OCTA image motion artifact and noise parallel removal method based on self-supervised learning

Abstract

An OCTA image motion artifact and noise parallel removal method based on self-supervised learning comprises the steps of constructing an OCTA image data set, learning a mapping relation between a low-quality image and a high-quality image by using a deep learning method, and further constructing an OCTA image motion artifact and noise parallel removal model. And inputting a low-quality image into the model to realize OCTA image motion artifact removal and noise reduction. According to the method, the OCTA image data set is constructed through an image processing and calculation method, the complexity of data set construction is reduced, and the diversity and representativeness of the data set are enhanced. Model construction is a self-supervision mode, and the processing capacity of the model on complex image features is enhanced by combining a window, a channel attention mechanism and a multi-scale noise reduction module. A composite loss function is constructed by using wavelet transform and a mean square error loss function, and the recognition and removal capability of the model on artifacts and noise is optimized.

Description

Technical Field

[0001] The present invention belongs to the field of image processing and imaging technology, and in particular relates to a method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning. Background Art

[0002] Optical coherence tomography angiography (OCTA) can image microvascular networks and has the advantages of high spatial resolution and high sensitivity, making it widely used in the medical field. OCTA imaging can provide detailed tissue information of the sample, which is indispensable for the accurate diagnosis of diseases and real-time monitoring of the treatment process. Its image quality plays a key role in the clinical visualization of early disease characteristics and can provide clinicians with clearer and more accurate diagnostic evidence. However, during the testing of living subjects, due to the presence of muscle tremors, heartbeats, and breathing of the tested subjects, a large number of motion artifacts will be generated in the imaging, resulting in a decrease in image quality, affecting clinical diagnosis and quantitative image analysis.

[0003] Hardware-based approaches offer an efficient solution for removing motion artifacts from OCTA images. This approach requires improvements and optimizations to the OCT system's imaging architecture, including increasing detection speed and introducing a motion monitoring module to compensate and correct for subtle movements of the subject in real time. This significantly reduces motion artifacts and produces high-quality images. However, this approach complicates the system architecture and increases equipment costs. Compared to hardware-based approaches, software-based approaches offer an alternative and flexible solution. For example, the registration-based method (Wei DW, Deegan AJ, Wang R K. Automatic motion correction for in vivo human skin optical coherence tomography angiography through combined rigid and nonrigid registration [J]. Journal of Biomedical Optics, 2017, 22 (6): 066013-066013.) (Camino A, Zhang M, Gao SS, et al. Evaluation of artifact reduction in optical coherence tomography angiography with real-time tracking and motion correction technology [J]. Biomedical optics express, 2016, 7 (10): 3905-3915.) uses repeated motion-artifact-free image information at the same position to replace motion artifacts on the image. However, this method requires a large number of repeated image acquisitions and a relatively long imaging time, which reduces the practicality of many clinical applications.

[0004] At the same time, speckle noise present in coherent imaging is one of the main sources of noise in OCTA images. Speckle noise primarily consists of randomly distributed granular textures within the image, which can easily obscure tissue features and boundaries, resulting in reduced image contrast. To mitigate the effects of speckle noise, researchers have employed various filtering methods or multiple sampling techniques for speckle noise reduction. However, filtering methods can blur features and edge information within the image, while multiple sampling increases imaging time and the likelihood of motion artifacts. Summary of the Invention

[0005] The primary objective of this invention is to overcome the shortcomings of some of the aforementioned existing technologies by proposing a method for simultaneously removing motion artifacts and noise from OCTA images based on self-supervised learning. This method not only effectively reduces noise in OCTA images but also simultaneously removes motion artifacts caused by object jitter. This method provides a more effective solution for removing motion artifacts and reducing noise in OCTA images, possessing broad application prospects and significant technical value.

[0006] To achieve the above objectives, the present invention adopts the following technical solutions:

[0007] A method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning comprises the following steps:

[0008] Step 1: Construct an OCTA image dataset;

[0009] The OCTA image dataset refers to a dataset consisting of pairs of low-quality OCTA simulated images and high-quality OCTA experimental images. All raw data in this dataset are acquired by an OCT system. First, the OCT system acquires OCT images, then calculates OCTA images using the spectral amplitude decorrelation angiography method, and then manually selects OCTA low-quality experimental images and OCTA high-quality experimental images.

