An OCT image segmentation method based on fast Fourier transform and SKAttention mechanism

Abstract

This invention discloses an OCT image segmentation method based on Fast Fourier Transform (FFT) and SKAttention mechanisms. By introducing an image preprocessing module into the initial stage of the UNet network (i.e., the backbone structure), this invention significantly improves the segmentation accuracy and efficiency of edema regions in OCT fundus images. The image preprocessing module, combining FFT and SKAttention mechanisms, is designed to perform preliminary transformations and feature weighting on the input OCT image data to enhance the network's ability to recognize edema features. This technical solution optimizes the feature processing flow and enhances the model's ability to recognize fundus lesion features, particularly demonstrating high efficiency and accuracy when processing complex or low-contrast edema regions. Furthermore, this method maintains high processing speed while possessing good adaptability and scalability, making it suitable for large-scale medical image processing applications.

Description

Technical Field

[0001] This invention belongs to the fields of medical image processing, deep learning, and signal processing technology, and specifically relates to an image processing technology based on deep learning. Background Technology

[0002] Current technologies for segmenting fundus edema in optical coherence tomography (OCT) primarily utilize traditional image processing techniques and basic deep learning models. Traditional methods include image segmentation using techniques such as thresholding and region growing. These methods rely on manually set parameters and thresholds, lacking flexibility and adaptability. With the development of deep learning technology, the UNet network model has been widely applied in medical image segmentation tasks due to its effective feature capture capabilities and good contextual information integration capabilities.

[0003] However, despite UNet's superior performance in various medical image processing tasks, existing techniques still face several technical challenges when processing OCT fundus images. First, OCT images often suffer from noise and low contrast due to light scattering and absorption, making it difficult for traditional UNet models to accurately identify and segment subtle edema areas. Second, existing UNet models lack attention mechanisms for local image regions when processing features, which limits the model's segmentation accuracy in detail.

[0004] This invention aims to address a major technical problem in existing technologies: improving the segmentation accuracy and contrast of edema regions in OCT fundus images. By introducing Fast Fourier Transform (FFT) to preprocess the original input image, this invention can enhance image features in the frequency domain, reduce the impact of noise, and simultaneously restore the spatial structure of the image through inverse transform. Furthermore, the introduction of the SelectKernelAttention (SKAttention) mechanism allows the network to focus more intently on key edema regions, thereby providing more accurate segmentation results. This method, combining FFT and SKAttention, offers a novel solution to address the limitations of existing technologies, significantly improving segmentation accuracy and efficiency. Summary of the Invention

[0005] Accurate edema segmentation is crucial for the diagnosis and treatment of ophthalmic diseases in the processing of optical coherence tomography (OCT) images of the fundus. Existing techniques often encounter technical challenges in handling such tasks, such as inconsistent image quality and difficulty in capturing details, which affect the accuracy and reliability of segmentation. To address these issues, this invention proposes an OCT image segmentation method based on Fast Fourier Transform (FFT) and SKAttention mechanisms, aiming to improve the accuracy and efficiency of edema region segmentation in OCT images.

[0006] The core of this invention is the integration of a specific image preprocessing module before the UNet network model. This module first uses FFT to transform the raw image data to the frequency domain. This step leverages the efficiency of FFT in image analysis, highlighting structural information in the image and providing a richer data foundation for subsequent processing. Next, adaptive SKattention is applied to recalibrate the features of the transformed data, dynamically adjusting the weights of each channel to enhance the response to key features. Finally, an inverse Fast Fourier Transform is performed to transform the processed data back to its original space, preserving the image's size and structure. Crucially, this preprocessing module demonstrates optimal performance only when applied to the initial part of the network; this finding provides important guidance for the module's design and application.

[0007] This invention effectively solves the problem of accurate segmentation of edema regions in complex and low-contrast OCT images. Experimental results show that compared with traditional processing methods, this invention significantly improves the accuracy and robustness of segmentation, especially in processing image details and low-contrast regions. Therefore, this invention not only improves the technical level of medical image processing but may also provide more reliable support for clinical diagnosis and patient management.

[0008] An OCT image segmentation method based on Fast Fourier Transform and SKAttention mechanism is described below:

[0009] Step 1: Data acquisition and preprocessing.

[0010] Step 2: Preprocess the original input image using Fast Fourier Transform (FFT);

[0011] Step 3: Apply adaptive SKAttention to recalibrate the features of the transformed data. By dynamically adjusting the weights of each channel, the response to key features is enhanced.

