A deep learning-based leopard-spot segmentation method for wide-angle fundus images

Abstract

This invention relates to the fields of medical image processing and artificial intelligence, specifically a deep learning-based method for leopard spot segmentation in wide-angle fundus images. The method includes: constructing a deep learning model for leopard spot segmentation suitable for wide-angle fundus images; performing multi-channel texture enhancement processing on the acquired wide-angle fundus images; inputting the enhanced image into the segmentation model to obtain local leopard spot prediction results; weighted reconstruction of the prediction results to generate a leopard spot probability mask for the entire fundus image, and completing binarization segmentation. The deep learning model uses a multi-source manually labeled fusion mask as a supervision signal and combines a pseudo-supervised optimization mechanism to expand the training data scale under limited manually labeled samples, thereby improving the model's segmentation accuracy. This invention can accurately extract leopard spot regions under complex texture backgrounds, effectively overcoming the problems of low efficiency and strong subjectivity in existing manual interpretation methods, as well as the insufficient segmentation performance of traditional models in wide-angle images.

Description

Technical Field

[0001] This invention relates to the fields of medical image processing and artificial intelligence, specifically to a method for segmenting leopard spots in wide-angle fundus images based on deep learning. Background Technology

[0002] Myopia in children and adolescents is showing a clear trend of occurring at younger ages, with the overall myopia rate among children and adolescents in my country exceeding 50% in 2020. The development of myopia is not only characterized by a shift in refractive state towards negative values, but also by a series of morphological changes in the eyeball, including axial elongation, choroidal thinning, and alterations in retinal structure. Fundus tessellation is one of the most common and easily visualized phenotypes among these changes in fundus imaging. It manifests as reddish, mottled, linear patterns formed by the large choroidal vessels protruding through the retina, commonly known as "leopard spots." The appearance of leopard spots is closely related to axial elongation, decreased choroidal thickness, and reduced retinal pigmentation, reflecting early structural changes in the posterior pole and possessing potential biomarker value in the early identification of myopia.

[0003] Currently, clinical identification of leopard spots mainly relies on manual image interpretation, which is time-consuming, subjective, and has poor repeatability, making it unsuitable for large-scale population studies or continuous follow-up. In recent years, with the development of artificial intelligence, researchers have begun to explore the use of deep learning to achieve automatic detection and segmentation of leopard spots. However, the training data for current models mainly comes from conventional fundus images with a standard field of view of 30°–50°. These images have a limited imaging range, low resolution, and difficulty in rendering fine structures, resulting in significant limitations in the model's ability to capture texture details and identify region boundaries.

[0004] With the upgrading of fundus imaging equipment, wide-angle fundus photography systems (covering a field of view of 100°–200°) have been widely used in clinical and population studies. Wide-angle images can provide higher resolution and a wider range of fundus structural information, making leopard spots appear as more fragmented, dense, and structurally complex texture features. However, the rich details also significantly increase the difficulty of segmentation: wide-angle images contain more small blood vessels, background brightness variations, pigment heterogeneity, and high-density textures. Leopard spots are more morphologically similar to surrounding tissues, and traditional models trained based on conventional fields of view generally suffer from insufficient recognition, blurred boundaries, and missed detections in such scenarios.

[0005] Therefore, there is an urgent need for a deep learning segmentation method that can fully adapt to the characteristics of wide-angle fundus images and achieve stable and high-precision extraction of leopard spot regions against complex texture backgrounds, so as to improve the accuracy and consistency of automated leopard spot analysis and provide reliable technical support for large-scale population studies and related clinical applications. Summary of the Invention

[0006] The purpose of this invention is to provide a deep learning-based method for segmenting leopard spots in wide-angle fundus images, in order to solve the problems mentioned in the background art.

[0007] To solve the above-mentioned technical problems, the present invention provides the following technical solution:

[0008] A deep learning-based method for leopard spot segmentation in wide-angle fundus images, comprising:

[0009] S100. The acquired wide-angle fundus images are labeled to construct data for supervised training of the subsequent segmentation model.

[0010] S200: The multi-channel texture enhancement module is used to preprocess and enhance the texture of the wide-angle fundus image to generate a standardized image, a CLAHE enhanced image, and a high-frequency texture enhanced image.

[0011] The multi-channel texture enhancement module inputs the unified original image, CLAHE image and high-frequency texture enhancement map together to enhance local contrast and high-frequency texture information, making the boundaries and fine structures of leopard spots more prominent, and providing multi-view texture features for the segmentation network.

[0012] S300. Using the UNet-based leopard spot segmentation module, the enhanced image is input into the segmentation model to obtain local leopard spot prediction results; the prediction results are weighted and reconstructed to generate a leopard spot probability mask for the entire fundus image, and binarization segmentation is completed.

