An image noise removal method and system based on as-octa
Noisy images were acquired by artificially simulating eye movements using AS-OCTA. Blood vessels and noise were separated using DCGAN and residual U-Net, and Otsu's thresholding method was employed to address the data scarcity problem in noise removal of AS-OCTA images. This resulted in efficient noise removal and blood vessel visualization, improving the accuracy of image analysis.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-07-25
- Publication Date
- 2026-04-07
AI Technical Summary
Existing techniques struggle to effectively remove stripe noise from AS-OCTA images, especially due to the scarcity of training datasets and the complexity of noise distribution. Traditional methods are ineffective, impacting the accuracy of image analysis.
Noisy images were acquired by artificially simulating eye movements using AS-OCTA, and noisy image data was generated using an improved generative adversarial network DCGAN. Blood vessels and noise were separated by a residual U-Net neural network, and the images were binarized using the Otsu thresholding method.
The training dataset was expanded, noise removal efficiency was improved, the integrity of blood vessels was maintained, image robustness was enhanced, and image quality was improved to facilitate quantitative analysis and diagnosis of blood vessels.
Smart Images
Figure CN115375628B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing, and specifically to an image noise removal method and system based on AS-OCTA. Background Technology
[0002] Optical coherence tomography (OCTA) provides a quantitative, objective, and non-invasive technique for assessing angiogenesis. In recent years, increasing research has focused on anterior segment OCTA (AS-OCTA), finding it superior to traditional anterior segment examination methods, particularly in its ability to detect deep vessels. Although AS-OCTA research is still in its early stages, its clinical application prospects are gradually gaining attention and recognition, including in the diagnosis, treatment, and prognosis of anterior segment diseases. However, it's important to note that current OCTA systems are designed for retinal angiography, which means eye-tracking functionality is not well-suited for AS-OCTA, leading to unavoidable noise and limiting its application. We believe noise can significantly impact the accuracy of image analysis results; therefore, to better promote and expand the application of AS-OCTA, the interference of noise on image quality must be addressed.
[0003] AS-OCTA noise exhibits a striped distribution, primarily caused by the rapid, extensive eye movements during imaging. Hardware-wise, increasing scanning speed or using eye-tracking can reduce striped noise. However, due to limitations in AS-OCTA imaging technology, these two auxiliary configurations are currently difficult to implement; therefore, noise removal can only be attempted through image processing. Existing traditional methods, such as those based on Fourier wavelet transform cluster filters, can effectively remove striped noise, but due to the dense distribution of AS-OCTA noise, the results are not ideal. With the widespread application of deep learning technology in various fields, several deep learning-based striped noise removal methods have been proposed, such as the Wavelet Transform Striped Noise Removal Network (SNRWDNN) and the Precise Noise Reduction Segmentation Network (Ant-UNet). However, due to the varying lengths and thicknesses of AS-OCTA noise, these methods cannot ideally remove the noise, and they require large amounts of data for training, while real-world medical imaging data is relatively scarce. The latter has fast training and forward inference speeds and good results, but its drawback is that the training set needs to be manually labeled first, which is almost impossible for AS-OCTA images with messy and irregular blood vessels. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides an image noise removal method and system based on AS-OCTA, which removes horizontal white stripe noise in images caused by the rapid movement of the eyeballs during the shooting process. This solves the problem of scarce training datasets and improves the efficiency of noise removal.
[0005] To achieve the above objectives, the present invention can adopt the following technical solutions:
[0006] An image noise removal method based on AS-OCTA includes the following steps:
[0007] By artificially simulating eye movements using AS-OCTA, several sets of first image data with pure noise and several sets of second image data without noise are obtained.
[0008] Each image from the first image dataset is input into the improved Generative Adversarial Network (DCGAN) model to generate several third image datasets that are similar to the first image dataset.
[0009] Each image from the second image dataset is fused with each image from the generated third image dataset and input into the residual U-Net neural network model to separate blood vessels from noise in the image data, thereby reconstructing a fourth image dataset.
