Clinical as-oct image restoration method and system based on conditional diffusion model

Abstract

The application provides a clinical AS-OCT image restoration method and system based on a conditional diffusion model, the method comprising: preprocessing an AS-OCT image dataset to obtain a degraded image and an edge image; training an initialized GAN model by using the degraded image and the edge image to obtain a trained GAN model; training an initialized U-Net model by using a clear image and the edge image to obtain a trained conditional diffusion model; and processing the degraded image and the edge image by using the trained GAN model and the trained conditional diffusion model to obtain a restored image. Through the two-stage cascade design of "edge prediction-edge guidance", the data consistency constraint term depending on iterative calculation in the traditional conditional diffusion model is abandoned, so that the long iterative sampling is avoided, higher image quality is obtained, the overall efficiency of the restoration process is improved, and the application feasibility in the clinical actual scene is enhanced.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a clinical AS-OCT image restoration method and system based on a conditional diffusion model. Background Technology

[0002] AS-OCT is a rapid in vivo imaging technique capable of high-resolution cross-sectional visualization of the anterior segment's physiological structures. This technology has been widely applied in clinical practice, with typical applications including providing valuable corneal morphological information for preoperative preparation of refractive surgery and quantifying the interaction mechanisms between the ciliary muscle and lens in myopic and presbyopic patients. However, AS-OCT has inherent imaging limitations, primarily manifested as global speckle noise caused by destructive interference of multiple scattered waves and frequent streaks of saturation artifacts at the corneal apex. Speckle noise obscures fine structural information and reduces image reliability, while saturation artifacts hinder structural visualization and interfere with accurate corneal diagnosis.

[0003] Traditionally, there are two main technical approaches to suppressing saturation artifacts and speckle noise in AS-OCT images. First, commercial systems generally rely on hardware-level solutions. For example, dual-channel detection technology can mitigate saturation artifacts, while composite scanning technology can suppress speckle noise. Although these hardware-oriented methods are widely used, they have inherent drawbacks: the required scanning process is often time-consuming, and the requirement for the patient to keep their eyes still is very strict, which is difficult to achieve accurately in clinical practice. Second, much research has focused on developing traditional post-processing algorithms to improve the quality of acquired OCT images. However, these software-level technologies have a fundamental limitation: their performance is highly dependent on manual parameter tuning, a shortcoming that severely restricts their practical application value in clinical scenarios.

[0004] Although existing studies have achieved good results in dealing with single degradation factors (such as speckle noise removal or saturation artifact restoration), there is currently no method to simultaneously solve the problems caused by multiple degradation factors in real AS-OCT clinical scenarios. This is mainly due to the following challenges: (1) Different degradation factors correspond to different image processing tasks. Speckle removal belongs to the denoising problem, while saturation artifact restoration belongs to the image completion task. (2) Clinical data acquisition is both time-consuming and costly, making it difficult to obtain large-scale labeled data that matches specific degradation factors. (3) The processing algorithm must ensure structural fidelity in order to be directly applied to clinical decision-making. Summary of the Invention

[0005] In view of the above, the main objective of this invention is to propose a clinical AS-OCT image restoration method and system based on a conditional diffusion model to solve the aforementioned technical problems.

[0006] This invention proposes a clinical AS-OCT image restoration method based on a conditional diffusion model, the method comprising the following steps:

[0007] Step 1: Preprocess the AS-OCT image dataset to obtain degraded images and edge images;

[0008] Step 2: Train the initialized GAN model using the degraded image and edge image to obtain the trained GAN model;

[0009] Step 3: Train the initialized U-Net model using clear images and edge images to obtain the trained conditional diffusion model;

[0010] Step 4: Use the trained GAN model and the trained conditional diffusion model to process the degraded image and edge image to obtain the restored image.

[0011] This invention also proposes a clinical AS-OCT image restoration system based on a conditional diffusion model, the system comprising:

[0012] Image preprocessing module, used for:

[0013] Preprocessing the AS-OCT image dataset yields degraded and edge images;

[0014] The GAN model training module is used for:

[0015] The initialized GAN model is trained using degraded and edge images to obtain a trained GAN model.

