An eye fundus OCT image segmentation method based on feature polarization and boundary regression

Abstract

The application discloses an eye fundus OCT image segmentation method based on feature polarization and boundary regression. First, the eye fundus OCT image is preprocessed, the eye fundus OCT image is input into a convolution backbone neural network and a transformer backbone network respectively to extract corresponding features, the extracted features are fused, and an output probability map is obtained by inputting the fused features into a decoder. Then, all the extracted features are subjected to feature polarization constraint, the output probability map is subjected to boundary regression constraint, an OCT segmentation model is trained based on back propagation by using a loss function, and finally, a new eye fundus OCT image is predicted based on the trained weight. The method of the application designs a segmentation model of a convolution network and a transformer network mixed as an encoder, can capture the shape and details of the retinal layer in the OCT, can enhance the robustness of the feature extraction of the encoder by using the proposed feature polarization loss, can optimize the accurate positioning of the retinal layer boundary by the decoder by using the proposed boundary regression loss, and is very suitable for the retinal layer segmentation of the eye fundus OCT image.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of medical image processing, and particularly relates to an eye fundus OCT image segmentation method based on feature polarization and boundary regression. BACKGROUND

[0002] In medical imaging, optical coherence tomography (OCT) is an important imaging method, which is often used in ophthalmology for retinal imaging. Based on OCT technology, ophthalmologists can observe features that cannot be observed by other imaging modes. For example, although fundus color photography can clearly show the optic cup, optic disc and fovea, it cannot reflect the thickness information of different retinal layers. OCT technology can conveniently perform three-dimensional imaging of the fundus, and OCT-B and OCT-A scanning are commonly used scanning modes, from which the distribution of multiple retinal layers can be seen. Doctors can measure the thickness of certain retinal layers in different parts (such as above or below the optic disc) on such OCT images, and then make judgments based on experience. For example, the clinical evaluation of optic neuropathy requires a comprehensive analysis of the retinal nerve fiber layer (RNFL) thickness by optical coherence tomography imaging. Ophthalmologists usually measure the thickness of the retinal layer to determine the loss of axon fiber bundles in various eye diseases, such as glaucoma, compressive optic neuropathy, optic neuritis and other diseases that can cause permanent blindness. Accurate segmentation of the retinal layer in the OCT image is a necessary prerequisite for measuring the geometric features of the OCT layer. Therefore, it is of great significance for intelligent imaging diagnosis of ophthalmic diseases to study an automatic, intelligent and efficient OCT layer segmentation algorithm.

[0003] The layer segmentation method of fundus OCT image has experienced the transformation from the existing edge positioning, contour search based on graph theory to automatic learning by machine learning and neural network. The early formed OCT image segmentation method is edge iteration method, such as filtering non-vascular area, and active contour method approximating retinal layer edge. There is also a relatively successful type of method based on graph theory to detect retinal layer boundary, which can segment two-dimensional and three-dimensional OCT surface according to graph theory knowledge, and can also be combined with other technologies such as machine learning method. Some prior knowledge such as local difference, statistical model and topological order between retinal layers can also be used to assist in judgment. However, these methods are heavily dependent on manual design of algorithm and adjustment of hyperparameters, and need to calibrate parameters when encountering new equipment or images collected in different environments, and the process is relatively cumbersome. In recent years, data-driven machine learning and even deep learning methods have become mainstream. Using machine learning to do OCT layer segmentation needs to manually extract image features, which is also called feature engineering. Usually, spatial features such as gray level statistical features, second-order gradient features and wavelet features of OCT images are extracted, and these features and the labels marked by doctors are sent into a machine learning classifier, so that a more robust layer classification method can be trained. The latest deep learning method does not need the cumbersome step of manually extracting features, but trains an end-to-end network. The design of network structure and loss function often determines the upper limit of learning on specific data. However, due to various imaging noises on the OCT image, the boundary is relatively fuzzy, and there are various tissue artifacts such as blood vessel artifacts, which will bring the problems of false positive within the layer and boundary offset between the layers to the neural network. SUMMARY

[0004] To solve the above technical problems, the application provides a fundus OCT image segmentation method based on feature polarization and boundary regression, which uses a double backbone hybrid network based on deep feature polarization and layer boundary regression for OCT image retinal layer segmentation, effectively reducing the problems of false positive and boundary offset.