[0010] Each image in the fourth image dataset is binarized using the Otsu thresholding method to obtain the fifth image dataset.
[0011] The image noise removal method based on AS-OCTA described above, further, in the improved Generative Adversarial Network (DCGAN) model,
[0012] The generator uses a five-layer deconvolution structure, and the loss is defined as:
[0013]
[0014] Where G(z) represents the output image of the generator, and D(G(z)) represents the input probability of the discriminator;
[0015] The discriminator uses a 5-layer convolutional structure, and its loss is defined as:
[0016]
[0017] Where D(x) represents the discriminator probability of the sample noise map,
[0018] The optimization objective of the training process is expressed as:
[0019]
[0020] The image noise removal method based on AS-OCTA described above, further, involves training the improved Generative Adversarial Network (DCGAN) model as follows:
[0021] Keeping the parameters of the generator unchanged, the network parameters of the discriminator are trained and updated. Random noise is input into the generator and then passed through a convolutional neural network to obtain the generated image. The label of the sample image is set to 1, and the label of the non-sample image is set to 0. The output image of the generator and the sample image are used as the training dataset of the discriminator for network training. The parameters of the discriminator are updated by calculating the gradient error between the two types of images.
[0022] Keeping the parameters of the discriminator unchanged, the generator is trained, wherein random noise is backpropagated through the parameters of the discriminator and generated by the generator. In order to make the output of the generator fit the direction of the sample image, the image label at this time is recorded as 1, and the parameters of the generator are updated.
[0023] The image noise removal method based on AS-OCTA described above further includes a step of training the residual U-Net neural network model. Specifically,
[0024] Each image in the second image dataset is fused with each image in the generated third image dataset to generate several noise-free and noise-only image pairs.
[0025] The noise-free to pure noise images are used as inputs to train the residual U-Net neural network model;
[0026] After training, the residual U-Net neural network model is validated using the remaining noise-free to pure noise images.
[0027] Repeat the above steps until verification is successful;
[0028] The residual U-Net neural network model includes an encoding part, a decoding part, a skip connection part, and a residual module part.
[0029] The image noise removal method based on AS-OCTA described above is further improved by binarization using the Otsu thresholding method. Specifically, the Otsu thresholding method is used to find the optimal threshold for the image data according to the following formula, and then the image is re-binarized according to the threshold:
[0030] G(t)=W0*(U0-U) 2 +W1*(U1-U) 2
[0031] Where G(t) is the objective function, U is the mean of the entire image, W0 and U0 refer to the proportion of foreground pixels in the image and the mean of foreground pixels, and W1 and U1 refer to the proportion of background pixels in the image and the mean of background pixels.
[0032] In addition, an image noise removal system based on AS-OCTA is provided, which includes:
[0033] The first processing unit is used to simulate eye movement via AS-OCTA to acquire a first set of image data with pure noise and a second set of image data without noise.
[0034] The second processing unit is used to input each image of the first image dataset into the improved generative adversarial network DCGAN model to generate several third image datasets that are similar to the first image dataset.
[0035] The third processing unit is used to fuse each image from the second image dataset with each image from the generated third image dataset, and input the fusion into a residual U-Net neural network model to separate blood vessels from noise in the image data, thereby reconstructing a fourth image dataset; and,
[0036] The fourth processing unit is used to binarize each image in the fourth image dataset using the Otsu thresholding method to obtain the fifth image dataset.
[0037] The image noise removal system based on AS-OCTA, as described above, further includes the improved Generative Adversarial Network (DCGAN) model.
