Retinal blood vessel segmentation method and system

Abstract

The present invention relates to a retinal vessel segmentation method and system. The method comprises the following steps: obtaining a retinal image dataset; processing the retinal image dataset of the present invention to obtain a processed retinal image dataset; constructing a bidirectional ConvLSTM residual UNet model; training the bidirectional ConvLSTM residual UNet model of the present invention to obtain a trained bidirectional ConvLSTM residual UNet model; and inputting the processed retinal image dataset of the present invention into the trained bidirectional ConvLSTM residual UNet model of the present invention to obtain a retinal vessel segmentation result. The method can distinguish small blood vessels from backgrounds in areas such as edges and low-contrast regions, thereby resolving overfitting and segmenting more complete blood vessels.

Description

Technical Field

[0001] The present invention relates to the field of blood vessel segmentation, and in particular to a retinal blood vessel segmentation method and system. Background Art

[0002] Because early-stage ophthalmic and cardiovascular diseases can cause changes in retinal vascular morphology, retinal vascular features play a crucial role in diagnosing these diseases. Diabetic retinopathy (DR) is a prime example. It is a retinal disease and a leading cause of blindness. Observing retinal vasodilation in the retinas of diabetic patients warrants particular attention. Furthermore, patients with hypertension may observe vascular tortuosity due to vascular stenosis or elevated arterial blood pressure, a condition known as hypertensive retinopathy (HR). Morphological changes in retinal vascular density, curvature, and thickness, among other factors, are crucial for diagnosing and detecting these diseases. Accurately segmenting retinal vascular images is crucial to assist physicians in making informed diagnoses. However, due to the significant variation in retinal vessel shape and size, the complex structures and regions in retinal images due to pathological areas, optic disc areas, hemorrhages, and exudates, and the low contrast in many edge regions, accurate retinal vessel segmentation is a challenging task. This also places higher demands on the automated and precise segmentation of retinal vessels. Consequently, numerous automatic retinal vascular segmentation methods have emerged.

[0003] Generally speaking, retinal vessel segmentation methods can be roughly divided into two categories: unsupervised and supervised. Unsupervised methods do not require manual annotation as a reference during training. Filter-based methods are typical unsupervised methods. For example, Zhang et al. proposed a filter-based method that uses two three-dimensional rotated frames for retinal vessel segmentation. For example, Azzoprardi et al. proposed a shift filter-response combination method that can automatically detect blood vessels. Zhang et al. used self-organizing maps for pixel clustering and used the Otsu method to divide each neuron in the output layer into retinal vessel neurons or non-retinal vessel neurons. In addition, the vascular tracking method is also an unsupervised method. However, due to the lack of manual annotation as a reference, the performance of unsupervised methods is generally lower than that of supervised methods.

[0004] In recent years, deep learning methods have been widely applied to retinal image segmentation, demonstrating superior performance due to their powerful data processing capabilities, ability to capture high-level semantic features, and robustness. Convolutional neural networks, in particular, have been widely used in numerous image processing tasks and have rapidly attracted considerable attention for retinal vessel segmentation. Ronneberger et al. proposed a well-known neural network architecture for biomedical image segmentation, known as U-Net, which was initially applied to cell segmentation. However, due to concerns about patient data privacy and ethical issues, medical images (such as retinal vascular images) are often difficult to obtain, resulting in relatively small datasets and a tendency for existing deep learning models to overfit. Consequently, researchers have focused on lightweight models, and the U-Net approach has demonstrated excellent performance on small datasets. Consequently, a large number of current retinal vessel segmentation methods are extensions of the U-Net approach. For example, Fu et al. improved vessel segmentation performance by combining a lateral output layer with a conditional random field model. Zhang et al. integrated an attention gate mechanism into a traditional guided filter to create an attention-guided filter. This model can merge structural information from feature maps of varying resolutions, while simultaneously recovering spatial information and removing complex background noise. Wang et al. proposed a DEU-Net model, in which the context path is used to capture more semantic information, while the spatial path is used to retain specific information. In order to obtain better segmentation performance, Zhang et al. proposed an edge-based mechanism based on U-Net. Hu et al. proposed a U-Net variant method combined with a saliency mechanism. Guo et al. introduced a dense residual network (DRNet) to achieve blood vessel segmentation in scanning laser ophthalmoscope (SLO) retinal images. Zhang et al. proposed a pyramid U-Net method, which proposed a pyramid scale aggregation block (PSAB) to aggregate multi-level features to achieve more accurate retinal blood vessel segmentation. Although the above-mentioned U-Net-based methods can achieve considerable results to a certain extent, there are still problems, such as: (1) it is difficult to distinguish small blood vessels from the background in areas such as edges and low contrast; (2) due to the small number of available samples, many models are prone to overfitting, and thus cannot segment more complete blood vessels. Summary of the Invention

