Intelligent retinal vessel segmentation method

Abstract

The application provides a retinal blood vessel intelligent segmentation method, characterized by comprising the following steps: acquiring a given retinal image sample set, preprocessing the retinal image sample set to obtain a pretreatment data set; dividing the pretreatment data into a training set and a test set according to a proportion; obtaining scale feature information corresponding to the retinal image in the training set through a feature extraction network, and further calculating the segmentation difficulty degree corresponding to the scale feature information; taking the retinal image and the segmentation difficulty degree as the network source input, taking the reference blood vessel segmentation image corresponding to the retinal image as the output, combining an adaptive optimization algorithm and a preset loss function, applying the retinal image, training the target neural network, and obtaining a retinal blood vessel segmentation network model; and segmenting the retinal image according to the retinal image and the retinal blood vessel segmentation network model.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image segmentation, in particular to a retinal blood vessel intelligent segmentation method. BACKGROUND

[0002] Accurate retinal blood vessel segmentation plays a crucial role in the early identification of clinical ophthalmic diseases. Researchers at home and abroad have proposed more and more intelligent segmentation methods based on deep learning, and have been committed to improving the performance of blood vessel segmentation. However, the complexity of retinal blood vessel structure and the lack of accurately labeled retinal images all pose great challenges to retinal blood vessel segmentation. Due to the particularity of retinal images, B-ultrasound, nuclear magnetic resonance and CT images do not show natural images as clearly, and the boundaries are very blurred and the contrast is poor. Secondly, there are relatively more noise points in retinal images, and the noise signals presented by different individuals are not stable.

[0003] Therefore, the retinal image widely uses the U-shaped convolutional neural network for image segmentation. Due to the characteristics of multi-scale information acquisition and skip connection, the U-shaped convolutional neural network continuously reduces the resolution in the process of continuous downsampling to obtain image information of different scales. The whole network completes the extraction and combination of "from fine to coarse" features, so that the information obtained by the U-shaped convolutional neural network is more comprehensive. At the same time, detailed information is added in the judgment of the target general region, which makes the U-shaped convolutional neural network obtain more accurate segmentation results. However, the traditional U-shaped convolutional neural network has a slow running speed for image segmentation, so the present application designs a retinal blood vessel intelligent segmentation method. SUMMARY

[0004] Therefore, it is necessary to provide a retinal blood vessel intelligent segmentation method aiming at the above technical problems, characterized by comprising:

[0005] A given retinal image sample set is obtained, and a pretreatment dataset is obtained by pretreating the retinal image sample set. The pretreatment dataset is divided into a training set of a retinal blood vessel segmentation network model and a test set for testing the retinal blood vessel segmentation network model according to a preset proportion. The retinal image sample set includes a retinal image and a reference blood vessel segmentation image.

[0006] Respectively for the retinal images in the training set, the scale feature information corresponding to the retinal images is obtained through a preset feature extraction network, and the respective segmentation difficulty degrees of the scale feature information are further calculated.

[0007] Each retina image and segmentation difficulty level is taken as a network source input, a benchmark blood vessel segmentation image corresponding to the retina image is taken as an output, an adaptive optimization algorithm and a preset loss function are combined, each retina image is applied, the target neural network is trained, and a retina blood vessel segmentation network model is obtained;

[0008] The retina image is segmented according to the retina image and the retina blood vessel segmentation network model.

[0009] In one embodiment, a retina blood vessel intelligent segmentation method further includes that the target neural network is a convolutional neural network model, including an encoding part and a decoding part, and the encoding part and the decoding part are connected by a cross-layer structure; wherein the encoding part includes a height-axis attention layer and a width-axis attention layer.

[0010] In one embodiment, pre-processing is performed on the retina image sample set to obtain a pre-processed data set, including:

[0011] The retina image sample set is subjected to gray scale transformation, standardization, adaptive histogram equalization, gamma correction and clipping to obtain the pre-processed data set.

[0012] In one embodiment, the scale feature information corresponding to the retina image is obtained by a preset feature extraction network for each retina image in the training set, including:

[0013] A random initial value is obtained, and the scale feature information corresponding to the retina image is obtained by the random initial value and the preset feature extraction network.