[0038] The generator uses a five-layer deconvolution structure, and the loss is defined as:
[0039]
[0040] Where G(z) represents the output image of the generator, and D(G(z)) represents the input probability of the discriminator;
[0041] The discriminator uses a 5-layer convolutional structure, and its loss is defined as:
[0042]
[0043] Where D(x) represents the discriminator probability of the sample noise map,
[0044] The optimization objective of the training process is expressed as:
[0045]
[0046] The image noise removal system based on AS-OCTA, as described above, further involves the following training process using the improved Generative Adversarial Network (DCGAN) model:
[0047] Keeping the parameters of the generator unchanged, the network parameters of the discriminator are trained and updated. Random noise is input into the generator and then passed through a convolutional neural network to obtain the generated image. The label of the sample image is set to 1, and the label of the non-sample image is set to 0. The output image of the generator and the sample image are used as the training dataset of the discriminator for network training. The parameters of the discriminator are updated by calculating the gradient error between the two types of images.
[0048] Keeping the parameters of the discriminator unchanged, the generator is trained, wherein random noise is backpropagated through the parameters of the discriminator and generated by the generator. In order to make the output of the generator fit the direction of the sample image, the image label at this time is recorded as 1, and the parameters of the generator are updated.
[0049] The AS-OCTA-based image noise removal system described above further includes a step of training the residual U-Net neural network model. Specifically,
[0050] Each image in the second image dataset is fused with each image in the generated third image dataset to generate several noise-free and noise-only image pairs.
[0051] The noise-free to pure noise images are used as inputs to train the residual U-Net neural network model;
[0052] After training, the residual U-Net neural network model is validated using the remaining noise-free to pure noise images.
[0053] Repeat the above steps until verification is successful;
[0054] The residual U-Net neural network model includes an encoding part, a decoding part, a skip connection part, and a residual module part.
[0055] The AS-OCTA-based image noise removal system described above further employs Otsu's thresholding method for binarization. Specifically, Otsu's thresholding method is used to find the optimal threshold for the image data according to the following formula, and then the image is re-binarized based on the threshold:
[0056] G(t)=W0*(U0-U) 2 +W1*(U1-U) 2
[0057] Where G(t) is the objective function, U is the mean of the entire image, W0 and U0 refer to the proportion of foreground pixels in the image and the mean of foreground pixels, and W1 and U1 refer to the proportion of background pixels in the image and the mean of background pixels.
[0058] Compared with the prior art, the advantages of this invention are as follows:
[0059] 1. A large number of real, pure noise images are obtained by artificially simulating eye movements using AS-OCTA. This method can obtain more realistic noise information, and the noise used in the next learning step is more reliable.
[0060] 2. DCGAN can reduce the workload of manual data collection, solve the current lack of such noise information, generate noise maps that are similar to the real noise distribution, effectively expand the dataset, and provide data support for subsequent work.
[0061] 3. Residual U-Net can effectively reduce noise caused by eye movement, remove most of the noise, and maintain the integrity of most blood vessels.
[0062] 4. The Otsu thresholding module can effectively highlight blood vessels, reduce white noise, and enhance image robustness, making blood vessel observation more intuitive and efficient, and making quantitative analysis and diagnosis of blood vessels more convenient. Attached Figure Description
[0063] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0064] Figure 1 The DCGAN generated artifact image and composite image in this embodiment of the invention are shown, wherein (a) is a clean image, (b) is a striped noise image generated by DCGAN, and (c) is a composite noise image;
[0065] Figure 2 The following is a comparison of the denoising effects of various methods in the embodiments of the present invention, wherein (a) is the original image, (b) is the residual Unet+OTSU, (c) is the residual Unet, (d) is the FFT bandpass filter+OSTU, and (e) is the SNRWDNN+OSTU.
[0066] Figure 3 This is a flowchart illustrating the implementation of the image noise removal method based on AS-OCTA according to an embodiment of the present invention.
[0067] Figure 4This invention presents the overall network framework and process of the image noise removal method based on AS-OCTA in this embodiment.
[0068] Figure 5 The following is an experimental diagram of the final output result of an embodiment of the present invention, wherein (a) is the original clinical image and (b) is the denoised image. Detailed Implementation
[0069] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.
[0070] Example:
[0071] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, in the embodiments of this invention are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.
