OCT retina image denoising method based on high-frequency enhanced diffusion model

Abstract

The invention discloses an OCT (Optical Coherence Tomography) retina image denoising method based on a high-frequency enhanced diffusion model, which is characterized in that a double-branch diffusion model network constructed based on frequency perception Fourier transform attention (FFTA) is used for separating different frequency domains, enhancing specific frequency domain features and fusing the specific frequency domain features into a spatial domain feature map to realize stronger frequency domain perception and processing capability; the invention relates to a time step T coding method for a module of frequency domain attention. The time step T of the diffusion model is used as a parameter to be coded into a frequency domain attention module to participate in self-attention matrix calculation, so that frequency domain feature processing is aligned with the iteration step number of the diffusion model, and different frequency domain information is pointedly processed at different time steps T; the frequency selective hopping mechanism uses pooling operation to obtain a high / low frequency characteristic pattern, so that the network has adaptive frequency domain retention capability for different local parts.

Description

Technical Field

[0001] This invention belongs to the field of medical image analysis technology, specifically relating to an OCT retinal image denoising method based on a high-frequency enhanced diffusion model. Background Technology

[0002] Optical coherence tomography (OCT) is a non-invasive, high-resolution imaging technique widely used in the diagnosis and monitoring of ophthalmology, dermatology, and cardiovascular diseases. With its high spatial resolution and real-time imaging capabilities, OCT has become the gold standard for retinal imaging. However, OCT images are susceptible to speckle noise, leading to blurred tissue details and posing challenges to clinical analysis. Therefore, speckle removal has become a core issue in OCT image processing.

[0003] Traditional denoising methods, such as multi-frame averaging, can reduce noise, but are limited by time consumption and eye motion artifacts. In recent years, deep learning methods have significantly improved the effectiveness and robustness of OCT denoising through end-to-end modeling, multi-scale feature extraction, and context fusion. For example, the multi-CNN residual framework uses block sampling and clustering to achieve noise estimation; the PSN framework combines physical priors to generate realistic noise; and SiameseGAN achieves high-quality reconstruction under unsupervised conditions.

[0004] In contrast, diffusion models (DM) demonstrate superior stability and interpretability in image generation and reconstruction.

[0005] The basic idea of ​​DM (Diffusion Model) is to learn the data distribution through a process of progressively adding noise and then reversing noise reduction, thereby achieving high-quality image restoration during the testing phase. The training and inference process of the diffusion model can be divided into two stages:

[0006] Forward Diffusion Process

[0007] In the forward process, given a clean image x0, the model gradually injects noise into the image over T time steps:

[0008]

[0009] Where, x t α represents the image state at time step t. t It is usually set to α t =1-β t ,β t ∈(0,1) controls the amount of noise injected at each step. N(·) represents a Gaussian distribution, and I is the identity matrix. Through iterative calculation, x can be finally obtained.T ≈N(0,I). The forward diffusion process of the overall distribution obtained from the initial data sampling is as follows.

[0010]

[0011] in α i This represents the cumulative noise factor. During the inverse generation process, the model aims to learn the inverse transformation from the noisy sample x. T Restore clean image x0:

[0012]

[0013] Here, μ θ (x t ,t) and Σ θ (x t ,t) represent the mean and variance, respectively, which are typically learned by a neural network. Most models assume the variance is constant and optimize only the mean term μ. θ The final generation process of the model involves iterative sampling. This process gradually removes noise to obtain the target image. There are many derivative models of the diffusion model. One notable variant, the Denoising Diffusion Implicit Model (DDIM), removes the Markov assumption, allowing DDIM to jump from time step t to time step t-1, thus achieving efficient sampling.

[0014] Previous studies have incorporated diffusion models and fractional prior iterative methods into OCT denoising, achieving good results. However, existing methods neglect the differences in frequency components at different time steps during the diffusion process, leading to insufficient recovery of high-frequency details and a smoother reconstruction result.

[0015] 1. Existing traditional denoising methods have limited denoising capabilities and timeliness. Traditional OCT denoising methods mainly rely on multi-frame averaging. These methods can reduce speckle noise to some extent, but they have two significant problems. First, the operation process is time-consuming, resulting in low efficiency in clinical applications. Second, they are easily affected by factors such as eye movement, which can introduce registration artifacts and reduce image quality.

