OCT Denoising Method Based on Self-Supervised Structure Representation Learning

Abstract

This invention proposes an OCT denoising method based on self-supervised structural representation learning. The method includes: constructing a self-supervised OCT denoising model using an encoder and two parallel coupled decoders; the encoder and the two parallel coupled decoders are stacked in an iterative manner with eight layers; the two parallel coupled decoders are a surrogate task decoder and a downstream task decoder, respectively; the surrogate task decoder and the downstream task decoder share the features output by the encoder and are coupled through feature concatenation. This invention, through an innovative multi-view synthetic degradation strategy, can automatically generate training samples with different noise characteristics using only easily obtainable noise-free OCT images, constructing the "input-target" pair required for self-supervised learning; the model is co-trained through parallel structural representation learning and image reconstruction tasks, without the need for any manually labeled noise-clear image pairs or tedious parameter tuning processes.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to an OCT denoising method based on self-supervised structural representation learning. Background Technology

[0002] Optical coherence tomography (OCT) is a non-invasive, high-resolution biomedical imaging technique that has become a core tool for diagnosing ophthalmic diseases. Its imaging principle is based on low-coherence interferometry, enabling visualization of cross-sections of biological tissues at micrometer-level resolution. However, the quality of OCT images is limited by its inherent speckle noise. This noise originates from random interference effects between the coherent light source and tissue scatterers, manifesting as granular textures covering the image. Speckle noise not only severely degrades the visual quality of the image, but more importantly, it can obscure subtle pathological structures (such as the boundaries of retinal layers and minute lesions), thereby impairing the reliability of clinical diagnosis and the accuracy of quantitative analysis.

[0003] In recent years, the rise of deep learning technology has brought revolutionary progress to speckle denoising in OCT. Supervised learning-based methods, by training large-scale, precisely paired "noisy-sharp" image pairs, enable deep networks to learn the complex mapping from noisy to sharp images, significantly outperforming traditional methods in denoising performance. However, obtaining large amounts of high-quality, pixel-level paired training data is extremely difficult and costly in clinical practice. A common alternative is to use multi-frame averaged images as "pseudo" sharp labels, but this process is susceptible to artifacts caused by patient eye movements, leading to inaccurate supervision signals and limiting the upper limit of model performance. Summary of the Invention

[0004] In view of the above situation, the main objective of this invention is to propose an OCT denoising method based on self-supervised structural representation learning to solve the above-mentioned technical problems.

[0005] This invention proposes an OCT denoising method based on self-supervised structural representation learning, the method comprising the following steps:

[0006] Step 1: Construct a self-supervised OCT denoising model using an encoder and two parallel coupled decoders; the encoder and the two parallel coupled decoders are stacked in an iterative manner with eight layers; the two parallel coupled decoders are the proxy task decoder and the downstream task decoder, respectively.

[0007] Step 2: Apply speckle noise of different intensities to the noise-free OCT image, and then stitch together the high-frequency features extracted using a high-pass filter to obtain high-frequency structural features.

[0008] Step 3: The high-frequency structural features are processed sequentially through the eight stacked layers of the encoder in an iterative manner to obtain eight encoder features.

[0009] Step 4: Input the encoder features into the proxy task decoder in an iterative manner for decoding processing to obtain eight proxy task decoder features respectively, and use the features of the eighth layer of the proxy task decoder as the reconstructed high-frequency structure feature map.

[0010] Step 5: Input the encoder features into the downstream task decoder in an iterative manner for processing, and make skip connections with the features of the proxy task decoder to obtain the reconstructed denoised image.

[0011] Step 6: Construct a focal frequency loss function using the reconstructed denoised image, and construct an L1 loss function using the reconstructed high-frequency structural feature map. Then, construct a cycle consistency loss function using the high-frequency features in the reconstructed denoised image and the reconstructed high-frequency structural feature map. Train the self-supervised OCT denoising model using the focal frequency loss function, L1 loss function, and cycle consistency loss function to obtain the trained self-supervised OCT denoising model. Use the downstream task decoder in the trained self-supervised OCT denoising model to perform denoising inference on the noisy OCT image to obtain the final denoised image.

