Boundary enhancement based segmentation method of choroid layer structure of OCT image and storage medium

Abstract

The application discloses an OCT image choroid layer structure segmentation method based on boundary enhancement and a storage medium, and the OCT image choroid layer structure segmentation method comprises the following steps: acquiring edge key points of an input image and generating a boundary enhancement soft map; constructing a boundary enhancement module, enhancing and extracting boundary information of a choroid layer in the image from three angles, namely, space, feature and channel; embedding the boundary enhancement module into a segmentation network of an encoding-decoding structure to obtain a predicted segmentation object with boundary information keeping characteristics; and processing the boundary enhancement soft map and the predicted segmentation object by using a boundary perception loss module to obtain a segmentation result of the choroid layer. The application generates a choroid layer boundary enhancement soft map, constructs a boundary enhancement module from multiple angles, and further optimizes the segmentation result by using a boundary perception loss module, so that an automatic choroid layer structure segmentation method is realized, manual intervention is not needed, the layer boundary blur problem is solved, and the segmentation effect is good.

Description

TECHNICAL FIELD

[0001] The application relates to a choroid layer structure segmentation method based on boundary enhancement of an OCT image and a storage medium, and belongs to the field of image processing, and can be used for assisting choroid-related disease analysis. BACKGROUND

[0002] Clinical medical research shows that the pathological process of high myopia will cause morphological changes of the choroid. The mechanical traction caused by the axial growth of high myopia leads to thinning of the retina and choroid, and causes various pathological changes in the fundus, and finally evolves into the most serious type of high myopia, i.e. pathological myopia. In recent years, China has also begun to carry out large-scale prevention and intervention on children's myopia or high myopia. However, the pathological mechanism of myopia is very complex, and there is no clear conclusion at present.

[0003] At present, the common means for diagnosis and treatment of high myopia is to analyze the changes in choroidal thickness. Usually, the optical coherence tomography (OCT) technology is used as a carrier to detect and analyze the choroidal thickness in the macular and optic cup disc regions. However, the overall thickness index of the choroid is difficult to fully reflect the characteristic changes of the choroid structure. With the development of optical coherence tomography technology, clinical experts have observed more detailed changes in the morphological structure of the choroid of high myopia patients, such as the thickness changes of the choroid layer structure. Precise extraction of the choroid structure in the OCT image for quantitative analysis is helpful for monitoring the occurrence of pathological changes in the fundus of high myopia patients. Therefore, the analysis of the morphological structure changes of the choroid based on the OCT image can be used as an important basis for the screening, diagnosis, course evaluation and prediction of high myopia.

[0004] At present, the quantitative analysis of the morphological structure of the choroid largely depends on the manual analysis of clinical experts on two-dimensional images. Since the manual analysis of clinical experts is time-consuming and laborious, there is an urgent need for automatic segmentation and analysis algorithms for the choroid. However, there are relatively few automatic segmentation and analysis methods for the choroid layer at present. The application for invention with publication number CN 112712520 A proposes a choroid layer segmentation method based on ARU-Net to segment the choroid layer and obtain the choroid layer boundary. The application for invention with publication number CN 110517235 A proposes an OCT image choroid automatic segmentation method based on GCS-Net. Although the above methods belong to the automatic segmentation methods for the choroid of the OCT image, there is still much room for improvement. The choroid layer structure segmentation task faces severe challenges: firstly, there are blood vessel artifacts and light attenuation effects on the OCT image, and the contrast of the choroid and sclera boundary is relatively low, making it difficult to accurately detect the choroid contour; secondly, the boundary information accounts for a small proportion of the global information of the layer structure and is easily ignored by the segmentation network. Therefore, it is urgent to find a choroid layer segmentation method based on boundary enhancement of the OCT image. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides an OCT image choroid layer segmentation method based on boundary enhancement and a storage medium, which is an important basis for high myopia screening, diagnosis, and disease course evaluation and prediction.

[0006] The technical solution of the present application to achieve the above-mentioned purposes is as follows:

[0007] (1) For the OCT image to be segmented, the expert annotation information is used to extract the choroid layer boundary, and the boundary enhancement points are obtained by using the enhancement point search method, and a boundary enhancement soft map is generated;

[0008] (2) A boundary enhancement module is constructed to enhance and extract the boundary information of the choroid layer in the image from three angles: space, feature, and channel;

[0009] (3) The boundary enhancement module is embedded into the segmentation network of the coding-decoding structure to obtain a predicted segmentation object with boundary information retention features;

[0010] (4) The boundary perception loss module is used to obtain the segmentation result of the choroid layer by using the boundary enhancement soft map and the predicted segmentation object.