[0016] 2. Limited denoising capabilities of diffusion models. While diffusion models (DDPM, DDIM, etc.) have demonstrated superior stability and generation quality compared to GANs in OCT denoising, their application still has shortcomings. First, most current diffusion models do not consider frequency attention differences that change with the diffusion time step; second, during reconstruction, the network tends to ignore high-frequency details, resulting in overly smooth outputs that reduce the structural fidelity and visual realism of the image.

[0017] 3. The lack of frequency domain methods in OCT. In the field of general image restoration, existing research has shown that combining Fourier transform for frequency domain modeling can more effectively capture low-frequency structures and high-frequency details; however, in OCT denoising tasks, relevant frequency domain separation mechanisms and modeling methods are still lacking, which limits the model's ability to differentiate between high and low frequency information. Summary of the Invention

[0018] In view of this, the purpose of this invention is to provide an OCT retinal image denoising method based on a high-frequency enhanced diffusion model. The invention proposes a time-step encoded high-frequency enhanced diffusion model THFN-OCT, which effectively improves high-frequency detail recovery and enhances the realism and diagnostic value of OCT images by strengthening frequency domain modeling in the network and combining it with time-step sensitivity.

[0019] A method for denoising OCT retinal images based on a high-frequency enhanced diffusion model includes:

[0020] For a given input image I, a convolutional layer is first applied for preliminary feature extraction and channel transformation; the processed features are then split into two branches: the first branch estimates the frequency domain attention weights, and the second branch preserves the initial feature map.

[0021] The first branch performs feature extraction and channel transformation through the downsampling pooling module (Down Avg operation) to obtain the feature tensor F. feature Then, the output of the Fuse_T module is fused with the encoded output of time step T.

[0022] The encoding process for time step T is as follows: For an input time step t∈T, the time embedding module TimeMLP(t) first generates the time feature e. t Then, it is transformed sequentially through a multilayer perceptron and dimensionality reshaping, and works together with the modulation parameters γ(t) and β(t) generated by the Adjust module to process the feature tensor F output by the downsampling pooling module Down Avg. feature The feature F′ is obtained. feature Specifically, it is expressed as:

[0023] F′ feature =F feature +γ(t)·MLP(e t )+β(t)

[0024] Among them, MLP(e t ) represents the multilayer perceptron (MLP) response to time features e t The processing;

[0025] Next, the feature map F′ output by the Fuse_T module feature Entering the FFTA module: The FFTA module first processes the input feature map F′feature Apply 3×3 convolution to expand its channels and obtain F in Then, two parallel links are used for processing:

[0026] In the first link of the FFTA module, the features are first processed using Fast Fourier Transform (FFT) to convert them from the spatial domain to the frequency domain; then the real part of the frequency features is extracted and passed to the learnable frequency enhancement (FE) module, which acts as a frequency domain filter and adaptively weights different frequency components to obtain enhanced features.

[0027] Finally, the enhanced features are transformed back to the spatial domain using IFFT, a process described as follows:

[0028] F = Φ ||IFFT(σ(FE(FFT(F)) in ))*FFT(F in )))||

[0029] Where Φ||·|| and σ represent modular arithmetic and element-wise multiplication, respectively;

[0030] The second link in the FFTA module executes a frequency domain attention mechanism: it employs a Transformer-based attention mechanism to map input features to queries, keys, and values; then, it weights and aggregates values ​​by calculating the similarity between queries and keys, thereby modeling global information; finally, it applies FFT again to obtain the three components required for attention computation, namely the query Q. f Key K f Sum V f :

[0031] Q f =FFT Q (Norm(F in )); K f =FFT K (Norm(F in ));V f =FFT V (Norm(F in ));

[0032] Norm() represents layer normalization;

[0033] Finally, embed the time step into F t =MLP(e t ), which is incorporated into the key value, namely K f *F t The initial attention map obtained is: Λ f =Q f *(K f *F t Extracting the real part and the virtual part Where conj(·) represents the conjugate complex number, the CCN module then activates and merges the real and imaginary parts to obtain the attention map, and the attention map is then passed through the Softmax activation function to obtain the frequency domain weights aΛ. f The processing procedure of the CCN module is as follows:

[0034]

[0035] Where Θ(·,·) represents the combination function that combines the imaginary and real parts into a complex number; Sof(·) represents the Softmax function;