[0012] Step 4: Fast Inverse Fourier Transform.

[0013] The frequency domain data processed by SKAttention, i.e. the final feature map Vc, is remapped to the spatial domain through inverse fast Fourier transform to restore the original spatial structure of the image.

[0014] Step 5: Network encoding and decoding processing.

[0015] Step 6: Train the OCT image segmentation model consisting of an image preprocessing module and a backbone structure.

[0016] Step 7: Use the trained model for inference, that is, predict the edema area on the new fundus OCT image.

[0017] The image to be segmented is preprocessed in step 1, and then segmented using a trained OCT image segmentation model. The final segmentation result is generated, and the image predicted to be the edema part is output.

[0018] Furthermore, the specific method for step 1 is as follows:

[0019] The raw OCT images are first standardized and resized using the OCT system to suit subsequent processing requirements. All images are uniformly processed to 512x512 pixels, with RGB three channels. Furthermore, retinal edema areas in the images are labeled to generate tag images; all tag images are stored in 512x512 pixel, single-channel format, with each raw OCT image corresponding to a labeled image.

[0020] Furthermore, the specific method for step 2 is as follows:

[0021] The standardized image data is converted to the frequency domain using the Fast Fourier Transform (FFT). The conversion formula is as follows:

[0022]

[0023] The symbols are represented as follows:

[0024] F(u,v): represents the image in the frequency domain, where u and v are the horizontal and vertical coordinates in the frequency domain, respectively;

[0025] f(x,y): is the pixel value at position (x,y) in the spatial domain;

[0026] M and N: are the dimensions of the image, namely the width and height of the image;

[0027] e: is the base of the natural logarithm, approximately equal to 2.71828;

[0028] i: is the imaginary unit, satisfying i 2 =-1;

[0029] ∑: Represents the summation symbol, used to accumulate the transformation results corresponding to all spatial domain pixels;

[0030] This part expresses the relationship between frequency and spatial location, in which and These represent the frequency components in the horizontal and vertical directions, respectively.

[0031] The application of Fast Fourier Transform allows the model to operate in the frequency domain in order to capture high-frequency and low-frequency features in the image.

[0032] Furthermore, the specific method for step 3 is as follows:

[0033] After the image is transformed into the frequency domain using Fast Fourier Transform, the input feature map X is processed through two different convolutional branches. In the splitting stage, the two convolutional branches use 3x3 depthwise convolution and 3x3 dilated convolution (with a dilation rate of 2 to replace the 5x5 kernel convolution), respectively. Then, the feature map is generated by batch normalization and ReLU activation. and

[0034] Next, in the fusion stage, the outputs of these two branches are fused element-wise to obtain the fused feature map U, and then channel-level statistics are generated through global average pooling. The c-th element is the feature vector Then through the dimension reduction fully connected layer F fc Generate compact feature vectors

[0035]

[0036] The meanings of each symbol are as follows:

[0037] F gp : Represents the global pooling operation;

[0038] U c : Represents the c-th channel in the fused feature map;

[0039] H, W: Represent the length and width of the image;

[0040] F fc : Represents a fully connected layer, a common type of layer in neural networks, typically used to learn non-linear combinations of input data;

[0041] This represents channel-level statistical information and serves as the input feature vector for the dimensionality-reducing fully connected layer. The c-th element is the feature vector This could be the output from any layer in the front part of the network, or some transformed form of the original input data;

[0042] W: This is the weight matrix in the fully connected layer. This matrix is ​​used to linearly transform the input features s to a new feature space, typically for dimensionality reduction or to adjust the feature representation to better suit the task.

[0043] Bn: This indicates the batch normalization operation applied after the fully connected layer. Batch normalization is a technique used to improve the training speed and stability of neural networks by normalizing the input of the layer to reduce the internal covariate bias.

[0044] Finally, in the selection phase, cross-channel soft attention is used to adaptively select information at different spatial scales, which utilizes compact feature vectors. This is implemented under guidance. Where A, B∈R C×d is a matrix defined in the real number field, used to guide the selection of different spatial scales of information. a and b represent... and The soft attention vector. C represents the feature map U. e and U b The number of channels, d represents the compact feature vector. The dimension. For each channel c, the calculation is as follows:

[0045]

[0046] Here, A c and B c These are the c-th rows of matrices A and B, respectively, while a c and b c These are the c-th elements of the soft attention vectors a and b, respectively. The final feature map is obtained by weighting and fusing the outputs of the two branches using these weights.