[0016] The diffusion model training module is used for:

[0017] The initialized U-Net model is trained using clear and edge images to obtain a trained conditional diffusion model;

[0018] Image restoration module, used for:

[0019] The degraded and edge images are processed using a trained GAN model and a trained conditional diffusion model to obtain the restored image.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0021] 1. This invention introduces edge structures as explicit constraints, which can effectively maintain the fidelity of key anatomical structures when restoring images.

[0022] 2. This invention uses a two-stage cascaded design of "edge prediction - edge guidance" to directly guide the reverse generation process of the diffusion model using the predicted edge map. This design eliminates the data consistency constraint term that relies on iterative calculation in the traditional conditional diffusion model, thereby avoiding lengthy iterative sampling. While obtaining higher image quality, it improves the overall efficiency of the restoration process and enhances its feasibility for application in clinical practice. Attached Figure Description

[0023] Figure 1 This is a flowchart of the clinical AS-OCT image restoration method based on the conditional diffusion model proposed in this invention;

[0024] Figure 2 This is a flowchart illustrating the overall workflow of the clinical AS-OCT image restoration method based on the conditional diffusion model proposed in this invention.

[0025] Figure 3 A pseudo-color comparison image and a schematic diagram of the corresponding residual image showing the speckle noise removal effect of a real AScasia image;

[0026] Figure 4 A pseudo-color comparison diagram of the speckle noise removal effect on the real CMCasia dataset;

[0027] Figure 5 A visual comparison of the saturation artifact restoration effects on real Corcasia dataset images;

[0028] Figure 6 A visual comparison of saturation artifact restoration effects based on simulated Corcasia dataset images;

[0029] Figure 7 For images containing both speckle noise and saturation artifacts;

[0030] Figure 8 This is a schematic diagram of the framework of the clinical AS-OCT image restoration system based on the conditional diffusion model proposed in this invention. Detailed Implementation

[0031] 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.

[0032] 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.

[0033] Please see Figure 1 and Figure 2 This invention proposes a clinical AS-OCT image restoration method based on a conditional diffusion model, which includes the following steps:

[0034] Step 1: Preprocess the AS-OCT image dataset to obtain degraded images and edge images;

[0035] In step 1, the AS-OCT image dataset is preprocessed to obtain degraded images and edge images. The specific steps are as follows:

[0036] For saturation artifact images in the AS-OCT image dataset, a degraded image is generated using a sharp image and a binary mask; for speckle noise images in the AS-OCT image dataset, a degraded image is obtained by pairing a sharp image with a noisy image.

[0037] The degraded image is processed using the Canny edge detection operator to obtain the edge image;

[0038] In the process of generating a degraded image from a sharp image and a binary mask for saturation artifact images in the AS-OCT image dataset, the following relationship exists:

[0039] ;

[0040] in, Indicates a degraded image. Represents a clear image; Represents a binary mask, when A value of 1 indicates a specific region. A value of 0 indicates a known background area; Indicates Adama's product;

[0041] In the process of pairing clear and noisy images to obtain degraded images from speckle noise images in the AS-OCT image dataset, the following relationship exists:

[0042] ;

[0043] in, This represents random speckle noise.

[0044] Step 2: Train the initialized GAN model using the degraded image and edge image to obtain the trained GAN model;

[0045] In step 2, the initialized GAN model is trained using the degraded image and the edge image to obtain the trained GAN model. The specific steps are as follows:

[0046] The network parameters of the generator and discriminator of a given GAN model are initialized to obtain an initialized GAN model.

[0047] The degraded image and the edge image are input into the initialized GAN model to obtain the predicted edge image;

[0048] An adversarial loss and a feature matching loss are constructed using predicted edge images;

[0049] An objective function is constructed using adversarial loss and feature matching loss.

[0050] The discriminator of the GAN model is fixed, and the network parameters of the generator of the GAN model are updated using the objective function. Then the generator is fixed, the network parameters of the discriminator are updated using adversarial loss, and the Adam optimizer is used for iterative optimization. Training stops after reaching a preset number of rounds, and a trained GAN model is obtained.

