An OCTA image retinal blood vessel segmentation method based on deep learning
By using a deep learning-based dual-branch axially compressed convolutional neural network model, the inaccuracy problem caused by retinal disease interference in OCTA image segmentation was solved, achieving high accuracy and fast retinal vessel segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANGHAI NORMAL UNIVERSITY
- Filing Date
- 2024-03-28
- Publication Date
- 2026-05-12
Smart Images
Figure CN118261924B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of image processing technology, and in particular to a deep learning-based method for retinal vessel segmentation in OCTA images. Background Technology
[0002] Optical coherence tomography (OCT) is one of the most important achievements in the field of retinal imaging. It captures three-dimensional structural data of the retina with micron-level resolution and in a non-contact manner, without requiring the injection of external vascular dyes for contrast. This technology provides efficient and concise information for clinical research and detection of retinal diseases. Based on the maturity of OCT technology, optical coherence tomography angiography (OCTA) has been developed into a highly efficient and accurate imaging method for providing functional information on retinal vessels and the small vascular system.
[0003] Compared to traditional fundus angiography methods, OCTA is non-invasive. It provides high-resolution, depth-resolution angiography, avoiding the need for intravenous injection of vascular dyes and overcoming the limitation of traditional OCT in accurately scanning blood vessels. This lays the foundation for the clinical diagnosis of retinal diseases and significantly improves the efficiency of physicians' medical decisions. Since retinal vessel (RV) segmentation and foveal ischemic zone (FAZ) segmentation are key criteria for assessing retinal diseases, accurate and efficient RV and FAZ segmentation of OCTA images is necessary and challenging.
[0004] In both segmentation tasks, OCTA provides 3D data, but the retinal indicators required for clinical assessment are primarily quantified on 2D projection maps, and 2D images are also more convenient for doctors to observe and diagnose. Therefore, the segmentation result we need should be a 2D image. To achieve the transformation from 3D input to 2D output, we introduce deep learning methods to allow the network to extract features from the 3D image and summarize them onto a 2D plane. Using image compression and dimensionality reduction strategies, the 3D image is gradually compressed into a 2D feature map. Deep learning has developed rapidly in recent years and has been widely applied in fields such as object detection and medical image processing. In this chapter, we consider using a CNN-based method to process OCTA images.
[0005] Improving the accuracy of OCTA image segmentation requires networks with strong spatial detail capture and feature extraction capabilities. Most CNN networks designed for image understanding consist of stacked convolutional layers. Due to computational cost considerations, the receptive field of feature maps gradually increases while the resolution of feature maps gradually decreases, allowing the encoder to learn more abstract semantic information. However, due to the limitation of the receptive field, traditional CNNs face challenges in extracting spatial information. Summary of the Invention
[0006] In view of the aforementioned shortcomings of the prior art, the technical problem to be solved by the present invention is that existing OCTA image segmentation methods only perform segmentation on the projection map of OCTA, which is easily affected by potential retinal diseases and may lead to inaccurate segmentation results. The present invention provides a deep learning-based OCTA image retinal vessel segmentation method, which features high accuracy and fast processing speed.
[0007] To achieve the above objectives, this invention provides a deep learning-based method for retinal vessel segmentation in OCTA images, comprising the following steps:
[0008] Construct a convolutional neural network segmentation model with dual-branch axial compression;
[0009] Train and optimize the parameters of the built model;
[0010] Using a trained model, the retinal vascular structure is rapidly located and accurately segmented based on OCTA to obtain a binary image of the retinal vascular segmentation result.
[0011] Furthermore, the specific steps include:
[0012] Obtain the dataset and preprocess it.
[0013] The labeled data in the preprocessed dataset is divided into training and test sets for the deep learning network.
[0014] Finally, a two-branch axial compression convolutional neural network segmentation model is constructed, which includes a two-branch structure, an axial compression module, and a cross-feature fusion module.