[0010] Filtering and transformation methods are used to extract artifact images and noise images from low-quality OCTA experimental images;

[0011] A set of weights is set for the artifact image and the noise image respectively. The artifact image and the noise image are added to the OCTA high-quality experimental image using a weighted method to obtain the OCTA low-quality simulated image. The OCTA low-quality simulated image and the corresponding OCTA high-quality experimental image are then paired to construct a dataset. The specific construction method is as follows:

[0012] Step 1.1: Scan the same position of the sample N times using the OCT system to obtain N OCT images, and calculate the OCTA experimental image I using the spectral amplitude decorrelation angiography method. OCTA , the calculation formula is:

[0013]

[0014] Among them, the value range of N is between 3 and 8; M is the number of spectrum divisions, and the value range of M is between 5 and 10; I im It represents the intensity map corresponding to the mth narrow-band spectrum after spectral segmentation of the i-th repeatedly acquired OCT image;

[0015] Step 1.2: Scan different samples or different locations of the sample to obtain a set of K OCTA experimental images I OCTA , the value range of K is a natural number greater than or equal to 500;

[0016] Step 1.3: K OCTA experimental images I obtained from step 1.2 OCTA Through observation, Q images without motion artifacts and noise were manually selected, and these images were named OCTA high-quality experimental images I. label , the value range of Q is a natural number greater than or equal to 100 and less than or equal to 200; at the same time, Q images with motion artifacts and noise are selected, and these images are named OCTA low-quality experimental images I raw ;

[0017] Step 1.4: Q OCTA low-quality experimental images I obtained in step 1.3 raw Perform image transformation and filtering to remove some motion artifacts and obtain Q artifact-reduced images I remove_artifact ; Wherein the image transformation is not limited to Fourier transform and wavelet transform, and the filtering is not limited to high-pass filtering, low-pass filtering and band-stop filtering;

[0018] Step 1.5: Subtract the artifacts from the Q images I obtained in step 1.4. remove_artifact Perform noise reduction processing to obtain Q noise-reduced and artifact-reduced images I remove_all ; The noise reduction processing includes but is not limited to mean filtering noise reduction, median filtering noise reduction and Gaussian filtering noise reduction;

[0019] Step 1.6: The Q OCTA low-quality experimental images I obtained in step 1.3 are respectively raw and the Q-th artifact-subtracted image I obtained in step 1.4 remove_artifact Subtract and get Q motion artifact images I artifact , the calculation formula is:

[0020] I artifact =I raw -I remove_artifact (2)

[0021] Step 1.7: respectively perform the Q artifact-reduced images I obtained in step 1.4 on the remove_artifact and the Q-reduced noise-reduced artifact-reduced image I obtained in step 1.5 remove_all Subtract and get Q noise images I noise , the calculation formula is as follows:

[0022] I noise =I remove_artifact -I remove_all (3)

[0023] Step 1.8: Take the Q motion artifact images I obtained in step 1.6 artifact and the Q noisy images I obtained in step 1.7 noise After multiplying the motion artifact image weight α and the noise image weight β respectively, randomly add the Q OCTA high-quality experimental images I obtained in step 1.3 label Add together to get n1×n2×Q OCTA low-quality simulated images I artifact_noise ;Each OCTA high-quality experimental image I label Can generate n1×n2 OCTA low-quality simulated images I artifact_noise ; The value range of α and β is a real number between 1 and 12, n1 and n2 are the number of values of α and β respectively, and the calculation formula is as follows:

[0024] I artifact_noise =I label +αI artifact +βI noise (4)

[0025] Step 1.9: Convert the OCTA high-quality experimental images described in step 1.3 into label and the corresponding OCTA low-quality simulated image I in step 1.8 artifact_noise Pairing, each OCTA high-quality experimental image I label Corresponding to n1×n2 OCTA low-quality simulated images I artifact_noise , generate n1×n2 image pairs; a total of n1×n2×Q image pairs of OCTA image datasets are generated;

[0026] Step 2: Construct a parallel model for removing motion artifacts and noise from OCTA images;

[0027] Using the OCTA image dataset constructed in step 1, a mapping relationship between low-quality simulated OCTA images and high-quality experimental images is learned through a deep learning method to construct the OCTA image motion artifact and noise parallel removal model; the model includes a feature extraction module, a parallel motion artifact and noise removal module, and a mapping module; the feature extraction module is used to extract image features; the parallel motion artifact and noise removal module is used to remove motion artifacts and noise in the image; and the mapping module is used to output a motion artifact removed and denoised image;

[0028] The motion artifact and noise parallel removal module includes a motion artifact removal submodule (MARSM) and an adaptive multi-scale denoising submodule (AMSDSM). The motion artifact removal submodule is used to remove motion artifacts in the image, including a window attention module and a channel attention module. The adaptive multi-scale denoising submodule is used to remove noise in the image.