[0005] The purpose of the present invention is to provide a retinal blood vessel segmentation method and system, which can distinguish small blood vessels from the background of edges and low-contrast areas, solve the overfitting phenomenon, and segment more complete blood vessels.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A retinal blood vessel segmentation method includes:

[0008] Obtain retinal image dataset;

[0009] processing the retinal image dataset to obtain a processed retinal image dataset;

[0010] Build a bidirectional ConvLSTM residual UNet model;

[0011] Training the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model;

[0012] The processed retinal image dataset is input into the trained bidirectional ConvLSTM residual UNet model to obtain retinal vessel segmentation results.

[0013] Optionally, processing the retinal image dataset to obtain a processed retinal image dataset specifically includes:

[0014] The retinal image dataset is subjected to horizontal, rotational, diagonal, and vertical flip enhancement processing to obtain a processed retinal image dataset.

[0015] Optionally, the bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model.

[0016] Optionally, constructing a bidirectional ConvLSTM residual UNet model specifically includes:

[0017] Design an encoder structure model to extract the most representative and discriminative high-level semantic information features of images. The design of the encoder structure model consists of three steps, each of which includes an SDRB and a maximum pooling operation.

[0018] Design a decoder structure model to restore the original image information. The decoder structure model includes three steps, each of which performs upsampling on the output of the previous step. The sampling process includes a transposed convolution with a stride of 2 and a normalization layer operation.

[0019] A bidirectional convolutional long short-term memory neural network model is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model.

[0020] Optionally, it also includes:

[0021] The accuracy, sensitivity, specificity, F1 score, area under the receiver operating characteristic curve, intersection over union ratio and Matthews correlation coefficient were used as evaluation indicators to evaluate the retinal vessel segmentation results.

[0022] A retinal blood vessel segmentation system comprises:

[0023] A retinal image dataset acquisition module is used to acquire a retinal image dataset;

[0024] a retinal image dataset processing module, configured to process the retinal image dataset to obtain a processed retinal image dataset;

[0025] Model building module, used to build a bidirectional ConvLSTM residual UNet model;

[0026] A model training module is used to train the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model;

[0027] The retinal blood vessel segmentation result determination module is used to input the processed retinal image dataset into the trained bidirectional ConvLSTM residual UNet model to obtain the retinal blood vessel segmentation result.

[0028] Optionally, the retinal image dataset processing module specifically includes:

[0029] The retinal image data set processing unit is used to perform horizontal, rotation, diagonal and vertical flip enhancement processing on the retinal image data set to obtain a processed retinal image data set.

[0030] Optionally, the bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model.

[0031] Optionally, the model building module specifically includes:

[0032] An encoder structure model design unit is used to design an encoder structure model to extract the most representative and discriminative high-level semantic information features of the image. The design of the encoder structure model consists of three steps, each of which includes an SDRB and a maximum pooling operation;

[0033] A decoder structure model design unit is used to design a decoder structure model to restore the original image information. The decoder structure model includes three steps, each of which performs upsampling on the output of the previous step. The sampling process includes a transposed convolution with a step size of 2 and a normalization layer (BN) operation;

[0034] A fusion unit is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model using a bidirectional convolutional long short-term memory neural network model.

[0035] Optionally, it also includes:

[0036] The segmentation result evaluation module is used to evaluate the retinal blood vessel segmentation result using accuracy, sensitivity, specificity, F1 score, area under the curve of the receiver operating characteristic curve, intersection-over-union ratio and Matthews correlation coefficient as evaluation indicators.