[0036] Then, the frequency domain weight aΛ f The result is transformed back to the spatial domain using Inverse Fourier Transform (IFFT) and concatenated with F. The concatenated result is then processed through convolution and further processed by parallel frequency and spatial branches: the spatial branch uses a depthwise separable convolution module (DWConv), while the frequency branch processes the result sequentially using a Fast Fourier Transform (FFT) frequency enhancement module (FE) and an Inverse Fourier Transform (IFFT). Finally, the results from both the spatial and frequency branches are convolved to obtain a frequency-weighted F guided by the fused time code t. out ;

[0037] The second branch executes the convolutional network processing flow, which consists of a downsampling pooling layer, a Fuse_T module, and a convolutional layer; the output of this branch is guided by the weights of the first branch. out The merging process employs matrix multiplication to obtain the first-layer downsampled feature map, which is then fed into the merging channel for processing, specifically including:

[0038] First, the first-layer downsampled feature map is sequentially processed by the downsampling pooling module Down Avg, the Fuse_T module, and the first convolutional layer to obtain the highest-level semantic encoding information;

[0039] Then, after encoding, a general decoding operation based on a convolutional network is performed. The decoding operation is implemented sequentially by the first upsampling convolutional module Up Conv, the Fuse_T module, the second convolutional layer, the second upsampling convolution, and the third convolutional layer.

[0040] Finally, a convolutional layer is used to adjust the channels to obtain the final denoising result output.

[0041] Preferably, the Adjust module is a shallow network consisting of three loops: a convolutional layer, a ReLU activation function, and a pooling layer, with the final convolutional layer used to adjust the channel output; its input is... x TThis is the noise map obtained in the forward diffusion T steps of the diffusion model, which is the initial noise map I input to the network. This is a denoised image recovered from a time step on the network; during the testing phase, Adjust is disabled, leaving only the time embedding to adjust the various stages of the network.

[0042] Preferably, the frequency enhancement module (FE) consists of a convolutional layer, an activation function, another convolutional layer, and a sigmoid layer, and the processing procedure is represented as follows:

[0043] FE(FFT(F in ))=Sigmoid(Conv(ReLU(Conv(FFT(F in )))))

[0044] Furthermore, it also includes the FreqSkip module, whose processing procedure is described as follows:

[0045] F skip =Fuse_T(Conv(F split (F in -Up(Pool(F in )))),F t )

[0046] Where Pool is the pooling operation, Up is the upsampling operation, and F... split Represents frequency domain decomposition, F split (F in -Up(Pool(F in That is, the features are divided into high-frequency and low-frequency parts in the graph by pooling.

[0047] The FreqSkip module outputs F skip The output of the first convolutional layer (Conv) in the merged channel is concatenated with the output of the first convolutional layer (Up Conv) in the channel dimension and then fed into the first upper convolutional layer (Up Conv) for subsequent upsampling processing.

[0048] The present invention has the following beneficial effects:

[0049] 1. A dual-branch diffusion model network based on frequency-aware Fourier transform attention (FFTA) is used to separate different frequency domains, enhance specific frequency domain features and fuse them into the spatial domain feature map, thereby achieving stronger frequency domain awareness and processing capabilities.

[0050] 2. Encoding method for time step T of frequency domain attention module. The time step T of the diffusion model is encoded as a parameter into the frequency domain attention module (FFTA) and participates in the matrix calculation of self-attention. This allows the frequency domain feature processing to be aligned with the iteration number of the diffusion model, and to process different frequency domain information in a targeted manner at different time steps T.

[0051] 3. The frequency-selective skipping mechanism (FreqSkip) uses pooling operations to obtain high / low frequency feature maps, which makes it easier for the network to have adaptive frequency domain preservation capabilities for different localities. Attached Figure Description

[0052] Figure 1 The THFN-OCT network structure of this invention is shown in (a) as the overall network structure; (b) as the frequency-based skip connection (FreqSkip) module; (c) as the Adjust network; and (d) as the Fast Fourier Transform Attention (FFTA) module.

[0053] Wherein, Conv is convolution; Up is upsampling; DWConv is depthwise convolution; Pool is average pooling; MLP is multilayer perceptron; Norm is layer normalization; FFT is fast Fourier transform; IFFT is inverse fast Fourier transform; CCN is custom complex domain normalization layer; Fuse_T is temporal embedding fusion module; and FE is frequency enhancement module.