[0072] See Figures 1 to 5 This invention first obtains a large number of real pure noise images and a small number of real noise-free images by artificially simulating eye movements using AS-OCTA. Then, it expands the available dataset using DCGAN. Next, it separates blood vessels from noise by training residual U-Net. Finally, it uses the Otsu thresholding method to re-binarize the images to enhance their robustness.
[0073] See Figure 3 and Figure 4 An image noise removal method based on AS-OCTA may include the following steps:
[0074] Step S0: Manually simulate eye movements using AS-OCTA to acquire a large number of real images with pure noise and a small number of real images without noise. In this embodiment, step S0 uses AS-OCTA to manually simulate eye movements to acquire a large number of real images with pure noise and a small number of real images without noise. This method of acquiring noise information is more consistent with the real denoised images, making the noise used in the next learning step more reliable.
[0075] Step S1: Input the purely noisy image into the improved Generative Adversarial Network (DCGAN) to generate an image that approximates the manually obtained purely noisy image. See also Figure 1 In this embodiment, the semi-supervised DCGAN mechanism reduces the labor of manual data collection, effectively expands the dataset, solves the problem of insufficient dataset quantity, and provides support for subsequent work.
[0076] Step S2: Fuse each noise-free image with the generated pure-noise image, input the result into the residual U-Net neural network, separate the blood vessels from the noise, and reconstruct a new image. In this embodiment, the residual U-Net network structure is relatively shallow, and the model can be trained to the ideal effect based on the dataset obtained in step S1, separating most blood vessels from noise with high efficiency and accuracy. The residual module in the model helps it extract features better and avoids gradient vanishing.
[0077] Step S3: Binarize the image output from the residual U-net using the Otsu thresholding method. In this embodiment, the denoised image is input into the Otsu thresholding module to find the optimal threshold, and the image is then re-binarized before being output. This helps to highlight blood vessels, reduce white noise, and enhance image robustness.
[0078] As an optional implementation, in some embodiments, step S1 may include:
[0079] S1-0: The input is 1000 collected 400*400 pure noise RGB images to train DCGAN.
[0080] The improved generative adversarial network used has a generator with a five-layer deconvolution structure, and the loss is defined as:
[0081]
[0082] Where G(z) represents the output image of the generator, and D(G(z)) represents the input probability of the discriminator.
[0083] The discriminator uses a 5-layer convolutional structure, and its loss is defined as:
[0084]
[0085] Where D(x) represents the discriminator probability of the sample noise map, and the optimization objective of the training process is expressed as:
[0086]
[0087] Specifically, the training process described above consists of two stages: In the first stage, keeping the generator parameters unchanged, the discriminator network parameters are trained and updated. Random noise is input into the generator G, and after passing through a convolutional neural network, generated images are obtained. The label for sample images is set to 1, and the label for non-sample images is set to 0. The generator output image and the sample images are used simultaneously as the training dataset for the discriminator. The network is trained by calculating the gradient error between the two classes of images to update the discriminator parameters. In the second stage, keeping the discriminator parameters unchanged, the generator is trained. Through backpropagation of the discriminator parameters, random noise passes through the generator to obtain generated images. To make the generator output image fit the direction of the sample image, the label of this image is set to 1, and the generator parameters are updated to improve the generator's image generation capability.
[0088] S1-1: After training is complete, the network outputs 6400 grayscale images of 400*400 pixels that are essentially photographed manually and contain only noise.
[0089] As an optional implementation, in some embodiments, step S2 includes:
[0090] S2-0: Merge the noise-free image with the generated pure noise image to produce a noise-free-pure noise RGB image pair.
[0091] S2-1: Retain a small portion of the data and use the remaining majority of noise-free to pure noise images to train the input residual U-Net.