[0051] The degraded image and the edge image are input into the initialized GAN model to obtain the predicted edge image. For the saturation artifact task, the following relationship applies:

[0052] ;

[0053] in, Represents the predicted edge image. This indicates processing via a generator. This indicates processing via the concatenation operator. Represents the edge image;

[0054] For speckle noise tasks, the following relationship holds:

[0055] ;

[0056] In constructing adversarial loss and feature matching loss using predicted edge images, the following relationship exists:

[0057] ;

[0058] in, Indicating resistance to loss, This indicates that the data is processed by a discriminator. Represents a true edge image. Indicates a pair of samples Expectations This expresses the expectation of a clear image. This indicates that the discriminator is effective against sample pairs. The judgment result, This indicates that the discriminator is effective against sample pairs. The judgment result, Represents feature matching loss, Represents the mathematical expectation. This indicates the total number of convolutional layers in the discriminator. Indicates the first The number of elements in the layer Indicates passing through the first In-layer discriminator processing, Represents the L1 norm;

[0059] In constructing the objective function using adversarial loss and feature matching loss, the following relationship exists:

[0060] ;

[0061] in, Both represent regularization parameters. Describe the objective function. This indicates that the generator is minimized for a fixed discriminator. This indicates that the discriminator is maximized for a fixed generator.

[0062] Furthermore, the role of feature matching loss is to minimize the difference between the intermediate activation maps of the discriminator for real and generated images, thereby stabilizing the network training process. This constraint can prompt the generator to learn and generate representations that are similar to the feature distribution of real data. Although feature matching loss is similar to the perceptual loss using a pre-trained VGG network, the former has a significant advantage in the early stages of training. Specifically, VGG networks are not designed for edge-aware tasks, so they cannot capture the structural information that is crucial at this stage.

[0063] Furthermore, to further improve training stability, this invention introduces spectral normalization, which normalizes the maximum singular value of the weight matrix to effectively constrain the Lipshitz constant of the network. Although spectral normalization was originally designed for the discriminator, its application to the generator can also alleviate abrupt fluctuations in parameters and gradients. Based on this, we apply spectral normalization to both the generator and the discriminator.

[0064] It should be noted that when the discriminator provides the activation map for the feature matching loss, it does not involve the gradient update process of that loss term. Therefore, only the adversarial loss needs to be maximized with respect to the discriminator. In this invention, the parameters are set... ;

[0065] In summary, in speckle noise removal, the edge generation process functions as a learnable Canny-like edge detection operator, while in saturation artifact removal, the edge generation process acts as an edge predictor.

[0066] Step 3: Train the initialized U-Net model using clear images and edge images to obtain the trained conditional diffusion model;

[0067] In step 3, the initialized U-Net model is trained using clear and edge images to obtain a trained conditional diffusion model. The specific steps are as follows:

[0068] Initialize the given U-Net model to obtain the initialized U-Net model;

[0069] Under the condition of a preset random sampling time step, a noisy image is generated using a clear image;

[0070] The noisy image and the edge image are stitched together along the channel dimension to obtain the stitched image;

[0071] The conditional diffusion loss is constructed using stitched images; the initial U-Net model is updated using the diffusion loss through the Adam optimizer, and training is stopped after reaching the preset sampling time step to obtain the trained conditional diffusion model.

[0072] Under the condition of a preset random sampling time step, the following relationship exists in the process of generating a noisy image from a clear image:

[0073] ;

[0074] in, Indicates time step Noisy images at that time, Indicates time step The cumulative coefficient of the diffusion process over time. Indicates Gaussian noise. Indicates the total number of sampling steps. Indicates time step The coefficients of the original signal are preserved at the same time. Indicates time step noise variance at time;

[0075] In the process of constructing conditional diffusion loss using stitched images, the following relationship exists:

[0076] ;

[0077] in, Indicates diffusion loss, Indicates time step Noise predicted by the network at that time This indicates processing via a noise estimator. It expresses a desire for joint expectation.