[0011] Preferably, in step (1), the boundary enhancement points are extracted and the boundary enhancement soft map is generated by the following method:

[0012] (1-1) Randomly select points on the choroid layer boundary, and calculate the overlap degree of the area formed by connecting each point to form a connected region with the choroid layer, and after n cycles, the boundary enhancement point with the highest overlap degree is retained.

[0013] (1-2) Take the extracted boundary enhancement point as the center, set the gray value of the center region to 1, and set the outermost position to 0, the gray value in the region is Gaussian distributed from the center to the boundary, and the boundary enhancement soft map is generated.

[0014] N is an integer, the value of which is set according to the size of the OCT image, and as a preferred, the value of n is 30000-40000.

[0015] Preferably, the boundary enhancement module in step (2) is designed by the following steps to obtain the ability to extract and enhance boundary information:

[0016] (2-1) Use different dilution rates of the empty convolution to extract choroid boundary information of different receptive field sizes for the feature map input into the module;

[0017] (2-2) Use a special activation function to activate the choroid layer boundary information of the image to strengthen the choroid boundary information;

[0018] (2-3) using a global average pooling operation, extracting a channel weighted vector of the feature map, and enhancing choroidal boundary information in different channels.

[0019] Further, the dilation rate sizes used in the step (2-1) are 1, 2, 4, and 6, respectively, and the final extracted choroidal boundary information which is expressed by the following formula:

[0020]

[0021] wherein represents a dilated convolution function with a dilation rate of r and a convolution kernel size of s x s, f i represents the input feature map of the i-th convolution block.

[0022] Further, the activation function in the step (2-2) adopts the following activation function:

[0023]

[0024] wherein f i represents the input feature map of the t-th convolution block, C2D1 represents a 1 x 1 convolution operation, and S represents a sigmoid function.

[0025] Further, the channel weighting in the step (2-3) adopts a one-dimensional convolution operation with a kernel size of 3.

[0026] Preferably, a U-Net network is used as the encoder-decoder architecture network in the step (3).

[0027] Preferably, the boundary enhancement module is embedded into the segmentation network of the encoder-decoder structure in the step (3) by the following method:

[0028] (3-1) selecting m boundary enhancement modules and embedding them into the encoder part and the decoder part of the U-Net in a certain proportion;

[0029] (3-2) the input image of each boundary enhancement module is the output feature map of the U-Net network in the previous stage.

[0030] m is set according to the size and data amount of the image, and as a preferred, the value of m is 4-6.

[0031] Preferably, the backbone network used in the boundary perception loss module in the step (4) is a VGG-19 network.

[0032] Preferably, the segmentation network adopts a joint loss function containing a segmentation map prediction loss, a boundary enhancement soft map prediction loss, and a boundary perception prediction loss.

[0033] Further, the segmentation map prediction loss takes cross-entropy loss, the boundary enhancement soft map prediction loss takes mean square error loss, and the boundary perception prediction loss function is as follows:

[0034]

[0035] wherein represents the output feature map after the i-th perception loss module, S GT represents the expert-labeled segmentation ground truth image, is a segmentation result image predicted by the segmentation network, M GT is a boundary point image ground truth, N i is a channel weighting vector, and represents a splicing operation.

[0036] Correspondingly, the application also provides a computer readable storage medium, which stores a computer program, and when the computer program is executed by a processor, any OCT image choroid layer structure segmentation method described above is realized.