[0014] In one embodiment, the segmentation difficulty level corresponding to the scale feature information is further calculated, including:

[0015] The similarity between the scale feature information of different random initial values is calculated by a similarity formula;

[0016] The average value of the scale feature information of different random initial values is calculated by the following formula:

[0017]

[0018] wherein, sim(x k ) represents the average value of the scale feature information of different random initial values, sim ij represents the similarity between the scale feature information of different random initial values, x k represents the kth retina image, and z represents z groups of random initial values.

[0019] The weight value for evaluating the difficulty level of the sample is calculated by a dispersion standardization formula;

[0020] The threshold value of the difficulty level of segmenting the retinal image is calculated by the following formula:

[0021]

[0022] Wherein, the threshold value of the difficulty level of segmenting the retinal image is represented by λ, the adjustment parameter is represented by α, which is a constant set by oneself, the preferred value is 0.9, e represents the current training round, m represents the total training round, and F represents the maximum value of the weight value of the evaluation sample difficulty level.

[0023] In one embodiment, the preset loss function is:

[0024]

[0025] Wherein, minE(ω,ν,λ) represents the preset loss function, L BCE (y i ,f(x i ,ω)) represents the binary cross-entropy loss function, n represents the total number of retinal images in the training set, x k represents the kth retinal image, ω represents the weight of the model, ν i represents the measurement value of the difficulty level of segmenting the retinal image, and λ represents the threshold value of the difficulty level of segmenting the retinal image.

[0026] In one embodiment, a retinal blood vessel intelligent segmentation method further comprises:

[0027] The convolutional neural network model is a U-shaped convolutional neural network model;

[0028] The adaptive optimization algorithm is an adam optimization algorithm.

[0029] Compared with the prior art, the advantages and beneficial effects of the present application are that the present application can overcome the above-mentioned shortcomings of the prior art, more intelligently input the retinal image for network learning according to the difficulty level of retinal image segmentation, design a local and global perception based on axis attention segmentation network, and improve the image segmentation quality. It can improve the running speed and overcome the defects of low calculation efficiency, lack of feature extraction and strengthening of the traditional U-shaped network model. Further enhance its ability to capture long-range dependent information, reduce the level of depth while effectively improve the accuracy. BRIEF DESCRIPTION OF DRAWINGS

[0030] Figure 1 It is an embodiment of the implementation process of a retinal blood vessel intelligent segmentation method;

[0031] Figure 2 It is a flowchart of a retinal blood vessel intelligent segmentation method in one embodiment.

[0032] Figure 3 A schematic diagram of a retinal blood vessel segmentation network model structure in one embodiment. DETAILED DESCRIPTION

[0033] Before the specific embodiment of the present application is described, the overall concept of the present application is described as follows:

[0034] The present application is mainly developed for image segmentation process. The current U-shaped convolutional neural network has too much redundancy because a window needs to be taken for each pixel point. The similarity of the windows of two adjacent pixel points is very high, and there is a lot of redundancy, which leads to slow network training. At the same time, classification accuracy and positioning accuracy cannot be achieved at the same time. When the selected perception field is relatively large, the classification accuracy will be reduced.

[0035] Therefore, the present application provides a retinal blood vessel intelligent segmentation method. First, a given retinal image sample set is obtained, and the retinal image sample set is preprocessed to obtain a preprocessed data set. The preprocessed data set is divided into a training set and a test set according to a proportion. The scale feature information corresponding to the retinal image is obtained by a feature extraction network for the retinal image in the training set, and the segmentation difficulty corresponding to the scale feature information is further calculated. The retinal image and the segmentation difficulty are used as the network source input, and the reference blood vessel segmentation image corresponding to the retinal image is used as the output. An adaptive optimization algorithm and a preset loss function are combined, and the retinal image is applied to train the target neural network to obtain a retinal blood vessel segmentation network model. The retinal image is segmented according to the retinal image and the retinal blood vessel segmentation network model.

[0036] After the overall concept of the present application is introduced, in order to make the purpose, technical scheme and advantages of the present application more clear and clear, the present application will be further described in detail through specific embodiments combined with the drawings. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0037] In one embodiment, the implementation process of a specific retinal blood vessel intelligent segmentation method is as shown in Figure 1 The given retinal image sample set is obtained, the retinal image sample set is preprocessed to obtain a preprocessed data set, the target neural network is trained according to the preprocessed data set, the retinal blood vessel segmentation network model is obtained, the retinal image in the obtained retinal image sample set is input into the retinal blood vessel segmentation network model, the retinal image is segmented, and finally the segmentation result is obtained.

