A visible light OCT image sharpening method based on generative adversarial network

Abstract

The present invention discloses a method for clearing visible light OCT images based on a generative adversarial network. The method is as follows: 1. Acquisition of training samples. Use a visible light OCT device to obtain a V-domain image. Use a near-infrared light OCT device to obtain an N-domain image. 2. Preprocess the V-domain image and the N-domain image obtained in step 1. 3. Establish a training sample data set based on the preprocessed V-domain image and the N-domain image. 4. Construct and train an OCT image clearing model. 5. Input the visible light OCT image into the generator G1 to obtain an OCT image with improved quality. The new unsupervised generative adversarial network proposed in the present invention constructs an OCT image clearing model with the help of the idea of ​​dual learning, and can obtain better results without using paired images for training. In addition, the present invention adds a network that can edit image attributes, which enhances the ability of the generator to learn clear structures during dual training.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a method for clarifying visible light OCT images based on a generative adversarial network. Background Art

[0002] Optical Coherence Tomography (OCT) uses the basic principle of weak coherent light interferometer to detect the back reflection or several scattered signals of the incident weak coherent light at different depths of biological tissues, and obtains two-dimensional or three-dimensional structural images of biological tissues by scanning. According to the different light sources used, it can be divided into visible light OCT (visible-light OCT) and near-infrared light OCT (near-infrared OCT). Compared with traditional near-infrared OCT, visible light OCT can not only realize structural imaging of tissues, but also functional imaging, which is especially important for the early detection of ophthalmic diseases such as glaucoma, retinopathy, etc. At the same time, this technology can also realize retinal blood oxygen measurement.

[0003] However, since visible light OCT uses visible light of about 500nm, repeated scanning causes the human eye to be constantly stimulated by the stroboscopic scanning beam, which causes visual fatigue and pupil contraction, resulting in poor image quality (part of the image cannot be collected because the pupil becomes smaller). Reducing the power of the light entering the eye can alleviate pupil contraction, but it will also reduce the light reflection signal received by the spectrometer.

[0004] The above problems cause visible light OCT images to contain a large amount of speckle noise and have low resolution, which is not conducive to the clinical analysis and diagnosis of retinal diseases. It is necessary to improve the quality of visible light OCT images. Improving the quality of OCT imaging from the hardware perspective will be limited by existing processes and increase system costs, so post-processing will be considered to improve image quality. Traditional image processing methods have problems such as over-smoothing or insufficient denoising, and cannot effectively retain retinal layering information, especially visible light OCT images. Methods based on deep learning can achieve better results. For example, the generative adversarial network model is used to output images that are very close to the real data during the continuous adversarial training of its generator G and discriminator D. By generating adversarial networks, the quality of visible light OCT images can be improved with the help of high-quality near-infrared OCT images. Patent application number 201811042548.7 discloses an OCT image speckle denoising method based on conditional generative adversarial networks (cGAN). The conditional generative adversarial network is used to train a mapping model of noisy OCT images and noise-free OCT images, and a constraint of maintaining edge details is introduced. The mapping model is used to eliminate speckle noise in retinal OCT images. Patent application number 201910515611.2 discloses an OCT image denoising method based on a ring generative adversarial network (CycleGAN), which avoids the limitation of previous methods such as conditional generative adversarial networks that require paired data, and can effectively convert high-noise images into clear images. Patent application number 201910665387.5 discloses an OCT image denoising method based on dense connections and generative adversarial networks. According to the characteristics of noise randomness, a multi-frame registration method is used to synthesize a reference image, so that the network can learn the mapping relationship between the noisy image and the reference image. Patent application number 201910732921.X discloses a SD-OCT denoising method based on an unsupervised adversarial neural network, which achieves migration from SD-OCT to EDI-OCT image domain through a cyclic generative adversarial network, and can generate SD-OCT images with EDI-OCT quality without paired data. Patent application number 202111293626.2 discloses a retinal OCT image sharpening method based on a double-layer CycleGAN model, which is trained twice by a double-layer CycleGAN network to further improve the sharpening effect.