[0012] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0013] 1. This invention effectively overcomes the reliance on manual parameter tuning in traditional methods and the strong demand for paired data in supervised learning, achieving high-quality, high-fidelity self-supervised denoising. Through an innovative multi-view synthetic degradation strategy, this invention automatically generates training samples with different noise characteristics using only readily available noise-free OCT images, constructing the "input-target" pairs required for self-supervised learning. The model is co-trained with the image reconstruction task through parallel structural representation learning, eliminating the need for any manually labeled noise-clear image pairs or tedious parameter tuning. This allows the model to automatically learn the optimal denoising mapping function from the data, significantly suppressing speckle noise while ensuring that the reconstructed image closely approximates the real, clear image in terms of pixel-level fidelity and structural similarity, thus solving the problems of data bottleneck in supervised learning and unstable performance in unsupervised learning.

[0014] 2. This invention overcomes the limitation of existing unsupervised methods that require exposure to or adaptation to the test domain data distribution during the training phase, endowing the model with excellent domain generalization and robustness for clinical deployment. The self-supervised structural representation learning framework proposed in this invention trains entirely based on noise-free images of the source domain and their synthetic degradation, without requiring any model fine-tuning or domain adaptation using data from the target domain (test domain). The network explicitly learns anatomical structural representations invariant to speckle noise (a proxy task) and couples and guides these robust representations to the denoising process (downstream task), enabling the trained model to directly and effectively process OCT images from unknown domains from different scanning devices, different patient groups, or different noise levels. This "training-as-deployment" characteristic greatly simplifies the clinical implementation process and enhances the practical value of the method.

[0015] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by means of embodiments of the invention. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of an OCT denoising method based on self-supervised structural representation learning proposed in this invention.

[0017] Figure 2 This is a diagram showing the overall structure of an OCT denoising method based on self-supervised structural representation learning proposed in this invention. Detailed Implementation

[0018] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0019] These and other aspects of the embodiments of the present invention will become clear from the following description and accompanying drawings. In these descriptions and drawings, some specific embodiments of the present invention are specifically disclosed to illustrate some ways of implementing the principles of the embodiments of the present invention; however, it should be understood that the scope of the embodiments of the present invention is not limited thereto.

[0020] Please see Figure 1 This embodiment provides an OCT denoising method based on self-supervised structural representation learning, the method comprising the following steps:

[0021] Step 1: Construct a self-supervised OCT denoising model using an encoder and two parallel coupled decoders; the encoder and the two parallel coupled decoders are stacked in an iterative manner with eight layers; the two parallel coupled decoders are a proxy task decoder and a downstream task decoder, respectively; the proxy task decoder and the downstream task decoder share the features output by the encoder and are coupled by feature concatenation.

[0022] Step 2: Apply speckle noise of different intensities to the noise-free OCT image, and then stitch together the high-frequency features extracted using a high-pass filter to obtain high-frequency structural features.

[0023] Please see Figure 2 In step 2, speckle noise of different intensities is applied to the noise-free OCT image, and high-frequency features are extracted using a high-pass filter and then stitched together to obtain high-frequency structural features. This process includes the following sub-steps:

[0024] Different intensities of speckle noise were applied to noise-free OCT images to obtain a set of multi-view degraded images. The following relationship exists in the correspondence process:

[0025] ;

[0026] in, Represents a collection of degraded images from multiple perspectives. Indicates the first Degraded images of a single view sample. Indicates the first In the degraded image of the first view sample The value of each pixel. Indicates the index of the view sample. Indicates the index of the pixel. Indicating the first noise-free OCT image The value of each pixel. The simulated synthesis of the first The first of the speckle noise samples The value of each pixel. This represents the total number of categories of speckle noise;

[0027] The speckle noise follows a gamma distribution, expressed as:

[0028] ;

[0029] in, Represents the speckle noise distribution function. A random sample representing the speckle noise distribution. This represents the distributed parameters used to control noise variance. Represents the gamma function. Represents the natural constant;

[0030] High-frequency features of all images in a multi-view degraded image set are extracted using a high-pass filter to obtain the high-frequency image set. The following relationship exists in the correspondence process:

[0031] ;

[0032] in, Represents a set of high-frequency images. Indicates the first High-frequency images of a sample view, This indicates that it has undergone high-frequency extraction processing. This indicates element-wise multiplication. Indicates a high-pass filter. Indicates the kernel size. Indicates standard deviation;

[0033] All images in the high-frequency image set are stitched together along the channel dimension to obtain high-frequency structural features.

