Automatic positioning method for optic disc atrophy area and optic disc area based on visual converter

Abstract

This invention discloses an automatic localization method for optic disc atrophy region and optic disc region based on a visual converter. The method includes constructing a dataset of ultra-wide-angle fundus color images and labeling the peripapillary atrophy region and optic disc region on each image; randomly selecting training and test sets; constructing a segmentation network, inputting the labeled data from the training set into the segmentation network for training, and training the network weights for the peripapillary atrophy region and the optic disc region respectively; inputting images from the test set into the trained segmentation network to obtain segmentation results and extracting the boundary lines of the peripapillary atrophy region and the optic disc region; mapping the boundary lines of the two regions to the corresponding original images through image mapping to complete the localization of the peripapillary atrophy region and the optic disc region; automatically and accurately locating the peripapillary atrophy region and the optic disc region in images and marking their boundaries, which can be used for clinical auxiliary diagnosis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to an automatic localization method for optic disc atrophy region and optic disc region based on a visual converter. Background Technology

[0002] High myopic macular degeneration (AMD) is a common fundus disease, often accompanied by peripapillary atrophy during its development. In the early stages, mild atrophy of the nerve fiber layer or pigmentation may appear in this area. As the disease progresses, the area of ​​peripapillary atrophy expands, especially its extension into the macula, significantly affecting vision and leading to visual impairment. The clarity of the atrophic area's margins is a crucial indicator of whether the disease has stopped progressing; clear margins usually signify that atrophy has ceased. It is important to note that when the optic disc atrophies completely, the resulting visual loss is usually irreversible. Therefore, comprehensively considering both peripapillary atrophy and the extent of optic disc atrophy is extremely important for the in-depth diagnosis and treatment of high myopic macular degeneration.

[0003] Currently, a radial line segmentation method based on grayscale changes has been developed for the peripapillary atrophy region. This method focuses on automatically segmenting the peripapillary atrophy arc in fundus images. Its main features include: Grayscale preprocessing: The method first converts the color fundus image into a grayscale image for further processing. Radial line sampling: Pixels in the optic disc and its lateral region are sampled using radial lines to select the two points with the largest grayscale changes as candidate boundary points for peripapillary atrophy and the optic disc. Outlier removal: The method includes a step to remove outliers caused by blood vessels or noise to improve segmentation accuracy. Ellipse fitting: For the selected candidate boundary points, the method uses ellipse fitting technology to determine the final segmentation result for peripapillary atrophy. This method performs well under certain conditions, but still has some limitations. For example, over-reliance on grayscale images may lead to reduced accuracy in complex environments. Furthermore, when the boundaries of the optic disc and the peripapillary atrophy region are unclear, traditional grayscale change methods may struggle to accurately identify the boundaries, resulting in inaccurate segmentation results. Furthermore, this method mainly focuses on the segmentation of atrophy around the optic disc, providing limited support for the comprehensive localization and analysis of the entire optic disc region.

[0004] In addition, a method and system for optic disc and optic cup segmentation based on a parallel cooperative diffusion model has been developed for the optic disc region. This method fuses a preliminary segmentation map with the image to be segmented and utilizes a parallel diffusion model to improve segmentation accuracy. It also introduces cross-attention blocks to bridge the optic disc and optic cup subnets, effectively promoting information exchange. Through these steps, the segmentation results of the optic disc and optic cup are aligned, thus significantly improving the accuracy of the segmentation results. However, this method primarily focuses on the segmentation of the optic disc and optic cup and may not fully consider the complexity of the atrophic area around the optic disc, which is crucial in certain pathological conditions. Furthermore, methods relying on parallel diffusion models and cross-attention blocks may exhibit varying degrees of sensitivity when processing different types of fundus images with lesions.