[0005] The technical scheme adopted by the application is as follows: a fundus OCT image segmentation method based on feature polarization and boundary regression, the specific steps are as follows:

[0006] S1, pre-processing the fundus OCT image;

[0007] S2, inputting the fundus OCT image into a convolutional backbone neural network to extract corresponding features;

[0008] S3, inputting the fundus OCT image into a deformator backbone network to extract corresponding features;

[0009] S4, fusing the features extracted in steps S2 and S3, and inputting the fused features into a decoder to obtain an output probability map;

[0010] S5, feature polarization constraint is performed on the features extracted in steps S2, S3 and S4, and feature polarization loss is calculated;

[0011] S6, boundary regression constraint is performed on the output probability map extracted in step S4, and boundary regression loss is calculated;

[0012] S7, the loss function is used to train the OCT segmentation model based on back propagation;

[0013] S8, based on the trained weight, a new fundus OCT image is predicted to obtain the segmentation result of the retinal layer in the fundus OCT image.

[0014] Further, the step S1 is specifically as follows:

[0015] The blank part not containing the retinal layer in the background is removed, the center of the fundus OCT image is cropped to only retain the part containing the nerve fiber layer, and the region of interest is obtained, then the matrix of the cropped image is divided by 255 to obtain a normalized image, and the pixel intensity distribution of the image is between 0 and 1.

[0016] Further, the step S2 is specifically as follows:

[0017] The normalized image obtained in step S1 is input into a convolution backbone neural network ResNeCt, which includes five stages of convolution network submodules Cross-Block that make the feature scale halved successively, wherein the input layer channel number is set to 3, the output layer channel number is set to the number of categories corresponding to the data set, and the other convolution layer channel numbers are all set to 32. In each Cross-Block, the input feature E is added to the GeLU layer, the 3*3 convolution layer, the LeakyReLU layer and the BatchNorm layer to obtain the output feature of the Cross-Block, after the features obtained in each stage are pooled, they are sent to the next stage, and five levels of convolution neural network features are obtained in total. cnn}.

[0018] Further, the step S3 is specifically as follows:

[0019] The normalized image obtained in step S1 is input to the deformer backbone network LHTran, which consists of Conv-Stem and four-stage deformer network submodules MP-TransBlock that successively halve the feature scale. The MP-TransBlocks are connected by a slice encoding module, and the embedding dimensions of the four MP-TransBlocks are set to [64, 96, 128, 160]. In the MP-TransBlock, the input feature V is first processed through a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer to obtain V1. It is then processed through a TransBlock and a residual connection to obtain V2. V1 and V2 are then superimposed and input to a 1×1 convolutional layer to obtain the output feature. Within the TransBlock, the input feature V is first added through a LayerNorm layer, a pooling layer, a Dropout layer, and a residual connection to obtain V3. V3 is then added through a LayerNorm layer, a Linear layer, a Dropout layer, and a residual connection to obtain the output feature. The features are re-fused after passing through the position encoding and transformer modules of each stage and then input into the next level, resulting in a total of five levels of transformer features {F vit}.

[0020] Furthermore, the step S4 is specifically as follows:

[0021] From deep to shallow, the convolutional neural network features {F cnn} and deformer features {F vit} to fuse and obtain the encoder features {F E}, features of the same level are directly added, and the deeper features of the adjacent level features are bilinearly interpolated and added to the shallower features, iterated in sequence and input to the decoder. The decoder contains a five-stage square convolution module Square-Block. In Square-Block, the input feature F is obtained through a 3*3 convolution layer, a BatchNorm layer, and a LeakyReLU layer to obtain F1. F+F1 is input to another 3*3 convolution layer to obtain the output decoder feature F D , F D After the 1*1 convolution layer and the Softmax layer, the output probability map P is obtained.

[0022] Furthermore, the step S5 is specifically as follows:

[0023] Initialize the prototype unit vector group {t} that is uniformly exclusive of each other, and calibrate the encoder features {F E} and decoder features {F D}Single-stage encoder features F in the set E and the single-stage decoder feature F DGrouping and sampling according to the confidence ranking, obtaining the loss L between the sampled features and the prototype vector FPL , for repelling features of different classes to enhance the segmentation robustness of the model; the feature polarization loss is obtained in three steps.