[0034] It should be noted that, in Figure 2 middle, Indicates the agent task decoder, This indicates the downstream task decoder.

[0035] Step 3: The high-frequency structural features are processed sequentially through the eight stacked layers of the encoder in an iterative manner to obtain eight encoder features.

[0036] In step 3, the high-frequency structural features are processed iteratively through the eight stacked layers of the encoder to obtain eight encoder features. This process includes the following sub-steps:

[0037] S301. The high-frequency structural features are processed sequentially through a convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the first convolutional feature.

[0038] S302. The first convolutional feature is processed sequentially through a convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the second convolutional feature.

[0039] S303. Perform residual connection between the second convolutional features and the high-frequency structural features to obtain the features of the first layer of the encoder;

[0040] By iteratively taking the features of the first layer of the encoder as input and repeating steps S301 to S303 seven times, the features of the second layer, third layer, fourth layer, fifth layer, sixth layer, seventh layer and eighth layer of the encoder are obtained respectively.

[0041] It should be noted that the kernel size of the convolutional layer is 3×3, and the stride is 1.

[0042] Step 4: Input the encoder features into the surrogate task decoder in an iterative manner for decoding processing to obtain eight surrogate task decoder features respectively, and use the features of the eighth layer of the surrogate task decoder as the reconstructed high-frequency structure feature map.

[0043] In step 4, the encoder features are iteratively input into the surrogate task decoder for decoding to obtain eight surrogate task decoder features. The features of the eighth layer of the surrogate task decoder are then used as the reconstructed high-frequency structure feature map. This process includes the following sub-steps:

[0044] S401. The features of the eighth layer of the encoder are processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the first transposed convolutional features.

[0045] S402. The first transposed convolutional feature is processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the second transposed convolutional feature.

[0046] S403. Perform residual connection between the second transposed convolutional features and the features of the eighth layer of the encoder to obtain the intermediate features of the first layer of the proxy task decoder.

[0047] S404. Perform skip connections between the intermediate features of the first layer of the proxy task decoder and the features of the corresponding layer of the encoder to obtain the features of the first layer of the proxy task decoder.

[0048] By iteratively using the features of the first layer of the proxy task decoder as input and repeating steps S401 to S404 seven times, the features of the second, third, fourth, fifth, sixth, seventh, and eighth layers of the proxy task decoder are obtained respectively. The following relationship exists in the correspondence process:

[0049] ;

[0050] in, Indicates the agent task decoder number Features of the layer This indicates that the feature splicing operation has been performed. This indicates that the data is processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer. Indicates the agent task decoder number Features of the layer Indicates the encoder's first... Features of the layer Index of the presentation layer;

[0051] The features of the eighth layer of the agent task decoder are used as the high-frequency structural feature map for reconstruction.

[0052] It should be noted that the kernel size of the transposed convolutional layer is 4×4, and the stride is 2.

[0053] Step 5: Input the encoder features into the downstream task decoder in an iterative manner for processing, and make skip connections with the features of the proxy task decoder to obtain the reconstructed denoised image.

[0054] In step 5, the encoder features are iteratively input into the downstream task decoder for processing, and skip connections are made with the features of the surrogate task decoder to obtain the reconstructed denoised image. This specifically includes the following sub-steps:

[0055] S501. Process the features of the eighth layer of the encoder through wavelet feature blocks to obtain the first wavelet feature;

[0056] S502. Perform a skip connection between the first wavelet feature and the feature of the corresponding layer of the proxy task decoder to obtain the feature of the first layer of the downstream task decoder.

[0057] By iteratively using the features of the first layer of the downstream task decoder as input and repeating steps S501 and S502 seven times, the features of the second, third, fourth, fifth, sixth, seventh, and eighth layers of the downstream task decoder are obtained respectively. The following relationship exists in the corresponding process:

[0058] ;

[0059] in, Indicates the downstream task decoder number Features of the layer This indicates that the signal has been processed using wavelet feature blocks. Indicates the downstream task decoder number Characteristics of the layer;

[0060] The features from the eighth layer of the downstream task decoder are used as the reconstructed denoised image.