[0005] In addition to the above, locating the peripapillary atrophy (PPA) and optic disc regions in ultra-wide-angle fundus color images presents three main challenges: First, the complex background. Unlike traditional wide-angle fundus images, ultra-wide-angle fundus images of patients with high myopia-induced macular degeneration contain rich information on retinal type and morphology, as well as the pervasive leopard-spot fundus (caused by retinal stretching in patients with high myopia-induced macular degeneration), making accurate extraction of PPA and optic disc features more difficult. Second, the blurred boundary. Because early PPA often presents as a crescent shape, growing close to the optic disc edge, the boundary between the optic disc and PPA becomes blurred. In this case, there is no clear, defined boundary between the two; this blurred boundary causes some areas of overlap between the optic disc and PPA, increasing the complexity and challenge of locating their boundary. Third, the morphological diversity of PPA and optic disc. The shape and size of PPA vary greatly among different patients and at different stages; the morphology of the optic disc also changes due to the erosion caused by PPA. This requires a positioning method that can effectively handle targets of different sizes and shapes. Therefore, there is an urgent need for an automatic positioning method for the optic disc atrophy region and optic disc region based on a vision converter to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an automatic localization method for the optic disc atrophy region and the optic disc region based on a vision converter, which can effectively solve the problems existing in the prior art.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: an automatic localization method for optic disc atrophy region and optic disc region based on a vision converter, comprising the following steps:

[0008] S1. Construct a dataset of ultra-wide-angle fundus color images, and mark the atrophic area around the optic disc and the optic disc region on each image. Randomly select the training set and the test set.

[0009] S2. Construct a segmentation network. Input the labeled data from the training set into the segmentation network for training. Train the network weights of the atrophic region around the optic disc and the network weights of the optic disc region separately.

[0010] S3. Input the ultra-wide-angle fundus color image from the test set into the trained segmentation network to obtain the segmentation results. After processing the results, extract the boundary lines of the atrophic area around the optic disc and the optic disc area.

[0011] S5. By mapping the boundary lines of the two regions to the corresponding original images, the localization of the atrophic area around the optic disc and the optic disc region is completed.

[0012] Preferably, in step S1, the image data in the training set and the test set are preprocessed, including:

[0013] The resolution of the ultra-wide-angle fundus color image is scaled to a preset value using center scaling.

[0014] Data augmentation was performed using random horizontal rotation, random vertical rotation, random translation and scaling rotation, random Gaussian noise, random Gaussian blur, and random height contrast.

[0015] Preferably, in step S2, the segmentation network includes a PVT encoder, a dilated convolution module, a selective boundary aggregation module, and a decoder, and the training is specifically as follows:

[0016] The pyramid visual converter in the PVT encoder part is pre-trained, and image features in ultra-wide-angle fundus color images are extracted through the trained PVT encoder part.

[0017] The top-level dilated convolution module consists of three convolutional layers. By setting different dilation coefficients and uniform kernel sizes for the three convolutional layers, receptive fields of different scales of feature maps are extracted.

[0018] The upsampled feature map is compared with the feature map output by the PVT encoder. Figure 1 The selective boundary aggregation module is used to capture spatial details.

[0019] The valid feature maps extracted from each part are restored into an image by the decoder, and then the segmentation result is output by bilinear upsampling with a magnification of 4.

[0020] Preferably, the PVT encoder part includes four stages, and each stage uses an image block embedding layer and a Transformer encoder module for the output feature map. The image block embedding layer is used to segment the ultra-wide-angle color fundus image into multiple small blocks and then convert them into serialized vectors for padding. The Transformer encoder module is used to process global and local information in the image.

[0021] Preferably, the dilation coefficients of the three convolutional layers of the dilated convolution module are set to 2, 3, and 5, respectively, the kernel size is 3×3, and ReLU is used as the activation function.

[0022] Preferably, the selective boundary aggregation module mainly consists of two parallel calibration attention units and one 3×3 convolutional layer. The calibration attention unit uses a 1×1 convolution with a stride of 1, employs batch normalization, and uses ReLU and Sigmoid as activation functions; and

[0023] Adaptive feature fusion is performed by calibrating attention units to compensate for the missing information in shallow and deep features. Then, the feature maps of the two calibrated attention units are stacked in channels to enhance the extraction of boundary information of the segmentation target.