[0005] The current OCT image quality improvement technology based on generative adversarial networks has the following main problems: (1) Many deep learning methods require paired noisy images and clear images, such as the method based on conditional generative adversarial networks. However, in actual processing, the two may not be completely aligned. If they are used in the training of supervised models, the generalization performance of the model may be affected; (2) The CycleGAN-based method does not require paired images, but the network implements image style transfer, which is reflected in the application of OCT images as reducing speckle noise, but cannot well preserve the detailed information of the retina; (3) The current OCT image quality improvement technology is mainly aimed at near-infrared OCT images. For visible light OCT images with more speckle noise and lower resolution, there is currently no visible light OCT image quality improvement technology that achieves ideal results. Summary of the invention

[0006] The purpose of the present invention is to provide a method for improving the quality of visible light OCT images in view of the problem that the prior art fails to effectively improve the quality of visible light OCT images; the method further optimizes the image detail information while denoising the image. The present invention is only used to improve the quality of OCT images and is not directly used for disease diagnosis.

[0007] In a first aspect, the present invention provides a method for improving the quality of visible light OCT images based on a generative adversarial network, which comprises the following steps:

[0008] Step 1: Acquisition of training samples. Use visible light OCT device to acquire V-domain images. Use near-infrared light OCT device to acquire N-domain images.

[0009] Step 2: preprocess the V domain image and N domain image obtained in step 1.

[0010] Step 3: Establish a training sample data set based on the preprocessed V-domain images and N-domain images.

[0011] Step 4: Build and train the OCT image sharpening model.

[0012] The OCT image sharpening model includes three generators G 1 , G 2 , G 3 And three discriminators D x , D y , D z . Generator G 1 Learn the nonlinear mapping relationship from V domain to N domain; Generator G 2 Learn the nonlinear mapping relationship from N domain to V domain. V domain image input generator G 1 , and obtain G with N domain characteristics 1 (v) Image; G1 (v) Image input generator G 2 , and get G 2 (G 1 (v)) Image; N-domain image input generator G 2 , and obtain G with V domain characteristics 2 (n) Image; G 2 (n) Image input generator G 1 , and get G 1 (G 2 (n)) image; discriminator D x For the generator G 2 The generated G 2 (n) Image discrimination; discriminator D y For the generator G 1 The generated G 1 (v) Image discrimination. Generator G 3 and the discriminator D z As a component of the F network. The F network decomposes the N domain image into target information u and background information A, and G 1 (G 2 (n)) The image is decomposed into target information ε and background information B, and recombined to form Bu image.

[0013] The objective function of constructing the OCT image sharpening model is as follows:

[0014] L(G 1 ,G 2 ,D X ,D Y )=L GAN (G 2 ,D X ,V,N)+L GAN (G 1 ,D Y ,V,N)+λL cyc (V,N)

[0015] Among them, L GAN (G 1 ,D Y ,V,N) and L GAN (G 2 ,D X ,V,N) are the generators G 1 , G 2 The corresponding adversarial loss function; L cyc (V,N) is the relationship between the N domain image and the Bu image output by the F network, and G 2 (G 1 (v) The cycle consistency loss between the image in domain V and the image in domain V. λ is the scaling factor of the cycle consistency loss.

[0016] Step 5: After the OCT image sharpening model training is completed, the visible light OCT image is input into the generator G 1 In the process, the quality of OCT images is improved.

[0017] As a preference, the adversarial loss function L described in step 4 GAN (G 1 ,D Y ,V,N) and L GAN (G 2 ,D X ,V,N) is as follows:

[0018]

[0019]

[0020] Among them, G 1 (·) is the generator G 1 Output of G 2 )·) is the generator G 2 Output; D X (·) is the discriminator D x Output; D Y (·) is the discriminator D Y The output of; v is the V domain image; n is the N domain image. The V domain image obeys the probability distribution p data (v) expectations; N domain images obey probability distribution p data (n) expectation. data (v) and p data (n) are the probability distributions of V domain images and N domain images respectively.

[0021] As a preference, the cycle-consistent loss L described in step 4 is cyc The expression of (V,N) is as follows:

[0022]

[0023] Among them, x Bu is the Bu image output by the F network. v is the V domain image; n is the N domain image. G 2 (G 1 (v)) is the V domain image input generator G 1 The obtained image is then input into the generator G 2 The obtained G 2 (G 1 (v))Image.

[0024] Preferably, the preprocessing of the V-domain image and the N-domain image in step 2 includes horizontal flipping, rotation, grayscale histogram adjustment and brightness adjustment.