[0092] Specifically, the residual U-Net used comprises four parts: an encoding part, a decoding part, a skip connection part, and a residual module part. The encoding part contains five stages, each including pooling layers and convolutional layers, with the convolutional layers also containing normalization and PReLU activation functions. The decoding part consists of seven stages; the first four stages are used for decoding. Correspondingly, each decoding stage includes deconvolution, normalization, and PReLU activation function steps, and finally restores the same number of channels as the input image through convolution. The skip connection part connects the low-level features of the encoding part with the high-level features of the decoding part.
[0093] The model uses the mean squared error loss function, defined as:
[0094]
[0095] Among them, Y m y represents the true value of the sample. m This represents the predicted value of the sample output by the network, where M is the number of samples.
[0096] S2-2: After training, use a subset of noise-free to pure noise RGB images for verification.
[0097] S2-3: After verifying successful network training, input a noisy image and output a reconstructed image with blood vessels separated from noise.
[0098] As an optional implementation, in some embodiments, step S3 includes:
[0099] A 400x400 denoised image is input into the Otsu thresholding module. The Otsu thresholding module recursively applies minimizing the intra-class variance to the image, automatically finding the globally and locally optimal threshold. Otsu aims to find the optimal threshold for the image and then re-binarize the image based on the threshold. The output is a 400x400 Otsu thresholding binarized denoised image. The objective function of the Otsu thresholding method is as follows:
[0100] G(t)=W0*(U0-U) 2 +W1*(U1-U) 2
[0101] Where U is the overall mean of the image, W0 and U0 refer to the proportion of foreground pixels in the image and the mean of foreground pixels, respectively; similarly, W1 and U1 refer to the proportion of background pixels in the image and the mean of background pixels, respectively. When the objective function reaches its maximum, t is the optimal threshold.
[0102] See Table 1 and Figure 2 Table 1 compares the calculated PSNR and SSMI values for various denoising methods. The ResNet, UNet+OTSU algorithm results are presented. PSNR measures the image distortion and noise levels between two images. SSIM is a metric that evaluates the structural consistency between a target image and a reference image by fusing information such as brightness, contrast, and structure. Higher PSNR and SSMI values indicate better denoising performance.
[0103] Table 1. Comparison of denoising results for AS-OCTA images using different algorithms and protocols
[0104]
[0105] In the table: FFT = Fast Fourier Transform; PSNR = Peak Signal-to-Noise Ratio; dB = Decibels; SSIM = Structural Similarity Index; WNN = Wavelet Neural Network.
[0106] All values are represented as medians (interquartile range). The rank-sum test was used to compare the differences in denoising results of different algorithms and protocols. The bolded content indicates P < 0.05.
[0107] See Figure 5 For example, the specific steps of the method of the present invention include:
[0108] Step S0: Manually collect 1000 pure noise AS-OCTA images and 30 noise-free images, all in RGB format with a resolution of 400*400.
[0109] Step S1: Input the purely noisy image into the improved Generative Adversarial Network (DCGAN) to generate an image that approximates the manually captured purely noisy image. This includes:
[0110] S1-0: The input is 1000 collected 400*400 pure noise RGB images. The DCGAN is trained to optimize the objective function F(G,D).
[0111] S1-1: After training is complete, the network outputs 6400 pure noise grayscale images of approximately 400*400 pixels, which are almost manually collected.
[0112] Step S2: In this embodiment, blood vessels are separated from noise using a residual U-Net neural network and reconstructed into a new image. This includes:
[0113] Step S2-0: Obtain 6400 pure noise grayscale images from step S1, and merge these images with 30 manually collected noise-free images to generate 192000 training image pairs.
[0114] Step S2-1: Retain 2000 data pairs, and input the remaining 190000 image pairs into the residual U-Net neural network to train the network. The label for the noise-free image is Y. m The network outputs the predicted label y m .
[0115] Step S2-2: After training is complete, input the 2000 pairs of data that have been retained to verify the network training effect.
[0116] Step S2-3: After verifying the effect, input the original data to obtain the reconstructed image after separating blood vessels and noise.