[0024] Preferably, the decoder includes four hierarchical upsampling layers and three parallel upsampling layers. The hierarchical upsampling layers use bilinear upsampling layers, with each upsampling layer having an expansion factor of 2. The parallel upsampling layers use bilinear upsampling layers, with each upsampling layer having an expansion factor of 8, 4, and 2, respectively.

[0025] Preferably, in step S2, during the training process, the joint loss function of the segmentation network is calculated, the parameters are iteratively updated, and training is stopped after a preset number of iterations.

[0026] Preferably, in step S3, extracting the boundary line specifically involves:

[0027] Perform image channel conversion on the segmentation results;

[0028] The morphological processing of the optic disc atrophy region was performed by creating a white mask.

[0029] Then, use a Gaussian kernel to smooth the edges of the region to reduce noise;

[0030] Next, the contour of the optic disc atrophy region is extracted using an edge detection filter;

[0031] Finally, the extracted optic disc atrophy region contour image is enlarged to the size of the corresponding original ultra-wide-angle fundus color image.

[0032] Preferably, in step S4, the mapping specifically involves mapping the contours of the atrophic region around the optic disc and the optic disc region onto the original ultra-wide-angle fundus color image using a geometric mapping method, and then using a rectangular kernel to perform dilation operations to emphasize the contours.

[0033] Beneficial effects: The method of this invention can automatically and accurately locate the peripapillary atrophy area and the optic disc area in fundus images of high myopia macular degeneration, and mark the boundaries of the peripapillary atrophy area and the optic disc area, which can be used for clinical auxiliary diagnosis. In actual clinical applications, only ultra-wide-angle fundus color images of patients need to be captured. There is no need for complicated positioning procedures or a lot of time to mark samples. It can achieve automatic positioning of the peripapillary atrophy area and the optic disc area under limited conditions with high accuracy and speed.

[0034] Furthermore, based on the visual converter method, accurate segmentation results of the peripapillary atrophy area and the optic disc area can be achieved with a small amount of labeled data, thereby improving the localization accuracy of the peripapillary atrophy area and the optic disc area and improving the clinical diagnostic efficiency of macular degeneration in high myopia. Attached Figure Description

[0035] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0036] In the attached diagram:

[0037] Figure 1 This is a framework diagram of the method for locating the atrophic area around the optic disc and the optic disc region based on the eye rotation angle of the visual converter in this invention;

[0038] Figure 2 This is a segmentation result of the peripapillary atrophy region and the optic disc region in the ultra-wide-angle fundus color image of the present invention;

[0039] Figure 3 This is a diagram showing the location results of the atrophic area around the optic disc and the optic disc area according to the present invention. Detailed Implementation

[0040] The embodiments of the present invention will now be described with reference to the accompanying drawings. The terminology used in the embodiments section of this invention is only for explaining specific embodiments of the invention and is not intended to limit the invention. The embodiments of this application will now be described with reference to the accompanying drawings. It will be understood by those skilled in the art that, with the development of technology and the emergence of new scenarios, the technical solutions provided in the embodiments of this application are also applicable to similar technical problems. The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such terms can be interchanged where appropriate; this is merely a way of distinguishing objects with the same attributes in the embodiments of this application. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion, so that a process, method, system, product, or device that comprises a series of units is not necessarily limited to those units, but may include other units not explicitly listed or inherent to these processes, methods, products, or devices.

[0041] Example: An automatic localization method for optic disc atrophy region and optic disc region based on a visual converter. First, the input ultra-wide-angle fundus color image is preprocessed to highlight the atrophy region around the optic disc and the optic disc region. Then, labeled data from the training set is input into a segmentation network for training, training the network weights for the atrophy region around the optic disc and the optic disc region separately. After a certain number of iterations, the network accuracy reaches a fitted level, and the two network weights are saved. A test ultra-wide-angle fundus color image is input to obtain two segmentation results. A series of image processing operations are then performed on the two segmentation results to extract the boundary lines of the atrophy region around the optic disc and the optic disc region. Finally, image mapping is used to map the boundary lines of the two regions onto the original image, completing the accurate localization of the atrophy region around the optic disc and the optic disc region. (Refer to...) Figure 1 As shown, the precise localization method for the peridiscal atrophy region and the optic disc region based on the visual converter mainly includes the following steps:

[0042] S1. Construct a dataset of ultra-wide-angle fundus color images, and mark the atrophic area around the optic disc and the optic disc region on each image. Randomly select training and test sets. Among them, the image data in the training and test sets are preprocessed, including: scaling the resolution of the ultra-wide-angle fundus color images to a preset value by center scaling; and performing data augmentation by random horizontal rotation, random vertical rotation, random translation scaling rotation, random Gaussian noise, random Gaussian blur, and random height contrast.