[0078] Step 4: Use the trained GAN model and the trained conditional diffusion model to process the degraded image and edge image to obtain the restored image;

[0079] In step 4, the degraded image and edge image are processed using the trained GAN model and the trained conditional diffusion model to obtain the restored image. The specific steps are as follows:

[0080] Input the degraded image and edge image into the trained GAN model to obtain the predicted edge image;

[0081] A noisy image is obtained by sampling from a standard Gaussian distribution; the predicted edge image is concatenated with the noisy image along the channel dimension and input into a trained conditional diffusion model to obtain the predicted noise;

[0082] By using the predicted noise, the noisy image of the previous time step is calculated through backsampling;

[0083] After completing the reverse sampling at the preset time step, the restored image is obtained;

[0084] In the process of using predicted noise to calculate the noisy image of the previous time step through inverse sampling, the following relationship exists:

[0085] ;

[0086] in, Indicates time step Noisy images at that time, Indicates time step Prediction noise at that time Indicates time step The standard deviation of the Gaussian distribution over time, This indicates the noise term.

[0087] It should be noted that, please refer to Figure 2 The framework of this invention will be achieved by... As guiding information, it effectively addresses the unique challenges of AS-OCT image restoration. The edge map is feature-stitched with the degraded image, thus providing conditional constraints for the backdiffusion sampling process.

[0088] Given a sample of a target conditional distribution The forward diffusion process is as follows: based on the variance sequence Gradually moving towards within T time steps Gaussian noise is added; this process generates a series of samples with increasing noise levels via a Markov chain. ;

[0089] It is important to note that all All with Sharing the same conditional edge information ( The relationship between ) and spatial dimension is as follows:

[0090] ;

[0091] in, Represents conditional distribution, This indicates that the distribution follows a Gaussian distribution with parameters. Represents the identity matrix;

[0092] make Therefore, we can conclude that: ;

[0093] In the reverse diffusion process, the core objective is to utilize conditional edge information ( From Gaussian noise samples China restoration However, directly estimating the conditional distribution It is usually quite difficult, but when the distribution is... When the condition is met, the solution becomes easier to handle, and the specific form is as follows:

[0094] ;

[0095] in, Represents conditional distribution, This represents the reparameterized mean;

[0096] Due to the generation process of new data samples, Essentially unknown, this invention trains a parameterized model for conditional distributions. An approximation is made, specifically in the following form:

[0097] ;

[0098] in, Indicates reverse diffusion distribution. Indicates the mean parameter;

[0099] Further by Represented as and The linear combination of these factors, the sampling steps can be formally expressed as:

[0100] ;

[0101] Among them, noise item Introduce randomness into the generation process at each time step;

[0102] In the model framework of this invention, the design purpose of the denoising U-Net is based on , as well as predict During the reverse diffusion process, and All will be with Perform feature splicing;

[0103] In summary, the GCDM model proposed in this invention is built upon two generative paradigms. First, it utilizes GAN to extract structured edge information from degraded images. Then, the predicted edge map is used as a guiding condition input to the conditional diffusion model to complete the image generation task. This design forms an end-to-end unified framework that seamlessly connects the feature extraction stage from "image to edge" with the structure generation stage from "edge to image".

[0104] Specifically, this invention also conducts a comprehensive experimental evaluation of the performance of the GCDM model in two typical AS-OCT image processing tasks (speckle noise removal and saturation artifact restoration).

[0105] The private real AS-OCT datasets used in the experiments of this invention are Corcasia, AScasia and CMCasia. The Corcasia dataset is used for the saturation artifact restoration task, while the CMCasia and AScasia datasets are used for the speckle noise removal task.

[0106] This invention uses four metrics to evaluate the performance of each method: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), Learned Perceptual Image Patch Similarity (LPIPS), and Perceptual Image Distortion (PID).