[0025] Preferably, the training process of the F network described in step 4 is as follows:

[0026] N-domain image and G 1 (G 2 (n)) images form a set of paired images as the input of the F network. The generator G in the F network 3 It includes an encoder and a decoder; the encoder includes three downsampling convolutional layers Conv connected in sequence; the decoder includes three upsampling deconvolutional layers Deconv connected in sequence.

[0027] The N-domain image is decomposed into background information A and target information u by the encoder; 1 (G 2 The (n)) image is decomposed into background information B and target information ε by the encoder. The empty information 0 is constructed with the same element dimension as the target information ε and all elements are 0.

[0028] The decoder recombines the information A, u, B, 0 to form a new image, and obtains the image Au, A0, Bu, B0. z The probability values ​​of judging that the image Au and the image Bu do not contain the target information u and the probability values ​​of judging that the image A0 and the image B0 contain the target information u are output.

[0029] Construct the discriminator D z The standard D loss L D :

[0030]

[0031] The expression is as follows:

[0032]

[0033]

[0034] Where z is random noise; Denotes that random noise z follows the probability distribution P of no target feature u 0 expectations; Indicates that the random noise z follows the probability distribution P with target feature u ≠0 Expectations; x Au 、x A0 、x Bu 、x B0 For images Au, A0, Bu, B0. Z (·) is the discriminator DZ Output.

[0035] Build generator G 3 The total loss function L G as follows:

[0036]

[0037] in, are two reconstruction losses; are the generator loss without target features and the generator loss with target features respectively; L 0 To zero loss.

[0038] Generator loss without target features and the generator loss with target features The expression is:

[0039]

[0040]

[0041] Reconstruction loss The expression is:

[0042]

[0043]

[0044] Zero loss L 0 The expression is:

[0045] L 0 =||ε|| 1 .

[0046] ε is the 1 (G 2 (n)) The target information is decomposed from the image.

[0047] Preferably, the process of establishing the training sample data set in step 3 is: sliding trimming the preprocessed V domain image and N domain image respectively to expand the training sample data set. The sliding trimming process is as follows: (1) setting the window size. (2) setting the step size according to the size of the image. (3) discarding the image blocks without retinal information to obtain two sets of training sample data sets of the same size. The two sets of training sample data sets are respectively a data set V consisting of V domain images and a data set N consisting of N domain images.

[0048] As a preferred method, the discriminator D x , D y The method of whether it converges is: judging by the discriminator D x and Dy Whether the output probability values ​​are respectively greater than their preset probability thresholds. If they are all greater than the preset probability thresholds, it is determined to be converged; otherwise, it is determined that the discriminator does not converge.

[0049] In a second aspect, the present invention provides a computer device comprising a memory and at least one processor; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor performs the aforementioned clarification method.

[0050] In a third aspect, the present invention provides a readable storage medium storing computer instructions; when the computer instructions are executed by a processor, they are used to implement the aforementioned clarification method.

[0051] The present invention has the following beneficial effects:

[0052] 1. The novel unsupervised generative adversarial network proposed in the present invention builds an OCT image sharpening model with the idea of ​​dual learning, and can obtain better results without using paired images for training. Given that it is currently difficult to obtain a large number of high-quality visible light OCT images that can be used for training, the present invention uses unpaired high-quality near-infrared light OCT images to improve the quality of visible light OCT images, solving the problem of difficulty in obtaining paired visible light OCT images for training.

[0053] 2. The present invention adds a network that can edit image attributes, which enhances the ability of the generator to learn clear structures during dual training, thereby achieving denoising of visible light OCT images while further optimizing image detail information, improving the quality of visible light OCT images, and facilitating clinical doctor image recognition and diagnosis as well as subsequent software analysis. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 It is a flow chart of the present invention.

[0055] Figure 2 This is a structural diagram of the OCT image sharpening model used in the present invention.

[0056] Figure 3 Schematic diagram of the training of the F network of the OCT image sharpening model used in the present invention.

[0057] Figure 4 This is a structural diagram of the F network of the OCT image clearing model used in the present invention. DETAILED DESCRIPTION

[0058] The present invention will be further described below in conjunction with the accompanying drawings.