[0043] In one specific embodiment:

[0044] The training and testing dataset consisted of 333 ultra-wide-angle fundus images (4328×4330 pixels) of patients with high myopia and macular degeneration, provided by Shenzhen Eye Hospital. 266 images were used for training, and 67 for testing. The images were annotated using the Labelme tool to label the peripapillary atrophy area and the optic disc region, a task completed collaboratively by three fundus disease experts. To adapt to the characteristics of the model and the segmentation target, the resolution of the ultra-wide-angle fundus color images input to the model was scaled to 512×512 using center scaling. To prevent overfitting due to limited data and to improve the model's generalization performance, data augmentation was performed using random horizontal rotation, random vertical rotation, random translation and scaling rotation, random Gaussian noise, random Gaussian blur, and random height contrast.

[0045] S2. Construct a segmentation network. Input the labeled data from the training set into the segmentation network for training. Train the network weights for the shrunk region around the optic disc and the optic disc region separately. Calculate the joint loss function of the segmentation network and iteratively update the parameters. After a certain number of training epochs, when the network's accuracy on the test set no longer improves, stop training. The segmentation network achieves its optimality through the joint loss function, which combines BCELoss (Binary Cross Entropy Loss) and DICELoss (Dice Loss), assigning weights of 0.5 and 1 to the two loss functions, respectively. The segmentation network includes a PVT encoder, a dilated convolutional module, a selective boundary aggregation module, and a decoder, and the training details are as follows:

[0046] To address complex background issues, the pyramid visual transformer in the PVT encoder is pre-trained to effectively balance global information processing and local feature extraction in ultra-wide-angle fundus images. Each layer of the pyramid visual transformer extracts rich image features while maintaining high efficiency. The pre-training process involves image classification training using multiple ultra-wide-angle fundus color images containing perioptic disc atrophy lesions and healthy images. The network weight parameters that have been fitted after training are then loaded into the PVT encoding of the segmentation network.

[0047] The PVT encoder consists of four stages, each comprising an image patch embedding layer and a Transformer encoder module. The number of channels in the output feature maps of stages 1, 2, 3, and 4 are 64, 128, 320, and 512, respectively, and the output feature map sizes are 128×128, 64×64, 32×32, and 16×16, respectively. Input data passes through each stage sequentially, progressively extracting and combining features. The image patch embedding layer segments the ultra-wide-angle color fundus image into multiple small patches, then converts them into serialized vectors for padding. The Transformer encoder module processes global and local information in the image.

[0048] To address the issue of diverse morphologies in the segmented optic disc periphery and optic disc region, a dilated convolutional module at the top is used to capture broader contextual information without losing crucial spatial information, accurately processing optic disc periphery and optic disc regions of different sizes and shapes. The dilated convolutional module consists of three convolutional layers with different dilation coefficients (2, 3, 5) and a uniform kernel size (3×3) to extract receptive fields of different scales from the feature maps. Subsequently, the extracted feature maps are channel-stacked, and feature fusion is achieved through a 1×1 convolutional layer, using ReLU as the activation function.

[0049] During the decoding stage, the upsampled feature map is compared with the feature map output by the encoder. Figure 1 The selective boundary aggregation module is fed into the system to accurately capture spatial details and enhance the ability to identify blurred boundaries. It can selectively aggregate features at different levels to enhance the extraction of boundary information of the segmentation target. It can also effectively suppress the feature redundancy problem caused by directly fusing low-level and high-level features. The selective boundary aggregation module mainly consists of two parallel calibration attention units and one 3×3 convolutional layer. The calibration attention unit uses a 1×1 convolution with a stride of 1, batch normalization, and ReLU and Sigmoid as activation functions. The calibration attention unit performs adaptive feature fusion to compensate for the missing information of shallow and deep features. Then, by stacking the feature maps of the two calibration attention units, the extraction of boundary information of the segmentation target is enhanced.