[0038] In one embodiment, as shown in Figure 2 A retinal blood vessel intelligent segmentation method is provided, which includes the following steps:

[0039] Step S101, a given retinal image sample set is obtained, the retinal image sample set is preprocessed to obtain a preprocessed data set; the preprocessed data set is divided into a training set of a retinal blood vessel segmentation network model and a test set for testing the retinal blood vessel segmentation network model according to a preset ratio; wherein the retinal image sample set includes a retinal image and a reference blood vessel segmentation image.

[0040] Specifically, a given retinal image sample set is obtained, the retinal image sample set includes a retinal image and a reference blood vessel segmentation image, the reference blood vessel segmentation image is a reference blood vessel segmentation image corresponding to the retinal image, and is used for training and testing of the retinal blood vessel segmentation network model. The retinal image sample set is preprocessed to obtain a preprocessed data set, and the preprocessed data set is divided into a training set of a retinal blood vessel segmentation network model and a test set for testing the retinal blood vessel segmentation network model according to a preset ratio. In actual application, based on the size of the entire preprocessed data set data, the division ratio of the training set data and the test set data can be 6:4, 7:3 or 8:2. For huge data, 9:1 or even 99:1 can be used. The specific test set division method is different.

[0041] In this embodiment, the given retinal image sample set is divided into a training set and a test set, the training set is a data sample used for model fitting, the training error is gradient descended in the training process, and learning is performed. The test set is a sample set left out of the model training process, which can be used to adjust the hyperparameters of the model and to preliminarily evaluate the ability of the model to evaluate the generalization ability of the final model.

[0042] On this basis, an intelligent retinal blood vessel segmentation method further includes:

[0043] The target neural network is a convolutional neural network model, which includes an encoding part and a decoding part, and the encoding part and the decoding part are connected by a cross-layer structure; wherein the encoding part includes a height-axis attention layer and a width-axis attention layer.

[0044] On this basis, the convolutional neural network model is a U-shaped convolutional neural network model, specifically, a retinal blood vessel segmentation network model structure is as follows: Figure 3As shown, the diagram includes an encoding section on the left and a decoding section on the right, connected by a cross-layer structure. The left encoding section comprises a left encoder and a right encoder. The left encoder includes convolutional layers and pooling layers, while the right encoder is an axis attention module. This module includes a height axis attention layer and a width axis attention layer. The height axis attention layer corresponds to the vertical axis of the image, and the width axis attention layer corresponds to the long axis of the image. The right decoding section uses the upsampled feature map to perform a concatenation operation with the feature map on the left.

[0045] In this embodiment, an axis attention mechanism is added to the U-shaped convolutional neural network model. This mechanism focuses on relevant information while ignoring irrelevant information, directly establishing the dependency relationship between input and output without loops. This enhances parallelism and significantly improves running speed. It overcomes some limitations of traditional neural networks, effectively models sequence data of variable length, further enhances its ability to capture long-range dependency information, and effectively improves accuracy while reducing layer depth.

[0046] Based on this, the retinal image sample set is preprocessed to obtain a preprocessed dataset, including:

[0047] The retinal image sample set is subjected to grayscale transformation, standardization, adaptive histogram equalization, gamma correction, and cropping to obtain a preprocessed dataset.

[0048] Specifically, the given retinal image sample set includes retinal images and baseline vessel segmentation images; the retinal image sample set is denoted as C = {(a i ,b i )|i∈[1,R]}, where a i Represents a retinal image, b i Let R represent the baseline blood vessel segmentation image, R represent the number of samples, and i represent the sample index; the entire set of retinal images C is copied and divided into a retinal image training set E = {(a i ,b i )|i∈[1,N]} and retinal image test set F={(a i ,b i), where N+M=R, and M and N represent the number of divided retinal images respectively; preprocessing refers to gray scale transformation, standardization, adaptive histogram equalization, gamma correction and cropping on the original color retinal image; cropping refers to image block cropping on the training set retinal image and the test set retinal image, for the training set retinal image, random coordinates are generated, and the image block is cropped with the random coordinates as the center point, the size of the image block is 128*128, and one retinal image is cropped to obtain multiple 128*128 retinal images; for the test set, the image block of 128*128 is cropped every 16 pixels in the long and wide directions, one retinal image is cropped to obtain multiple 128*128 retinal images, and the last block is padded with 0 if it is less than 128; then the new training set obtained after cropping is Ep={(pa i ,pb i )|i∈[1,Mp]} and the new test set obtained after cropping is Fp={(pa i ,pb i )|i∈[1,Np]} where Mp and Np are the number of retinal images after cropping.