[0037] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0038] The present invention provides a retinal vessel segmentation method, comprising: obtaining a retinal image dataset; processing the retinal image dataset of the present invention to obtain a processed retinal image dataset; constructing a bidirectional ConvLSTM residual UNet model; training the bidirectional ConvLSTM residual UNet model of the present invention to obtain a trained bidirectional ConvLSTM residual UNet model; and inputting the processed retinal image dataset of the present invention into the trained bidirectional ConvLSTM residual UNet model of the present invention to obtain a retinal vessel segmentation result. The above method can distinguish small blood vessels from backgrounds in areas such as edges and low-contrast regions, resolve overfitting, and segment more complete blood vessels. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0040] Figure 1 It is the structured Dropout residual block diagram;

[0041] Figure 2 It is a bidirectional ConvLSTM structure diagram;

[0042] Figure 3 This is the BCR-UNet network architecture diagram;

[0043] Figure 4 This is the segmentation visualization effect diagram;

[0044] Figure 5 Visualization example diagrams of segmentation results for four datasets;

[0045] Figure 6 This is a flow chart of the retinal blood vessel segmentation method of the present invention;

[0046] Figure 7 This is a structural diagram of the retinal blood vessel segmentation system of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] The purpose of the present invention is to provide a retinal blood vessel segmentation method and system, which can distinguish small blood vessels from the background of edges and low-contrast areas, solve the overfitting phenomenon, and segment more complete blood vessels.

[0049] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] High-precision segmentation of retinal blood vessels from retinal images is an important step for doctors to diagnose various diseases such as glaucoma and cardiovascular disease. However, for the perivascular area, the previous U-Net-based segmentation method failed to significantly retain the small blood vessels with low contrast. To address this problem, the present invention designs a bidirectional convolutional long-short-term memory residual U-type network (BCR-UNet), which fully utilizes and integrates the advantages of traditional U-Net, Dropblock, residual convolution (Residualconvolution) and bidirectional convolutional long-short-term memory network (BConvLSTM). First, in order to improve the robustness and generalization ability of the network model, the present invention designs a new structured Dropout residual block (SDRB) and uses it to replace the convolution block in U-Net to construct the network skeleton. Then, in order to improve the network's ability to discriminate small blood vessels and retain more semantic information, the present invention jumps the BConvLSTM model to the first residual block and the last upper convolution layer in a nonlinear manner to capture high-level semantic feature maps. Figure 6 This is a flow chart of the retinal blood vessel segmentation method of the present invention. Figure 6 As shown, specific retinal blood vessel segmentation methods include:

[0051] Step 101: Acquire a retinal image dataset, which includes DRIVE, CHASE DB1, STARE, and IOSTAR. Specific information is shown in Table 1.

[0052] Table 1 Detailed information of the four databases DRIVE, CHASE DB1, STARE, and IOSTAR

[0053]

[0054] Step 102: Processing the retinal image dataset to obtain a processed retinal image dataset, specifically including:

[0055] The retinal image dataset is subjected to horizontal, rotational, diagonal, and vertical flip enhancement processing to obtain a processed retinal image dataset.

[0056] Step 103: Construct a bidirectional ConvLSTM residual UNet model. The bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model. Step 103 specifically includes:

[0057] Design an encoder structure model to extract the most representative and discriminative high-level semantic information features of images. The design of the encoder structure model consists of three steps, each of which includes an SDRB and a maximum pooling operation.

[0058] Design a decoder structure model to restore the original image information. The decoder structure model includes three steps. Each step upsamples the output of the previous step. The sampling process includes a transposed convolution with a stride of 2 and a normalization layer (BN) operation.

[0059] A bidirectional convolutional long short-term memory neural network model is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model.

[0060] Before proceeding to step 103, in order to avoid the overfitting problem of deep neural networks, the present invention introduces the Dropblock module. Compared with the Dropout method, the Dropblock module discards continuous areas in the feature map instead of randomly discarding independent units. Dropblock has two important parameters s and y, which represent the size of the discarded control block and the number of discarded activation units, where y can be calculated as:

[0061]

[0062] Among them, p represents the probability of keeping a unit in an activated state, and f represents the size of the feature map at that position.

[0063] Based on the idea of ​​Dropout module and residual network, the present invention designs a structured Dropout residual block (SDRB), whose structure is as follows Figure 1Among them, Rectified Linear Unit (ReLU), Convolutional layer (Conv), and Batch Normalization layer (BN) represent activation function, convolution and normalization layer respectively.