[0037] The technical scheme of the application realizes a complete automatic choroid layer segmentation method by generating a choroid layer boundary enhancement soft map, constructing a boundary enhancement module from multiple angles, and further optimizing the segmentation result by using a boundary perception loss module, without human intervention, and solves the problem of blurred layer boundaries, and the segmentation effect is good. Compared with the prior art, the application has the advantages that: by screening and analyzing the enhancement points of the blurred boundary, and integrating the information of the enhancement points into the deep learning network framework, the segmentation accuracy is improved; the boundary enhancement module is embedded in the segmentation network, so that the information of the boundary enhancement points can be combined in the process of predicting the segmentation map to obtain a predicted segmentation result with preserved boundary information; through the perception loss module, expert knowledge can be embedded into the fully automatic segmentation model to further optimize the segmentation result. The automatic choroid layer structure segmentation method provided by the application provides an effective tool for quantitatively evaluating choroid morphological structure, and also provides a possibility for early screening of high myopia and other diseases. BRIEF DESCRIPTION OF DRAWINGS

[0038] Figure 1-a 、 1-b , 1-c is an enhanced boundary map and choroid enhancement soft map schematic diagram obtained based on an OCT choroid image according to an embodiment of the application, and the input image, layer boundary image, and boundary enhancement soft map are sequentially shown as follows:

[0039] Figure 2 is a schematic diagram of a choroid layer structure segmentation method framework based on boundary enhancement of an OCT image according to an embodiment of the application;

[0040] Figure 3is a boundary enhancement module structure diagram proposed by an embodiment of the present application;

[0041] Figure 3-a 、 3-b , 3-c are feature extraction branch (FEB), boundary channel enhancement branch (CEB), boundary information activation branch (BAB) structure schematic diagram included by the boundary enhancement module in turn;

[0042] Figure 4 is an experimental result comparison diagram of the OCT image choroid layer structure segmentation method based on boundary enhancement proposed by an embodiment of the present application, from left to right are: input image, gold standard, segmentation result. DETAILED DESCRIPTION

[0043] In order to make the purpose, technical scheme and advantages of the present application more clear, the present application is further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0044] The OCT image shown in Figure 1-a is taken as an example to illustrate the OCT image choroid layer structure segmentation method based on boundary enhancement, and the OCT input image to be segmented is shown in 1-a. The size of the OCT image is 378x379 pixels. Due to the light attenuation effect of the OCT device and the retinal blood vessel artifact, the upper and lower boundary structures of the choroid in the image are blurred.

[0045] The OCT image is segmented by using the OCT image choroid layer segmentation method based on boundary enhancement, and the overall framework of the segmentation method is shown in Figure 2 The segmentation framework mainly includes M1 segmentation network and M2 perception loss network. The M1 segmentation network is an encoding-decoding network embedded with a plurality of boundary enhancement modules (BEM), which is used to predict the segmentation of the input image and obtain a predicted segmentation image with boundary information keeping features. Loss Seg is a loss function of the predicted segmentation. The boundary enhancement module is composed of a boundary feature enhancement branch (FEB), a boundary channel enhancement branch (CEB) and a boundary information activation branch (BAB), which respectively enhance and extract the boundary information of the choroid layer in the OCT image from the spatial, feature and channel dimensions. M10 is a convolution module used by the M1 segmentation network encoder / decoder, which is used to perform convolution operation on the boundary features output by the boundary enhancement module. The M2 perception loss module is used to enhance the ability to identify the boundary. Loss GFThis is the loss function of the perceptual loss network. Before using this segmentation framework for image segmentation, it is necessary to automatically extract boundary enhancement points and generate a boundary enhancement softmap using expert-annotated choroidal boundaries (gold standard) in the OCT image, which is used to optimize the segmentation effect. The input of the M2 perceptual loss module is the boundary enhancement softmap (i.e., Figure 2 (Middle boundary point map) and prediction results (i.e.) Figure 2 The images (predicted segmentation map) and the gold standard image, after undergoing channel concat operations respectively, are processed by a perceptual loss network to output the final segmentation result.

[0046] Specifically, the OCT image choroid layer segmentation method based on boundary enhancement provided in this embodiment includes at least the following steps:

[0047] S1. Using the choroidal boundaries annotated by experts in the OCT image, boundary enhancement points are automatically extracted and a boundary enhancement soft map is generated for optimizing the segmentation effect.

[0048] The choroidal layer boundary is extracted, and boundary enhancement points are obtained using an enhancement point search method, while a boundary enhancement soft map is generated. The enhancement point search method includes: based on a boundary key point selection method, randomly selecting points on the choroidal boundary annotated by experts in the OCT image, calculating the overlap between the area of ​​the connected region formed by connecting these points and the choroidal layer, and after several iterations, retaining the boundary enhancement point with the highest overlap. A Gaussian model is then used to generate the boundary soft map. This process includes at least the following steps:

[0049] S11. Obtain the labeled segmentation map S using the Canny operator edge detection method. GT The target region, i.e., the layer boundary image of the choroid layer, is shown in the image. Figure 1-b As shown; the label segmentation image is the gold standard image annotated by experts.