[0050] The decoder includes four hierarchical upsampling layers and three parallel upsampling layers. The hierarchical upsampling layers use bilinear upsampling layers with an expansion factor of 2 for each layer. The parallel upsampling layers use bilinear upsampling layers with expansion factors of 8, 4, and 2 for each layer. The decoder restores the effective feature maps extracted from each part into an image, and then outputs the segmentation result through a bilinear upsampling with a factor of 4.

[0051] In one specific embodiment, the training was conducted on a workstation equipped with an NVIDIA-RTX4070 GPU. The programming language was Python, the PyTorch deep learning framework was used, and AdamW was employed as the optimizer for model training. The batch size was 2, the weight decay was 0.0001, and the learning rate was 0.0001. Ultra-wide-angle fundus color images from the training data were input into the segmentation network for training. After two training iterations of 200 epochs, the model gradually fitted the model. During each iteration, the parameters of the model with the best training performance were saved. To verify the model's performance, testing was conducted on a test set. The ultra-wide-angle fundus color images from the test set were also preprocessed and resolution-scaled before being fed into two weighted networks to output the predicted segmentation results. The segmentation results for the optic disc atrophy region and the optic disc region are shown below. Figure 2 As shown;

[0052] S3. Input the ultra-wide-angle fundus color images from the test set into the trained segmentation network, and after processing the results, extract the boundary lines of the peridiscal atrophy area and the optic disc area.

[0053] The specific steps for extracting the boundary lines are as follows: image channel conversion is performed on the segmentation results; the morphological processing of the optic disc atrophy region is performed by creating a white mask; the edges of the region are then smoothed using a Gaussian kernel to reduce noise; next, the contour of the optic disc atrophy region is extracted using an edge detection filter; finally, the extracted contour image of the optic disc atrophy region is enlarged to the size of the corresponding original ultra-wide-angle fundus color image.

[0054] In one specific embodiment, the optic disc atrophy region segmentation result output by the segmentation network undergoes image channel conversion. Next, the optic disc atrophy region is extracted and morphologically processed by creating a white mask; specifically, an 80×80 rectangular kernel is used for erosion to reduce noise in the highlight areas of the optic disc atrophy region. Then, a 7×7 Gaussian kernel is used to smooth the region edges to reduce noise while better preserving image details. Next, an edge detection filter is used to extract the contour of the optic disc atrophy region. Finally, the extracted optic disc atrophy region contour image is enlarged from 512×512 pixels to the size of the corresponding original ultra-wide-angle fundus color image. The above operations are repeated with the optic disc region segmentation result output by the segmentation network to complete the extraction of the optic disc region contour.

[0055] S4. After processing the segmentation results, the boundary lines of the two regions can be mapped to the corresponding original images through image mapping to complete the localization of the atrophic region around the optic disc and the optic disc region. The mapping is specifically as follows: the contours of the atrophic region around the optic disc and the optic disc region are mapped to the original ultra-wide-angle fundus color image using geometric mapping, and then a rectangular kernel is used to perform dilation operations to emphasize the contours. The processing also includes image scaling, using bilinear interpolation to reduce the corresponding resolution, as well as grayscale conversion, image erosion, image dilation, and thresholding.

[0056] In one specific embodiment: the contours of the peridiscal atrophic region and the optic disc region are geometrically mapped onto the original ultra-wide-angle fundus color image. Then, a dilation operation is performed using a 90×90 rectangular kernel to emphasize the contours, thereby enhancing key image features while reducing noise. The final result is as follows: Figure 3 As shown.

[0057] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. For those skilled in the art, after learning the contents described in the present invention, several equivalent changes and substitutions can be made without departing from the principle of the present invention. These equivalent changes and substitutions should also be considered to fall within the protection scope of the present invention.