[0054] Figure 2 The denoising performance of various denoising methods on the Duke-2013 dataset is presented.

[0055] Figure 3 The denoising performance of different denoising methods on the Duke-2012 dataset is presented.

[0056] Figure 4 The denoising performance of different denoising methods on our dataset is presented. Detailed Implementation

[0057] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0058] To address the core challenges in OCT retinal image denoising, such as the complexity of speckle noise, the difficulty in simultaneously preserving high-frequency details and maintaining the low-frequency substrate, and the insufficient dynamic adaptability of existing methods to frequency domain features, this invention proposes a high-frequency enhanced diffusion model, THFN-OCT, based on the ColdDiffusion backbone structure. Unlike existing methods that suppress noise in a single intensity domain or across the entire frequency range, this invention, for the first time, combines the time-step driven characteristics of the diffusion model with a frequency domain feature decoupling mechanism. The main strategies are as follows: First, frequency domain decoupling decouples high-frequency details (edges / textures) from low-frequency substrates (tissue morphology) into independent subspaces, avoiding detail blurring caused by global processing. Second, a frequency domain attention module incorporating time-step awareness is designed, encoding the time step of the diffusion process into dynamic weights to guide the network to perform differentiated reconstruction of high and low frequency components at different restoration stages—prioritizing substrate structure restoration in the early stages and gradually enhancing detail depiction in later stages. Third, different frequency components are distinguished and preserved across connection parts, enabling the network to adaptively preserve the basic structure and detailed edges of image features.

[0059] The overall architecture of THFN-OCT is as follows: Figure 1 As shown in (a). For a given input image First, a convolutional layer (Conv) is applied for initial feature extraction and channel transformation. The processed features are then split into two branches: one branch estimates the frequency domain attention weights, and the other branch preserves the initial feature map.

[0060] The first branch performs feature extraction and channel transformation through downsampling pooling (Down Avg) to obtain the feature tensor F. feature Then, it is fused with the encoded output of time step T through the Fuse_T module. The encoding process of time step T is as follows: Figure 1 (a) As shown on the right, specifically, for an input time step t∈T, the time embedding module TimeMLP(t) first generates the time feature e. t Then, it is transformed sequentially through a multilayer perceptron (MLP) and a dimensionality reshaping (Rearrange) module, and works together with the modulation parameters γ(t) and β(t) generated by the Adjust module to process the feature tensor F output by downsampling pooling. feature The feature F′ is obtained. feature .

[0061] F′ feature =F feature +γ(t)·MLP(e t )+β(t)

[0062] Among them, MLP(e t ) represents the multilayer perceptron (MLP) response to time features et The processing.

[0063] The Adjust module is a shallow network. Figure 1 (c) illustrates the network structure of Adjust. It consists of three loops: convolutional layers, ReLU activation function, and pooling layers, with the final convolutional layer adjusting the output channels. Its input is... x T This is the noise map obtained in the forward diffusion T steps of the diffusion model, which is the initial noise map I input to the network. This refers to the denoised image recovered at a given time step from the network. The role of Adjust is to assist in training, specifically by embedding the MLP (e...) at the time step. t Feature modulation is performed to correct the misalignment between input features and time-step features, and to improve model convergence and feature alignment. During the testing phase, Adjust is disabled, leaving only temporal embeddings to adjust the various stages of the network.

[0064] Next, the feature map F′ output by the Fuse_T module feature Enter the FFTA module proposed in this invention. For example... Figure 1 As shown in (d), the FFTA module first processes the input feature map F′ feature Apply 3×3 convolution to expand its channels and obtain F in Then, two parallel links are used for processing:

[0065] In the first link of the FFTA module (see...) Figure 1 (d) Bottom), firstly, the features are processed using Fast Fourier Transform (FFT) to convert them from the spatial domain to the frequency domain; then, the real part of the frequency features is extracted and passed to a learnable frequency enhancement (FE) module. This module acts as a frequency domain filter, adaptively weighting different frequency components. The frequency enhancement module FE consists of convolution, activation function, convolution, and sigmoid layers in sequence, and the processing procedure is represented as follows:

[0066] FE(FFT(F in ))=Sigmoid(Conv(ReLU(Conv(FFT(F in )))))

[0067] Finally, the enhanced features are transformed back to the spatial domain using IFFT. This process can be described as follows:

[0068] F = Φ ||IFFT(σ(FE(FFT(F)) in ))*FFT(F in )))||

[0069] Where Φ||·|| and σ represent modular arithmetic and element-wise multiplication, respectively.