[0050] S12. Randomly select n points (x, y) on the boundary of the target region to form a point set, and use... Let be the number of iterations, where t represents the number of iterations. Preferably, t takes a value between 30000 and 40000. In this embodiment, t = 40000 and n = 6;

[0051] S13. Will Connecting the points in the diagram forms a boundary region, using... express;

[0052] S14. Select the boundary region in iteration t (t = 40000). With label segmentation map S GT The set of points with the maximum intersection (IOU) is designated as the key point set of the structural boundary, and these key points are represented by the following formula:

[0053]

[0054] Boundary key point set The position of the point in the graph is represented, in order to be more suitable for the convolutional network training of processing the image, a two-dimensional boundary point graph is used to represent the points, and the generated boundary enhanced soft graph is as shown in Figure 1-c .

[0055] Specifically, the boundary point graph is centered on the points, the center region is set to 1, the outermost position is set to 0, and the grayscale value in the region is Gaussian distribution, and the specific formula of the boundary point distribution is as follows:

[0056]

[0057] K is the size of the two-dimensional Gaussian convolution kernel, σ is the standard deviation, and (i, j) is the center of the Gaussian template.

[0058] S2. Construct a boundary enhancement module BEM to enhance and extract the boundary information of the choroid layer in the image from the spatial, feature and channel dimensions. The structure of the boundary enhancement module is as shown in Figure 3 , a feature extraction branch (FEB) is constructed in the spatial enhancement angle, a dilated convolution method is used to extract choroid boundary information of different receptive field sizes; a boundary channel enhancement branch (CEB) is constructed in the channel enhancement angle, a pooling operation is used to extract the channel weighting vector of the feature map, and the choroid boundary information of different channels is enhanced; in the feature enhancement angle, a boundary information activation branch (BAB) is constructed, an activation function is used to activate the boundary information, and the choroid boundary information is strengthened; the input feature (w: image width, h: image height, c: image depth (channel number)) and the output of the three branches are element-wise multiplied, and then the operation result and the input feature f i are added to obtain the structure boundary information maintaining feature as the output feature. Loss map represents the loss function of the boundary point graph.

[0059] The construction process of the boundary enhancement module is as follows:

[0060] S21. Construct a feature extraction branch (FEB) to extract boundary information of different receptive fields in the input feature, and the structure is as shown in Figure 3-a . FEB uses dilated convolution with different expansion rates to obtain features of different receptive fields. After the input feature is convolved, the extracted choroid boundary information generates a boundary point graph The formula is expressed as follows:

[0061]

[0062] wherein, is an input feature map, represents a dilated convolution function with a dilation rate of r and a convolution kernel size of s x s. As shown in Figure 3-a represents a dilated convolution function with a dilation rate of b and a convolution kernel size of a, and represents a paste operation. In this embodiment, and a dilated convolution function with a dilation rate of 1, a convolution kernel of 1, and a dilation size of 1, 2, 4, and 6, respectively, and a convolution kernel of 3 are used, that is, input features After convolution feature extraction, an HW*HW feature map is obtained.

[0063] S22. A boundary channel enhancement branch (CEB) is constructed to obtain a channel weighting vector of the features to strengthen the boundary information based on different channels, and the structure is as shown in Figure 3-b The CEB generates channel statistical information of the input feature map using a global average pool to generate an input feature map , and applies a one-dimensional convolution, a sigmoid function transposition processing, and finally obtains a channel weighting vector Ni with a size of C*HW. In this embodiment, the convolution kernel size is 3, and Ni is represented as:

[0064] N i = S(C1D3(f gap (f i ))

[0065] wherein f gap represents a global average pool, C1D3 represents a one-dimensional convolution with a convolution kernel size of 3, and S represents a sigmoid function, is an input feature map.

[0066] S23. A boundary information activation branch (BAB) is constructed to further optimize and extract boundary information using an activation function for boundary information, and the structure is as shown in Figure 3-c Specifically, the input feature map is down-sampled, the number of channels of the input feature map is changed to 1 using a 1 x 1 convolution operation, and the boundary information is further optimized through an activation function. The activation function in this embodiment is as follows:

[0067]

[0068] wherein f i represents an input feature map, C2D1 represents a 1 x 1 convolution operation, and S represents a sigmoid function. After the input feature is subjected to a boundary information activation operation, an HW*HW size feature map is output.