[0015] Furthermore, a dual-branch axial compression convolutional neural network segmentation model is constructed, specifically including the following steps: The dual-branch axial compression convolutional neural network segmentation model is set to use the 3D-Unet framework as the basic structure, integrating multiple consecutive axial compression modules into the network to accumulate effective information from three-dimensional space to two-dimensional plane; the cross-feature fusion module extracts features from OCTA images by using images of different scales as input; the dual-branch structure uses images of different scales as input to extract features from OCTA images, uses a convolutional neural network as the backbone network, and the output of the network is the final global segmentation result.
[0016] Furthermore, when accumulating effective information from three-dimensional space to a two-dimensional plane, it is necessary to reduce the depth of the volumetric image to 1 and extract the features of its projection direction.
[0017] Furthermore, the axial compression module is used for backbone feature extraction at different scales, including three 3D convolutional layers, one generalized mean pooling layer, and one batch normalization layer; wherein the kernel size of the 3D convolution is set to 3×3×3.
[0018] Furthermore, the built model is trained and its parameters are optimized. Specifically, the deep learning network is initialized using Nvidia's CUDNN convolutional layers and standard normal initialization methods.
[0019] To prevent deep learning networks from overfitting, the Dice coefficients of the validation set are calculated during training, and the optimal model is saved when the Dice coefficients of the validation set reach their maximum value.
[0020] During the testing phase, output the target probability images for the validation set and the test set;
[0021] Finally, the stochastic gradient descent algorithm was used to train the deep learning network.
[0022] Furthermore, the labeled data in the preprocessed dataset is divided into training and testing sets for the deep learning network. Specifically, the preprocessed and online-enhanced retinal vessel dataset with retinal vessel segmentation labels is divided into training, validation, and testing sets in a 7:2:1 ratio, and ten-fold cross-validation is used to pre-train the network segmentation model.
[0023] Furthermore, the cross-entropy loss function is used as the optimization objective of the segmentation network, and the objective function is defined as follows:
[0024]
[0025] in, It's a real label. These are the predicted values from the model.
[0026] Furthermore, the deep learning network is finally trained using the stochastic gradient descent algorithm, and its parameter formula is shown below:
[0027]
[0028] in, This represents the hyperparameter, which is the initially set learning rate. These are network parameters. Indicates the number of iterations of the model. The gradient of the parameter.
[0029] Furthermore, using the trained model, the retinal vascular structure based on OCTA is rapidly located and accurately segmented to obtain a binary image of the retinal vascular segmentation result. This specifically includes the following steps:
[0030] The fundus images based on OCTA are sequentially preprocessed and enhanced to obtain the processed images;
[0031] The image after online enhancement is fed into an axial compression module consisting of convolutional layers, pooling layers, and normalization layers as a four-channel input for feature extraction, and the reconstructed retinal vascular image feature map is automatically located and output.
[0032] The reconstructed retinal blood vessel image feature map is input into the classifier, and the pixel of the feature map is predicted one by one using a sliding window method to generate two pixel label prediction score maps with the same size as the original image.
[0033] Use the ReLU function to transform the predicted scores into a probability distribution;
[0034] By taking the index component with the highest probability of each pixel as the pixel category label, we can achieve rapid localization of blood vessel structures and obtain a binary image of retinal blood vessel segmentation results.
[0035] Technical effect
[0036] This invention provides a deep learning-based method for retinal vessel segmentation in OCTA images. It employs a dual-scale axial compression network, an enhanced 3D-to-2D medical image segmentation network designed for RV and FAZ segmentation in OCTA images, achieving higher accuracy. A novel axial compression module is used to reduce the height axis of the 3D image by adjusting the generalized mean pooling layer, while maintaining the horizontal dimension, thus achieving dimensionality reduction. A dual-scale feature fusion strategy is introduced to better learn feature information from feature maps at different scales, capturing complete vascular structures and avoiding discontinuities or breaks in retinal vessel segmentation, thereby improving segmentation accuracy.
[0037] The following will further explain the concept, specific structure, and technical effects of the present invention in conjunction with the accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Attached Figure Description
[0038] Figure 1 This is a schematic diagram of a preferred embodiment of the OCTA image retinal vessel segmentation method based on deep learning.
[0039] Figure 2 This is a flowchart of the generalized mean pooling training process for a deep learning-based OCTA image retinal vessel segmentation method, which is a preferred embodiment of the present invention.