[0070] The second link in the FFTA module performs a frequency domain attention mechanism. In this invention, a Transformer-based attention mechanism is employed to map input features to a query (Q), a key (K), and a value (V). Then, we model global information by weighted aggregation of V based on the similarity between Q and K. Compared to convolutional operations that only capture local receptive fields, this mechanism effectively captures long-range dependencies and enhances feature representation. We again apply FFT to obtain the three components required for attention computation: query Q... f Key K f Sum V f :

[0071] Q f =FFT Q (Norm(F in )); K f =FFT K (Norm(F in ));V f =FFT V (Norm(F in )).

[0072] Norm() represents layer normalization, which is used to stabilize the training process and accelerate network convergence.

[0073] Finally, in order to control the frequency domain information recovery by the time step t, we embed the time step into F. t =MLP(e t This is incorporated into the key value K, i.e., K f *F t The resulting preliminary attention map is: Λ f =Q f *(K f *F t Since the attention map contains both real and imaginary parts, it cannot be directly activated using the Softmax function. Therefore, the real part is extracted here. and the virtual part Where conj(·) represents the conjugate complex number, the CCN module then activates and merges the real and imaginary parts to obtain the attention map, and the attention map is then passed through the Softmax activation function to obtain the frequency domain weights aΛ. f The processing procedure of the CCN module is as follows:

[0074]

[0075] Where Θ(·,·) represents the combination function that combines the imaginary and real parts to form a complex number. Sof(·) represents the Softmax function.

[0076] Then, the frequency domain weight aΛ f The result is transformed back to the spatial domain using an inverse Fourier transform (IFFT) and concatenated with F. The concatenated result is then processed through convolution and further processed by parallel frequency and spatial branches: the spatial branch uses depthwise separable convolution (DWConv), and the frequency branch processing steps are similar, consisting of a Fast Fourier Transform (FFT) frequency enhancement module and an inverse Fourier Transform (IFFT); both processes introduce residual connections. Finally, the results from the spatial and frequency branches are convolved (Conv) to obtain a frequency-weighted F controlled by the fusion time code t. out .

[0077] The second branch of the overall architecture of THFN-OCT is a traditional convolutional network processing flow, consisting of downsampling pooling layers, Fuse_T modules, and convolutional layers. The output of this branch is related to the weights of the first branch. out The merging process employs matrix multiplication to obtain the first-layer downsampled feature map, which is then fed into the merging channel for processing, specifically including:

[0078] First, the first-layer downsampled feature map is sequentially processed through downsampling pooling (Down Avg), the Fuse_T module, and the first convolutional layer to obtain the highest-level semantic encoding information.

[0079] Then, after encoding, a general decoding operation based on a convolutional network is performed. The decoding operation is implemented sequentially by the first upsampling convolution (Up Conv), Fuse_T, the second convolutional layer, the second upsampling convolution, and the third convolutional layer.

[0080] Finally, a convolutional layer (Conv_out) is used to adjust the channels to obtain the final denoising result output.

[0081] The above describes the THFN-OCT backbone structure. In addition, considering that pooling and upsampling operations are a major cause of high-frequency detail loss, and that skip connections establish a direct information transmission path between the encoder and decoder to overcome feature loss and enhance detail recovery, we propose the FreqSkip module to implement a frequency-selective skipping mechanism.