[0047]

[0048] To maintain consistency and standardization of feature weighting, the design ensures that the a value of each channel is... c and b c The sum is 1. This ensures that the information from the merging of the two branches is balanced, preventing one branch from contributing excessively to the final output.

[0049] This attention mechanism allows the model to adaptively adjust its receptive field size based on the input content, optimizing information flow and improving feature processing performance. By using the SKattention mechanism to dynamically assess and adjust the importance of frequency components, feature representation is optimized, enhancing the model's sensitivity to key regions such as edema.

[0050] Furthermore, step 5 is detailed below:

[0051] First, the image data after inverse fast Fourier transform (IFFT) undergoes basic preprocessing, including normalization and denoising, to ensure the quality and consistency of the input data. The preprocessed image is then input into the subsequent backbone structure. It first passes through a shrinking path, consisting of a four-layer encoder structure, which progressively reduces the spatial dimension while increasing the feature depth through convolutional and pooling layers. Then, the image passes through an expanding path, consisting of a four-layer decoder structure, which includes upsampling and convolutional operations to gradually restore it to the original image size. In the expanding path, features from the shrinking path are combined with those from the expanding path through skip connections to preserve key image information. At the network output, the probability that each pixel belongs to an edema region is obtained. This output is a segmentation map of the same size as the input image, where the value of each pixel represents the probability that the pixel belongs to an edema region.

[0052] Furthermore, step 6 is specifically implemented as follows: During training, an OCT image segmentation model consisting of an image preprocessing module and a backbone structure is trained using the training data labeled in step 1 to identify and predict edema regions in previously unseen fundus OCT images. The training process selects a two-dimensional cross-entropy loss function to evaluate the difference between the model output and the actual annotations, and continuously optimizes the model weights through backpropagation, gradient descent, dynamic learning rate adjustment, and regularization techniques. The image preprocessing module includes Fast Fourier Transform (FFT), SKAttention, and Inverse Fast Fourier Transform.

[0053] The beneficial effects of this invention are as follows:

[0054] This invention significantly improves the segmentation accuracy and efficiency of edema regions in OCT fundus images by introducing an image preprocessing module at the initial stage of the UNet network, i.e., the backbone structure. This technical solution optimizes the feature processing flow and enhances the model's ability to identify fundus lesion features, particularly demonstrating high efficiency and accuracy when processing complex or low-contrast edema regions. Furthermore, this method maintains high processing speed while possessing good adaptability and scalability, making it suitable for large-scale medical image processing applications. Attached Figure Description

[0055] Figure 1 This diagram illustrates the network locations applicable to the module.

[0056] Figure 2 This describes the workflow of the OCT image segmentation model in an embodiment of the present invention.

[0057] Figure 3 The following describes the specific working process of the image preprocessing module in an embodiment of the present invention. Detailed Implementation

[0058] The technical solution of the present invention will be further described below with reference to the accompanying drawings and embodiments.

[0059] Glossary

[0060] OCT (Optical Coherence Tomography) is an imaging technique that uses the principle of light reflection to obtain high-resolution three-dimensional images of the tissue by measuring the time delay and intensity of light reflected from it. OCT is widely used in ophthalmology, dermatology, and other fields, especially in fundus examinations, where it can provide detailed imaging of ocular structures such as the retina.

[0061] FFT (Fast Fourier Transform) is an efficient algorithm used to convert time or space signals into frequency or wavenumber signals. FFT is a fast implementation of the Fourier Transform and is commonly used in signal processing, image analysis, speech processing, and other fields to analyze the existence and characteristics of different frequency components.

[0062] UNet is a convolutional neural network architecture specifically designed for biomedical image segmentation. The UNet architecture employs a symmetrical "U" shape, containing a contracting path for capturing contextual information and a symmetrical expanding path for precise localization. This design allows UNet to effectively process images of different scales while preserving image details, making it widely used in medical image processing, particularly excelling in the precise segmentation of tissues and cells.

[0063] Specifically, this invention provides an improved OCT image segmentation model that integrates a special preprocessing module at its beginning to enhance the automatic segmentation of edema regions in fundus OCT images. This technique is particularly suitable for medical image processing tasks requiring high-precision and high-efficiency fundus edema detection.