[0059] like Figure 1As shown, a method for improving the quality of visible light OCT images based on a generative adversarial network is used to denoise the image while further optimizing the image detail information, which includes the following steps:

[0060] S1, acquisition of training samples. Use near-infrared light OCT device to obtain high-quality image samples; use visible light OCT device to obtain low-quality image samples. The specific process is as follows:

[0061] Near-infrared OCT images and visible light OCT images were obtained for people without retinal lesions and patients with glaucoma and macular degeneration. The collected retinal OCT images include healthy retinas and retinas with different types of lesions. Since different types of lesions can cause different retinal deformations, different lesion images are used to establish a training sample data set, so that the trained model can adapt to OCT images with different pathologies.

[0062] S2, preprocessing of training samples. To achieve better training results, after obtaining the image samples of the two modes, they are preprocessed using software. The specific process is as follows:

[0063] The near-infrared OCT image and the visible light OCT image obtained in step S1 are preprocessed including horizontal flipping, rotation, grayscale histogram adjustment and brightness adjustment.

[0064] Horizontal flipping simulates symmetrical left and right eyes, so that the model trained with the processed dataset can adapt to the left and right eyes.

[0065] The rotation is used to simulate the different tilts of the retina in the OCT images, so that the model trained with the processed dataset can robustly denoise OCT images with different tilts.

[0066] Adjusting the grayscale histogram and brightness is used to make the OCT image more suitable for human eye observation.

[0067] S3, establishing a training sample data set. To establish a training sample data set, the two groups of image samples pre-processed in step S2 need to be subjected to sliding trimming to expand the training sample data set.

[0068] The process of sliding trimming is as follows:

[0069] (1) Set the appropriate window size.

[0070] (2) Set an appropriate step size according to the size of the OCT image.

[0071] (3) Image blocks without retinal information are discarded to obtain two sets of training sample data sets of the same size. The two sets of training sample data sets are respectively a data set V consisting of visible light OCT images and a data set N consisting of near-infrared light OCT images.

[0072] S4, model training. The OCT image sharpening model is trained based on the training sample data set. The entire OCT image sharpening model includes three generators and three discriminators. With unpaired images as input, the OCT image sharpening model can retain the detail information of the retina while denoising.

[0073] like Figure 2 and Figure 3 As shown, the OCT image sharpening model includes three generators G 1 , G 2 , G 3 And three discriminators D x , D y , D z The OCT image sharpening model takes the input image from the dataset V and generates it through the generator G 1 Learn the nonlinear mapping relationship from V domain (visible light OCT image) to N domain (near infrared light OCT image), that is, G 1 :V→N, generator G 1 Output G with N domain features 1 (v) Image, the noise in the V domain image is removed.

[0074] G 1 (v) Then as the generator G 2 Input; Generator G 2 Learn the nonlinear mapping relationship from N domain to V domain, that is, G 2 :N→V, G again 1 (v) Converted to G with the original V domain characteristics 2 (G 1 (v))Image.

[0075] At the same time, the generator G 2 Also using N domain image as input, generate G with V domain features 2 (n) Image; G 2 (n) The image is then used as the generator G 1 The input is G with N domain features. 1 (G 2 (n))image.

[0076] In the above transformation, the discriminator D x For the generator G 2 The generated G 2 (n) Image is discriminated; by the discriminator Dy For the generator G 1 The generated G 1 (v) Image discrimination; during the confrontation process, the discriminator will continuously improve its discrimination ability, and the images generated by the generator will continue to approach the target domain.

[0077] The discrimination process is: compare the V domain image with the generator G 2 The generated image G 2 (n) Input discriminator D x , discriminator D x Output judgment image G 2 (n) is the probability value of the V domain image; the N domain image is combined with the generator G 1 The generated image G 1 (v) Input discriminator D y , discriminator D y Output judgment image G 1 (v) is the probability value of the N-domain image.

[0078] Next, we determine the discriminator D x , D y Whether it converges; if it converges, it will be generated by the generator G 1 The generated image is used as the output image after the visible light OCT image quality is improved; if it does not converge, the parameters of the generated network are continuously updated before the discriminator converges.

[0079] Judgement Discriminator D x , D y The method of whether it converges is: judging by the discriminator D x and D y Whether the output probability values ​​are respectively greater than their preset probability thresholds. If they are all greater than the preset probability thresholds, it is determined to be converged; otherwise, it is determined that the discriminator does not converge.