[0024] In the first step, a set of prototype vectors {t} of C = 32 classes and length 32 is generated, and the prototype vector set is subjected to a feature construction constraint , that is, it becomes a vector with equal distances.

[0025] where t i ,t j ∈{t}, and i≠j, T represents the transpose of the matrix.

[0026] In the second step, F D and F E are first added to obtain Fs, that is, Fs is mapped to the shape of (B*H*W)*L, and the output probability map P is divided into C = 32 groups according to the pixel intensity, and the features corresponding to the pixel positions in each group are taken out as a feature group {v}.

[0027] where N = 3 represents the number of feature levels to be fused, ||*|| c represents the Euclidean norm of the feature in the channel dimension, B represents the training batch, H and W represent the image height and width respectively, and L represents the number of feature channels.

[0028] In the third step, the consistency loss L is calculated for the 32 feature groups and the corresponding 32 prototype vectors

[0029] where ||*||2 represents the mean square error of the matrix, t i ∈{t}, and v i ∈{v}.

[0030] Further, the step S6 is specifically as follows:

[0031] Based on the differentiable soft-argmax algorithm, the coordinates of the retinal layer boundaries in each layer of the output probability map P are calculated, and the difference between the predicted coordinates and the true coordinates is obtained, denoted as the boundary regression loss L BRL .

[0032] The boundary regression loss includes boundary coordinate regression loss and boundary gradient regression loss. On the one hand, first, the longitudinal weighting coefficient

[0033] where η represents noise disturbance, h ∈ [1, H], H represents the height of the output probability map P, and P i,h represents the data of the output probability map P in the i-th channel and the h-th row.

[0034] Then the edge prediction value is calculated The boundary coordinate regression loss is obtained according to the gold standard G

[0035] Wherein, k∈[1,K], K represents the number of fundus OCT image boundaries, represents the predicted kth OCT layer edge position, G k represents the real kth OCT layer edge position.

[0036] On the other hand, define▽ v is the vertical image gradient, abs is the absolute value function, σ is the Sigmoid function, CC is the two convolution layers, and the boundary gradient regression loss L brl-grad =||σ(CC(P))-abs(▽ v (G))||2.

[0037] Finally, the total boundary regression loss is L BRL = L brl-pos + L brl-grad .

[0038] Further, the step S7 is specifically as follows:

[0039] The OCT segmentation model includes the convolution backbone neural network in step S2, the deformater backbone network in step S3, and the decoder in step S4.

[0040] The segmentation loss L SEG is obtained, and the total loss L SUM = L SEG + L FPL +0.1*L BRL is obtained by summing the three loss functions, gradient back propagation is performed, the learnable parameters of the two backbone networks and the decoder are adjusted, and the training is stopped after 100 rounds;

[0041] Wherein, the segmentation loss adopts the form of Dice function, that is, ∩ represents the intersection of two sets.

[0042] Further, the step S8 is specifically as follows:

[0043] On the specified OCT data set, steps S1 to S7 are performed to forward inference, loss calculation and back propagation on the OCT segmentation model, the corresponding model weight is obtained after the training is completed, the normalized fundus OCT image is input into the trained ResNeCt and LHTran respectively, the features extracted by the two are fused and input into the decoder to obtain the segmentation probability map, and the threshold method is used to obtain the binary segmentation result.

[0044] The beneficial effects of the present invention are as follows: the method of the present invention first preprocesses the fundus OCT image, inputs the fundus OCT image into a convolutional backbone neural network and a deformer backbone network respectively to extract corresponding features, and fuses the extracted features, inputs the decoder to obtain an output probability map, then performs feature polarization constraints on all extracted features, and then performs boundary regression constraints on the output probability map, uses a loss function to train the OCT segmentation model based on backpropagation, and finally predicts a new fundus OCT image based on the trained weights. The method of the present invention designs a hybrid of a convolutional network and a deformer network as the encoder segmentation model, which can capture the shape and details of the retinal layer in OCT, proposes a feature polarization loss to enhance the robustness of the encoder feature extraction, and proposes a boundary regression loss to optimize the decoder's precise positioning of the retinal layer boundary, which is very suitable for retinal layer segmentation of fundus OCT images. BRIEF DESCRIPTION OF THE DRAWINGS

[0045] Figure 1 This is a flow chart of a fundus OCT image segmentation method based on feature polarization and boundary regression of the present invention.