[0064] The core of this invention lies in introducing a module combining FFT and SKAttention mechanisms. This module is designed to perform preliminary transformations and feature weighting on the input OCT image data to enhance the network's ability to recognize edema features. The specific implementation process of the embodiment is as follows:

[0065] Preprocessing stage: The public dataset OIMHS was used, which contains 10,908 pairs of images as training data and 2,700 pairs of images as validation test data. The raw OCT images were first standardized and resized to suit the requirements of subsequent processing. All images were uniformly processed to 512x512 pixels, RGB three channels.

[0066] Fast Fourier Transform: In this step, the standardized image data is transformed into the frequency domain. The transformation formula is as follows: The symbols are as follows: F(u,v): represents the image in the frequency domain, where u and v are the horizontal and vertical coordinates in the frequency domain, respectively; f(x,y): is the pixel value at position (x,y) in the spatial domain; M and N: are the dimensions of the image, i.e., the width and height; e: is the base of the natural logarithm, approximately equal to 2.71828; i: is the imaginary unit, satisfying i 2 =-1; ∑: represents the summation symbol, used to accumulate the transformation results corresponding to all spatial domain pixels; This part expresses the relationship between frequency and spatial location, in which and These represent the frequency components in the horizontal and vertical directions, respectively. The application of the Fast Fourier Transform allows the model to operate in the frequency domain, enabling it to capture high-frequency and low-frequency features in the image.

[0067] SKAttention mechanism: After the image is transformed into the frequency domain by Fast Fourier Transform, the input feature map X is processed through two different convolutional branches. In the Split stage, 3x3 grouping or depthwise convolution and 3x3 dilated convolution (with a dilation rate of 2 to replace 5x5 convolution) are used respectively to generate feature maps U and U.

[0068] Then, in the Fuse phase, the outputs of these two branches are fused through element-wise addition. Then, a single feature vector is obtained through global average pooling. Then through the dimension reduction fully connected layer F fc Generate compact feature vector Z = F fc (s)Bn(σ(Ws)), where the meanings of each symbol are as follows: F gp : Represents global pooling operation; U c : Represents the c-th channel in the feature map; H, W: Represent the length and width of the image; F fc: Represents the function of the fully connected layer, a common layer in neural networks, typically used to learn nonlinear combinations of input data; s: Represents the input feature vector of the fully connected layer, which can be the output of any layer in the preceding parts of the network, or some transformed form of the original input data; W: The weight matrix in the fully connected layer. This matrix is ​​used to linearly transform the input features s to a new feature space, usually for dimensionality reduction or to adjust the feature representation to better suit the task; Bn: Represents the batch normalization operation applied after the fully connected layer. Batch normalization is a technique used to improve the training speed and stability of neural networks by reducing internal covariate bias by normalizing the layer's input.

[0069] Finally, in the Select phase, compact feature vectors are used. Calculate the soft attention weights for each branch and The meanings of each symbol are as follows: where A, B∈R C×d is a matrix defined in the real number field, used to guide the selection of different spatial scales of information. a and b represent... and The soft attention vector. C represents the feature map U. e and U b The number of channels, d represents the compact feature vector. The dimension. A c and B c These are the c-th rows of matrices A and B, respectively, while a c and b c These are the c-th elements of the soft attention vectors a and b, respectively. The final feature map is obtained by weighting and fusing the outputs of the two branches using these weights. This attention mechanism allows the model to adaptively adjust its receptive field size based on the input content, optimizing information flow and improving feature processing performance. By using the SKAttention mechanism to dynamically assess and adjust the importance of frequency components, feature representation is optimized, enhancing the model's sensitivity to key regions such as edema.

[0070] Inverse Fast Fourier Transform: Frequency domain data processed by SKAttention is remapped to the spatial domain through inverse transform to restore the original spatial structure of the image.

[0071] Network encoding and decoding processing: The transformed image data first undergoes basic preprocessing, including normalization and denoising, to ensure the quality and consistency of the input data. The preprocessed image is then input into the backbone structure. First, it passes through a shrinking path, which consists of a four-layer encoder structure. Through convolutional and pooling layers, the spatial dimension is progressively reduced while the feature depth is increased. Then, the image passes through an expanding path, which consists of a four-layer decoder structure, including upsampling and convolutional operations, gradually restoring it to the original image size. In the expanding path, features from the shrinking path are combined with features from the expanding path through skip connections to preserve key image information. At the network output, the probability that each pixel belongs to an edema region is obtained. This output is typically a segmentation map of the same size as the input image, where the value of each pixel represents the probability that the pixel belongs to an edema region.