[0107] It's important to note that PSNR quantifies the low-level pixel-level fidelity of an image, while SSIM assesses the structural similarity of an image, providing a semantic integrity metric related to human visual perception. LPIPS and PID further supplement the evaluation dimensions, both used to quantify high-level perceptual differences in images. PID, trained based on human visual perception, exhibits strong robustness against various distortion types. In result evaluation, higher PSNR and SSIM values ​​indicate better image restoration quality, while lower LPIPS and PID values ​​represent better perceptual consistency.

[0108] This invention compares and evaluates the speckle noise removal performance of various methods based on real clinical AS-OCT images from the AScasia and CMCasia datasets. The quantitative evaluation results are summarized in Tables 1 and 2. Specifically, the GCDM model achieved the highest values ​​for both PSNR and SSIM, while achieving the lowest values ​​for both LPIPS and PID. This comprehensive leading performance demonstrates that the GCDM model achieves the optimal balance between efficiently removing speckle noise and accurately preserving image structural details, thereby significantly improving image quality.

[0109] Table 1. Comparison of speckle noise removal results based on real AScasia images

[0110]

[0111] Table 2. Comparison of speckle noise removal results based on real CMCasia images

[0112]

[0113] Furthermore, visual quality assessment is a key criterion for evaluating speckle noise removal performance. In practical applications, a successfully denoised image needs to retain clear geometric structure while minimizing the introduction of visible artifacts. To conduct this qualitative analysis, this invention... Figure 3 and Figure 4 The paper presents representative experimental results based on the AScasia and CMCasia datasets, including pseudo-color visualizations and their corresponding residual plots.

[0114] It should be noted that, Figure 3 The middle (orange and pink) area is a magnified local area used to visually highlight the model's edge preservation capabilities; Figure 4 The corresponding residual image and downstream ciliary muscle segmentation results are also listed, with the red area being the labeled area and the yellow area being the segmentation output of the pre-trained U-Net model.

[0115] pass Figure 3 and Figure 4 The magnified visual results demonstrate that the GCDM model excels in preserving fine structural details. As shown in the magnified pink and orange areas, the model not only effectively maintains clear tissue boundaries but also achieves superior speckle suppression in uniform background regions. In contrast, the comparative methods reveal significant shortcomings. The ILVR and MCG methods still retain a large amount of noise, while the outputs of the DDRM, DPS, and DiffPIR methods exhibit excessive smoothing, leading to a loss of structural fidelity in the images.

[0116] To evaluate the performance of the GCDM model proposed in this invention, a variety of image restoration methods were selected for comparative experiments. The selected methods were divided into three categories: the first category is dedicated techniques designed for the restoration of saturation artifacts in AS-OCT images, including DualSC and SCGAN; the second category is general-purpose image restoration models, such as LGNet and EdgeConnect; the third category is diffusion model-based restoration algorithms that have emerged in recent years. Repaint was introduced for comparison. In addition, considering that saturation artifact restoration belongs to a specific image restoration task, the current leading diffusion model DPS was also included as a supplementary baseline method.

[0117] Saturation artifacts in AS-OCT images can severely distort corneal structures, and the core task of artifact restoration is to fill in the missing tissue structures within the saturation regions. Figure 5 This paper presents qualitative results of saturation artifact restoration based on real AS-OCT images. Key corneal structures are highlighted and magnified in the images for detailed examination. Figure 5 In the image, the pink area in the first row has been enlarged to the second row; the third row shows the corneal segmentation results: white arrows indicate missing areas, green arrows indicate distorted areas, orange arrows highlight redundant segmentation areas, and blue arrows represent overly smoothed areas.

[0118] Experimental results show that DualSC, EdgeConnect and other methods result in corneal geometric distortion after restoration, while RePaint and LGNet methods have the defect of blurred structural boundaries. In contrast, DPS, SCGAN and the GCDM model proposed in this invention can successfully restore artifacts and generate images with clear and distinguishable structural boundaries.

[0119] Given the lack of corresponding real labels for artifact regions in real clinical images, this invention constructs experimental data by manually adding saturation masks to complete AS-OCT images to conduct quantitative evaluation. As shown in Table 1 above, the GCDM model achieves the best restoration effect, not only achieving the highest values ​​in both PSNR and SSIM, but also the lowest values ​​in both LPIPS and PID.