[0046] Figure 2 This is a structural diagram of the OCT segmentation model in an embodiment of the present invention.

[0047] Figure 3 Flowchart of the characteristic polarization method in an embodiment of the present invention.

[0048] Figure 4 This is a diagram showing the segmentation results on the OCT image in an embodiment of the present invention.

[0049] Figure 5 This figure shows the segmentation results for challenging areas such as blur and effusion in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0051] like Figure 1 As shown in FIG, a flow chart of a fundus OCT image segmentation method based on feature polarization and boundary regression of the present invention is shown, and the specific steps are as follows:

[0052] S1. Preprocess fundus OCT images;

[0053] S2, input the fundus OCT image into the convolutional backbone neural network to extract the corresponding features;

[0054] S3, input the fundus OCT image into the deformer backbone network to extract the corresponding features;

[0055] S4, fusing the features extracted in step S2 and step S3, and inputting them into the decoder to obtain an output probability map;

[0056] S5, performing feature polarization constraint on the features extracted in steps S2, S3 and S4 to obtain feature polarization loss;

[0057] S6, performing boundary regression constraint on the output probability map extracted in step S4 to obtain boundary regression loss;

[0058] S7, training the OCT segmentation model based on back propagation of the loss function;

[0059] S8, predicting a new fundus OCT image based on the trained weight to obtain the segmentation result of the retinal layer in the fundus OCT image.

[0060] In the embodiment, the step S1 is specifically as follows:

[0061] Two existing open-source OCT image segmentation datasets are downloaded for training and testing the model proposed in the application:

[0062] Duke DME (download address is https: / / people.duke.edu / ~sf59 / Chiu_BOE_2014_ dataset.htm ), and the number of segmentation categories is 8. GOALS (download address is https: / / aistudio.baidu.com / aistudio / competition / detail / 230 ), and the number of segmentation categories is 4.

[0063] Each dataset is divided into a training set and a test set in a 1:1 ratio, and the fundus OCT images in the training set are randomly cropped to a size of 256*256 for training the model proposed in the application. The images in the test set are not cropped and are used to test the model proposed in the application.

[0064] The blank part not containing the retinal layer in the background is removed, and the fundus OCT image is center cropped to only retain the part containing the nerve fiber layer to obtain the region of interest. The cropped region of interest is a fixed size rectangle. In the embodiment, when the GOALS dataset is used, the size of the cropped image is 512*608, and when the Duke DME dataset is used, the size of the cropped image is 224*500. Then, the matrix of the cropped image is divided by 255 to obtain a normalized image, and the pixel intensity distribution of the image is between 0 and 1.

[0065] In the embodiment, the step S2 is specifically as follows:

[0066] As Figure 2As shown, the normalized image obtained in step S1 is input into the convolutional backbone neural network ResNeCt, which contains five stages of convolutional network submodules Cross-Block that successively reduce the feature scale by half. The number of input layer channels is set to 3, the number of output layer channels is set to the number of categories corresponding to the data set, and the number of other convolutional layer channels is set to 32. In each Cross-Block, the input feature E is passed through two layers of 3*3 square convolutional layers, LeakyReLU layers, and BatchNorm layers to obtain E1. On the other hand, it is passed through 1*k and k*1 cross convolutional layers, LeakyReLU layers, and BatchNorm layers to obtain E2. E1 and E2 are added and input into the GeLU layer, 3*3 convolutional layer, LeakyReLU layer, and BatchNorm layer to obtain the output features of the Cross-Block. The features obtained in each stage are pooled and sent to the next stage, resulting in a total of five levels of convolutional neural network features {F cnn}.

[0067] In this embodiment, the square convolution uses 3*3 convolution, the cross convolution uses two layers of 1*k and k*1 convolution, and the number of channels of all convolution layers except the input and output layers is set to 32, k∈[3,5,7,9,11].