[0069] S24. The outputs of the three branches described in steps S21, S22, and S23 are combined.​ N i 、Q i and input feature f i element-wise multiplication, and then add the multiplication result and input feature f i to obtain structure boundary information preserving feature as output feature.

[0070]

[0071] wherein represents element-wise addition operation, represents channel-wise multiplication operation.

[0072] Each boundary enhancement module has corresponding output feature.

[0073] S3. Embed the boundary enhancement module into a segmentation network of an encoding-decoding structure, so that the input image is processed by the segmentation network to obtain a predicted segmented object with boundary information preserving feature;

[0074] Preferably, a U-Net network can be used as the segmentation network of the encoding-decoding structure.

[0075] Preferably, one or more boundary enhancement modules can be embedded into the segmentation network of the encoding-decoding structure.

[0076] Specifically, the boundary enhancement module can be embedded into the U-Net network by the following method:

[0077] S31. Select m boundary enhancement modules, and embed them into the encoder part and the decoder part of the U-Net in a certain proportion;

[0078] S32. The input image of each boundary enhancement module is the output feature map of the previous stage, i.e. the feature enhancement map after the output of the previous boundary enhancement module is subjected to convolution operation.

[0079] m is an integer, which is set according to the size and data amount of the image. As a preferred, the value of m is 4-6. For example, Figure 2 As shown in the encoding-decoding structure segmentation network M1, four boundary enhancement modules are embedded into the segmentation network in the embodiment of the present application; the parameters of the convolution module used in the structure encoder / structure decoder in the segmentation network are shown in M10, the structure encoder includes 3x3 convolution, normalization, ReLU activation function for down-sampling; the structure decoder includes up-sampling convolution, feature splicing, and then outputs the enhanced feature through 3x3 convolution layer.

[0080] S4. The boundary enhanced soft map obtained in step S1 is respectively channel spliced with the gold standard and the predicted segmentation object obtained in step S3, and then the result is input into the boundary-aware loss module, and finally the segmentation result of the choroid layer structure of the OCT image is obtained.

[0081] Preferably, the boundary-aware loss module in the present embodiment at least includes:

[0082] S41. The VGG-19 network is used as the backbone network in the boundary-aware loss module;

[0083] S42. The input of the boundary-aware loss module is the image obtained after the splicing operation of the boundary enhanced soft map, the segmentation prediction result and the gold standard respectively;

[0084] S43. The boundary-aware prediction loss function is as follows:

[0085]

[0086] Wherein S represents the output feature map of the i-th after the perception loss module, S GT represents the gold standard image, represents the predicted segmentation result image, M GT is the boundary enhanced point map obtained in step 1, and represents the splicing operation.

[0087] The segmentation method described in the present embodiment adopts a joint loss function including segmentation map prediction loss, boundary enhanced soft map prediction loss and boundary-aware prediction loss as the overall loss function. The segmentation map prediction loss adopts cross-entropy loss, and the boundary enhanced soft map prediction loss adopts mean square error loss. The overall loss function is as follows:

[0088]

[0089] Wherein, Loss Seg is the loss function of the predicted segmentation map, used to reduce the difference between the segmentation map and the predicted segmentation map, and the formula is as follows:

[0090]

[0091] represents the loss function of the predicted boundary point map, used to train the dilated convolutional network of the feature extraction branch (FEB) described in step 21, and i represents the i-th enhanced module. The formula is as follows:

[0092]

[0093] Based on the above-described segmentation method, the automatic segmentation of the choroid layer structure of the OCT image can be realized. The experimental results after the automatic segmentation by the method described in the present embodiment are as follows:Figure 4 As shown, from left to right, A: original drawing, B: gold standard, C: segmentation result.

[0094] The technical scheme provided by the present application improves the segmentation accuracy by screening and analyzing the enhanced points of the fuzzy boundary, generating a boundary enhancement soft map, and integrating the information of the enhanced points into the deep learning network framework; the boundary enhancement module is embedded in the segmentation network to enhance and extract boundary information from three angles of space, features and channels; finally, through the boundary perception loss module, expert knowledge can be embedded into the fully automatic segmentation model to further optimize the segmentation result. The automatic segmentation method of the choroid layer structure provided by the present application provides an effective tool for quantitatively evaluating the choroid sub-morphological structure, and also provides a possibility for early screening of high myopia and other diseases.