[0072] During training, this stage utilizes labeled training data obtained in the preprocessing phase. This data includes fundus OCT images and their corresponding edema region annotations. Using this data, you can train the model to identify and predict edema regions in previously unseen fundus OCT images. Training typically employs a two-dimensional cross-entropy loss function to evaluate the difference between the model output and the actual annotations, and continuously optimizes the model weights through backpropagation, gradient descent, dynamic learning rate adjustment, and regularization techniques. Specifically, in this dataset, background pixels have a value of 0, and edema pixels have a value of 192. At the start of training, each edema pixel is assigned a probability value of 1, and the background pixel is assigned a probability value of 0. Forward propagation of the model outputs the corresponding probability values, which are then used to obtain the loss for each pixel using the two-dimensional cross-entropy loss function. The losses calculated for all pixels in the entire image or batch are accumulated to obtain the final loss value. Backpropagation calculates the gradient of the loss with respect to each model parameter. This gradient represents the sensitivity of the loss function to each parameter under the current model parameters. The optimizer adjusts the values ​​of each parameter in the model based on the calculated gradient, optimizing the model weights. Finally, by continuously iterating and updating the parameters, the minimum value of the loss function is gradually approached, thereby optimizing the segmentation effect.

[0073] The trained model can be used for inference, i.e., predicting edema regions on new fundus OCT images. This process involves performing the same preprocessing on the new image, then segmenting it using the trained model, and finally generating a segmentation result that outputs the image predicted as the edema region.

[0074] The core technology of this invention lies in introducing a Fast Fourier Transform (FFT) to process the original input data. Through this transformation, the image data is converted to the frequency domain, enabling the model to analyze and process image features at different frequency levels. This processing step is particularly important for improving the network's sensitivity to image details and segmentation accuracy, especially when processing medical image data with complex textures and noise. Compared to existing technologies, this method utilizes information from the frequency domain, which helps to better capture and utilize structural features in the image, thereby improving segmentation performance. Furthermore, experimental results demonstrate that using this technique at the beginning of the network yields more significant results than using it in other parts of the network.

[0075] This invention applies the SKAttention mechanism after FFT processing. SKAttention is an adaptive attention control mechanism that automatically adjusts attention weights based on different frequency responses. In deep learning, typical attention mechanisms are directly added to the model to process the input data. The method described in this invention allows the model to better extract information from the image's frequency domain, increasing the model's processing dimensionality. Furthermore, compared to ordinary attention mechanisms, SKAttention not only focuses on detailed information in local regions but also adjusts and optimizes based on global information. This strategy significantly enhances the model's ability to recognize important features in images, especially in identifying edema regions. After applying this module, the segmentation model can more accurately locate and identify lesion areas.

[0076] The above description, in conjunction with specific / preferred embodiments, provides a further detailed explanation of the present invention. It should not be construed that the specific implementation of the present invention is limited to these descriptions. Those skilled in the art can make various substitutions or modifications to these described embodiments without departing from the inventive concept, and all such substitutions or modifications should be considered within the scope of protection of the present invention.

[0077] The parts of this invention not described in detail are well-known to those skilled in the art.