[0040] Figure 3 This is a schematic diagram of a dual-branch network structure for a deep learning-based OCTA image retinal vessel segmentation method according to a preferred embodiment of the present invention.
[0041] Figure 4 This is a schematic diagram of the axial compression module structure of a deep learning-based OCTA image retinal vessel segmentation method according to a preferred embodiment of the present invention.
[0042] Figure 5 This is a schematic diagram of the cross-feature fusion module structure of a deep learning-based OCTA image retinal vessel segmentation method according to a preferred embodiment of the present invention.
[0043] Figure 6 This is a schematic diagram of the original image and network segmentation results of a deep learning-based OCTA image retinal vessel segmentation method according to a preferred embodiment of the present invention. Detailed Implementation
[0044] To make the technical problems to be solved, the technical solutions, and the beneficial effects of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the present invention and are not intended to limit the present invention.
[0045] In the following description, specific details, such as particular internal procedures and techniques, are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of the invention. However, those skilled in the art will appreciate that the invention may be practiced in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of the invention with unnecessary detail.
[0046] like Figure 1 As shown, the present invention provides a deep learning-based OCTA image retinal vessel segmentation method, comprising the following steps:
[0047] Step 100: Construct a convolutional neural network segmentation model with dual-branch axial compression;
[0048] Step 200: Train and optimize the parameters of the built model;
[0049] Step 300: Using the trained model, the retinal vascular structure based on OCTA is rapidly located and accurately segmented to obtain a binary image of the retinal vascular segmentation result.
[0050] Step 100 involves constructing a dual-branch axial compression convolutional neural network segmentation model. This model includes a dual-branch structure, an axial compression module, and a cross-feature fusion module. In each branch, a series of 3D data blocks after segmentation are input into the axial compression module. Each axial compression module consists of three 3D convolutional layers, a unidirectional pooling layer, and a normalization layer. Unidirectional pooling is performed on the activations of the last convolutional layer to ensure that the height of the 3D image decreases by the same proportion after each pooling operation. This pooling method predicts a weight that represents the discriminative power of each activation at each pixel location for image matching and retrieval.
[0051] A two-branch axial compression convolutional neural network (CNN) segmentation model is constructed, specifically including the following steps: The model uses the 3D-Unet framework as its basic structure, integrating multiple consecutive axial compression modules into the network to accumulate effective information from 3D space to the 2D plane. When accumulating this information, the depth of the volumetric image needs to be reduced to 1, while simultaneously extracting features from its projection direction. The cross-feature fusion module extracts features from the OCTA image by using images of different scales as input. The two-branch structure uses images of different scales as input to extract features from the OCTA image, using a convolutional neural network as the backbone network. The network output is the final global segmentation result. Figure 4 As shown, in this embodiment, OCT and OCTA images are used as inputs to the 3D-to-2D method. In the FAZ segmentation task, we use an introduced distance map as additional input. The OCTA image size in the large-scale branch is 640px × 400px × 400px, and the slice size is 160px × 100px × 100px. The OCTA image size in the small-scale branch is 640px × 304px × 304px, and the slice size is 160px × 76px × 76px. Specifically, the cropping and slicing operations on the input 3D data, and the pixel value conversion between the input 3D data and the 2D image, include:
[0052] In the vertical direction, the 640px dimension is cropped, retaining only pixels with vertical coordinates in the range [161, 480].
[0053] In the horizontal direction, the entire 3D data is divided into 16 small blocks, and the size of each block is 19px×19px×320px or 25px×25px×320px.
[0054] Transform the pixel value range of the input 3D data and 2D image from [0,255] to [0,1].
[0055] In this embodiment of the invention, image compression along the height axis is achieved by adjusting the kernel size of the generalized mean pooling layer, and downsampling is performed only in the height direction. In the single-branch network, the input 3D data block is unidirectionally pooled by the generalized mean pooling layer, and the attention weights of the 3D data block are calculated and weighted.