[0068] In this embodiment, step S3 is specifically as follows:

[0069] like Figure 2 As shown in Figure 1, the normalized image obtained in step S1 is input into the deformer backbone network LHTran, which consists of Conv-Stem and four-stage deformer network submodules MP-TransBlock that successively halve the feature scale. The MP-TransBlocks are connected by a slice encoding module, and the embedding dimensions of the four MP-TransBlocks are set to [64, 96, 128, 160]. In the MP-TransBlock, the input feature V is first processed through a 1×1 convolutional layer, a 3×3 convolutional layer, and a 1×1 convolutional layer to obtain V1. It is then processed through a TransBlock and a residual connection to obtain V2. V1 and V2 are then superimposed and input to a 1×1 convolutional layer to obtain the output feature. Within the TransBlock, the input feature V is first added through a LayerNorm layer, a pooling layer, a Dropout layer, and a residual connection to obtain V3. V3 is then added through a LayerNorm layer, a Linear layer, a Dropout layer, and a residual connection to obtain the output feature. The features are re-fused after passing through the position encoding and transformer modules of each stage and then input into the next level, resulting in a total of five levels of transformer features {F vit}.

[0070] In this embodiment, step S4 is specifically as follows:

[0071] From deep to shallow, the convolutional neural network features {F cnn} and deformer features {F vit} to fuse and obtain the encoder features {F E}, same-level feature F i cnn and F i vit Direct addition, the deeper feature F in the adjacent level feature i E After bilinear interpolation and shallower features The shallowest features are mapped to the same number of layers as the number of categories. The decoder contains a five-stage square convolution module Square-Block. In Square-Block, the input feature F is passed through a 3*3 convolution layer, a BatchNorm layer, and a LeakyReLU layer to obtain F1. F+F1 is then input to another 3*3 convolution layer to obtain the output decoder feature F. D , F D After the 1*1 convolution layer and the Softmax layer, the output probability map P is obtained. For the Duke DME dataset, the number of channels in the output layer and the output probability map P are both 8. For the GOALS dataset, the number of channels in the output layer and the output probability map P are both 4.

[0072] In this embodiment, step S5 is specifically as follows:

[0073] like Figure 3 As shown, initialize the prototype unit vector group {t} that are uniformly exclusive of each other, and calibrate the encoder features {F E} and decoder features {F D}Single-stage encoder features F in the set E and the single-stage decoder feature F D Group and sample according to the confidence order, and calculate the loss L between the sampled features and the prototype vector FPL , which is used to exclude features of different categories to enhance the segmentation robustness of the model.

[0074] In this embodiment, the characteristic polarization loss is obtained in three steps:

[0075] In this embodiment, the first step is to generate a prototype vector group {t} with a category number C=32 and a length of 32. The prototype vector group is constrained by the feature construction. That is, they become vectors with equal distance.

[0076] Among them, t i ,t j∈ {t}, and i≠j, T denotes the transpose of a matrix.

[0077] Second step, F D and F E Firstly, Fs is obtained by adding up F and F Fs is mapped to the shape of (B*H*W)*L, and the output probability map P is equally divided into C=32 groups according to the pixel intensity, such as the pixel intensity interval of the first group is The features corresponding to the pixel positions of each group are taken out as a feature group {v}.

[0078] wherein N=3 represents the number of fused feature levels, ||*||2 c represents the Euclidean norm of the feature in the channel dimension, B represents the training batch, H and W represent the image height and width respectively, and L represents the number of feature channels.

[0079] Third step, the consistency loss is calculated for 32 feature groups and 32 corresponding prototype vectors

[0080] wherein ||*||2 represents the mean square error of a matrix, t i ∈ {t}, v i ∈ {v}.

[0081] In the embodiment, the step S6 is specifically as follows:

[0082] Based on the differentiable soft-argmax algorithm, the coordinates of the retinal layer boundaries in each layer of the output probability map P are calculated, and the difference between the predicted coordinates and the true coordinates is obtained, denoted as boundary regression loss L BRL , which is used to constrain the retinal boundaries of each layer and improve the segmentation accuracy.