[0061] Furthermore, the processing of the wavelet feature blocks specifically includes the following steps:

[0062] Discrete wavelet transform is applied to the input features to obtain low-frequency approximate components. High-frequency detail information in the horizontal direction High-frequency detail information in the vertical direction High-frequency detail information in the diagonal direction ;

[0063] The wavelet channel attention mechanism is used to aggregate low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction to obtain channel-level descriptors. The following relationship exists in the correspondence process:

[0064] ;

[0065] in, Indicates a channel-level descriptor. Represents the pixel height of the wavelet domain feature map. This represents the pixel width of the wavelet domain feature map. This indicates element-wise addition. Indicates the low-frequency approximate component at position The value, This indicates the high-frequency detail information in the horizontal direction at the location. The value, This indicates the high-frequency detail information in the vertical direction at the location. The value, This indicates the high-frequency detail information in the diagonal direction at the position. The value, Indicates pixel position;

[0066] The channel-level descriptors are processed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid function to obtain the channel attention map. The following relationship exists in the corresponding process:

[0067] ;

[0068] in, This represents a channel attention map. This indicates that the signal has been processed by the Sigmoid function. This represents the weight parameters of the second fully connected layer. This indicates that the device has undergone ReLU activation function processing. This represents the weight parameters of the first fully connected layer;

[0069] The wavelet spatial attention mechanism is used to aggregate low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction to obtain spatial feature descriptors. The following relationship exists in the correspondence process:

[0070] ;

[0071] in, Represents spatial feature descriptors;

[0072] The spatial feature descriptors are processed sequentially through a third fully connected layer, a ReLU activation function, a fourth fully connected layer, and a Sigmoid function to obtain the spatial attention map. The following relationship exists in the corresponding process:

[0073] ;

[0074] in, Representing a spatial attention map, This represents the weight parameters of the fourth fully connected layer. This represents the weight parameters of the third fully connected layer;

[0075] The input features are reweighted using channel attention maps and spatial attention maps to obtain the output features of wavelet feature blocks.

[0076] Step 6: Construct a focal frequency loss function using the reconstructed denoised image, and construct an L1 loss function using the reconstructed high-frequency structural feature map. Then, construct a cycle consistency loss function using the high-frequency features in the reconstructed denoised image and the reconstructed high-frequency structural feature map. Train the self-supervised OCT denoising model using the focal frequency loss function, L1 loss function, and cycle consistency loss function to obtain the trained self-supervised OCT denoising model. Use the downstream task decoder in the trained self-supervised OCT denoising model to perform denoising inference on the noisy OCT image to obtain the final denoised image.

[0077] In step 6, a focal frequency loss function is constructed using the reconstructed denoised image, and an L1 loss function is constructed using the reconstructed high-frequency structure feature map. Then, a cycle consistency loss function is constructed using the high-frequency features in the reconstructed denoised image and the reconstructed high-frequency structure feature map. The self-supervised OCT denoising model is trained using the focal frequency loss function, the L1 loss function, and the cycle consistency loss function to obtain the trained self-supervised OCT denoising model. The downstream task decoder in the trained self-supervised OCT denoising model is used to perform denoising inference on the noisy OCT image to obtain the final denoised image. The expression for the focal frequency loss function is:

[0078] ;

[0079] in, Represents the focal frequency loss function. Represented in frequency coordinates Spatial frequency weighting at that location This represents the corresponding coordinates in the frequency spectrum. Representing an image In frequency coordinates Complex frequency domain value at that point, Indicates high-frequency structural features, Representing an image In frequency coordinates Complex frequency domain value at that point, Represents the high-frequency structural feature map of the reconstruction;

[0080] The expression for the L1 loss function is:

[0081] ;

[0082] in, Describes the L1 loss function. This represents the reconstructed, denoised image. This represents a noise-free OCT image. Represents the L1 norm;

[0083] The expression for the cycle consistency loss function is as follows:

[0084] ;

[0085] in, Represents the cycle consistency loss function. Represents the denoised image from the reconstruction High-frequency features extracted from them.