[0013] S400: The pseudo-supervised optimization module generates leopard spot pseudo-labels on unlabeled images based on the initially trained model, and further improves the model performance through pseudo-label optimization and retraining.

[0014] The pseudo-supervised optimization module uses the pre-trained model to generate leopard spot pseudo-labels on unlabeled images. After manually correcting local noise, small false positive areas, or irregular edge structures, these pseudo-labels are used as new training samples for retraining. This effectively expands the scale of training data under limited manual labeling conditions and significantly improves the segmentation performance of the model.

[0015] The S500, based on the collaborative work of three core modules—a multi-channel texture enhancement module, a leopard spot segmentation module, and a pseudo-supervised optimization module—constructs a dedicated deep learning model for leopard spot segmentation in wide-angle fundus images. This model can achieve stable and high-precision leopard spot region segmentation in wide-angle fundus images with complex textures, uneven illumination, and diverse structures.

[0016] Preferably, step S100 involves constructing labeled data from the acquired wide-angle fundus images, including:

[0017] Wide-angle fundus images were acquired, and three professionally trained human annotators independently annotated the leopard spot regions of the same batch of wide-angle fundus images pixel by pixel. The annotation tool used was Labelme to accurately depict the fine boundaries and irregular shapes of the leopard spots, ensuring that the annotation results have sufficient spatial resolution and structural details.

[0018] After obtaining three independent labeled masks, a majority voting strategy is used to fuse them at the pixel level. This strategy compares the annotation results of the three annotators at each pixel location: if at least two annotators identify the pixel as a leopard spot, then the pixel is labeled as a leopard spot in the fused mask; otherwise, it is labeled as background. The majority voting mechanism effectively reduces noise caused by individual annotation differences, eliminates occasional misclassifications, and preserves the region structure with high consistency among the three annotators. This fused mask serves as the main supervision signal in subsequent model training, providing a data foundation for the deep learning model to learn the texture features and region distribution of leopard spots.

[0019] Preferably, the multi-channel texture enhancement module in S200 includes:

[0020] S201. Perform preprocessing on the original image, including brightness correction, contrast unification, and size standardization, to reduce the impact of factors such as lighting differences, imaging noise, and inconsistent field of view.

[0021] S202, Construct the CLAHE channel and the high-frequency texture enhancement channel respectively;

[0022] The CLAHE channel enhances the local brightness difference of the fundus image by performing contrast-limited adaptive histogram equalization on the standardized original image, making the potential leopard spot texture in the low contrast area more visible, thereby improving the model's recognition ability in weak texture scenes.

[0023] The high-frequency texture enhancement channel is implemented based on frequency domain filtering: the three channels of the RGB image are converted to the frequency domain respectively, and a circular high-pass filter mask is constructed at the center of the spectrum to suppress low-frequency brightness and large-scale background components, retaining only high-frequency detail textures; then, inverse Fourier transform and amplitude normalization are performed to obtain an enhanced image containing rich high-frequency structural information. This high-frequency image can highlight the fine strip-like textures in the fundus, which helps to enhance the expression of the edges and fragmented structures of leopard spots.

[0024] S203. The unified original image, CLAHE image, and high-frequency texture enhancement map are jointly input into the segmentation module in a multi-channel format, providing the model with three complementary features: original color structure, local contrast enhancement, and high-frequency texture enhancement. While preserving the original color and structural information, it emphasizes local contrast and texture details, providing a more comprehensive and stable multi-view feature representation for subsequent leopard spot segmentation.

[0025] Preferably, the leopard spot segmentation module in S300 includes:

[0026] S301. After completing multi-channel image enhancement and obtaining the supervised mask, the wide-angle fundus image is segmented into local image patches of size 512×512 pixels with a stride of 64 pixels to obtain high-resolution samples for model training. Subsequently, all local image patches are divided into training set, validation set, and test set in a ratio of 50%:20%:30%. The training set is used for parameter learning of the segmentation model; the validation set is used for hyperparameter tuning and model selection; and the test set is used for independent performance evaluation of the final model.

[0027] S302. Using sliced ​​multi-channel image patches (normalized original image, CLAHE image, and high-frequency texture enhancement image) as input, pixel-by-pixel prediction of the leopard spot region is achieved through a U-Net-based encoder-decoder structure model:

[0028] The U-Net encoder extracts multi-scale semantic features through progressive convolution and downsampling operations, enabling the model to capture the texture patterns of leopard spots at different scales. The decoder recovers spatial details through progressive upsampling and skip connections, allowing the model to accurately reconstruct the local boundaries of leopard spots in the output layer. Since the input includes contrast information from the original image, CLAHE local enhancement information, and high-frequency texture information after denoising, the model can integrate multi-dimensional features for discrimination, thereby improving its ability to recognize weak texture regions and small structures.