Claims

1. An OCT image segmentation method based on Fast Fourier Transform and SKAttention mechanism, characterized in that, The specific implementation process is as follows: Step 1: Data acquisition and preprocessing; Step 2: Preprocess the original input image using Fast Fourier Transform; Step 3: Apply adaptive SKAttention to recalibrate the features of the transformed data, and enhance the response to key features by dynamically adjusting the weights of each channel; Step 4: Inverse Fast Fourier Transform; The frequency domain data processed by SKAttention is the final feature map. The image's original spatial structure is restored by remapping it to the spatial domain using the inverse fast Fourier transform. Step 5: Network encoding and decoding processing; Step 6: Train the OCT image segmentation model consisting of an image preprocessing module and a backbone structure; Step 7: Use the trained model for inference, that is, predict the edema area on the new fundus OCT image; The image to be segmented is preprocessed in step 1, and then segmented using the trained OCT image segmentation model. Finally, the segmentation result is generated, and the image predicted to be the edema part is output. The specific method for step 3 is as follows: After the image is transformed into the frequency domain using Fast Fourier Transform, the input feature map X is processed through two different convolutional branches. In the segmentation stage, the two convolutional branches use 3x3 depthwise convolution and 3x3 dilated convolution, respectively. Then, the feature map is generated by batch normalization and ReLU function processing. ; Next, in the fusion stage, the outputs of these two branches are fused element-wise to obtain the fused feature map. Channel-level statistics are generated through global average pooling. , The c-th element is the feature vector Then through a dimensionality reduction fully connected layer Generate compact feature vectors : ， ， ； The meanings of each symbol are as follows: : Represents global average pooling operation; : Represents the first feature in the fused feature map One channel; H, W: Represent the length and width of the image; : Represents the function of a fully connected layer; Channel-level statistical information is used as the input feature vector for the dimensionality reduction fully connected layer. The c-th element is the feature vector ; : is the weight matrix in the fully connected layer; : Indicates the batch normalization operation applied after the fully connected layer; Finally, in the selection phase, cross-channel soft attention is used to adaptively select information at different spatial scales, which utilizes compact feature vectors. Provided with guidance for implementation; where A, B∈ is a matrix defined in the real number field, used to guide the selection of different spatial scales of information; a and b represent respectively The soft attention vector; This represents the feature map. and The number of channels, Represents compact feature vectors The dimension; for each channel c, it is calculated as follows: and ; here, and These are the c-th rows of matrices A and B, respectively. and These are the c-th elements of the soft attention vectors a and b, respectively; the final feature map is obtained by weighting and fusing the outputs of the two branches using these weights. ； To maintain consistency and standardization in feature weighting, the design ensures that each channel... and The sum is 1; Step 5 is detailed below: First, the image data after inverse fast Fourier transform (IFFT) undergoes basic preprocessing, including normalization and denoising, to ensure the quality and consistency of the input data. The preprocessed image is then input into the subsequent backbone structure. It first passes through a shrinking path, consisting of a four-layer encoder structure, which progressively reduces the spatial dimension and increases the feature depth through convolutional and pooling layers. Next, the image passes through an expanding path, consisting of a four-layer decoder structure, which includes upsampling and convolutional operations to gradually restore it to the original image size. In the expanding path, features from the shrinking path are combined with those from the expanding path through skip connections to preserve key image information. At the network output, the probability of each pixel belonging to an edema region is obtained. This output is a segmentation map of the same size as the input image, where the value of each pixel represents the probability that the pixel belongs to an edema region.

2. The OCT image segmentation method based on Fast Fourier Transform and SKAttention mechanism according to claim 1, characterized in that, The specific method for step 1 is as follows: The original OCT images are first standardized and resized using the OCT system to suit subsequent processing requirements. All images are uniformly processed to 512x512 pixels with RGB three channels. The edema in the fundus of the image is then labeled to generate label images. All label images are stored in 512x512 pixel, single-channel format, with each original OCT image corresponding to a labeled image.

3. The OCT image segmentation method based on Fast Fourier Transform and SKAttention mechanism according to claim 1, characterized in that, Step 2 is explained in the following steps: The standardized image data is converted to the frequency domain using the Fast Fourier Transform (FFT). The conversion formula is as follows: ； The symbols are represented as follows: : Represents the image in the frequency domain, where and These are the horizontal and vertical coordinates in the frequency domain, respectively. f(x,y): is a value at position y in the spatial domain. ; : This refers to the dimensions of the image, specifically its width and height; : is the base of the natural logarithm; : is the imaginary unit, satisfying =−1; ∑: Represents the summation symbol, used to accumulate the transformation results corresponding to all spatial domain pixels; This part expresses the relationship between frequency and spatial location, where and These represent the frequency components in the horizontal and vertical directions, respectively.

4. The OCT image segmentation method based on Fast Fourier Transform and SKAttention mechanism according to claim 1, characterized in that, Step 6 is specifically implemented as follows: During the training process, an OCT image segmentation model consisting of an image preprocessing module and a backbone structure is trained using the training data labeled in Step 1 to identify and predict edema regions in unseen fundus OCT images; the training process selects a two-dimensional cross-entropy loss function to evaluate the difference between the model output and the actual annotation, and continuously optimizes the model weights through backpropagation, gradient descent, dynamic learning rate adjustment, and regularization techniques; the image preprocessing module includes Fast Fourier Transform, SKAttention, and Inverse Fast Fourier Transform.

Images