[0117] Step S3: In this embodiment, the 400*400 image obtained from S2-3, after separating the blood vessel noise, is re-binarized using the Otsu thresholding method to enhance the blood vessels and suppress white noise. Thus, the final denoised 400*400 image is obtained.
[0118] Meanwhile, the present invention also provides an image noise removal system based on AS-OCTA, which includes:
[0119] The first processing unit is used to simulate eye movement via AS-OCTA to acquire a first set of image data with pure noise and a second set of image data without noise.
[0120] The second processing unit is used to input each image of the first image dataset into the improved generative adversarial network DCGAN model to generate several third image datasets that are similar to the first image dataset.
[0121] The third processing unit is used to fuse each image from the second image dataset with each image from the generated third image dataset, and input the fusion into a residual U-Net neural network model to separate blood vessels from noise in the image data, thereby reconstructing a fourth image dataset; and,
[0122] The fourth processing unit is used to binarize each image in the fourth image dataset using the Otsu thresholding method to obtain the fifth image dataset.
[0123] This invention directly acquires real pure artifact images by simulating eye movements using AS-OCTA and extracts the unique noise characteristics of the images; it learns the noise distribution through an improved Generative Adversarial Network (DCGAN) and expands the dataset; it uses U-Net to segment noise and blood vessels; and it uses the Otsu Thresholding Unit (OSTU) module to enhance the image and suppress noise, thereby obtaining clear blood vessel images that meet the requirements of clinical analysis.
[0124] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the present invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples. Moreover, without contradiction, those skilled in the art can combine and integrate the different embodiments or examples described in this specification, as well as the features of different embodiments or examples.
[0125] The above embodiments are merely illustrative of the technical concept and features of the present invention, and are intended to enable those skilled in the art to understand the content of the present invention and implement it accordingly. They should not be construed as limiting the scope of protection of the present invention. All equivalent changes or modifications made based on the essence of the content of the present invention should be covered within the scope of protection of the present invention.
Claims
1. An image noise removal method based on AS-OCTA, characterized in that, Includes the following steps: By artificially simulating eye movements using AS-OCTA, several sets of first image data with pure noise and several sets of second image data without noise are obtained. Each image from the first image dataset is input into an improved Generative Adversarial Network (DCGAN) model to generate several third image datasets that are similar to the first image dataset; wherein, in the improved DCGAN model, The generator uses a five-layer deconvolution structure, and the loss is defined as: Where G(z) represents the output image of the generator, and D(G(z)) represents the input probability of the discriminator; The discriminator uses a 5-layer convolutional structure, and its loss is defined as: Where D(x) represents the discriminator probability of the sample noise map, The optimization objective of the training process is expressed as: Each image from the second image dataset is fused with each image from the generated third image dataset and input into the residual U-Net neural network model to separate blood vessels from noise in the image data, thereby reconstructing a fourth image dataset. Each image in the fourth image dataset is binarized using the Otsu thresholding method to obtain the fifth image dataset.
2. The image noise removal method based on AS-OCTA according to claim 1, characterized in that, The training process using the improved Generative Adversarial Network (DCGAN) model is as follows: Keeping the parameters of the generator unchanged, the network parameters of the discriminator are trained and updated. Random noise is input into the generator and then passed through a convolutional neural network to obtain the generated image. The label of the sample image is set to 1, and the label of the non-sample image is set to 0. The output image of the generator and the sample image are used as the training dataset of the discriminator for network training. The parameters of the discriminator are updated by calculating the gradient error between the two types of images. Keeping the parameters of the discriminator unchanged, the generator is trained, wherein random noise is backpropagated through the parameters of the discriminator and generated by the generator. In order to make the output of the generator fit the direction of the sample image, the image label at this time is recorded as 1, and the parameters of the generator are updated.