Claims

1. A method for automatic localization of optic disc atrophy region and optic disc region based on a vision converter, characterized in that, Includes the following steps: S1. Construct a dataset of ultra-wide-angle fundus color images, and mark the atrophic area around the optic disc and the optic disc region on each image. Randomly select the training set and the test set. S2. Construct a segmentation network. Input the labeled data from the training set into the segmentation network for training. Train the network weights of the atrophic region around the optic disc and the network weights of the optic disc region separately. The segmentation network consists of a PVT encoder, a dilated convolutional module, a selective boundary aggregation module, and a decoder, and the training process is as follows: The pyramid visual converter in the PVT encoder part is pre-trained, and image features in ultra-wide-angle fundus color images are extracted through the trained PVT encoder part. The top-level dilated convolution module consists of three convolutional layers. By setting different dilation coefficients and uniform kernel sizes for the three convolutional layers, receptive fields of different scales of feature maps are extracted. The upsampled feature map and the feature map output by the PVT encoder are fed into the selective boundary aggregation module to capture spatial details. The valid feature maps extracted from each part are restored into an image by the decoder, and then the segmentation result is output by bilinear upsampling with a magnification of 4. The decoder includes four hierarchical upsampling layers and three parallel upsampling layers. The hierarchical upsampling layers use bilinear upsampling layers, with each upsampling layer having an expansion factor of 2. The parallel upsampling layers use bilinear upsampling layers, with each upsampling layer having an expansion factor of 8, 4, and 2, respectively. S3. Input the ultra-wide-angle fundus color image from the test set into the trained segmentation network to obtain the segmentation results. After processing the results, extract the boundary lines of the atrophic area around the optic disc and the optic disc area. S4. By mapping the boundary lines of the two regions to the corresponding original images, the localization of the atrophic area around the optic disc and the optic disc region is completed.

2. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: In step S1, the image data in the training and test sets are preprocessed, including: The resolution of the ultra-wide-angle fundus color image is scaled to a preset value using center scaling. Data augmentation was performed using random horizontal rotation, random vertical rotation, random translation and scaling rotation, random Gaussian noise, random Gaussian blur, and random height contrast.

3. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: The PVT encoder part includes four stages, and each stage uses an image block embedding layer and a Transformer encoder module for the output feature map. The image block embedding layer is used to segment the ultra-wide-angle color fundus image into multiple small blocks, and then convert them into serialized vectors for padding. The Transformer encoder module is used to process global and local information in the image.

4. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: The three convolutional layers of the dilated convolution module have dilation coefficients of 2, 3, and 5, respectively, a kernel size of 3×3, and ReLU is used as the activation function.

5. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: The selective boundary aggregation module mainly consists of two parallel calibration attention units and one 3×3 convolutional layer. The calibration attention unit uses a 1×1 convolution with a stride of 1, employs batch normalization, and uses ReLU and Sigmoid as activation functions. as well as Adaptive feature fusion is performed by calibrating attention units to compensate for the missing information in shallow and deep features. Then, the feature maps of the two calibrated attention units are stacked in channels to enhance the extraction of boundary information of the segmentation target.

6. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: In step S2, during the training process, the joint loss function of the segmentation network is calculated, the parameters are iteratively updated, and training is stopped after a preset number of iterations.

7. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: In step S3, the extraction of the boundary line specifically involves: Perform image channel conversion on the segmentation results; The morphological processing of the optic disc atrophy region was performed by creating a white mask. Then, use a Gaussian kernel to smooth the edges of the region to reduce noise; Next, the contour of the optic disc atrophy region is extracted using an edge detection filter; Finally, the extracted optic disc atrophy region contour image is enlarged to the size of the corresponding original ultra-wide-angle fundus color image.

8. The method for automatic localization of optic disc atrophy region and optic disc region based on a visual converter according to claim 1, characterized in that: In step S4, the mapping specifically involves mapping the contours of the atrophic region around the optic disc and the optic disc region onto the original ultra-wide-angle fundus color image using a geometric mapping method, and then using a rectangular kernel to perform dilation operations to emphasize the contours.

Images