[0029] In some embodiments, the feature extraction module uses three serially connected convolutional layers, normalization layers, and activation function layers to extract shallow image information;

[0030] In some embodiments, the motion artifact removal submodule first uses a normalization layer to normalize the image, the normalization layer is followed by an attention mechanism module, and then the normalization layer is connected and a densely connected network is used to further extract image features; the densely connected network is followed by a normalization layer and a channel attention mechanism module, and finally the normalization layer and the densely connected network are connected, and residual connections are used between each module.

[0031] In some embodiments, the channel attention module first uses a convolution layer, a normalization layer, and an activation function layer to extract features of different channels; after the activation function layer, the convolution layer is connected again to extract image detail features; then it is divided into two parallel paths and an adaptive average pooling layer and an adaptive maximum pooling layer are used to further extract image features, and after the adaptive average pooling layer and the adaptive maximum pooling layer, the convolution layer, the activation function layer, and the convolution layer are connected respectively, and then the two feature extraction graphs are fused; after the two branches are fused, the activation function is connected to remove unnecessary image features, and finally a residual connection is used.

[0032] In some embodiments, the adaptive multi-scale denoising submodule stacks three convolutional layers in sequence to capture the diversity of local features through convolution kernels of different scales; then a self-attention layer is introduced, followed by a gated convolution to dynamically adjust the importance of features, and then another convolutional layer is used to further compress the channel, and then an activation function is connected; the network introduces another convolutional layer at the end, and the residual layer extracts the final image features.

[0033] In some embodiments, the mapping module first connects a deconvolution layer, a normalization layer, and an activation function layer to preliminarily restore the spatial resolution of the image, then connects a deconvolution layer, a normalization layer, and an activation function layer again to further restore the image details, then connects a convolution layer, and finally connects a residual layer to output a motion artifact-free denoised image.

[0034] The OCTA image motion artifact and noise parallel removal model uses a composite loss function consisting of a weighted wavelet transform loss function based on the mean square error loss function to update the model parameters;

[0035] The loss function of the model is the mean square error loss function L MSE Weighted wavelet transform loss function L wavelet The total loss function L full , the calculation formula is as follows:

[0036]

[0037]

[0038] L full =λL MSE +(1-λ)L wavelet (7)

[0039] where x k represents the kth OCTA low-quality simulated image, y k represents the kth OCTA high-quality experimental image, R represents the number of OCTA low-quality simulated images input into the network model, f represents the motion artifact and noise parallel removal network model, DWT is wavelet transform, and λ is the weight factor.

[0040] Step 3: OCTA image motion artifact removal and noise reduction;

[0041] The OCTA images with motion artifacts and noise collected in the experiment are input into the OCTA image motion artifact and noise parallel removal model constructed in step 2, and the OCTA images with motion artifacts and noise removed in parallel are output.

[0042] Beneficial effects of the present invention:

[0043] 1. This paper proposes a scheme for constructing an OCTA image dataset based on image processing and computational methods, which reduces the complexity of dataset acquisition and construction and can effectively improve the diversity and representativeness of the dataset.

[0044] 2. This paper designs an OCT image motion artifact removal and denoising model. By combining a window attention mechanism, a channel attention mechanism, and a multi-scale denoising module, this model significantly improves the model's ability to process complex image features. The window attention mechanism efficiently extracts local features and enhances the model's perception of details; the channel attention mechanism further optimizes the feature extraction process, highlights important channel information, and improves the model's grasp of global features; the multi-scale denoising module effectively removes image noise by processing features at different scales in parallel, while preserving image edges and texture details.

[0045] 3. The present invention utilizes the multi-resolution analysis and time-frequency localization advantages of wavelet transform, combined with the mean square error loss function, to construct a composite loss function. This not only optimizes the model's ability to identify and remove motion artifacts and noise in OCTA images, but also enhances the model's capture of high-frequency details through the wavelet transform loss function, thereby improving the accuracy and efficiency of image restoration.