[0120] Please refer to Figure 6 , Figure 6 The visualization comparison focuses on the key corneal vertex region, which has been highlighted and magnified to aid analysis. The comparison results show that EdgeConnect and LGNet methods cannot restore structurally coherent tissue content; RePaint and DPS methods can restore artifacts, but introduce additional artifacts in the boundary regions; SCGAN and the GCDM model proposed in this study can generate visually reasonable restoration results. It is worth noting that the GCDM model shows superior performance in restoring fine texture details and preserving clear anatomical structure boundaries.

[0121] It should be noted that, Figure 6 The pink area in the first row is enlarged to the second row for display. The third row shows the corneal segmentation results, where white arrows indicate missing areas, green arrows indicate distorted areas, and orange arrows indicate redundant segmentation areas.

[0122] To further evaluate the downstream application value of the GCDM model in the saturation artifact restoration task, this invention conducts corneal segmentation experiments based on the Corcasia dataset, and the network architecture and training configuration used in the experiments are as follows:

[0123] The quantitative results of the downstream corneal segmentation task are shown on the right side of Table 3. The various indicators verify that the proposed GCDM model can significantly improve the performance of the downstream automated analysis task, and achieve the best values ​​in both the F1 score and IoU, the two core indicators.

[0124] Table 3. Comparison of simulated saturation artifact restoration effects based on the Corcasia dataset

[0125]

[0126] Representative segmentation visualization results of real and simulated saturation artifact images are shown in [the figures]. Figure 5 and Figure 6The last line; Visual analysis of segmentation results based on the real Corcasia dataset shows that existing restoration methods have significant limitations; specifically, images processed by EdgeConnect, LGNet, and DualSC methods lose key tissue structures, thus affecting the performance of subsequent segmentation tasks; the SCGAN method introduces obvious structural distortions; the RePaint method produces redundant oversegmented regions; and the output of the DPS method suffers from over-smoothing, missing important details.

[0127] In contrast, the restored images generated by the GCDM model proposed in this invention possess both structural coherence and visual superiority, preserving the anatomical fidelity required for accurate automated analysis to the greatest extent. Furthermore, qualitative evaluation results on the simulated Corcasia dataset show that several baseline restoration methods, including EdgeConnect, LGNet, SCGAN, RePaint, and DPS, share common defects. Specifically, these methods generate images with significant edge loss; compared to the true labels, key structural boundaries appear fragmented or even completely disappeared. The restoration results generated by the GCDM model effectively preserve these key edge information, and the final segmentation results outperform all the compared methods in both structural accuracy and visual effect.

[0128] This result confirms that the GCDM model has the potential to improve the reliability of edge feature extraction, thereby improving the overall performance of computer-aided corneal diagnosis.

[0129] To explore the effect of edge prior guidance on conditional diffusion models, this invention introduces the edge map generated by the Canny edge detector into each contrast restoration model by concatenating it with the input image during the training and testing phases.

[0130] As shown in Table 4, all methods incorporating edge guidance exhibit performance improvements, specifically increased Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index (SSIM), and decreased Perceptual Patch Similarity (LPIPS) and Perceptual Image Distortion (PID). This indicates that incorporating prior structural information helps optimize the model's generation process. It is noteworthy that the core difference between DDPM+ (Canny) and GCDM lies in the fact that GCDM uses the edge map predicted in its first stage, while DDPM+ (Canny) directly uses the edge map generated by the Canny operator.

[0131] The GCDM model proposed in this invention achieves superior image restoration quality through a dedicated two-step process of "edge prediction - edge-aware guidance," outperforming all baseline models that incorporate edge guidance.

[0132] Table 4. Comparative Study of Different Speckle Removal Methods Based on Canny Edge Guidance on the AScasia Dataset

[0133]

[0134] In the edge prediction stage of the GCDM framework, considering that the Canny edge detector has the advantages of high computational efficiency, strong robustness and simple operation, the edge map generated by it is used as the training label in this invention.