[0086] Furthermore, the total loss function can be defined as:

[0087] ;

[0088] in, Represents the total loss function. Indicates the first Focus frequency loss per view sample Indicates the first L1 loss for each view sample Indicates the first Cyclic consistency loss under a single view sample.

[0089] It should be understood that although the steps in the flowcharts of the various embodiments of the present invention are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the various embodiments may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least a portion of the sub-steps or stages of other steps.

[0090] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0091] In the description of this specification, 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 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.

[0092] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A self-supervised structural representation learning-based OCT denoising method, characterized in that, The method includes the following steps: Step 1: Construct a self-supervised OCT denoising model using an encoder and two parallel coupled decoders; the encoder and the two parallel coupled decoders are stacked in an iterative manner with eight layers; the two parallel coupled decoders are the proxy task decoder and the downstream task decoder, respectively. Step 2: Apply speckle noise of different intensities to the noise-free OCT image, and then stitch together the high-frequency features extracted using a high-pass filter to obtain high-frequency structural features. Step 3: The high-frequency structural features are processed sequentially through the eight stacked layers of the encoder in an iterative manner to obtain eight encoder features. Step 4: Input the encoder features into the proxy task decoder in an iterative manner for decoding processing to obtain eight proxy task decoder features respectively, and use the features of the eighth layer of the proxy task decoder as the reconstructed high-frequency structure feature map. Step 5: Input the encoder features into the downstream task decoder in an iterative manner for processing, and make skip connections with the features of the proxy task decoder to obtain the reconstructed denoised image. Step 6: Construct a focal frequency loss function using the reconstructed denoised image, and construct an L1 loss function using the reconstructed high-frequency structure feature map. Then, construct a cycle consistency loss function using the high-frequency features in the reconstructed denoised image and the reconstructed high-frequency structure feature map. Train the self-supervised OCT denoising model using the focal frequency loss function, L1 loss function, and cycle consistency loss function to obtain the trained self-supervised OCT denoising model. Use the downstream task decoder in the trained self-supervised OCT denoising model to perform denoising inference on the noisy OCT image to obtain the final denoised image. The proxy task decoder and the downstream task decoder share the features output by the encoder and are coupled through feature concatenation. In step 2, speckle noise of different intensities is applied to the noise-free OCT image, and high-frequency features are extracted using a high-pass filter and then stitched together to obtain high-frequency structural features. This process includes the following sub-steps: Different intensities of speckle noise were applied to noise-free OCT images to obtain a set of multi-view degraded images; High-frequency features of all images in a multi-view degraded image set are extracted using a high-pass filter to obtain a high-frequency image set. All images in the high-frequency image set are stitched together along the channel dimension to obtain high-frequency structural features; In step 5, the encoder features are iteratively input into the downstream task decoder for processing, and skip connections are made with the features of the proxy task decoder to obtain the reconstructed denoised image. Specifically, this includes the following sub-steps: S501. The features of the eighth layer of the encoder are processed sequentially through wavelet feature blocks, batch normalization layer and ReLU activation layer to obtain the first wavelet features. S502. Perform a skip connection between the first wavelet feature and the feature of the corresponding layer of the proxy task decoder to obtain the feature of the first layer of the downstream task decoder.

2. The OCT denoising method based on self-supervised structural representation learning according to claim 1, characterized in that, In step 2, speckle noise of different intensities is applied to the noise-free OCT image, and high-frequency features are extracted using a high-pass filter and then stitched together to obtain high-frequency structural features. This process includes the following sub-steps: Different intensities of speckle noise were applied to noise-free OCT images to obtain a set of multi-view degraded images. ; in, Represents a collection of degraded images from multiple perspectives. Indicates the first Degraded images of a single view sample. Indicates the index of the view sample. This represents the total number of categories of speckle noise; High-frequency features of all images in a multi-view degraded image set are extracted using a high-pass filter to obtain the high-frequency image set. ; in, Represents a set of high-frequency images. Indicates the first High-frequency images of a single view sample; All images in the high-frequency image set are stitched together along the channel dimension to obtain high-frequency structural features.