[0029] Preferably, the pixel-by-pixel prediction of the leopard spot region in S302 using a U-Net-based encoder-decoder structure model includes:

[0030] The various predicted local image patches output by the model are stitched together according to their spatial location and fused into a complete leopard spot prediction mask:

[0031] First, for the same source image, record the coordinates of the top-left corner of each prediction patch. and its corresponding prediction probability map ,in, For patch index, The coordinates are the internal coordinates of the patch; based on the maximum coordinate value of all patches and the preset patch size. The original full image size is inferred to be and initialize the reconstructed image at that size. With weight accumulation matrix Let be a zero matrix, where, These are the pixel coordinates of the entire image;

[0032] Subsequently, a two-dimensional Gaussian weight kernel is constructed using a one-dimensional Gaussian function, specifically within the interval... Take evenly from the top sampling points ,calculate ; and on Perform maximum value normalization, and then use the outer product to obtain the two-dimensional Gaussian weight matrix: ;

[0033] This weighting kernel gives a larger weight to the center region of the patch and a smaller weight to the edge region, which is beneficial for achieving a smooth transition in the overlapping areas of the patch.

[0034] During the weighted write phase, for each patch, its internal coordinates are traversed. Predicting probability Multiply by Gaussian weights Accumulate to the corresponding position in the whole image Simultaneously, the corresponding weights are accumulated:

[0035] ;

[0036] ;

[0037] After all patches are written, pixel-by-pixel weight normalization is performed on the entire image; to avoid division by zero, when a certain pixel... Setting it to 1, we obtain the final leopard spot probability map:

[0038] ;

[0039] Normalized probability graph Then through the preset threshold (e.g., 0.5) is binarized when If a pixel is identified as a leopard spot, it is identified as background. Finally, a binary segmentation mask of the leopard spot in the entire wide-angle fundus image is obtained. This reconstruction process based on Gaussian weighting can achieve continuous and smooth fusion in the overlapping area of ​​multiple patches, effectively reducing the boundary breaks and block artifacts caused by the sliding window.

[0040] Preferably, the pseudo-supervised optimization module in S400 includes:

[0041] After the initial training is completed, the trained UNet model is applied to unlabeled wide-angle fundus images to automatically generate corresponding leopard spot pseudo-labels. Since the pseudo-labels are automatically predicted by the model, they may contain problems such as missing labels, mislabeling, or inaccurate boundaries. Manual quality control is performed on these pseudo-labels.

[0042] Specifically, operators manually inspect the pseudo-labels generated by the model and manually correct obviously unreasonable areas, including filling in missing areas, deleting obvious false positive areas, and adjusting discontinuous or large-deviation boundaries, so that the pseudo-labels can more accurately reflect the real leopard spot structure in the fundus image.

[0043] The pseudo-labels that have undergone manual quality control are regarded as "expanded labeled samples" and are used together with the original manually labeled training data for model retraining. By introducing manually corrected pseudo-labels, this invention can effectively expand the scale of training data under limited manual labeling conditions, enabling the model to learn richer texture representations and imaging conditions, and improving the segmentation accuracy and stability of the model on complex wide-angle images.

[0044] Compared with the prior art, the beneficial effects achieved by the present invention are:

[0045] First, this invention constructs more stable, reliable, and highly consistent supervisory labels by independently labeling three annotators and using a voting method for pixel-level fusion. This effectively reduces random errors caused by subjective differences among individual annotators, providing high-quality and reliable ground truth for subsequent model training, thereby improving the model's ability to learn the boundaries and structural details of leopard spots.

[0046] Secondly, by introducing a multi-channel input method that includes the original image, CLAHE image, and high-frequency texture enhancement image, this invention enables the model to simultaneously acquire key information under different brightness conditions, different contrast ranges, and multi-scale texture spaces. This significantly improves the model's sensitivity to segmentation of weak contrast regions, fine textures, and high-frequency choroidal structures, overcoming the limitation of traditional single-channel input in accurately identifying subtle leopard spots.

[0047] Furthermore, this invention introduces a pseudo-supervised training mechanism. This mechanism uses the model obtained from the initial training to generate pseudo-labels on unlabeled images, and then corrects obvious errors through manual quality control, allowing these pseudo-labels to be used as new training samples in model retraining. This mechanism effectively expands the scale of training data without significantly increasing the cost of manual annotation, enabling the model to learn more diverse texture patterns and imaging conditions, thereby significantly improving the model's generalization ability and stability.