[0082] Unlike existing skip connections that rely on direct connections, we split the feature maps of skip connections into high-frequency and low-frequency components. This design allows the network to selectively preserve the image's basis (low-frequency) and edge details (high-frequency) based on the specific features of the input. Notably, instead of using Fourier transform to obtain different frequency bands, we use pooling operations to extract the high-frequency and low-frequency components, such as... Figure 1 As shown in (b). We believe this is more suitable for compensating for the information loss caused by pooling and upsampling in the network, because the residual generated by pooling directly captures the fine details discarded during downsampling and maintains a one-to-one correspondence with the spatial structure; while Fourier decomposition mainly separates global frequency components, which may not effectively meet the requirement of preserving local details required for skip connections. This process is also affected by the time-step embedding F. t Modulation. The processing procedure of the FreqSkip module can be described as follows:

[0083] F skip =Fuse_T(Conv(F split (F in -Up(Pool(F in )))),F t )

[0084] Where Pool is the pooling operation, Up is the upsampling operation, and F... split Represents frequency domain decomposition, F split (F in -Up(Pool(F in This involves using pooling to divide features into "High" and "Low" components, as shown in the graph. The output F of the FreqSkip module... skip The output of the first convolutional layer (Conv) in the merged channel is concatenated with the output of the first convolutional layer (Cat) in the channel dimension and then fed into the first upper convolutional layer (Up Conv) for subsequent upsampling processing.

[0085] We evaluated the proposed method on two widely used public OCT retinal datasets. The first dataset is Duke-2013, which contains SD-OCT data from 28 patients. Each patient's image size is 900×450 (width×height) pixels, with an axial resolution of approximately 45 μm / pixel. We used 18 images as the training set and 10 images as the test set. We cropped them into 256×256 pixel blocks with a stride of 20 pixels as network input.

[0086] The second dataset is Duke-2012, containing 17 pairs of OCT B-scan images. Each pair includes a noisy original B-scan and a clean B-scan obtained by averaging. These images were acquired using a Bioptigen SD-OCT imaging system (Durham, North Carolina) and are 1000×500 pixels in size. Since this study uses anonymized data, no Institutional Review Board (IRB) approval was required. Due to artifact bias at the edges of some images in this dataset, we removed these artifacts and processed the data in the same manner as Duke-2013.

[0087] In addition, we constructed other systems and collected datasets. The imaging system employed SS-OCT with a center wavelength of 1035.7 nm and a bandwidth of 137.03 nm. The A-scan axial field of view was 3.62 nm. Each Bscan consisted of 1024 Ascans, with each Ascan containing 1024 depth acquisition points, forming a 1024×1024 image size. Each voxel consisted of 512 Bscans. The A-scan speed was 100.16 kHz. We imaged isolated pig eyes, acquiring data once per eye. The maximum imaging field of view for the pig fundus was 16 mm × 2.72 mm × 16 mm. The system's lateral resolution and sampling step size were both 15.6 μm.

[0088] To evaluate denoising performance, we used two standard image quality metrics: Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM). PSNR measures the similarity between the denoised image and the ground truth image based on Mean Squared Error (MSE). A higher PSNR value indicates better reconstruction quality. It is defined as follows:

[0089]

[0090] Where MAX represents the maximum possible pixel value of the image, and the mean squared error (MSE) is as follows:

[0091]

[0092] Where I(i,j) and K(i,j) represent the pixel intensity at position (i,j) in the reconstructed image and the reference image, respectively, and M and N are the width and height of the image.

[0093] SSIM evaluates the perceptual similarity between two images, taking into account brightness, contrast, and structural information. It is more consistent with human visual perception. The formula for SSIM is as follows:

[0094]

[0095] Where, μ I and μ KThese are the means of images I and K, respectively. and These are the corresponding variances, σ IK This is the covariance between the two images. The constant C1 = (0.01L) 2 And C2 = (0.03L) 2 Used to stabilize division when the denominator is close to zero, where L represents the dynamic range of pixel values.

[0096] To validate the effectiveness of our denoising method, we compared it with various state-of-the-art denoising methods. These methods include the traditional BM3D method, advanced OCT and CT neural network denoising methods such as TCFL, LIT-former, CoreDiff, and SiameseGAN, and state-of-the-art (SOTA) models in the field of natural image denoising, such as NAFNet and image inpainting RDDM. We used the same training and testing data for all methods, and their performance on the Duke-2013 dataset is shown in Table 1.

[0097] Table 1. PSNR and SSIM performance of various methods on the Duke-2013 dataset.

[0098]

[0099] The proposed method achieves a PSNR of 31.553 dB and an SSIM of 0.7294, demonstrating the best denoising performance. The visualized denoising results are shown below. Figure 2 As shown.