[0135] To quantitatively verify the applicability of the detector, this invention conducted a comparative analysis of the restoration performance of different detectors extracting edge features within the basic GCDM framework. As shown in Table 5, compared with the Prewitt, Sobel and HED operators, the Canny detector can significantly improve the image restoration quality and is therefore selected as the edge label generation tool in the initial stage.

[0136] It is important to emphasize that although the Canny edge detector demonstrates superior performance among traditional methods, its results are still inferior to the GCDM model proposed in this study when directly applied to guided diffusion models. This indicates that the edge generation model dedicated to the first stage of GCDM has a stronger ability to capture edge information. Furthermore, unlike simple edge extraction methods, the GCDM edge prediction model can infer missing structural information in saturation artifact restoration tasks—a capability with significant value in clinical analysis.

[0137] Table 5. Performance comparison of different edge detection operators based on GCDM in speckle removal task on the AScasia dataset.

[0138]

[0139] Please see Figure 7 Analysis results show that in images containing speckle noise, the edge prediction model proposed in this invention performs comparably to the Canny detector; while in complex regions with artifacts, the model outperforms the model in extracting key anatomical contours such as the corneal apex. This performance is crucial for downstream tasks such as corneal segmentation and clinical assessment.

[0140] Depend on Figure 7 The red arrow in the middle highlights the ability of the GCDM model to recover missing edge information in the predicted edge results output in the first stage.

[0141] Please see Figure 8 This invention also provides a clinical AS-OCT image restoration system based on a conditional diffusion model, the system comprising:

[0142] Image preprocessing module, used for:

[0143] Preprocessing the AS-OCT image dataset yields degraded and edge images;

[0144] The GAN model training module is used for:

[0145] The initialized GAN model is trained using degraded and edge images to obtain a trained GAN model.

[0146] The diffusion model training module is used for:

[0147] The initialized U-Net model is trained using clear and edge images to obtain a trained conditional diffusion model;

[0148] Image restoration module, used for:

[0149] The degraded and edge images are processed using a trained GAN model and a trained conditional diffusion model to obtain the restored image.

[0150] 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.

[0151] 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.

[0152] 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 clinical AS-OCT image restoration method based on a conditional diffusion model, characterized in that, The method includes the following steps: Step 1: Preprocess the AS-OCT image dataset to obtain degraded images and edge images. The specific steps are as follows: For saturation artifact images in the AS-OCT image dataset, a degraded image is generated using a sharp image and a binary mask; for speckle noise images in the AS-OCT image dataset, a pairing process is performed between a sharp image and a noisy image to obtain a degraded image. The degraded image is processed using the Canny edge detection operator to obtain the edge image; Step 2: Train the initialized GAN model using the degraded image and edge image to obtain the trained GAN model. The specific steps are as follows: The network parameters of the generator and discriminator of a given GAN model are initialized to obtain an initialized GAN model. The degraded image and the edge image are input into the initialized GAN model to obtain the predicted edge image; An adversarial loss and a feature matching loss are constructed using predicted edge images; An objective function is constructed using adversarial loss and feature matching loss. The discriminator of the fixed-initialization GAN model is used to update the network parameters of the generator of the initialized GAN model using the objective function; Then fix the generator, update the network parameters of the discriminator using adversarial loss, and use the Adam optimizer for iterative optimization. Stop training after reaching the preset number of rounds to obtain the trained GAN model. Step 3: Train the initialized U-Net model using clear and edge images to obtain the trained conditional diffusion model. The specific steps are as follows: Initialize the given U-Net model to obtain the initialized U-Net model; Under the condition of a preset random sampling time step, a noisy image is generated using a clear image; The noisy image and the edge image are stitched together along the channel dimension to obtain the stitched image; Conditional diffusion loss is constructed using stitched images; Using the diffusion loss, the initialized U-Net model is updated through the Adam optimizer. After reaching the preset sampling time step, training is stopped, and the trained conditional diffusion model is obtained. Step 4: Use the trained GAN model and the trained conditional diffusion model to process the degraded image and edge image to obtain the restored image. The specific steps are as follows: Input the degraded image and edge image into the trained GAN model to obtain the predicted edge image; A noisy image is obtained by sampling from a standard Gaussian distribution; the predicted edge image is concatenated with the noisy image along the channel dimension and input into a trained conditional diffusion model to obtain the predicted noise; By using the predicted noise, the noisy image of the previous time step is calculated through backsampling; After completing the reverse sampling at the preset time step, the restored image is obtained.