[0049] In this embodiment, the gray scale transformation can increase the image dynamic range, expand the contrast, make the image clear and the features obvious, the standardization can make the distribution of the input data more consistent with the assumption of the deep learning model, improve the generalization ability and robustness of the model, and is beneficial to speed up the training speed of the model and improve the accuracy of the model, the adaptive histogram equalization can be used to increase the global contrast of many images, and the gamma correction can facilitate the retinal image to more effectively save the image brightness information. The effectiveness and efficiency of storage are improved.

[0050] In step S102, the scale feature information corresponding to the retinal image is obtained through the preset feature extraction network for the retinal image in the training set, and the respective segmentation difficulty degrees corresponding to the scale feature information are further calculated.

[0051] Specifically, the scale feature information corresponding to the retinal image is obtained through the preset feature extraction network, and the feature extraction network can contain N layers of convolution. For example, the feature extraction network can contain three layers of convolution.

[0052] In this embodiment, the feature extraction network contains multiple layers of convolution, the basic unit is simple, and multiple basic units can be expanded to a very complex nonlinear function. It is easy to construct, and the model has strong expression ability, and the parallelism of training and testing calculation is very good.

[0053] On this basis, the scale feature information corresponding to the retinal image is obtained through the preset feature extraction network for the retinal image in the training set, including:

[0054] obtaining scale feature information corresponding to the retinal image through the random initial value and a preset feature extraction network.

[0055] Specifically, S sets of different random initial values are input into the feature extraction network, and one scale feature information is obtained each time, so that S sets of scale feature information can be obtained. For example, S can be 1, 2, 3, 4, etc. Three sets of random initial values are input into the feature extraction network to obtain three sets of scale feature information.

[0056] On this basis, and further, the scale feature information is calculated to obtain the respective segmentation difficulty levels, including:

[0057] The similarity between the scale feature information of different random initial values is calculated through a similarity formula.

[0058] The average value of the scale feature information of different random initial values is calculated through the following formula:

[0059]

[0060] wherein, sim(x k ) represents the average value of the scale feature information of different random initial values, sim ij represents the similarity between the scale feature information of different random initial values, x k represents the kth retinal image, and z represents the zth set of random initial values.

[0061] The weight value of the evaluation sample difficulty level is calculated through a dispersion standardization formula.

[0062] The threshold value of the segmentation retinal image difficulty level is calculated through the following formula:

[0063]

[0064] wherein, λ represents the threshold value of the segmentation retinal image difficulty level, a represents an adjustment parameter, which is a constant set by the user, and the preferred value is 0.9, e represents the current training round number, m represents the total training round number, and F represents the maximum value of the weight value of the evaluation sample difficulty level.

[0065] Specifically, in one embodiment, the similarity formula can be:

[0066]

[0067] wherein, sim ij represents the similarity between the scale feature information of different random initial values, b i and b j represent the scale feature information of the ith and jth sets of random initial values, respectively.

[0068] The similarity formula is not limited to the above formula. Any formula that can calculate the similarity using scale feature information of two pairs of different random initial values ​​is within the scope of protection of this invention.

[0069] The standardization formula for deviations can be:

[0070]

[0071] Where, sim(x) k ) represents the average value of the scale feature information with different random initial values, v k The weights represent the difficulty of segmenting retinal images, where n represents the total number of retinal images in the training set, and min represents the weights. i∈[1,n] sim(x i ) represents taking the middle sim(x) i The minimum value in ), max i∈[1,n] sim(x i ) represents taking sim(x) i The maximum value in ).

[0072] The threshold for the difficulty of segmenting a retinal image is calculated using the following formula:

[0073]

[0074] Where λ represents the threshold for the difficulty of segmenting the retinal image, α represents an adjustment parameter, which is a user-defined constant with an optimal value of 0.9, e represents the current training round number, m represents the total number of training rounds, and F represents the maximum or minimum value of the weights used to evaluate the difficulty of the samples. F = max k∈[1,n] v k Indicates taking v k The maximum value.