[0056] like Figure 3 As shown, the dual-branch structure uses images of different scales as input to extract features from OCTA images, uses a convolutional neural network as the backbone network, and the network output is the final global segmentation result. The input and output can be represented as:
[0057]
[0058] The overall process can be represented as:
[0059]
[0060] Where I is the network input and O is the network output. L, W, and H are the length, width, and height of the original image; l and w are the length and width of the segmented image. For the output of a single-branch network, This is the output after feature fusion of the two branches.
[0061] The axial compression module is used for backbone feature extraction at different scales, and includes three 3D convolutional layers, one generalized mean pooling layer, and one batch normalization layer. The convolutional layers are responsible for extracting image features, while the generalized mean pooling layer is used to effectively select features along the height direction. The kernel size of the 3D convolutions is set to 3×3×3. Figure 2 As shown, the descriptor for generalized mean pooling is represented as follows: in It is given by the following formula:
[0062]
[0063] The feature map produced by the last convolutional layer is used as the output, represented as... Where k represents the number of channels. Each feature map generated by each module is represented as... ,correspond The k-th feature map in the dataset. We use pooling operations to generate the representation. of This operation allows the input to be represented as a vector. This pooling method uses... and This represents the depth of pooling and the maximum activation of convolution.
[0064] The sampling process of generalized mean pooling on the height axis can be represented as:
[0065]
[0066] in, Indicates the height of the input image. Let k be the height of the output image, k be the pooling kernel size, and d be the stride. In this chapter, we set k equal to d. Therefore, the above equation becomes:
[0067]
[0068] Through this process, the height axis of the 3D image is reduced while the length and width dimensions remain unchanged. After several consecutive ACMs, the height of the 3D image eventually becomes 1, becoming a 2D feature map.
[0069] like Figure 5 As shown, the cross-feature fusion module extracts features from OCTA images by using images at different scales as input. First, the outputs from the two branch networks at different scales are used as input to this module, and supervised learning features are applied to capture local spatial features. Second, the output features extracted through the large-scale receptive field are upsampled and merged with the output features extracted through the small-scale receptive field. The dimension of the feature map is set to... Where l is the length, h is the height, w is the width, and c is the number of channels. The dimension of the feature map extracted through the large-scale receptive field is set to... Defined as Set the dimension of the feature map extracted from the small-scale receptive field to 1. Defined as The final fusion features are calculated using the following formula:
[0070]
[0071] In this process This indicates the upsampling process, which directly uses the nearest neighbor interpolation method for calculation. Nearest neighbor interpolation is a basic grayscale interpolation method that aims to ensure the converted pixel's grayscale value equals the grayscale value of its nearest neighbor input pixel. Achieving small-scale features It has large-scale characteristics Same length and width properties.
[0072] Simultaneously, no sampling is performed on the height axis to ensure that the output scale is the same for both branches at the same stage. The two feature maps are merged along the channel dimension using CFFM, resulting in a final fused feature dimension of [...]. ].
[0073] Step 200 involves training and optimizing the parameters of the constructed model. Among these steps...
[0074] Step 201: Obtain the dataset and preprocess the obtained dataset. Divide the labeled data in the preprocessed dataset into training set and test set for the deep learning network.
[0075] Step 202: Finally, the training set is used as input to train and optimize the network.
[0076] Specifically, this includes initializing the deep learning network using Nvidia's CUDNN convolutional layers and standard normal initialization methods; in this embodiment, the cross-entropy loss function is used, the batch size is set to 4, the maximum number of iterations is 20k, and the initial learning rate is... The network was initialized with a variance of 0.02 using the standard normal initialization method. To prevent overfitting of the deep learning network, the Dice coefficients of the validation set were calculated during training, and the optimal model was saved when the Dice coefficients of the validation set reached their maximum value.
[0077] During the testing phase, output the target probability images for the validation set and the test set;
[0078] Finally, the stochastic gradient descent algorithm was used to train the deep learning network.
[0079] Step 202 includes step 2021, which divides the preprocessed and online-enhanced retinal vessel dataset with retinal vessel segmentation labels into training set, validation set and test set in a 7:2:1 ratio, and uses ten-fold cross-validation to pre-train the network segmentation model.