[0100] To verify the generalization ability of our proposed method, we applied it, along with all the comparison methods, to another dataset, Duke-2012 for training and testing. Quantitative results are shown in Table 2. Compared to the Duke-2013 dataset, the Duke-2012 dataset is more challenging due to its more complex noise artifacts and smaller sample size. Therefore, the performance of all methods is lower compared to the results in Table 1. Our proposed method achieves a peak signal-to-noise ratio (PSNR) of 28.104 dB and an SSIM of 0.6924, both higher than the other methods. The denoising results are visualized as follows: Figure 3 As shown.

[0101] Table 2. PSNR and SSIM performance of various methods on the Duke-2012 dataset.

[0102]

[0103] Figure 4This demonstrates the denoising performance of different methods on our actual acquired images. Similar to the previous two datasets, our proposed method outperforms other methods in both image performance and metric evaluation, achieving the highest PSNR (29.160 dB) and SSIM (0.5977). Compared to the previous two datasets, all methods achieve relatively low SSIM metrics on this dataset (as shown in Table 3).

[0104] Table 3. PSNR and SSIM performance of various methods on our dataset.

[0105]

[0106] To verify the impact of the various modules proposed in this paper on denoising capabilities, we conducted ablation experiments based on the cold diffusion model and evaluated the PSNR and SSIM metrics on the Duke-2013 dataset (results are shown in Table 4). Overall, the baseline model achieved a performance of 31.168 dB / 0.7225. After removing the frequency domain attention module (THFN-OCT w / o FFTA), PSNR and SSIM only slightly improved (31.396 dB / 0.7231), indicating that explicit frequency domain modeling has a positive impact on image restoration. When the time step encoding T was further added to FFTA, the peak signal-to-noise ratio (PSNR) increased from 31.469 dB to 31.553 dB, and the SSIM increased from 0.7231 to 0.7294. This shows that the time step modulation mechanism can guide the model to more rationally prioritize low-frequency structures and high-frequency details at different diffusion stages, thereby improving the overall reconstruction quality.

[0107] Regarding skip connections, removing them has a devastating impact on the network (THFN-OCT without skip connections), causing a sharp drop in performance to 18.350 dB / 0.1534. This demonstrates the crucial role of skip connections in maintaining structural consistency and preserving detailed features. Replacing the traditional skip with the proposed FreqSkip improves performance from 31.482 dB / 0.7222 to 31.553 dB / 0.7294, indicating that FreqSkip can more effectively balance structural consistency and detail fidelity of images through high- and low-frequency decomposition and selective preservation.

[0108] Table 4. Ablation studies of THFN-OCT on the Duke-2013 dataset.

[0109] “√” indicates that the module is used, and “×” indicates that the module is removed.

[0110]