[0064] In order to solve the problem that the standard Long Short-Term Memory (LSTM) network only uses full connection to transform input to state and state to state, but ignores spatial correlation, a ConvLSTM model is proposed to use convolution operation for input to state and state to state transformation. t , forget gate f t , memory gate m t , and output gate o t The input gate, forget gate, and output gate function as control gates for accessing, clearing, and updating memory cells. ConvLSTM can be expressed as:

[0065]

[0066] in, and ⊙ represent the convolution function and Hadamard function respectively, X t is the input tensor, h t is the hidden state tensor, and The convolution kernels corresponding to the input and hidden states, b i ,b f ,b m and b o is the bias term.

[0067] Although ConvLSTM is an improvement over LSTM, it only processes forward dependencies and does not fully consider all the information in the sequence. Therefore, considering backward dependencies at the same time and analyzing both forward and backward dependencies is beneficial to improving prediction accuracy. Therefore, the present invention uses a bidirectional convolutional long short-term memory neural network model (BConvLSTM), whose structure is as follows Figure 2 As shown in the figure, the model uses two ConvLSTMs to process input data in the forward and backward directions respectively, and then analyzes the dependency relationship between the data in these two directions to make decisions about the current input. The BConvLSTM output formula is:

[0068]

[0069] in, represents the hidden state tensor of the forward state, and represents the hidden state tensor of the backward state, b represents the bias term, Yt ∈R F×W×H represents the final output taking into account bidirectional spatiotemporal information. In addition, tanh stands for hyperbolic tangent, which is used to combine the outputs of the forward and backward states in a nonlinear way.

[0070] According to the U-Net design concept, the present invention designs a bidirectional ConvLSTM residual UNet (BCR-UNet) model, that is, step 103. Its structure is as follows Figure 3 As shown: BCR-UNet consists of two parts, an encoder and a decoder, which can achieve end-to-end learning. The encoder is designed to extract the most representative and discriminative high-level semantic information features of the image. In BCR-UNet, the encoder consists of three steps. Each step consists of an SDRB and a maximum pooling operation. The decoder is mainly used to restore the initial image information, and also includes three steps. Each step is to perform upsampling on the output of the previous step. The sampling process includes a transposed convolution with a step size of 2 and a normalization layer (BN) operation. Unlike U-Net, which copies the feature map matched by the autoencoder to the decoder and connects it with the upsampled output, the present invention designs a complex and efficient way to fuse the two feature maps, that is, by using BConvLSTM to combine the output of the first SDRB in the encoder and the output of the last upsampled step in the decoder.

[0071] Step 104: Train the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model.

[0072] During the training phase, the learning rate is set to 0.001, the Adam algorithm is selected as the optimization method, and binary cross entropy is used as the loss function.

[0073] Step 105: Input the processed retinal image dataset into the trained bidirectional ConvLSTM residual UNet model to obtain retinal vessel segmentation results.

[0074] The above-mentioned retinal blood vessel segmentation method further includes:

[0075] Step 106: Use accuracy, sensitivity, specificity, F1 score, area under the curve of the receiver operating characteristic curve, intersection-over-union ratio and Matthews correlation coefficient as evaluation indicators to evaluate the retinal vessel segmentation results. The above steps can quantitatively evaluate the performance of BCR-UNet. Among them, the accuracy (ACC), sensitivity (SEN), specificity (SPE), F1 score (F1-score), area under the curve (AUC) of the receiver operating characteristic curve (ROC), intersection-over-union ratio (IOU) and Matthews correlation coefficient (MCC) are defined as follows:

[0076]

[0077]

[0078]

[0079]

[0080]

[0081]

[0082] Among them, T p and T n They represent true positives and true negatives, respectively, which are indicators that the predicted pixels are accurately classified as vessels and non-vessels when compared with the pixels at the same position in the ground-truth reference. p is a false positive, meaning that a pixel in the segmented image is classified as a vessel, while the corresponding pixel at the same position in the ground truth is a non-vessel pixel. n A false negative means that a pixel is classified as a non-vessel in the predicted image, while the corresponding pixel with the same position in the reference real image is a vessel pixel. AUC is the area under the receiver operating characteristic curve (ROC), which measures the segmentation performance based on recall and precision and is not affected by unbalanced data such as retinal vascular images. IOU is an evaluation of the degree of overlap between two regions (i.e., the true region and the detected region). F1 is defined as the weighted average of precision and recall, where precision is expressed as T p Divide the value by T p and F p The sum of , and the recall rate is defined as T p Divide the value by T p and F pMCC is a very effective evaluation indicator and is often used to test the performance of classification models in two-class imbalanced situations.