[0046] 4. The present invention uses a digital method to remove motion artifacts and noise in OCTA images in parallel, which can effectively improve the quality of OCTA images. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 A flow chart of a method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning provided by the present invention;

[0048] Figure 2 A schematic diagram of the training of the OCTA image motion artifact and noise parallel removal model constructed by the present invention;

[0049] Figure 3 This is a schematic diagram of the structure of the motion artifact removal module constructed in the present invention;

[0050] Figure 4 Schematic diagram of the channel attention module structure constructed for the present invention;

[0051] Figure 5 A schematic diagram of the structure of the adaptive multi-scale noise reduction module constructed in the present invention;

[0052] Figure 6 The low-quality OCTA model image of the human conjunctiva obtained by the present invention contains motion artifacts and noise;

[0053] Figure 7 The motion artifact image obtained by the present invention;

[0054] Figure 8 The noise image obtained by the present invention;

[0055] Figure 9 This is the high-quality experimental image of human conjunctiva OCTA obtained by the present invention;

[0056] Figure 10 This is the optimized OCTA image of the human conjunctiva. DETAILED DESCRIPTION

[0057] The realization of the purpose, functional features and advantages of the present invention are further described below with reference to the accompanying drawings.

[0058] A method for parallel removal of motion artifacts and noise in OCTA images based on self-supervised learning, the flow chart of which is as follows: Figure 1As shown, the following steps are included:

[0059] Step 1: Construct an OCTA image dataset;

[0060] The OCTA image dataset is composed of pairs of low-quality simulated OCTA images and high-quality experimental OCTA images. All raw data in this dataset were acquired using an OCT system. First, the OCT system acquired OCT images, then spectral amplitude decorrelation angiography was used to calculate OCTA images. Subsequently, manual selection was used to obtain low-quality experimental OCTA images and high-quality experimental OCTA images.

[0061] Filtering and transformation methods are used to extract artifact images and noise images from low-quality OCTA experimental images;

[0062] A set of weights is set for the artifact image and the noise image respectively. The artifact image and the noise image are added to the OCTA high-quality experimental image using a weighted method to obtain the OCTA low-quality simulated image. The OCTA low-quality simulated image and the corresponding OCTA high-quality experimental image are then paired to construct a dataset. The specific construction method is as follows:

[0063] Step 1.1: Scan the same position of the sample N times using the OCT system to obtain N OCT images, and calculate the OCTA experimental image I using the spectral amplitude decorrelation angiography method. OCTA , the calculation formula is:

[0064]

[0065] Among them, the value range of N is between 3 and 8; M is the number of spectrum divisions, and the value range of M is between 5 and 10; I im represents the intensity map corresponding to the mth narrow-band spectrum after spectral segmentation of the i-th repeatedly acquired OCT image; in this embodiment, N is 5 and M is 8;

[0066] Step 1.2: Scan different samples or different locations of the sample to obtain a set of K OCTA experimental images I OCTA , the value range of K is a natural number greater than or equal to 500. In this embodiment, K is 500;

[0067] Step 1.3: K OCTA experimental images I obtained from step 1.2 OCTA Through observation, Q images without motion artifacts and noise were manually selected, and these images were named OCTA high-quality experimental images I. label, the value range of Q is a natural number greater than or equal to 100 and less than or equal to 200; at the same time, Q images with motion artifacts and noise are selected, and these images are named OCTA low-quality experimental images I raw ; In this embodiment, Q is 100;

[0068] Step 1.4: Q OCTA low-quality experimental images I obtained in step 1.3 raw Perform image transformation and filtering to remove some motion artifacts and obtain Q artifact-reduced images I remove_artifact ; Wherein image transformation includes but is not limited to Fourier transform and wavelet transform, filtering includes but is not limited to high-pass filtering, low-pass filtering and band-stop filtering; in this embodiment, image transformation adopts Fourier transform, and filtering adopts band-stop filtering;

[0069] Step 1.5: Subtract the artifacts from the Q images I obtained in step 1.4. remove_artifact Perform noise reduction processing to obtain Q noise-reduced and artifact-reduced images I remove_all ; The noise reduction processing includes but is not limited to mean filtering noise reduction, median filtering noise reduction and Gaussian filtering noise reduction; in this embodiment, the noise reduction adopts mean filtering noise reduction;

[0070] Step 1.6: The Q OCTA low-quality experimental images I obtained in step 1.3 are respectively raw and the Q-th artifact-subtracted image I obtained in step 1.4 remove_artifact Subtract and get Q motion artifact images I artifact , the calculation formula is:

[0071] I artifact =I raw -I remove_artifact (2)

[0072] Step 1.7: respectively perform the Q artifact-reduced images I obtained in step 1.4 on the remove_artifact and the Q-reduced noise-reduced artifact-reduced image I obtained in step 1.5 remove_all Subtract and get Q noise images I noise , the calculation formula is as follows:

[0073] I noise =I remove_artifact -I remove_all (3)