[0083] In the embodiment, the boundary regression loss includes boundary coordinate regression loss and boundary gradient regression loss. On the one hand, the longitudinal weighting coefficient

[0084] wherein η represents noise disturbance, h∈[1,H], H represents the height of the output probability map P, and P i,h represents the data of the output probability map P in the i-th channel and the h-th row.

[0085] Then, the edge prediction value is calculated According to the gold standard G, the boundary coordinate regression loss is obtained

[0086] wherein k∈[1,K], K represents the number of fundus OCT image boundaries, represents the predicted edge position of the k-th OCT layer, and G krepresents the true edge position of the kth OCT layer.

[0087] On the other hand, the definition v is the vertical image gradient, abs is the absolute value function, σ is the Sigmoid function, CC is the two-layer convolution layer, and the boundary gradient regression loss L is calculated. brl-grad =||σ(CC(P))-abs(▽ v (G))||2.

[0088] Finally, the total boundary regression loss is L BRL =L brl-po s+L brl-grad .

[0089] In this embodiment, step S7 is specifically as follows:

[0090] The OCT segmentation model includes a convolutional backbone neural network in step S2, a deformer backbone network in step S3, and a decoder in step S4.

[0091] Calculate the segmentation loss L SEG , sum up the three loss functions to get the total loss L SUM =L SEG +L FPL +0.1*L BRL , perform gradient backpropagation, adjust the learnable parameters of the two backbone networks and the decoder, and stop training after 100 rounds;

[0092] Among them, the segmentation loss adopts the form of Dice function, that is, ∩ represents the intersection of two sets.

[0093] In this embodiment, step S8 is specifically as follows:

[0094] On the GOALS and Duke DME datasets, steps S1 to S7 were performed on the OCT segmentation model to perform forward inference, loss calculation, and backpropagation. After training, the corresponding model weights were obtained. Normalized fundus OCT images were input to the trained ResNeCt and LHTran datasets, respectively. The features extracted from the two were fused and input to the decoder to obtain a segmentation probability map. A threshold method was used to obtain a binary segmentation result. In this example, a threshold of 0.5 was used to obtain the binary segmentation result.

[0095] The test results are as follows Figure 4 and Figure 5The left-to-right columns in the figure are, respectively, the original fundus OCT image (Source), the expert annotation result (Label), the result of the method of the present application (Ours), and the results of other methods (MedT and TransUNet based on a transformer structure, MGU and ReLayNet based on a CNN structure, and the classic medical image segmentation models UNet and UNet++). Figure 4 The figure is a comparison of the overall segmentation results of the method of the present application on fundus OCT images. As can be seen from the figure, the method of the present application can obtain more uniform and continuous retinal layers when performing layer segmentation on fundus OCT images, while other networks, especially TransUNet, MGU and ReLayNet, are prone to large-area false segmentation. Figure 5 The above various algorithms are shown in the segmentation of challenging cases such as more noise and the influence of artifacts. The false positive area generated by the method of the present application is significantly less than that of other methods, and fewer errors are generated in the boundary area. Whether from the comparison of the overall segmentation results of the above Figure 4 The comparison of the overall segmentation results of the above Figure 5 The segmentation performance of the method of the present application is significantly better than that of other methods, whether from the comparison of the overall segmentation results of the above

[0096] In summary, the method of the present application uses the idea of combining convolutional neural networks and transformer backbone networks in network structure design. It not only adopts the advantage of convolutional neural networks in handling local details, but also incorporates the feature of transformer networks in capturing long-distance dependencies. The combination of the two structures enables the network to control the overall shape of the OCT nerve fiber and handle local details well. The feature polarization loss designed in the method of the present application can define the goal of high-dimensional feature learning for each class in advance, uniformly maximize the feature distance of each class during the training phase, enhance the robustness of feature extraction of the encoder, and reduce the classification pressure of the classifier. The boundary regression loss designed in the method of the present application not only constrains the boundary topology of the OCT segmentation target from the perspective of coordinate regression, but also locates the spatial position of the segmentation from the perspective of spatial gradient, which can well solve the boundary offset problem.

[0097] Those skilled in the art will realize that the embodiments described herein are for the purpose of helping the reader understand the principles of the present application and should be understood as not limiting the scope of protection of the present application to such specific statements and embodiments. Those skilled in the art can make various other specific modifications and combinations of the present application without departing from the spirit of the present application, and these modifications and combinations are still within the scope of protection of the present application.

