AUC MPC-EM 96.54 81.38 98.07 - 98.50 HAnet 96.70 82.39 98.13 81.91 98.71 MMDC-Net 96.46 84.40 97.28 95.41 BCU-Net 97.31 84.63 98.17 79.91 98.14 RVS-FDSC 97.43 83.56 98.56 80.50 98.67 FSE-Net 97.43 77.89 98.74 98.71 DSFM-Net 97.64 85.80 98.45 82.06 99.13

[0078] Table 5 Comparative experimental results based on STARE dataset

[0079] Method ACC SE SP F1 AUC MPC-EM 96.95 81.62 98.69 - 98.98 DUNet 96.41 75.95 98.78 - 98.32 SID2Net 96.20 - - 82.33 98.24 NFN+ 96.72 79.63 98.63 - 98.75 HAnet 96.73 81.86 98.44 83.79 98.87 MMDC-Net 95.91 85.09 96.89 - 96.87 BCU-Net 97.01 85.00 98.07 82.23 98.07 RVS-FDSC 96.92 75.45 98.82 78.75 98.08 FSE-Net 97.10 68.10 99.16 98.76 DSFM-Net 97.66 87.07 98.52 84.78 98.90

[0080] Example 2

[0081] The only difference from the above embodiment is that a semantic network-based fundus vessel segmentation network system is also provided, comprising a collection module, a processing module, a construction module, a training module, a segmentation module and an evaluation module connected in sequence;

[0082] The acquisition module is used to obtain fundus retinal vascular datasets; specific fundus retinal vascular datasets include DRIVE, CHASE_DB1, STARE and HRF.

[0083] The processing module is used to perform preliminary processing on the fundus retinal vascular dataset acquired by the acquisition module to obtain a processed fundus retinal vascular dataset; specifically, the fundus retinal vascular dataset is subjected to channel color transformation, data normalization, contrast-limited adaptive histogram equalization and gamma correction processing.

[0084] The building module is used to construct a two-layer semantic fusion network model with enhanced convolution;

[0085] The training module is used to train the enhanced convolutional double-layer semantic fusion network model to obtain the trained enhanced convolutional double-layer semantic fusion network model;

[0086] The segmentation module is used to input the retinal image data set processed by the processing module into the trained enhanced convolutional double-layer semantic fusion network model to obtain the fundus retinal blood vessel segmentation result;

[0087] The evaluation module is used to evaluate the fundus retinal blood vessel segmentation result using accuracy (ACC), sensitivity (SE), F1 score (F1-Score) and area under the receiver operating characteristic curve (ROC) (AUC) as evaluation indicators.

[0088] The above is only an embodiment of the present invention, and the common knowledge such as the known specific structure and / or characteristics in the scheme is not described in detail here. It should be pointed out that for those skilled in the art, several deformations and improvements can be made without departing from the structure of the present invention, which should also be regarded as the protection scope of the present invention, and these will not affect the effect of the implementation of the present invention and the practicality of the patent. The scope of protection required by this application shall be based on the content of its claims, and the specific implementation methods and other records in the specification can be used to interpret the content of the claims.

Claims

1. A method for segmenting fundus blood vessels based on semantic networks, characterized in that: The following steps are involved: Acquire fundus retinal blood vessel dataset; Preliminarily processing the acquired fundus retinal blood vessel dataset to obtain a processed fundus retinal blood vessel dataset; Construct a two-layer semantic fusion network model with enhanced convolution; Training the enhanced convolutional double-layer semantic fusion network model to obtain a trained enhanced convolutional double-layer semantic fusion network model; The processed fundus retinal vascular dataset is input into the trained enhanced convolutional double-layer semantic fusion network model to obtain the fundus retinal vascular dataset segmentation result.

2. The method for segmenting fundus blood vessels based on semantic network according to claim 1, characterized in that: The preliminary processing includes the following steps: performing channel color transformation, data normalization, contrast-limited adaptive histogram equalization and gamma correction processing on the acquired fundus retinal blood vessel data set to obtain a processed fundus retinal blood vessel data set.