[0074] Step 1.8: Take the Q motion artifact images I obtained in step 1.6 artifact and the Q noisy images I obtained in step 1.7 noise After multiplying the motion artifact image weight α and the noise image weight β respectively, randomly add the Q OCTA high-quality experimental images I obtained in step 1.3 labelAdd together to get n1×n2×Q OCTA low-quality simulated images I artifact_noise ;Each OCTA high-quality experimental image I label Can generate n1×n2 OCTA low-quality simulated images I artifact_noise ; The value range of α and β is a real number between 1 and 12, n1 and n2 are the number of values of α and β respectively, and the calculation formula is as follows:

[0075] I artifact_noise =I label +αI artifact +βI noise (4)

[0076] In this embodiment, α and β take the same value, which is 1, 2, 3, 4, 5, 6, 7, 8, 9, and 10 respectively; n1 and n2 are both 10;

[0077] Step 1.9: Convert the OCTA high-quality experimental images from step 1.3 into label and the corresponding OCTA low-quality simulated image I in step 1.8 artifact_noise Pairing, each OCTA high-quality experimental image I label Corresponding to n1×n2 OCTA low-quality simulated images I artifact_noise , generate n1×n2 image pairs; a total of n1×n2×Q image pairs of OCTA image datasets are generated;

[0078] In this embodiment, α and β take the same value, that is, α=β. Since the parameters α and β change synchronously, each OCTA high-quality experimental image I label Generate n1 OCTA low-quality simulated images I artifact_noise , generating a total of n1×Q pairs of images. In this example, Q is 100 and n1 is 10. Finally, a total of 1000 pairs of images with a pixel size of 300×300 are obtained, of which 900 pairs are used for model training and 100 pairs are used for model testing.

[0079] Step 2: Construct a parallel model for removing motion artifacts and noise from OCTA images;

[0080] Using the OCTA image dataset constructed in step 1, the mapping relationship between low-quality simulated OCTA images and high-quality experimental images is learned through deep learning methods to construct a parallel removal model for OCTA image motion artifacts and noise. The model includes a feature extraction module, a parallel motion artifact and noise removal module, and a mapping module. The feature extraction module is used to extract image features; the parallel motion artifact and noise removal module (such as Figure 2 ) is used to remove motion artifacts and noise in an image; the mapping module is used to output a motion artifact removed and denoised image;

[0081] The motion artifact and noise parallel removal module includes a motion artifact removal submodule (MARSM) and an adaptive multi-scale denoising submodule (AMSDSM). The MARSM submodule is used to remove motion artifacts from images and includes a window attention module and a channel attention module. The AMSDSM submodule is used to remove noise from images.

[0082] The feature extraction module uses three convolutional layers, normalization layers and activation function layers in series to extract shallow image information. The feature extraction module structure of this embodiment is as follows: Figure 2 As shown in , it includes three convolutional layers Conv, each of which is followed by a normalization layer IN and an activation function Relu layer;

[0083] The motion artifact removal submodule first uses a normalization layer to normalize the image, and the normalization layer is followed by an attention mechanism module; then the normalization layer is connected and a densely connected network is used to further extract image features; the densely connected network is followed by a normalization layer and a channel attention mechanism module, and finally the normalization layer and the densely connected network are connected, and residual connections are used between each module.

[0084] The structure of the motion artifact removal submodule in this embodiment is as follows: Figure 3As shown, it includes four conditional normalization layers CLN. The first CLN is followed by a window multi-head attention mechanism module W-MSA, which is used to extract image features within the window. The window size is set to 3, and the two layers use residual connections; the second CLN layer is followed by a dense connection Dense to further extract image features, and the two layers are connected using residual connections; the third CLN layer is followed by a channel attention module CAM, which captures different features through multiple channels, and the two layers use residual connections; the fourth CLN layer is followed by a Dense to comprehensively summarize the extracted image features, and the two layers also use residual connections.

[0085] The channel attention mechanism module first uses the convolution layer, normalization layer, and activation function layer to extract the features of different channels; after the activation function layer, the convolution layer is connected again to extract the image detail features; then it is divided into two parallel paths and uses the adaptive average pooling layer and the adaptive maximum pooling layer to further extract the image features. After the adaptive average pooling layer and the adaptive maximum pooling layer, the convolution layer, the activation function layer, and the convolution layer are connected respectively, and then the two feature extraction graphs are fused; after the two branches are fused, the activation function is connected to remove unnecessary image features, and finally the residual connection is used.