3. The OCT denoising method based on self-supervised structural representation learning according to claim 2, characterized in that, Different intensities of speckle noise are applied to noise-free OCT images to obtain a multi-view degraded image set. The following relationship exists in the steps: ； in, Indicates the first In the degraded image of the first view sample The value of each pixel. Indicates the index of the pixel. Indicating the first noise-free OCT image The value of each pixel. The simulated synthesis of the first In the first speckle noise sample The value of each pixel; The speckle noise follows a gamma distribution, expressed as: ； in, Represents the speckle noise distribution function. A random sample representing the speckle noise distribution. This represents the distributed parameters used to control noise variance. Represents the gamma function. Represents the natural constant; High-frequency features of all images in a multi-view degraded image set are extracted using a high-pass filter to obtain a high-frequency image set. The following relationship exists in the steps: ； in, This indicates that it has undergone high-frequency extraction processing. This indicates element-wise multiplication. Indicates a high-pass filter. Indicates the kernel size. It represents the standard deviation.

4. The OCT denoising method based on self-supervised structural representation learning according to claim 3, characterized in that, In step 3, the high-frequency structural features are processed iteratively through the eight stacked layers of the encoder to obtain eight encoder features, specifically including the following sub-steps: S301. The high-frequency structural features are processed sequentially through a convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the first convolutional feature. S302. The first convolutional feature is processed sequentially through a convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the second convolutional feature. S303. Perform residual connection between the second convolutional features and the high-frequency structural features to obtain the features of the first layer of the encoder; By iteratively taking the features of the first layer of the encoder as input and repeating steps S301 to S303 seven times, the features of the second layer, third layer, fourth layer, fifth layer, sixth layer, seventh layer and eighth layer of the encoder are obtained respectively.

5. The OCT denoising method based on self-supervised structural representation learning according to claim 4, characterized in that, In step 4, the encoder features are iteratively input into the surrogate task decoder for decoding to obtain eight surrogate task decoder features. The features of the eighth layer of the surrogate task decoder are then used as the reconstructed high-frequency structure feature map. Specifically, this includes the following sub-steps: S401. The features of the eighth layer of the encoder are processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the first transposed convolutional features. S402. The first transposed convolutional feature is processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer to obtain the second transposed convolutional feature. S403. Perform residual connection between the second transposed convolutional features and the features of the eighth layer of the encoder to obtain the intermediate features of the first layer of the proxy task decoder. S404. Perform skip connections between the intermediate features of the first layer of the proxy task decoder and the features of the corresponding layer of the encoder to obtain the features of the first layer of the proxy task decoder. By iteratively using the features of the first layer of the proxy task decoder as input and repeating steps S401 to S404 seven times, the features of the second, third, fourth, fifth, sixth, seventh, and eighth layers of the proxy task decoder are obtained respectively. The following relationship exists in the correspondence process: ； in, Indicates the agent task decoder number Features of the layer This indicates that the feature splicing operation has been performed. This indicates that the data is processed sequentially through a transposed convolutional layer, a batch normalization layer, and a ReLU activation layer. Indicates the agent task decoder number Features of the layer Indicates the encoder's first... Features of the layer Index of the presentation layer; The features of the eighth layer of the agent task decoder are used as the high-frequency structural feature map for reconstruction.

6. The OCT denoising method based on self-supervised structural representation learning according to claim 5, characterized in that, In step 5, the encoder features are iteratively input into the downstream task decoder for processing, and skip connections are made with the features of the proxy task decoder to obtain the reconstructed denoised image. Specifically, this includes the following sub-steps: S501. The features of the eighth layer of the encoder are processed sequentially through wavelet feature blocks, batch normalization layer and ReLU activation layer to obtain the first wavelet features. S502. Perform a skip connection between the first wavelet feature and the feature of the corresponding layer of the proxy task decoder to obtain the feature of the first layer of the downstream task decoder. By iteratively using the features of the first layer of the downstream task decoder as input and repeating steps S501 and S502 seven times, the features of the second, third, fourth, fifth, sixth, seventh, and eighth layers of the downstream task decoder are obtained respectively. The following relationship exists in the corresponding process: ； in, Indicates the downstream task decoder number Features of the layer This indicates that the signal has been processed using wavelet feature blocks. Indicates the downstream task decoder number Characteristics of the layer; The features from the eighth layer of the downstream task decoder are used as the reconstructed denoised image.