2. The clinical AS-OCT image restoration method based on the conditional diffusion model according to claim 1, characterized in that, In the process of generating a degraded image from a sharp image and a binary mask for saturation artifact images in the AS-OCT image dataset, the following relationship exists: ； in, Indicates a degraded image. Indicates a clear image. Represents a binary mask. Indicates Adama's product; In the process of pairing clear and noisy images to obtain degraded images from speckle noise images in the AS-OCT image dataset, the following relationship exists: ； in, This represents random speckle noise.

3. The clinical AS-OCT image restoration method based on the conditional diffusion model according to claim 2, characterized in that, The degraded image and the edge image are input into the initialized GAN model to obtain the predicted edge image. For the saturation artifact task, the following relationship applies: ； in, Represents the predicted edge image. This indicates processing via a generator. This indicates processing via the concatenation operator. Represents the edge image; For speckle noise tasks, the following relationship holds: ； In constructing adversarial loss and feature matching loss using predicted edge images, the following relationship exists: ； in, Indicating resistance to loss, This indicates that the data is processed by a discriminator. Represents a true edge image. Indicates a pair of samples Expectations This expresses the expectation of a clear image. This indicates that the discriminator is effective against sample pairs. The judgment result, This indicates that the discriminator is effective against sample pairs. The judgment result, Represents feature matching loss, Represents the mathematical expectation. This indicates the total number of convolutional layers in the discriminator. Indicates the first The number of elements in the layer Indicates passing through the first In-layer discriminator processing, Represents the L1 norm; In constructing the objective function using adversarial loss and feature matching loss, the following relationship exists: ； in, Both represent regularization parameters. Describe the objective function. This indicates that the generator is minimized for a fixed discriminator. This indicates that the discriminator is maximized for a fixed generator.

4. The clinical AS-OCT image restoration method based on the conditional diffusion model according to claim 3, characterized in that, Under the condition of a preset random sampling time step, the following relationship exists in the process of generating a noisy image from a clear image: ； in, Indicates time step Noisy images at that time Indicates time step The cumulative coefficient of the diffusion process over time, Indicates Gaussian noise. Indicates the total number of sampling steps. Indicates time step The coefficients of the original signal are preserved at the same time. Indicates time step noise variance at time; In the process of constructing conditional diffusion loss using stitched images, the following relationship exists: ； in, Indicates diffusion loss, Indicates time step Noise predicted by the network at that time This indicates processing via a noise estimator. It expresses a desire for joint expectation.

5. The clinical AS-OCT image restoration method based on the conditional diffusion model according to claim 4, characterized in that, In the process of using predicted noise to calculate the noisy image of the previous time step through inverse sampling, the following relationship exists: ； in, Indicates time step Noisy images at that time, Indicates time step Prediction noise at that time Indicates time step The standard deviation of the Gaussian distribution over time, This indicates the noise term.

6. A clinical AS-OCT image restoration system based on a conditional diffusion model, characterized in that, The system employs the clinical AS-OCT image restoration method based on the conditional diffusion model as described in any one of claims 1 to 5, and the system comprises: Image preprocessing module, used for: Preprocessing the AS-OCT image dataset yields degraded and edge images; The GAN model training module is used for: The initialized GAN model is trained using degraded and edge images to obtain a trained GAN model. The diffusion model training module is used for: The initialized U-Net model is trained using clear and edge images to obtain a trained conditional diffusion model; Image restoration module, used for: The degraded and edge images are processed using a trained GAN model and a trained conditional diffusion model to obtain the restored image.

Images