3. The image noise removal method based on AS-OCTA according to claim 1, characterized in that, It also includes the step of training the residual U-Net neural network model, specifically, Each image in the second image dataset is fused with each image in the generated third image dataset to generate several noise-free and noise-only image pairs. The noise-free to pure noise images are used as inputs to train the residual U-Net neural network model; After training, the residual U-Net neural network model is validated using the remaining noise-free to pure noise images. Repeat the above steps until verification is successful; The residual U-Net neural network model includes an encoding part, a decoding part, a skip connection part, and a residual module part.
4. The image noise removal method based on AS-OCTA according to claim 1, characterized in that, Binarization is performed using Otsu's thresholding method. Specifically, Otsu's thresholding method is used to find the optimal threshold for the image data according to the following formula, and then the image is re-binarized based on the threshold: G(t)=W0*(U0-U) 2 +W1*(U1-U) 2 Where G(t) is the objective function, U is the mean of the entire image, W0 and U0 refer to the proportion of foreground pixels in the image and the mean of foreground pixels, and W1 and U1 refer to the proportion of background pixels in the image and the mean of background pixels.
5. An image noise removal system based on AS-OCTA, characterized in that, include: The first processing unit is used to simulate eye movement via AS-OCTA to acquire a first set of image data with pure noise and a second set of image data without noise. The second processing unit is used to input each image from the first image dataset into an improved generative adversarial network (DCGAN) model to generate several third image datasets that are similar to the first image dataset; wherein, in the improved DCGAN model, The generator uses a five-layer deconvolution structure, and the loss is defined as: Where G(z) represents the output image of the generator, and D(G(z)) represents the input probability of the discriminator; The discriminator uses a 5-layer convolutional structure, and its loss is defined as: Where D(x) represents the discriminator probability of the sample noise map, The optimization objective of the training process is expressed as: The third processing unit is used to fuse each image from the second image dataset with each image from the generated third image dataset, and input the fusion into a residual U-Net neural network model to separate blood vessels from noise in the image data, thereby reconstructing a fourth image dataset; and, The fourth processing unit is used to binarize each image in the fourth image dataset using the Otsu thresholding method to obtain the fifth image dataset.
6. The image noise removal system based on AS-OCTA according to claim 5, characterized in that, The training process using the improved Generative Adversarial Network (DCGAN) model is as follows: Keeping the parameters of the generator unchanged, the network parameters of the discriminator are trained and updated. Random noise is input into the generator and then passed through a convolutional neural network to obtain the generated image. The label of the sample image is set to 1, and the label of the non-sample image is set to 0. The output image of the generator and the sample image are used as the training dataset of the discriminator for network training. The parameters of the discriminator are updated by calculating the gradient error between the two types of images. Keeping the parameters of the discriminator unchanged, the generator is trained, wherein random noise is backpropagated through the parameters of the discriminator and generated by the generator. In order to make the output of the generator fit the direction of the sample image, the image label at this time is recorded as 1, and the parameters of the generator are updated.
7. The image noise removal system based on AS-OCTA according to claim 5, characterized in that, It also includes the step of training the residual U-Net neural network model, specifically, Each image in the second image dataset is fused with each image in the generated third image dataset to generate several noise-free and noise-only image pairs. The noise-free to pure noise images are used as inputs to train the residual U-Net neural network model; After training, the residual U-Net neural network model is validated using the remaining noise-free to pure noise images. Repeat the above steps until verification is successful; The residual U-Net neural network model includes an encoding part, a decoding part, a skip connection part, and a residual module part.
8. The image noise removal system based on AS-OCTA according to claim 5, characterized in that, Binarization is performed using Otsu's thresholding method. Specifically, Otsu's thresholding method is used to find the optimal threshold for the image data according to the following formula, and then the image is re-binarized based on the threshold: G(t)=W0*(U0-U) 2 +W1*(U1-U) 2 Where G(t) is the objective function, U is the mean of the entire image, W0 and U0 refer to the proportion of foreground pixels in the image and the mean of foreground pixels, and W1 and U1 refer to the proportion of background pixels in the image and the mean of background pixels.