[0086] The CAM structure of this embodiment is as follows Figure 4 As shown, it first includes two convolutional layers. The first convolutional layer is followed by a normalization layer LN and an activation function ReLU layer; the second convolutional layer is followed by two pooling layers in parallel. The pooling layers in this example are the adaptive average pooling layer Adaptive AvgPool and the adaptive maximum pooling layer Adaptive MaxPool. These two layers extract important features in the image based on the average and maximum values in the convolution kernel, respectively. The size of the convolution kernel is 3; the two pooling layers are followed by two convolutional layers Conv and an activation function layer ReLU in the middle. The output results are then merged and connected to an activation function Sigmoid layer. At the same time, the beginning and end of the CAM are connected using a residual connection.

[0087] The adaptive multi-scale denoising submodule stacks three convolutional layers in sequence, capturing the diversity of local features through convolution kernels of different scales; then a self-attention layer is introduced, followed by gated convolution to dynamically adjust the importance of features, and then another convolutional layer is used to further compress the channel, and then an activation function is connected; the network introduces another convolutional layer at the end, and the residual layer extracts the final image features.

[0088] The structure of the adaptive multi-scale noise reduction module in this embodiment is as follows: Figure 5As shown in the figure, it first includes three convolutional layers with different convolution kernel sizes. These three convolutional layers are connected in parallel to extract image features of different scales. Then they are connected to the self-attention module, which captures the global dependencies based on the self-attention mechanism to extract important features of the image. A gated convolutional layer Gate Conv is further connected to dynamically adjust the contribution of different features and improve the expression ability of the network. Then, two convolutional layers Conv and a middle activation function layer ReLU are connected. At the same time, the beginning and end of the adaptive multi-scale denoising module are connected using residual connections.

[0089] The mapping module first connects the deconvolution layer, normalization layer and activation function layer to initially restore the spatial resolution of the image, then connects the deconvolution layer, normalization layer and activation function layer again to further restore the image details, then connects a convolution layer, and finally connects the residual layer to output a motion artifact-free denoised image.

[0090] The mapping module structure of this embodiment is as follows Figure 2 As shown in , it includes two deconvolution layers DeConv connected in series, and these two DeConvs are followed by a normalization layer IN and an activation function layer ReLU respectively. These two DeConvs restore the low-resolution feature map to a high-resolution feature map; finally, a convolution layer is connected, and the output feature map of the convolution layer and the initial input image form a residual connection;

[0091] The OCTA image motion artifact and noise parallel removal model uses a composite loss function consisting of a weighted wavelet transform loss function based on the mean square error loss function to update the model parameters;

[0092] The loss function of the model is the mean square error loss function L MSE Weighted wavelet transform loss function L wavelet The total loss function is calculated as follows:

[0093]

[0094]

[0095] L full =λL MSE +(1-λ)L wavelet (7)

[0096] where x k represents the kth OCTA low-quality simulated image, y k represents the kth OCTA high-quality experimental image, R represents the number of low-quality OCTA simulated images input to the network model, f represents the motion artifact and noise parallel removal network model, DWT is wavelet transform, and λ is the weight factor; in this embodiment, R is 4 and λ is 0.3;

[0097] The convolution kernel size (k), number (n), and step size (s) corresponding to each convolution layer in the embodiment network are marked in the figure.

[0098] This example uses the Adam algorithm to optimize a model for concurrently removing motion artifacts and noise from OCTA images. The network is first trained for 100 epochs at a learning rate of 0.0002, then linearly decays the learning rate to zero over the next 900 epochs, for a total of 1000 epochs. Training and testing of this model for concurrently removing motion artifacts and noise from OCTA images are performed using the deep learning framework Pytorch.

[0099] Step 3: OCTA image motion artifact removal and noise reduction;

[0100] The OCTA images with motion artifacts and noise collected in the experiment are input into the OCTA image motion artifact and noise parallel removal model constructed in step 2, and the OCTA images with motion artifacts and noise removed in parallel are output.

[0101] As attached Figure 6 、 Figure 7 、 Figure 8 , Attachment Figure 9 , Attachment Figure 10 As shown, the optimization results of the OCTA image of the present invention are given. Figure 6 The low-quality simulated OCTA image of the human conjunctiva obtained by the present invention contains motion artifacts and noise. Figure 7 To obtain the motion artifact image, Figure 8 To obtain the noise image, Figure 9 This is the high-quality experimental image of human conjunctiva OCTA obtained by the present invention. Figure 10 This is the optimized OCTA image of the human conjunctiva. The blood vessel structure in the eye can be clearly seen from the optimized image. Compared with the low-quality experimental OCTA image ( Figure 6 ), optimized OCTA images ( Figure 10 ) is closer to the image quality of OCTA high-quality experimental images ( Figure 9 ), which shows that through the method of the present invention, the motion artifacts of the optimized OCTA images are removed, the noise is suppressed, and the image quality is significantly improved.