[0080] The OCTA image of the same retinal vessel sample is input into two branches of the network, and the retinal vessel segmentation result is generated through the forward computation of the network. The forward computation of the network includes convolution operation, nonlinear excitation, probability value transformation, etc.
[0081] Step 2022: The cross-entropy loss function is used as the optimization objective of the segmentation network. The objective function is defined as follows:
[0082]
[0083] in, It's a real label. These are the predicted values from the model.
[0084] Step 2023: Optimize the function: Use the stochastic gradient descent algorithm to train and optimize the network. The parameter formulas are shown below:
[0085]
[0086] in, This represents the hyperparameter, which is the initially set learning rate. These are network parameters. Indicates the number of iterations of the model. The gradient of the parameter.
[0087] The nonlinear activation is achieved by using a rectified linear unit (ReLU) as the nonlinear activation function to nonlinearly transform each value in the output feature map generated by the convolution kernel. The ReLU is defined as: f(x) = max(0, x).
[0088] Here, f(x) represents the rectified linear unit function, max represents finding the maximum value, and x is an input value.
[0089] Step 300: Using the trained model, the retinal vascular structure based on OCTA is rapidly located and accurately segmented to obtain a binary image of the retinal vascular segmentation result. This specifically includes the following steps:
[0090] The fundus images based on OCTA are sequentially preprocessed and enhanced to obtain the processed images;
[0091] The image after online enhancement is fed into an axial compression module consisting of convolutional layers, pooling layers, and normalization layers as a four-channel input for feature extraction, and the reconstructed retinal vascular image feature map is automatically located and output.
[0092] The reconstructed retinal blood vessel image feature map is input into the classifier, and the pixel of the feature map is predicted one by one using a sliding window method to generate two pixel label prediction score maps with the same size as the original image.
[0093] Use the ReLU function to transform the predicted scores into a probability distribution;
[0094] By taking the index component with the highest probability of each pixel as the pixel category label, we can achieve rapid localization of blood vessel structures and obtain a binary image of retinal blood vessel segmentation results.
[0095] The following is a specific example illustrating a deep learning-based OCTA image retinal vessel segmentation method according to an embodiment of the present invention, which specifically includes the following steps:
[0096] Step 1) Construction of a dual-scale axial compression convolutional neural network segmentation model:
[0097] Step 11) The two-branch axial compression convolutional neural network segmentation model consists of a two-branch structure ( Figure 3 ), axial compression module ( Figure 4 ) and cross-feature fusion module ( Figure 5 The network is composed of a 400×400×640 three-dimensional OCTA image as input, which outputs a 400×400×1 2D image after successive ACMs. Axial compression modules of different scales are used to downsample the retinal vascular feature map based on the OCTA fundus image five times along the height axis, gradually changing the height to 640, 128, 32, 8, 2, 1. This integrates the spatial information of the three-dimensional OCTA image onto a two-dimensional plane, requiring the depth of the volumetric image to be reduced to 1 while extracting features of its projection direction.
[0098] Step 12) Further, vascular detail features are extracted from the two vascular feature maps of different scales, and the features of the two branches are fused and enhanced through a feature fusion module. The input image and the output image have the same two-dimensional size, enabling end-to-end semantic segmentation.
[0099] Step 13) A 400×400×640 3D OCTA image is input into the first axial compression module, which consists of three 3D convolutional layers, one generalized mean pooling layer, and one batch normalization layer. The convolutional layers are responsible for extracting image features, while the generalized mean pooling layer is used to effectively select features along the height direction. The kernel size of the 3D convolutions is set to 3×3×3, and the kernel size of the pooling layer is adjusted to 3×1 for unidirectional pooling. After passing through the first ACM, a feature map of size 400×400×128 is output.
[0100] Step 14) The feature fusion module fuses the feature information extracted by each ACM in the two branches of the network, and applies supervised learning features to capture local spatial features. The dimension of the feature map is set to... Where l is the length, h is the height, w is the width, and c is the number of channels. The dimension of the feature map extracted through the large-scale receptive field is set to... Defined as Set the dimension of the feature map extracted from the small-scale receptive field to 1. Defined as During this process, This indicates the upsampling process, which directly uses the nearest neighbor interpolation method for calculation. Nearest neighbor interpolation is a basic grayscale interpolation method that aims to ensure the converted pixel's grayscale value equals the grayscale value of its nearest neighbor input pixel. Achieving small-scale features It has large-scale characteristics The same length and width attributes are used. No sampling is performed on the height axis to ensure that the output scale is the same for both branches at the same stage. The two feature maps are merged along the channel dimension using CFFM, resulting in a final fused feature with dimensions of [...]. ].