[0080] Discriminator D x Determine whether the image is a V domain image or a G domain image. 2 The generated G 2 (n) image, which can form an adversarial loss function L GAN (G 2 ,D X ,V,N) are as follows:

[0081]

[0082] Among them, p data (v) and p data are the probability distributions of V domain images and N domain images, G 2 (·) is the generator G 2 Output; D X (·) is the discriminator D xThe output of ; v is the visible light OCT image that has not passed through the generator; The V domain image obeys the probability distribution p data (v) expectations; N domain images obey probability distribution p data (n) expectation.

[0083] For the generator G 1 The same is true for the conversion of . It can form an adversarial loss function L GAN (G 1 ,D Y ,V,N):

[0084]

[0085] Among them, G 1 (·) is the generator G 1 Output; D Y (·) is the discriminator D Y The output of ; n is the near-infrared OCT image that has not passed through the generator.

[0086] In the above process, the V domain image and the N domain image are transformed by the generator once, that is, G 1 (v) Image and G 2 (n) image; these two sets of images need to be input into the generator G separately 2 With the generator G 1 , generating G 2 (G 1 (v)) Image and G 1 (G 2 (n)) image. In order to achieve better conversion effect, we require that the image converted by the generator twice should be as similar as possible to its original image, that is, v→G 1 (v) →G 2 (G 1 (v))≈v, and n→G 2 (n)→G 1 (G 2 (n))≈n, and define a loss function to achieve this.

[0087] The above two generators and two discriminators can well achieve style transfer from V domain to N domain, and the speckle noise of V domain images can be suppressed to a certain extent. However, the preservation of image detail information after conversion is poor, which can easily lead to structural loss of generated images. When processing retinal OCT image tasks, the unclear retinal layered structure can easily lead to the loss of key detail features.

[0088] In order to improve the degree of preservation of the retinal hierarchical structure in the converted image, the present invention provides a novel unsupervised generative adversarial network method to enhance the detail information of the generated image.

[0089] In the above steps, the N domain image is transformed twice by the generator G 1 The output image G 1 (G 2 Although (n)) is restored to the N domain, the reconstructed image has certain differences from the original image.

[0090] In order to further preserve the retinal layering information in the visible light OCT image, the present invention introduces the F network based on the above network to optimize the image detail information. 3 and the discriminator D z It is a component of the F network. The F network can extract features from the two input images respectively, decouple the images into background feature parts and target feature parts, and transplant the target features of one image to another image.

[0091] In the present invention, the target feature is the retinal structure part in the N-domain image.

[0092] Bu image and N domain image output by F network, and G 2 (G 1 (v)) The image and the V domain image form a cycle-consistent loss L cyc (V,N):

[0093]

[0094] Among them, x Bu is the Bu image output by the F network, which decomposes the N domain image into target information u and background information A through the F network, and G 1 (G 2 (n) After the image is decomposed into target information ε and background information B, the background information B is combined with the target information u.

[0095] Control cycle consistency loss L cyc (V,N) can ensure that the image after being transformed by multiple generators is as similar as possible to the original input image.

[0096] The above generator and discriminator constitute the objective function as follows:

[0097] L(G 1 ,G 2 ,D X ,D Y )=L GAN (G 2 ,D X,V,N)+L GAN (G 1 ,D Y ,V,N)+λL cyc (V,N)

[0098] Where λ is the scaling factor of the cycle consistency loss and is a hyperparameter.

[0099] The training process of the F network is as follows:

[0100] like Figure 3 and Figure 4 As shown, the N domain image and the G domain image obtained after two transformations 1 (G 2 (n)) images constitute a set of paired images as the input of the F network. The paired images have the same retinal structure information, but the clarity of the detail information is different.

[0101] Generator G in F network 3 It includes an encoder and a decoder; the encoder includes three downsampling convolutional layers Conv connected in sequence; the decoder includes three upsampling deconvolutional layers Deconv connected in sequence.

[0102] The N-domain image is decomposed into background information A and target information u by the encoder; 1 (G 2 The (n)) image is decomposed into background information B and target information ε by the encoder; ε is used as the zeroing loss, which reflects the degree of separation between the target feature and the background feature. The empty information 0 is constructed with the same element dimension as the information ε and all elements are 0; u and 0 respectively represent the presence or absence of the target object.