[0102] When implementing the present invention, the sample is first scanned and imaged by the OCT system to obtain the OCT image of the sample, and then the OCTA low-quality experimental image and the OCTA high-quality experimental image are obtained by calculation and manual selection. The motion artifact image ( Figure 7 ) and the noisy image ( Figure 8), motion artifact images and noise images are added to the high-quality OCTA experimental images at magnifications of 1x, 2x, 3x, 4x, 5x, 6x, 7x, 8x, 9x, and 10x. The resulting low-quality simulated OCTA images and the original high-quality OCTA experimental images form image pairs to construct an OCTA image dataset. A parallel model for removing motion artifacts and noise from OCTA images is then constructed. This model learns the mapping relationship between the low-quality simulated OCTA images and the high-quality experimental OCTA images, and ultimately optimizes the OCTA images using this model. Once the parallel model for removing motion artifacts and noise from OCTA images is constructed, the present invention can achieve motion artifact removal and noise reduction in OCTA images without increasing the hardware complexity of the OCT system, with high speed and portability.

Claims

1. A method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning, characterized in that: The following steps are involved: Step 1: Construct an OCTA image dataset; The OCTA image dataset is composed of pairs of low-quality simulated OCTA images and high-quality experimental OCTA images. All raw data in this dataset were acquired using an OCT system. First, the OCT system acquired OCT images, then spectral amplitude decorrelation angiography was used to calculate OCTA images. Subsequently, manual selection was used to obtain low-quality experimental OCTA images and high-quality experimental OCTA images. Filtering and transformation methods are used to extract artifact images and noise images from low-quality OCTA experimental images; A set of weights is set for the artifact image and the noise image respectively. The artifact image and the noise image are added to the OCTA high-quality experimental image using a weighted method to obtain the OCTA low-quality simulated image. The OCTA low-quality simulated image and the corresponding OCTA high-quality experimental image are then paired to construct a dataset. Step 2: Construct a parallel model for removing motion artifacts and noise from OCTA images; The model includes a feature extraction module, a motion artifact and noise parallel removal module and a mapping module; the feature extraction module is used to extract image features; the motion artifact and noise parallel removal module is used to remove motion artifacts and noise in the image; The mapping module is used to output a motion artifact removed and denoised image; The motion artifact and noise parallel removal module includes a motion artifact removal submodule and an adaptive multi-scale denoising submodule; The motion artifact removal submodule is used to remove motion artifacts in images, including the window attention module and the channel attention module; The adaptive multi-scale denoising submodule is used to remove noise from the image; The OCTA image motion artifact and noise parallel removal model updates the model parameters using a composite loss function consisting of a mean square error loss function and a weighted wavelet transform loss function. Step 3: OCTA image motion artifact removal and noise reduction; The OCTA images with motion artifacts and noise collected in the experiment are input into the OCTA image motion artifact and noise parallel removal model constructed in step 2, and the OCTA images with motion artifacts and noise removed in parallel are output.

2. The method for parallel removal of motion artifacts and noise in OCTA images based on self-supervised learning according to claim 1, characterized in that: The method for constructing the OCTA image dataset described in step 1 is as follows: Step 1.1: Scan the same position of the sample N times using the OCT system to obtain N OCT images, and calculate the OCTA experimental image I using the spectral amplitude decorrelation angiography method. OCTA , the calculation formula is: Among them, the value range of N is between 3 and 8; M is the number of spectrum divisions, and the value range of M is between 5 and 10; I im It represents the intensity map corresponding to the mth narrow-band spectrum after spectral segmentation of the i-th repeatedly acquired OCT image; Step 1.2: Scan different samples or different locations of the sample to obtain a set of K OCTA experimental images I OCTA , the value range of K is a natural number greater than or equal to 500; Step 1.3: K OCTA experimental images I obtained from step 1.2 OCTA Through observation, Q images without motion artifacts and noise were manually selected, and these images were named OCTA high-quality experimental images I. label , the value range of Q is a natural number greater than or equal to 100 and less than or equal to 200; at the same time, Q images with motion artifacts and noise are selected, and these images are named OCTA low-quality experimental images I raw ; Step 1.4: Q OCTA low-quality experimental images I obtained in step 1.3 raw Perform image transformation and filtering to remove some motion artifacts and obtain Q artifact-reduced images I remove_artifact ; Step 1.5: Subtract the artifacts from the Q images I obtained in step 1.