[0101] Step 15) After consecutive ACM and CFFM, the final output is a 2D result image with a size of 400×400×1.
[0102] Step 2) Divide the labeled data in the preprocessed dataset into training and test sets for the deep learning network. See Table 1 below:
[0103] Table 1 Dataset Distribution
[0104] Table 1
[0105]
[0106] Step 21) In this embodiment, the image data used comes from the OCTA-500 dataset, which includes two subsets with different field-of-view types. The first subset, called OCTA_6M, consists of OCTA images of 300 normal individuals or patients with a 6mm × 6mm FOV, with a file size of 400 pixels (px) × 400px × 640px. Its label image size is 400px × 400px. The second subset, called OCTA_3M, consists of OCTA images of 200 normal individuals with a 3mm × 3mm FOV, with a file size of 304px × 304px × 640px. Its label image size is 304px × 304px.
[0107] Step 22) In the height direction, the 640px dimension is cropped, keeping only the pixels with height coordinates in the range [161, 480]. In the horizontal direction, the entire 3D data is divided into 16 small blocks, and the size of each block is 19px×19px×320px or 25px×25px×320px.
[0108] Step 23) Convert the pixel value range of the input 3D data and 2D image from [0, 255] to [0, 1].
[0109] Step 3) Model training and parameter optimization
[0110] Step 31) Use the CUDNN convolutional layer from Nvidia. We use the cross-entropy loss function, set the batch size to 4, the maximum number of iterations to 20k, and the initial learning rate to... The network was initialized with a variance of 0.02 using the standard normal initialization method.
[0111] Step 32) To prevent the network from overfitting, the Dice coefficients of the validation set are calculated during training. The optimal model is saved when the Dice coefficients of the validation set reach their maximum value. During the testing phase, target probability images of the validation and test sets are output. The final threshold is determined by collecting Dice coefficients on the validation set, and this threshold is used to obtain the best trained model.
[0112] Step 33) The loss functions used in this invention are cross-entropy loss and Dice loss.
[0113] The cross-entropy loss function is commonly used for classification tasks, and its formula is shown in Figure 2-8:
[0114]
[0115] in, It's a real label. This represents the model's predicted value. The loss for the aforementioned dual-branch feature cross-fusion task is implemented using cross-entropy loss.
[0116] Dice loss is a type of loss function that appears very frequently in medical image segmentation tasks, and its formula is shown in 2-9:
[0117]
[0118] Here, X and Y represent the true labels and the model's predicted values, respectively. In Chapters 3 and 4 of this paper, labeled OCTA images are used for training, and Dice is used to calculate the loss. The smaller the Dice loss, the more similar the two images are, indicating better segmentation. The best Dice iteration is used during network training to retain the model and ensure optimal network performance.
[0119] Step 34) Train the network using the stochastic gradient descent algorithm. The parameter formulas are as follows:
[0120]
[0121] in, This represents the hyperparameter, which is the initially set learning rate. These are network parameters. Indicates the number of iterations of the model. The gradient of the parameter.
[0122] Step 4) Use the optimized OCTA image segmentation network to segment the OCTA image and obtain the segmentation results of retinal vessels and central ischemic area.
[0123] The preferred embodiments of the present invention have been described in detail above. It should be understood that those skilled in the art can make numerous modifications and variations based on the concept of the present invention without creative effort. Therefore, all technical solutions that can be obtained by those skilled in the art based on the concept of the present invention through logical analysis, reasoning, or limited experimentation on the basis of existing technology should be within the scope of protection defined by the claims.