[0103] The decoder converts the N domain image and the G 1 (G 2 (n)) The four information decomposed from the image are recombined to form a new image, and four sub-combined images are obtained: Au, A0, Bu, and B0.

[0104] Discriminator D z The probability values ​​of judging that the image Au and the image Bu do not contain the target information u and the probability values ​​of judging that the image A0 and the image B0 contain the target information u are output.

[0105] The trained generator G 3 It is required that Au and Bu cannot be distinguished, and A0 and B0 cannot be distinguished. In other words, all target object information is encoded in u as much as possible, so as to ensure that the separated A does not contain attribute u information as much as possible, thereby achieving a good separation of background information and target information, forming a discriminator D z The standard D loss L D :

[0106]

[0107] The expression is as follows:

[0108]

[0109]

[0110] Where z is random noise; It means that the random noise z follows the probability distribution P that the target feature u does not exist in the image 0 expectations; Indicates that the random noise z follows the probability distribution P of the target feature u in the image ≠0 Expectations; x Au 、x A0 、x Bu 、x B0 For images Au, A0, Bu, B0; D Z (·) is the discriminator D Z Output.

[0111] Generator G 3 The total loss of is composed of three types of losses: standard G loss, reconstruction loss and zeroing loss.

[0112] The standard G loss contains the generator loss without target features. and the generator loss with target features

[0113]

[0114]

[0115] Among them, x A0 is an image with N domain image as background without target features generated by the generator after encoding and decoding, x Bu The target feature is generated by the generator after encoding and decoding. 1 (G 2 (n)) is the background image.

[0116] At the same time, the network needs to reconstruct the loss To ensure that the original image and the restored image after encoding and decoding have the greatest similarity:

[0117]

[0118]

[0119] Generator G3 The total loss function L G The expression is:

[0120]

[0121] Among them, L 0 To zero the loss, that is:

[0122] L 0 =||ε|| 1

[0123] From the N domain image and G 1 (G 2 After the paired images composed of (n) images pass through the F network, the retinal part of the N domain image is transplanted to the G domain as the target feature. 1 (G 2 (n)) image, that is, to generate 1 (G 2 (n)) image is the background and Bu image with retinal information of N domain image.

[0124] The objective function L(G 1 ,G 2 ,D X ,D Y ), the total loss function L G Minimize the goal and train three generators G 1 , G 2 , G 3 And three discriminators D x , D y , D z , and obtain the trained OCT image sharpening model.

[0125] S5, use of the model. The visible light OCT image is input into the generator G of the trained OCT image sharpening model. 1 In the generator G 1 Output the OCT image with improved quality. Generator G 1 The image detail information can be further optimized while denoising the image.

Claims

1. A method for clearing visible light OCT images based on a generative adversarial network, characterized in that: The following steps are involved: Step 1: Acquisition of training samples; using visible light OCT images as V-domain images and near-infrared light OCT images as N-domain images; using visible light OCT devices to acquire V-domain images; using near-infrared light OCT devices to acquire N-domain images; Step 2: preprocessing the V domain image and the N domain image obtained in step 1; Step 3: Establish a training sample data set based on the preprocessed V domain image and N domain image; Step 4: Build and train the OCT image sharpening model; The OCT image sharpening model includes three generators G1, G2, G3 and three discriminators D x , D y , D z ; Generator G1 learns the nonlinear mapping relationship from V domain to N domain; Generator G2 learns the nonlinear mapping relationship from N domain to V domain; V domain image is input into generator G1, and G1(v) image with N domain features is obtained; G1(v) image is input into generator G2, and G2(G1(v)) image is obtained; N domain image is input into generator G2, and G2(n) image with V domain features is obtained; G2(n) image is input into generator G1, and G1(G2(n)) image is obtained; discriminator D x Discriminate the G2(n) image generated by the generator G2; discriminator D y Discriminate the G1(v) image generated by the generator G1; the generator G3 and the discriminator D z As a component of the F network; the F network decomposes the N domain image into target information u and background information A, decomposes the G1(G2(n)) image into target information ε and background information B, and recombines to form B u image; The objective function of constructing the OCT image sharpening model is as follows: L(G1,G2,D X ,D Y )=L GAN (G2,D X ,V,N)+L GAN (G1,D Y ,V,N)+λL cyc (V,N) Among them, L GAN (G1, D Y , V, N) and L GAN (G2, D X , V, N) are the adversarial loss functions corresponding to generators G1 and G2 respectively; L cyc (V, N) is the cycle consistency loss between the N domain image and the Bu image output by the F network, and between the G2(G1(v)) image and the V domain image; λ is the scaling factor of the cycle consistency loss; Step 5: After the OCT image sharpening model training is completed, the visible light OCT image is input into the generator G1 to obtain an OCT image with improved quality.

2. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: The adversarial loss function L described in step 4 GAN (G1, D Y , V, N) and L GAN (G2, D X , V, N) are as follows: Where G1(·) is the output of generator G1; G2(·) is the output of generator G2; D X (·) is the discriminator D x Output; D Y (·) is the discriminator D Y The output of; v is the V domain image; n is the N domain image; The V domain image obeys the probability distribution p data (v) expectations; N domain images obey probability distribution p data (n) expectation; p data (v) and p data (n) are the probability distributions of V domain images and N domain images respectively.

3. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: The cycle-consistent loss L described in step 4 cyc The expression of (V, N) is as follows: Among them, x Bu is the Bu image output by the F network; v is the V domain image; n is the N domain image; G2(G1(v)) is the V domain image input into the generator G1 and then input into the generator G2 to obtain the G2(G1(v)) image.

4. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: The training process of the F network described in step 4 is as follows: The N domain image and the G1(G2(n)) image form a pair of paired images as the input of the F network; the generator G3 in the F network includes an encoder and a decoder; the encoder includes three downsampling convolutional layers Con connected in sequence v ; The decoder consists of three upsampling deconvolution layers connected in sequence. v ; The N-domain image is decomposed into background information A and target information u by the encoder; the G1(G2(n)) image is decomposed into background information B and target information ε by the encoder; the empty information 0 is constructed with the same element dimension as the target information ε and all elements have the same value of 0; The decoder recombines the information A, u, B, and 0 to form a new image, and obtains image A. u , A0, B u , B0; Discriminator D z Output the probability value of determining that image Au and image Bu do not contain target information u, and the probability value of determining that image A0 and image B0 contain target information u; Construct the discriminator D z The standard D loss L D : The expression is as follows: Where z is random noise; It means that the random noise z follows the expectation of the probability distribution P0 that the target feature u does not exist in the image; Indicates that the random noise z follows the probability distribution P of the target feature u in the image ≠0 Expectations; x Au 、x A0 、x Bu 、x B0 For images Au, A0, Bu, B0; D Z (·) is the output of the discriminator Dz; Construct the total loss function L of generator G3 G as follows: in, are two reconstruction losses; They are respectively the generator loss without target features and the generator loss with target features; L0 is the zero loss; Generator loss without target features and the generator loss with target features The expression is: Reconstruction loss The expression is: The expression of zero loss L0 is: L0=old|1; ε is the target information decomposed from the G1(G2(n)) image.

5. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: The preprocessing of the V-domain image and the N-domain image in step 2 includes horizontal flipping, rotation, grayscale histogram adjustment and brightness adjustment.

6. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: The process of establishing the training sample data set in step three is: sliding trimming the preprocessed V domain images and N domain images respectively to expand the training sample data set; the sliding trimming process is as follows: (1) setting the window size; (2) setting the step size according to the size of the image; (3) discarding the image blocks without retinal information to obtain two sets of training sample data sets of the same size; the two sets of training sample data sets are respectively a data set V consisting of V domain images and a data set N consisting of N domain images.

7. The method for clearing visible light OCT images based on a generative adversarial network according to claim 1, characterized in that: Judgement Discriminator D x , D y The method of whether it converges is: judging by the discriminator D x and D y Whether the output probability values ​​are respectively greater than their preset probability thresholds; if they are all greater than the preset probability thresholds, it is determined to be converged; otherwise, it is determined that the discriminator does not converge.

8. A computer device comprising a memory and at least one processor; characterized in that: The memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes a method for clearing visible light OCT images based on a generative adversarial network as described in any one of claims 1 to 7.

9. A readable storage medium storing computer instructions; characterized in that: When the computer instructions are executed by a processor, they are used to implement a method for clearing visible light OCT images based on a generative adversarial network as described in any one of claims 1 to 7.

Images