[0083] The BCR-UNet model was then implemented on an experimental platform with 32GB of memory and a Tesla V100 graphics card, using the Keras advanced framework and TensorFlow as the backend. Random horizontal, rotational, diagonal, and vertical flipping augmentations were applied, and 10% of the augmented images were randomly selected as the validation set. The batch size was set to 2, with 100 epochs for all datasets except the STARE dataset, which was trained 300 times. Furthermore, the dropblock size was uniformly set to 7 and the dropout rate to 0.2.

[0084] Next, different experiments were performed. The first set of ablation experiments was mainly to verify the effectiveness of the structured Dropout residual block and Bi-ConvLSTM module proposed in this invention. The present invention uses a U-shaped network stacked with SDRB modules without Dropblock as a benchmark method (called Baseline method). First, the BConvLSTM module and Dropblock were added to the Baseline (i.e., Baseline+BConvLSTM, Baseline+Dropblock), and the results are shown in Table 2. The experimental results show that the introduction of BConvLSTM and Dropblock helps to improve the segmentation performance of the method. Figure 4 The segmentation visualization effect is given, and the experimental results show that the method designed by the present invention can segment more complete blood vessels.

[0085] Table 2 Ablation experiments on four datasets

[0086]

[0087] The second set of comparative experiments with existing methods further validates the performance of the proposed method. This set of experiments compares the proposed method with existing methods, including M-Net, AG-Net, RSAN, Pyramid U-Net, U-Net, AttentionUNet, SD-UNet, MultiResUNet, and DRNe. The results on four datasets are shown in Tables 3 to 6. As can be seen from the results, the BCR-UNet designed by the proposed method achieves the highest AUCs of 0.9866, 0.9898, 0.9873, and 0.9882; the highest F1 values ​​of 0.8246, 0.8118, 0.8302, and 0.8234; the highest IOUs of 0.7015, 0.6832, 0.7103, and 0.6999; and the highest MCC values ​​of 0.8075, 0.7992, 0.8168, and 0.8091. The other three metrics also achieve comparable performance. Figure 5 Visualizations of segmentation results from four datasets are provided. Compared to other algorithms, the experimental results clearly demonstrate that the BCR-UNet method designed by the present invention can predict the majority of large and small blood vessels (indicated by red and green arrows). This further demonstrates that the combination of the SDRB module and the BConvLSTM module designed by the present invention makes the network more robust and effectively preserves small blood vessels in low-contrast vessel terminal regions.

[0088] Table 3 Results of BCR-UNET and other methods on the DRIVE dataset

[0089]

[0090] Table 4 Results of BCR-UNet and other methods on the CHASE DB1 dataset

[0091]

[0092] Table 5 Results of BCR-UNet and other methods on the STARE dataset

[0093]

[0094]

[0095] Table 6 Results of BCR-UNet and other methods on the IOSTAR dataset

[0096]

[0097] The present invention also provides a retinal blood vessel segmentation system. Figure 7 This is a structural diagram of the retinal blood vessel segmentation system of the present invention, as shown in FIG. Figure 7As shown, a retinal blood vessel segmentation system includes:

[0098] A retinal image dataset acquisition module 201 is used to acquire a retinal image dataset;

[0099] a retinal image dataset processing module 202, configured to process the retinal image dataset to obtain a processed retinal image dataset;

[0100] The model construction module 203 is used to construct a bidirectional ConvLSTM residual UNet model. The bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model.

[0101] A model training module 204 is used to train the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model;

[0102] The retinal blood vessel segmentation result determination module 205 is configured to input the processed retinal image dataset into the trained bidirectional ConvLSTM residual UNet model to obtain a retinal blood vessel segmentation result.

[0103] The retinal image dataset processing module 202 specifically includes:

[0104] The retinal image data set processing unit is used to perform horizontal, rotation, diagonal and vertical flip enhancement processing on the retinal image data set to obtain a processed retinal image data set.