Claims

1. A fundus OCT image segmentation method based on feature polarization and boundary regression, the specific steps being as follows: S1, preprocessing the fundus OCT image; S2, inputting the fundus OCT image into a convolution backbone neural network to extract corresponding features; S3, inputting the fundus OCT image into a transformer backbone network to extract corresponding features; S4, fusing the features extracted in steps S2 and S3, and inputting the fused features into a decoder to obtain an output probability map; From deep to shallow, the convolutional neural network features {F cnn} and the transformer features {F vit} obtained in steps S2 and S3 are fused to obtain the encoder features {F E}, the features of the same level are directly added, the features of the deeper level among the adjacent level features are added to the features of the shallower level after being bilinearly interpolated, and the iteration is sequentially performed and input to the decoder; the decoder includes five-stage square convolution modules Square-Block; in the Square-Block, the input features F are subjected to a 3*3 convolution layer, a BatchNorm layer and a LeakyReLU layer to obtain F1, F+F1 is input to another 3*3 convolution layer to obtain the output decoder features F D , and the F D is subjected to a 1*1 convolution layer and a Softmax layer to obtain the output probability map P; S5, performing feature polarization constraint on the features extracted in steps S2, S3 and S4, and calculating a feature polarization loss; Initialize a set of prototype unit vectors {t} that are uniformly repelling each other, compute the loss L E between the sampled features and prototype vectors based on the output probability map P and the ground truth G D ; Group and sample the one-stage encoder features {F E} and one-stage decoder features {F D} according to the confidence ranking; Compute the loss L FPL between the sampled features and prototype vectors. S6, performing boundary regression constraint on the output probability map extracted in step S4, and calculating a boundary regression loss; Based on the differentiable soft-argmax algorithm, the predicted coordinates of the retinal layer boundaries in each layer of the output probability map P are calculated, and the difference between the predicted coordinates and the true coordinates is obtained, denoted as the boundary regression loss L BRL ; The boundary regression loss includes two parts: boundary coordinate regression loss and boundary gradient regression loss. On the one hand, first, the vertical weighting coefficient is calculated for the output probability map P wherein η represents a noise disturbance, h e [1, H], H represents a height of the output probability map P, P i,h represents data of the output probability map P at the i-th channel and the h-th row; The edge prediction value is then calculated The boundary coordinate regression loss is obtained according to the gold standard G wherein k ∈ [1, K], K represents the number of boundaries of the fundus OCT image, represents the predicted kth OCT layer edge position, G k represents the true kth OCT layer edge position; On the other hand, define ▽ v is the vertical image gradient, abs is the absolute value function, σ is the Sigmoid function, CC is two convolutional layers, and the boundary gradient regression loss L brl-grad = ||σ(CC(P))-abs(▽ v (G))||2; Finally, the total boundary regression loss, i.e. L BRL = L brl-pos + L brl-grad ; S7, training an OCT segmentation model based on back propagation of a loss function; S8, predicting a new fundus OCT image based on the trained weight, and obtaining a segmentation result of a retinal layer in the fundus OCT image.

2. The method of claim 1, wherein the method is characterized by, The step S1 is specifically as follows: The blank part not containing the retinal layer in the background is removed, the fundus OCT image is center cropped to only retain the part containing the nerve fiber layer, an area of interest is obtained, then the matrix of the cropped image is divided by 255 to obtain a normalized image, and the pixel intensity of the image is distributed between 0 and 1. 3.The fundus OCT image segmentation method based on feature polarization and boundary regression according to claim 1, characterized in that, The step S2 is specifically as follows: The normalized image obtained in step S1 is input into a convolutional backbone neural network ResNeCt, which includes five stages of convolutional network submodules Cross-Block that make the feature scale halve successively, wherein the input layer channel number is set to 3, the output layer channel number is set to the number of categories corresponding to the data set, and the channel numbers of other convolutional layers are all set to 32. In each Cross-Block, the input feature E is input into a GeLU layer, a 3*3 convolutional layer, a LeakyReLU layer and a BatchNorm layer to obtain the output feature of the Cross-Block, on one hand, the input feature E is input into two layers of 3*3 square convolutional layers, LeakyReLU layers and BatchNorm layers to obtain E1, on the other hand, the input feature E is input into a cross convolutional layer of 1*k and k*1, a LeakyReLU layer and a BatchNorm layer to obtain E2, and E1 and E2 are added to be input into a GeLU layer, a 3*3 convolutional layer, a LeakyReLU layer and a BatchNorm layer to obtain the output feature of the Cross-Block. The features obtained in each stage are sent to the next stage after being pooled, and five levels of convolutional neural network features {F cnn} are obtained in total.