Claims
1. A deep learning-based method for retinal vessel segmentation in OCTA images, characterized in that, Includes the following steps: A two-branch axial compression convolutional neural network (CNN) segmentation model is constructed. This model comprises a two-branch structure, an axial compression module, and a cross-feature fusion module. Specifically, the model includes the following steps: The two-branch axial compression CNN segmentation model uses the 3D-Unet framework as its basic structure, integrating multiple consecutive axial compression modules into the network to accumulate effective information from three-dimensional space to a two-dimensional plane. The cross-feature fusion module extracts features from OCTA images using images of different scales as input. The two-branch structure uses images of different scales as input to extract features from OCTA images, using a convolutional neural network as the backbone network, with the network output being the final global segmentation result. When accumulating effective information from three-dimensional space to a two-dimensional plane, the depth of the volumetric image needs to be reduced to 1, while simultaneously extracting features from its projection direction. The axial compression module is used for backbone feature extraction at different scales, including three 3D convolutional layers, one generalized mean pooling layer, and one batch normalization layer. The kernel size of the 3D convolution is set to 3×3×3. Train and optimize the parameters of the built model; Using a trained model, the retinal vascular structure is rapidly located and accurately segmented based on OCTA to obtain a binary image of the retinal vascular segmentation result.
2. The deep learning-based OCTA image retinal vessel segmentation method as described in claim 1, characterized in that, Specifically, the following steps are included: Construct a convolutional neural network segmentation model with dual-branch axial compression; Obtain the dataset and preprocess it. The labeled data in the preprocessed dataset is divided into training and testing sets for the deep learning network. The training set is used as the input to the network for training and optimization. Finally, the optimized OCTA image segmentation network was used to segment the OCTA image to obtain the segmentation results of retinal vessels and central ischemic area.
3. The method for retinal vessel segmentation in OCTA images based on deep learning as described in claim 1, characterized in that, The built model is trained and its parameters are optimized. Specifically, the deep learning network is initialized using Nvidia's CUDNN convolutional layers and standard normal initialization methods. To prevent deep learning networks from overfitting, the Dice coefficients of the validation set are calculated during training, and the optimal model is saved when the Dice coefficients of the validation set reach their maximum value. During the testing phase, output the target probability images for the validation set and the test set; Finally, the stochastic gradient descent algorithm was used to train the deep learning network.
4. The deep learning-based OCTA image retinal vessel segmentation method as described in claim 2, characterized in that, The labeled data in the preprocessed dataset is divided into training and testing sets for the deep learning network. Specifically, the preprocessed and online-enhanced retinal vessel dataset with retinal vessel segmentation labels is divided into training, validation, and testing sets in a 7:2:1 ratio, and ten-fold cross-validation is used to pre-train the network segmentation model.
5. The deep learning-based OCTA image retinal vessel segmentation method as described in claim 3, characterized in that, The cross-entropy loss function is used as the optimization objective of the segmentation network. The objective function is defined as follows: ， in, It's a real label. These are the predicted values from the model.
6. The deep learning-based OCTA image retinal vessel segmentation method as described in claim 5, characterized in that, Finally, the stochastic gradient descent algorithm is used to train the deep learning network, and its parameter formula is shown below: ， in, This represents the hyperparameter, which is the initially set learning rate. These are network parameters. Indicates the number of iterations of the model. The gradient of the parameter.
7. The deep learning-based OCTA image retinal vessel segmentation method as described in claim 1, characterized in that, Using a trained model, the retinal vascular structure is rapidly located and accurately segmented based on OCTA to obtain a binary image of the retinal vascular segmentation result. The specific steps include: The fundus images based on OCTA are sequentially preprocessed and enhanced to obtain the processed images; The image after online enhancement is fed into an axial compression module consisting of convolutional layers, pooling layers, and normalization layers as a four-channel input for feature extraction, and the reconstructed retinal vascular image feature map is automatically located and output. The reconstructed retinal blood vessel image feature map is input into the classifier, and the pixel of the feature map is predicted one by one using a sliding window method to generate two pixel label prediction score maps with the same size as the original image. Use the ReLU function to transform the predicted scores into a probability distribution; By taking the index component with the highest probability of each pixel as the pixel category label, we can achieve rapid localization of blood vessel structures and obtain a binary image of retinal blood vessel segmentation results.