[0095] It should be understood that the above embodiments are only to illustrate the technical concept and characteristics of the present application, and its purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and cannot limit the protection scope of the present application. Any equivalent changes or modifications made in accordance with the spirit and essence of the present application shall be covered within the protection scope of the present application.

Claims

1. A method for choroid layer structure segmentation of an OCT image based on boundary enhancement, characterized by, At least comprising the following steps: S1. Using the choroid boundary labeled by experts in the OCT image, automatically extracting boundary enhancement points and generating a boundary enhancement soft map; step S1 automatically extracts boundary enhancement points by the following method: Randomly selecting points on the choroid layer boundary labeled by experts, and calculating the overlap degree of the area formed after connecting each point with the choroid layer, after a certain number of cycles, the point with the highest overlap degree is retained as the boundary enhancement point; The generation of the boundary enhancement soft map includes: Taking the boundary enhancement point as the center, the gray value of the center area is set to 1, the gray value of the outermost position is set to 0, and the gray value in the area is Gaussian distributed from the center to the boundary; S2. Constructing a boundary enhancement module from the perspectives of space, features, and channels respectively to enhance and extract the boundary information of the choroid layer in the OCT image, and step S2 includes: From the perspective of spatial enhancement, a hole convolution method is used to extract choroid boundary information of different receptive field sizes; From the perspective of feature enhancement, an activation function for boundary information is used to activate the boundary information and strengthen the choroid boundary information; From the perspective of channel enhancement, a pooling operation is used to extract the channel weighting vector of the feature map and enhance the choroid boundary information of different channels, the pooling operation is global average pooling, and the channel weighting uses one-dimensional convolution with a kernel size of 3; Element-wise multiplication is performed on the choroid boundary information, the channel weighting vector, the activation function, and the input features, and then the multiplication result and the input features are added to obtain the structural boundary information retention feature; In the hollow convolution method, the extracted choroid boundary information The formula is expressed as shown below: ； wherein represents the expansion rate r , the convolution kernel size s × s of the empty convolution function, represents the input feature map of the th convolution block; The activation function formula is as follows: ; wherein represents the input feature map of the convolutional block, represents a 1x1 convolution operation, and S represents a sigmoid function. The boundary perception prediction loss function is as follows: ; wherein represents the output feature map after the perceptual loss module, is the expert-annotated segmentation ground truth image, is the predicted segmentation map output by the segmentation network, is the boundary enhanced point map, is the channel weighting vector, represents the collage operation; S3. Embedding the boundary enhancement module of step S2 into a segmentation network with an encoding-decoding structure to output a predicted segmented object with boundary information retention features according to the input image features; S4. Constructing a boundary perception loss module to enhance the boundary of the boundary enhancement soft map and the predicted segmented object, and finally outputting the segmentation result of the choroid layer structure.

2. The boundary enhancement-based OCT image choroid layer structure segmentation method according to claim 1, characterized in that, The segmentation network with an encoding-decoding structure of step S3 is a U-Net network including an encoder and a decoder.

3. The boundary enhancement based OCT image choroid layer structure segmentation method according to claim 1, characterized in that, The backbone network used by the boundary perception loss module of step S4 is a VGG-19 network.

4. The boundary enhancement-based OCT image choroid layer structure segmentation method according to claim 1, characterized in that, A joint loss function including a segmentation map prediction loss, a boundary enhancement soft map prediction loss, and a boundary perception prediction loss is used in the segmentation process.

5. The boundary enhancement based OCT image choroid layer structure segmentation method according to claim 4, characterized in that, The segmentation map prediction loss uses a cross-entropy loss.

6. The boundary enhancement-based OCT image choroid layer structure segmentation method according to claim 4, characterized in that, The boundary enhancement soft map prediction loss uses a mean square error loss.

7. A computer readable storage medium characterized in that, The computer readable storage medium stores a computer program, when the computer program is executed by the processor, the boundary enhancement based OCT image choroid layer structure segmentation method of any one of claims 1-6 is realized.

Images