[0111] In summary, the above are merely preferred embodiments of the present invention and are not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for denoising OCT retinal images based on a high-frequency enhanced diffusion model, characterized in that, include: For a given input image I, a convolutional layer is first applied to perform preliminary feature extraction and channel transformation; The processed features are then split into two branches: the first branch estimates the frequency domain attention weights, and the second branch preserves the initial feature map. The first branch performs feature extraction and channel transformation through the downsampling pooling module (Down Avg operation) to obtain the feature tensor F. feature Then, the output of the Fuse_T module is fused with the encoded output of time step T. The encoding process for time step T is as follows: For an input time step t∈T, the time embedding module TimeMLP(t) first generates the time feature e. t Then, it is transformed sequentially through a multilayer perceptron and dimensionality reshaping, and works together with the modulation parameters γ(t) and β(t) generated by the Adjust module to process the feature tensor F output by the downsampling pooling module Down Avg. feature The feature F′ is obtained. feature Specifically, it is expressed as: F′ feature =F feature +γ(t)·MLP(e t )+β(t) Among them, MLP(e t ) represents the multilayer perceptron (MLP) response to time features e t The processing; Next, the feature map F′ output by the Fuse_T module feature Entering the FFTA module: The FFTA module first processes the input feature map F′ feature Apply 3×3 convolution to expand its channels and obtain F in Then, two parallel links are used for processing: In the first link of the FFTA module, the features are first processed using Fast Fourier Transform (FFT) to convert them from the spatial domain to the frequency domain; then the real part of the frequency features is extracted and passed to the learnable frequency enhancement (FE) module, which acts as a frequency domain filter and adaptively weights different frequency components to obtain enhanced features. Finally, the enhanced features are transformed back to the spatial domain using IFFT, a process described as follows: F=Φ||IFFT(σ(FE(FFT(F in ))*FFT(F in )))|| Where Φ||·|| and σ represent modular arithmetic and element-wise multiplication, respectively; The second link in the FFTA module executes a frequency domain attention mechanism: it employs a Transformer-based attention mechanism to map input features to queries, keys, and values; then, it weights and aggregates values ​​by calculating the similarity between queries and keys, thereby modeling global information; finally, it applies FFT again to obtain the three components required for attention computation, namely the query Q. f Key K f Sum V f : Q f =FFT Q (Norm(F in ));K f =FFT K (Norm(F in ));V f =FFT V (Norm(F in )); Where Norm() represents layer normalization; Finally, embed the time step into F t =MLP(e t ), which is incorporated into the key value, namely K f ·F t The initial attention map obtained is: Λ f =Q f *(K f *F t Extracting the real part and the virtual part Where conj(·) represents the conjugate complex number, the CCN module then activates and merges the real and imaginary parts to obtain the attention map, and the attention map is then passed through the Softmax activation function to obtain the frequency domain weights aΛ. f The processing procedure of the CCN module is as follows: Where Θ(·,·) represents the combination function that combines the imaginary and real parts into a complex number; Sof(·) represents the Softmax function; Then, the frequency domain weight aΛ f The result is transformed back to the spatial domain using Inverse Fourier Transform (IFFT) and concatenated with F. The concatenated result is then processed through convolution and further processed by parallel frequency and spatial branches: the spatial branch uses a depthwise separable convolution module (DWConv), while the frequency branch processes the result sequentially using a Fast Fourier Transform (FFT) frequency enhancement module (FE) and an Inverse Fourier Transform (IFFT). Finally, the results from both the spatial and frequency branches are convolved to obtain a frequency-weighted F guided by the fused time code t. out ; The second branch executes the convolutional network processing flow, which consists of a downsampling pooling layer, a Fuse_T module, and a convolutional layer; the output of this branch is guided by the weights of the first branch. out The merging process employs matrix multiplication to obtain the first-layer downsampled feature map, which is then fed into the merging channel for processing, specifically including: First, the first-layer downsampled feature map is sequentially processed by the downsampling pooling module Down Avg, the Fuse_T module, and the first convolutional layer to obtain the highest-level semantic encoding information; Then, after encoding, a general decoding operation based on a convolutional network is performed. The decoding operation is implemented sequentially by the first upsampling convolutional module Up Conv, the Fuse_T module, the second convolutional layer, the second upsampling convolution, and the third convolutional layer. Finally, a convolutional layer is used to adjust the channels to obtain the final denoising result output.

2. The OCT retinal image denoising method based on a high-frequency enhanced diffusion model as described in claim 1, characterized in that, The Adjust module is a shallow network consisting of three loops of convolutional layers, ReLU activation function, and pooling layers, and finally uses convolutional layers to adjust the channel output. Its input is x T This is the noise map obtained in the forward diffusion T steps of the diffusion model, which is the initial noise map I input to the network. This is a denoised image recovered from a time step on the network; during the testing phase, Adjust is disabled, leaving only the time embedding to adjust the various stages of the network.

3. The OCT retinal image denoising method based on a high-frequency enhanced diffusion model as described in claim 1, characterized in that, The frequency enhancement module (FE) consists of convolution, activation function, convolution, and sigmoid layers, and the processing procedure is as follows: FE(FFT(F in ))=Sigmoid(Conv(ReLU(Conv(FFT(F in ))))) 4. The OCT retinal image denoising method based on a high-frequency enhanced diffusion model as described in claim 1, characterized in that, It also includes the FreqSkip module, whose processing procedure is described as follows: F skip =Fuse_T(Conv(F split (F in -Up(Pool(F in )))),F t ) Where Pool is the pooling operation, Up is the upsampling operation, and F... split Represents frequency domain decomposition, F split (F in -Up(Pool(F in That is, the features are divided into high-frequency and low-frequency parts in the graph by pooling. The FreqSkip module outputs F skip The output of the first convolutional layer (Conv) in the merged channel is concatenated with the output of the first convolutional layer (Up Conv) in the channel dimension and then fed into the first upper convolutional layer (Up Conv) for subsequent upsampling processing.

Images