3. The method for segmenting fundus blood vessels based on semantic network according to claim 1, characterized in that: The method of constructing a double-layer semantic fusion network model with enhanced convolution includes the following steps: Design a surface network structure model, which simplifies basic semantic information processing and improves the efficiency of basic feature extraction; the surface network structure model includes an enhanced convolution module, a maximum pooling operation module, and a sampling process module; the sampling process module is used to restore the feature map to the original input size through a transposed convolution operation with a step size of 2, and finally use 1×1 convolution and Sigmoid function to output the fundus retinal blood vessel segmentation probability map; A deep network structure model is designed, which is used to extract high-order semantic information vascular feature information; the deep network structure model includes a CLAFM module, an enhanced convolution module and an HLFFM module; the HLFFM module uses a maximum pooling operation to reduce the output feature map of the first three layers of the surface network encoder to a size of 32×32 to reduce computational complexity and concentrate key information; the number of channels of the feature map is adjusted to 512 through a 1×1 convolution to unify the feature expression; these processed feature maps are fused and connected in series with the output of the underlying encoder of the deep network to further optimize the feature expression.

4. The method for segmenting fundus blood vessels based on semantic network according to claim 3, characterized in that: The CLAFM module includes a dual attention mechanism, each branch includes a spatial attention module and a channel module; the first attention module is used to strengthen the relevance of important channels in the feature map; The second attention module enhances the expression of key features by adjusting channel weights; The feature fusion strategy uses sophisticated concatenation and upsampling operations; the feature maps extracted by encoders at different levels are first upsampled and 1×1 convolved, and then concatenated.

5. The method for segmenting fundus blood vessels based on semantic network according to claim 4, characterized in that: The method also includes an evaluation step: using accuracy, sensitivity, F1 score and area under the receiver operating characteristic curve as evaluation indicators to evaluate the fundus retinal blood vessel segmentation result.

6. The semantic network-based fundus blood vessel segmentation method according to claim 5, characterized in that: Before constructing the enhanced convolutional double-layer semantic fusion network model, first construct the enhanced convolutional double-layer semantic fusion network.

7. The semantic network-based fundus blood vessel segmentation method according to claim 6, characterized in that: In constructing the enhanced convolution double-layer semantic fusion network model, the specific steps of constructing the enhanced convolution include: replacing the 3×3 convolution layer in the convolution in the deep learning network with a 1×1+3×3+1×1 convolution layer, adding FRN normalization and Mish activation function for pre-activation before the convolution, and adding a Dropout layer after the 3×3 convolution layer.

8. The semantic network-based fundus blood vessel segmentation method according to claim 7, characterized in that: The specific steps of constructing the two-layer semantic fusion network include: weighting the vascular probability map output by the surface network and the vascular probability map output by the deep network to generate a final prediction probability map.

9. A semantic network-based fundus vessel segmentation system, characterized by: It includes a collection module, a processing module, a construction module, a training module, a segmentation module and an evaluation module connected in sequence; The acquisition module is used to obtain the fundus retinal blood vessel data set; The processing module is used to perform preliminary processing on the fundus retinal blood vessel data set acquired by the acquisition module to obtain a processed fundus retinal blood vessel data set; The building module is used to build a two-layer semantic fusion network model with enhanced convolution; The training module is used to train the enhanced convolutional double-layer semantic fusion network model to obtain the trained enhanced convolutional double-layer semantic fusion network model; The segmentation module is used to input the retinal image data set processed by the processing module into the trained enhanced convolutional double-layer semantic fusion network model to obtain the fundus retinal blood vessel segmentation result; The evaluation module is used to evaluate the fundus retinal blood vessel segmentation result using accuracy, sensitivity, F1 score and area under the receiver operating characteristic curve as evaluation indicators.

Images