4. remove_artifact Perform noise reduction processing to obtain Q noise-reduced and artifact-reduced images I remove_all ; Step 1.6: respectively perform the Q OCTA low-quality experimental images I obtained in step 1.

3. raw and the Q-th artifact-subtracted image I obtained in step 1.4 remove_artifact Subtract and get Q motion artifact images I artifact , the calculation formula is: I artifact =I raw -I remove_artifact (2) Step 1.7: respectively perform the Q artifact-reduced images I obtained in step 1.4 on the remove_artifact and the Q-reduced noise-reduced artifact-reduced image I obtained in step 1.5 remove_all Subtract and get Q noise images I noise , the calculation formula is as follows: I noise =I remove_artifact -I remove_all (3) Step 1.8: Take the Q motion artifact images I obtained in step 1.6 artifact and the Q noisy images I obtained in step 1.7 noise After multiplying the motion artifact image weight α and the noise image weight β respectively, randomly add the Q OCTA high-quality experimental images I obtained in step 1.3 label Add together to get n1×n2×Q OCTA low-quality simulated images I artifact_noise ;Each OCTA high-quality experimental image I label Can generate n1×n2 OCTA low-quality simulated images I artifact_noise ; The value range of α and β is a real number between 1 and 12, n1 and n2 are the number of values of α and β respectively, and the calculation formula is as follows: I artifact_noise =I label +αI artifact +βI noise (4) Step 1.9: Convert the OCTA high-quality experimental images described in step 1.3 into label and the corresponding OCTA low-quality simulated image I in step 1.8 artifact_noise Pairing, each OCTA high-quality experimental image I label Corresponding to n1×n2 OCTA low-quality simulated images I artifact_noise , generating n1×n2 image pairs; a total of n1×n2×Q OCTA image datasets are generated.

3. The method for parallel removal of motion artifacts and noise in OCTA images based on self-supervised learning according to claim 2, characterized in that: The image transformation in step 1.4 includes Fourier transform and wavelet transform, and the filtering includes high-pass filtering, low-pass filtering and band-stop filtering; The noise reduction process described in step 1.5 includes mean filtering noise reduction, median filtering noise reduction and Gaussian filtering noise reduction.

4. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 1, characterized in that: The feature extraction module described in step 2 uses three concatenated convolutional layers, normalization layers, and activation function layers to extract shallow image information.

5. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 1, characterized in that: The motion artifact removal submodule in step 2 first normalizes the image using a normalization layer, followed by an attention mechanism module. Then connect the normalization layer and use the dense connection network to further extract image features; The densely connected network is followed by a normalization layer and a channel attention mechanism module, and finally the normalization layer and the densely connected network are connected. Residual connections are used between the modules.

6. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 5, characterized in that: The channel attention mechanism module first uses convolutional layers, normalization layers, and activation function layers to extract features of different channels; After the activation function layer, the convolution layer is connected again to extract image detail features; then it is divided into two parallel paths and the adaptive average pooling layer and the adaptive maximum pooling layer are used to further extract image features. After the adaptive average pooling layer and the adaptive maximum pooling layer, the convolution layer, the activation function layer, and the convolution layer are connected respectively, and then the two feature extraction graphs are fused; after the two branches are fused, the activation function is connected to remove unnecessary image features, and finally the residual connection is used.

7. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 1, characterized in that: The adaptive multi-scale denoising submodule described in step 2 stacks three convolutional layers in sequence to capture the diversity of local features through convolution kernels of different scales; Then a self-attention layer is introduced, followed by gated convolution to dynamically adjust the importance of features, and then another convolution layer is used to further compress the channel, and then an activation function is connected; the network introduces another convolution layer at the end, and the residual layer extracts the final image features.

8. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 1, characterized in that: In step 2, the mapping module first connects the deconvolution layer, the normalization layer, and the activation function layer to initially restore the spatial resolution of the image, then connects the deconvolution layer, the normalization layer, and the activation function layer again to further restore the image details, then connects a convolution layer, and finally connects the residual layer to output a motion artifact-free denoised image.

9. The method for concurrently removing motion artifacts and noise from OCTA images based on self-supervised learning according to claim 1, characterized in that: The loss function of the model described in step 2 is the mean square error loss function L MSE Weighted wavelet transform loss function L wavelet The total loss function L full , the calculation formula is as follows: THE full =λL MSE +(1-λ)L wavelet (7) where x k represents the kth OCTA low-quality simulated image, y k represents the kth OCTA high-quality experimental image, R represents the number of OCTA low-quality simulated images input to the network model, f represents the motion artifact and noise parallel removal network model, DWT is wavelet transform, λ is the weight factor.

Images