4. The method of claim 1, wherein the method is characterized by, The step S3 is specifically as follows: The normalized image obtained in step S1 is input into a transformer backbone network LHTran, the network includes a Conv-Stem and four-stage transformer network submodules MP-TransBlock that make the feature scale halve successively, the MP-TransBlock are connected by slice encoding modules, and the embedding dimensions of the four MP-TransBlock are set as [64, 96, 128, 160]; in the MP-TransBlock, the input feature V is input into a 1*1 convolution layer, a 3*3 convolution layer and a 1*1 convolution layer to obtain V1, and is input into a TransBlock and a residual connection to obtain V2, V1 and V2 are superimposed and input into a 1*1 convolution layer to obtain an output feature; inside the TransBlock, the input feature V is input into a LayerNorm layer, a pooling layer, a Dropout layer and a residual connection to obtain V3, and V3 is input into a LayerNorm layer, a Linear layer, a Dropout layer and a residual connection to obtain an output feature; the features After being re-fused after each stage of positional encoding and transformer module, it is then input to the next level, and a total of five levels of transformer features {F vit} are obtained.

5. The method of claim 1, wherein the method is characterized by, The step S5 is specifically as follows: Initialize prototype unit vector set {t} that repel each other uniformly, encode the encoder feature {F E} and decoder feature {F D} set according to the output probability map P and the gold standard G E , and the single-stage decoder feature F D Group and sample according to the confidence ranking, and obtain the loss L FPL between the sampled feature and the prototype vector, which is used to repel features of different categories to enhance the segmentation robustness of the model; the feature polarization loss is obtained in three steps; First, generate a prototype vector set {t} of C = 32 categories and length 32, and the prototype vector set is subjected to feature construction constraints That is, it becomes a vector with equal distances; where t i ,t j ∈{t}, and i≠j, T denotes the transpose of a matrix. Second step, F D and F E Fs are first added to get Fs, i.e. Fs are mapped to the shape of (B*H*W)*L, and the output probability map P is equally divided into C=32 groups according to the pixel intensity, and the features corresponding to the pixel positions of each group are taken out as a feature group {v}. where N = 3 represents the number of fused feature levels, ||*| c represents the Euclidean norm of the features in the channel dimension, B represents the training batch, H and W represent the image height and width, respectively, and L represents the number of feature channels; In a third step, a consistency loss is computed for the 32 feature groups and the corresponding 32 prototype vectors where ||*||2 denotes the mean square difference for matrices, t i ∈{t},v i ∈{v}. 6.The fundus OCT image segmentation method based on feature polarization and boundary regression according to claim 1, characterized in that, The step S7 is specifically as follows: The OCT segmentation model includes the convolution backbone neural network in step S2, the transformer backbone network in step S3 and the decoder in step S4; The segmentation loss L is obtained SEG The three loss functions are summed to obtain the total loss L SUM = L SEG + L FPL + 0.1 * L BRL Gradient back propagation is performed, the learnable parameters of the two backbone networks and the decoder are adjusted, and the training is stopped after 100 rounds. Wherein, the segmentation loss adopts the form of Dice function, that is ∩ represents the intersection of two sets.

7. The method of claim 1, wherein the method is based on feature polarization and boundary regression. The step S8 is specifically as follows: On a specified OCT dataset, steps S1 to S7 are performed to forward infer the OCT segmentation model, calculate a loss and perform back propagation, the corresponding model weight is obtained after the training is completed, the normalized fundus OCT image is input into the trained ResNeCt and LHTran respectively, the features extracted by the two are fused and input into the decoder to obtain a segmentation probability map, and a threshold method is used to obtain a binary segmentation result.

Images