[0075] In one embodiment, three scale feature information can be generated from three sets of random initial values ​​through a feature extraction network with three convolutional layers. Similarity is calculated between each pair of the three feature information using a similarity formula, and the average of the results is taken. A weight value for evaluating the difficulty of segmenting the retinal image is calculated based on this average value. A threshold λ representing the difficulty of segmenting the retinal image is then calculated for each training round based on this weight value. Finally, the segmentation difficulty v of the retinal image is... k Retinal images x with a segmentation difficulty threshold greater than λ k The data is input into the target neural network. The threshold λ for the difficulty of segmenting retinal images is adjusted after each training round.

[0076] In the embodiment, the retinal images greater than the threshold of the retinal image segmentation difficulty are calculated, the threshold of the retinal image segmentation difficulty is adjusted by training, the accuracy of the training model is improved, the difficulty of the retinal image segmentation is calculated, and the image segmentation is performed according to the difficulty of the retinal image segmentation, so that the retinal image segmentation is faster and the calculation efficiency is higher.

[0077] In step S103, each retinal image and the segmentation difficulty are taken as a network source input, the reference blood vessel segmentation image corresponding to the retinal image is taken as an output, the adaptive optimization algorithm and the preset loss function are combined, each retinal image is applied, the target neural network is trained, and a retinal blood vessel segmentation network model is obtained.

[0078] Specifically, the adaptive optimization algorithm can be an adam optimization algorithm. First, the target neural network is initialized to obtain initial parameters of the network, and related training hyperparameters are set. The initialization can be performed in multiple ways, and the Xavier initialization method is selected here. In the adam algorithm, the related training hyperparameters can be set as follows.

[0079] (1) The learning rate is e-4.

[0080] (2) betas (beta1, beta2): Two momentum parameters in the adam algorithm. The default values are (0.9, 0.999).

[0081] (3) eps (epsilon): A very small value used to maintain numerical stability. The default value is 1e-8.

[0082] Other parameters in the training can be as follows: the number of iterations is 300 rounds, and the batch size is 8.

[0083] In the embodiment, when the target neural network is trained, reasonable weight initialization directly determines whether the model can converge. If the weight is too small, the information of the input layer will gradually disappear as the network deepens. If the weight is too large, the input value to the activation layer is too large, the gradient is close to saturation, and the model learns nothing. Using the Xavier initialization method for initialization can effectively avoid the problems of gradient disappearance or gradient explosion, and can accelerate the convergence speed of the neural network. The adam algorithm is used for training and calculation of the target neural network, which is very fast in training speed and can be applied to unstable objective functions, and has good interpretability for hyperparameters.

[0084] On this basis, the preset loss function is as follows.

[0085]

[0086] wherein, minE (ω, v, λ) represents a preset loss function, L BCE (y i , f(x i , ω)) represents a binary cross-entropy loss function, n represents the total number of retinal images in the training set, x k represents the kth retinal image, ω represents the weight of the model, v i represents a measure of the difficulty of segmenting the retinal image, and λ represents a threshold of the difficulty of segmenting the retinal image.

[0087] In this embodiment, the preset loss function can measure the goodness of the model prediction, evaluate the accuracy of the model, and guide the optimization of the model.

[0088] Step S104: segmenting the retinal image according to the retinal image and the retinal blood vessel segmentation network model.

[0089] Specifically, the to-be-tested retinal image in the given retinal image sample set is input into the trained retinal blood vessel segmentation network model for inference to obtain a segmented retinal blood vessel image.

[0090] The present application provides a kind of retinal blood vessel intelligent segmentation method, more intelligently according to the retinal image segmentation difficulty degree input retinal image carries out network learning, designs local and global perception's based on axis attention's retinal blood vessel segmentation network model, improves image segmentation quality.

[0091] It should be understood by those skilled in the art that the above discussion of any embodiment is merely exemplary and is not intended to suggest that the scope of the present disclosure (including the claims) is limited to these examples; under the idea of the present disclosure, the technical features of the above embodiments or different embodiments can also be combined, the steps can be implemented in any order, and there are many other changes of different aspects of the embodiments of the present disclosure as described above. In order to be brief, they are not provided in detail.