[0105] The model building module 203 specifically includes:

[0106] An encoder structure model design unit is used to design an encoder structure model to extract the most representative and discriminative high-level semantic information features of the image. The design of the encoder structure model consists of three steps, each of which includes an SDRB and a maximum pooling operation;

[0107] A decoder structure model design unit is used to design a decoder structure model to restore the original image information. The decoder structure model includes three steps, each of which performs upsampling on the output of the previous step. The sampling process includes a transposed convolution with a step size of 2 and a normalization layer (BN) operation;

[0108] A fusion unit is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model using a bidirectional convolutional long short-term memory neural network model.

[0109] The above-mentioned retinal blood vessel segmentation system further includes:

[0110] The segmentation result evaluation module 206 is used to evaluate the retinal vessel segmentation result using accuracy, sensitivity, specificity, F1 score, area under the curve of the receiver operating characteristic curve, intersection over union ratio and Matthews correlation coefficient as evaluation indicators.

[0111] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0112] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A retinal blood vessel segmentation method, characterized in that: include: Obtain retinal image dataset; processing the retinal image dataset to obtain a processed retinal image dataset; Build a bidirectional ConvLSTM residual UNet model; Training the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model; Inputting the processed retinal image dataset into the trained bidirectional ConvLSTM residual UNet model to obtain retinal blood vessel segmentation results; The construction of the bidirectional ConvLSTM residual UNet model specifically includes: Design an encoder structure model to extract the most representative and discriminative high-level semantic information features of the image. The design of the encoder structure model consists of three steps, each of which includes a structured Dropout residual block and a maximum pooling operation; Design a decoder structure model to restore the original image information. The decoder structure model includes three steps, each of which performs upsampling on the output of the previous step. The sampling process includes a transposed convolution with a stride of 2 and a normalization layer operation. A bidirectional convolutional long short-term memory neural network model is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model.

2. The retinal blood vessel segmentation method according to claim 1, wherein: The processing of the retinal image dataset to obtain a processed retinal image dataset specifically includes: The retinal image dataset is subjected to horizontal, rotational, diagonal, and vertical flip enhancement processing to obtain a processed retinal image dataset.

3. The retinal blood vessel segmentation method according to claim 1, wherein: The bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model.

4. The retinal blood vessel segmentation method according to claim 1, wherein: Also includes: The accuracy, sensitivity, specificity, F1 score, area under the receiver operating characteristic curve, intersection over union ratio and Matthews correlation coefficient were used as evaluation indicators to evaluate the retinal vessel segmentation results.

5. A system based on the retinal blood vessel segmentation method according to any one of claims 1 to 4, characterized in that: include: A retinal image dataset acquisition module is used to acquire a retinal image dataset; a retinal image dataset processing module, configured to process the retinal image dataset to obtain a processed retinal image dataset; Model building module, used to build a bidirectional ConvLSTM residual UNet model; A model training module is used to train the bidirectional ConvLSTM residual UNet model to obtain a trained bidirectional ConvLSTM residual UNet model; a retinal vessel segmentation result determination module, configured to input the processed retinal image dataset into the trained bidirectional ConvLSTM residual UNet model to obtain a retinal vessel segmentation result; The model building module specifically includes: An encoder structure model design unit is used to design an encoder structure model to extract the most representative and discriminative high-level semantic information features of the image. The design of the encoder structure model consists of three steps, each of which includes a structured Dropout residual block and a maximum pooling operation; A decoder structure model design unit is used to design a decoder structure model to restore the original image information. The decoder structure model includes three steps, each of which performs upsampling on the output of the previous step. The sampling process includes a transposed convolution with a stride of 2 and a normalization layer operation; A fusion unit is used to combine the output of the first structured Dropout residual block in the encoder structure model and the output of the last upsampling step in the decoder structure model using a bidirectional convolutional long short-term memory neural network model.

6. The retinal blood vessel segmentation system according to claim 5, characterized in that: The retinal image dataset processing module specifically includes: The retinal image data set processing unit is used to perform horizontal, rotation, diagonal and vertical flip enhancement processing on the retinal image data set to obtain a processed retinal image data set.

7. The retinal blood vessel segmentation system according to claim 5, characterized in that: The bidirectional ConvLSTM residual UNet model is an encoder-decoder structure model.

8. The retinal blood vessel segmentation system according to claim 5, characterized in that: Also includes: The segmentation result evaluation module is used to evaluate the retinal blood vessel segmentation result using accuracy, sensitivity, specificity, F1 score, area under the curve of the receiver operating characteristic curve, intersection-over-union ratio and Matthews correlation coefficient as evaluation indicators.

Images