7. The OCT denoising method based on self-supervised structural representation learning according to claim 6, characterized in that, The processing of the wavelet feature blocks specifically includes the following steps: Discrete wavelet transform is applied to the input features to obtain low-frequency approximate components. High-frequency detail information in the horizontal direction High-frequency detail information in the vertical direction High-frequency detail information in the diagonal direction ; The wavelet channel attention mechanism is used to aggregate low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction to obtain a channel-level descriptor. The channel-level descriptors are processed sequentially through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid function to obtain the channel attention map; The wavelet spatial attention mechanism is used to aggregate low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction to obtain spatial feature descriptors. The spatial feature descriptors are processed sequentially through a third fully connected layer, a ReLU activation function, a fourth fully connected layer, and a Sigmoid function to obtain a spatial attention map. The input features are reweighted using channel attention maps and spatial attention maps to obtain the output features of wavelet feature blocks.

8. The OCT denoising method based on self-supervised structural representation learning according to claim 7, characterized in that, In the step of aggregating low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction using the wavelet channel attention mechanism to obtain the channel-level descriptor, the following relationship exists: ； in, Indicates a channel-level descriptor. Represents the pixel height of the wavelet domain feature map. This represents the pixel width of the wavelet domain feature map. This indicates element-wise addition. Indicates the low-frequency approximate component at position The value, This indicates the high-frequency detail information in the horizontal direction at the location. The value, This indicates the high-frequency detail information in the vertical direction at the location. The value, This indicates the high-frequency detail information in the diagonal direction at the position. The value, Indicates pixel position; In the process of sequentially processing the channel-level descriptors through the first fully connected layer, the ReLU activation function, the second fully connected layer, and the Sigmoid function to obtain the channel attention map, the following relationship exists: ； in, This represents a channel attention map. This indicates that the signal has been processed by the Sigmoid function. This represents the weight parameters of the second fully connected layer. This indicates that the device has undergone ReLU activation function processing. This represents the weight parameters of the first fully connected layer; In the step of aggregating low-frequency approximate components, high-frequency detail information in the horizontal direction, high-frequency detail information in the vertical direction, and high-frequency detail information in the diagonal direction using the wavelet spatial attention mechanism to obtain a spatial feature descriptor, the following relationship exists: ； in, Represents spatial feature descriptors; In the process of processing the spatial feature descriptors sequentially through the third fully connected layer, the ReLU activation function, the fourth fully connected layer, and the Sigmoid function to obtain the spatial attention map, the following relationship exists: ； in, Representing a spatial attention map, This represents the weight parameters of the fourth fully connected layer. This represents the weight parameters of the third fully connected layer.

9. The OCT denoising method based on self-supervised structural representation learning according to claim 8, characterized in that, In step 6, a focal frequency loss function is constructed using the reconstructed denoised image, and an L1 loss function is constructed using the reconstructed high-frequency structure feature map. Then, a cycle consistency loss function is constructed using the high-frequency features in the reconstructed denoised image and the reconstructed high-frequency structure feature map. The self-supervised OCT denoising model is trained using the focal frequency loss function, the L1 loss function, and the cycle consistency loss function to obtain the trained self-supervised OCT denoising model. The downstream task decoder in the trained self-supervised OCT denoising model is used to perform denoising inference on the noisy OCT image to obtain the final denoised image. The expression for the focal frequency loss function is: ； in, Represents the focal frequency loss function. Represented in frequency coordinates Spatial frequency weighting at that location This represents the corresponding coordinates in the frequency spectrum. Representing an image In frequency coordinates Complex frequency domain value at that point, Indicates high-frequency structural features, Representing an image In frequency coordinates Complex frequency domain value at that point, Represents the high-frequency structural feature map of the reconstruction; The expression for the L1 loss function is: ； in, Describes the L1 loss function. This represents the reconstructed, denoised image. This represents a noise-free OCT image. Represents the L1 norm; The expression for the cycle consistency loss function is as follows: ； in, Represents the cycle consistency loss function. Represents the denoised image from the reconstruction High-frequency features extracted from them.

Images