[0092] Additionally, to simplify the description and discussion, and so as not to obscure the embodiments of the disclosure with details that are well known to those skilled in the art, some conventional attributes of integrated circuit (IC) chips and other components can or can not be shown in the drawings or discussed below. Furthermore, devices can be shown in block diagram form in order to avoid obscuring the embodiments of the disclosure, and this also acknowledges the fact that the details in regard to how such block devices are implemented are highly dependent on the platform within which an embodiment of the disclosure is being implemented (i.e., such details should be completely evident to those skilled in the art). Where specific details are set forth in order to describe an illustrative embodiment of the disclosure, it will be apparent to one skilled in the art that the embodiment of the disclosure can be practiced without, or with variations of, these specific details. Thus, the description is to be considered as illustrative and not restrictive, and the scope of the disclosure is to be determined not with the assistance of the foregoing description alone.

[0093] While the disclosure has been described in connection with specific embodiments thereof, many alternatives, modifications and variations will be apparent to those skilled in the art in light of the foregoing description. For example, other memory architectures (e.g., dynamic RAM (DRAM)) can use the embodiments discussed.

[0094] Embodiments of the disclosure are intended to cover all such alternatives, modifications and variations as falling within the scope of the appended claims. Accordingly, any omission, modification, equivalent replacement, improvement, etc. made within the spirit and principle of the embodiments of the disclosure should be included in the protection scope of the disclosure.

Claims

1. A method for intelligent segmentation of retinal blood vessels, the method comprising: The method comprises the following steps: obtaining a given retinal image sample set, and preprocessing the retinal image sample set to obtain a preprocessed data set; dividing the preprocessed data set into a training set of a retinal blood vessel segmentation network model and a test set for testing the retinal blood vessel segmentation network model according to a preset ratio; wherein the retinal image sample set comprises a retinal image and a reference blood vessel segmentation image; obtaining scale feature information corresponding to the retinal image through a preset feature extraction network for each retinal image in the training set, and further calculating the segmentation difficulty degree corresponding to the scale feature information; training the target neural network by applying each retinal image and combining an adaptive optimization algorithm and a preset loss function, taking each retinal image and the segmentation difficulty degree as the network source input and taking the reference blood vessel segmentation image corresponding to the retinal image as the output, to obtain the retinal blood vessel segmentation network model; segmenting the retinal image according to the retinal image and the retinal blood vessel segmentation network model; obtaining a random initial value, and obtaining the scale feature information corresponding to the retinal image through the random initial value and the preset feature extraction network; calculating the similarity between the scale feature information of different random initial values through a similarity formula; calculating the average value of the scale feature information of different random initial values through the following formula: calculating the weight value of the evaluation sample difficulty degree through a deviation standardization formula; calculating the threshold value of the segmentation retinal image difficulty degree through the following formula: ， wherein, represents an average of the scale feature information of different random initial values, represents a similarity between the scale feature information of different random initial values, represents the th retinal image, represents a group of random initial values; The target neural network is a convolutional neural network model, which comprises an encoding part and a decoding part connected by a cross-layer structure; wherein the encoding part comprises a height-axis attention layer and a width-axis attention layer. The preprocessing of the retinal image sample set to obtain the preprocessed data set comprises: ， Wherein, Threshold value representing the difficulty level of the segmented retinal image, The adjustment parameter is a constant set by itself, the value is 0.9, Indicates the current training round, Indicates the total number of training rounds, and F represents the maximum value of the weight value of the evaluation sample difficulty level.

2. The method of claim 1, wherein, performing grayscale transformation, standardization, adaptive histogram equalization, gamma correction and clipping on the retinal image sample set to obtain the preprocessed data set. The preset loss function is: 3.The method of claim 1, wherein, The convolutional neural network model is a U-shaped convolutional neural network model; The adaptive optimization algorithm is an adam optimization algorithm.

4. The method of claim 1, wherein the step of segmenting the retinal blood vessels is performed by using a convolutional neural network. ​ ， in, Indicates the preset loss function. This represents the binary cross-entropy loss function. This represents the total number of retinal images in the training set. Indicates the first A retinal image, Represents the model's weights. A measure of the difficulty of segmenting a retinal image. A threshold representing the difficulty of segmenting a retinal image.

5. The method of claim 2, wherein, ​ ​ ​

Images