[0048] In summary, the wide-angle fundus leopard spot segmentation framework constructed in this invention, which combines multi-channel enhancement, supervised training, and pseudo-supervised optimization, can achieve high-precision, stable, and globally consistent segmentation results under limited annotation conditions. This provides a reliable technical foundation and application value for subsequent fundus structure analysis, pediatric myopia research, and quantitative clinical image assessment. Attached Figure Description

[0049] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0050] Figure 1 This is a schematic diagram of the deep learning model structure for wide-angle fundus image leopard spot segmentation according to the present invention;

[0051] Figure 2 This is a schematic diagram of the leopard-spot three-source annotation fusion mask of the present invention;

[0052] Figure 3 This is a schematic diagram of the multi-channel texture enhancement module of the present invention;

[0053] Figure 4 This is a schematic diagram of the dataset used to develop the leopard spot segmentation algorithm of this invention;

[0054] Figure 5 A schematic diagram of the leopard spot segmentation module structure of the present invention;

[0055] Figure 6 This is a schematic diagram of the leopard spot prediction mask reconstruction of the present invention;

[0056] Figure 7 This is a segmentation result of the leopard spot segmentation pre-trained model of this invention on the test set;

[0057] Figure 8 This is a schematic diagram of the pseudo-supervised optimization module of the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] Please see Figures 1-8 The present invention provides the following technical solution:

[0060] Example 1: This invention provides a deep learning-based method for leopard spot segmentation in wide-angle fundus images, constructing a segmentation model framework suitable for wide-angle images. This model, through the collaborative work of three core modules—multi-channel texture enhancement, deep learning segmentation, and pseudo-supervised optimization—achieves stable and high-precision leopard spot region segmentation in wide-angle fundus images with complex textures, uneven illumination, and diverse structures.

[0061] like Figure 1As shown, the overall structure of the model in this invention includes a multi-channel texture enhancement module, a UNet-based leopard spot segmentation module, and a pseudo-supervised optimization module.

[0062] The multi-channel texture enhancement module is responsible for preprocessing and texture enhancement of wide-angle fundus images. By generating standardized images, CLAHE enhancement maps, and high-frequency texture enhancement maps, it achieves joint expression of leopard spot features under different lighting conditions, different contrasts, and different texture scales.

[0063] The UNet-based leopard spot segmentation module uses a mask fused by three annotators as a supervision label to extract cross-scale semantic features from multi-channel input, achieving preliminary leopard spot region prediction.

[0064] The pseudo-supervised optimization module uses the initial segmentation model to generate pseudo-labels on unlabeled data, and further improves model performance through pseudo-label optimization and retraining.

[0065] This invention achieves a complete segmentation process from multi-channel feature enhancement and master-supervised training to pseudo-supervised optimization through multi-module collaboration. It can effectively improve the accuracy and stability of leopard spot region extraction in wide-angle fundus images and is suitable for large-scale cohort studies and clinical analysis tasks.

[0066] like Figure 2 This invention employs a three-source annotation fusion method to generate a supervisory mask. First, for a selected wide-angle fundus image, three trained human annotators independently annotate the leopard spot region pixel-by-pixel, resulting in three independent binary masks. Because the leopard spots in the wide-angle image exhibit fine, striped textures, and there are blurred areas between the structural boundaries and background color variations, subjective interpretation differences may exist between different annotators within a local range. Therefore, directly using any annotator's result as the model training label easily introduces unstable noise, affecting the model's convergence and generalization ability. To address these issues, this invention uses a majority voting fusion strategy to construct highly consistent supervisory labels. Specifically, the three annotated masks are compared pixel-by-pixel. If a pixel is identified as a leopard spot by at least two of the three annotators, then that pixel is marked as a leopard spot in the final fused mask; pixels identified as leopard spots by only one or no annotator are marked as background areas. This pixel-level majority voting mechanism effectively eliminates noise caused by individual annotators' local misjudgments, fatigue, or subjective biases, while preserving a stable structure that is mutually recognized by all three parties. This results in the final fused mask being significantly superior to any single annotation result in terms of boundary coherence, texture integrity, and regional consistency.

[0067] like Figure 3As shown, in the multi-channel texture enhancement module, the input wide-angle fundus image first undergoes preprocessing operations such as brightness correction, contrast unification, and size standardization to reduce the impact of different shooting devices, lighting variations, and original imaging differences on subsequent feature learning. After obtaining the standardized original image, this invention further constructs two enhancement channels to strengthen the local contrast and texture details of the leopard spot structure.

[0068] The CLAHE channel enhances the local brightness differences in the fundus image by applying contrast-limited adaptive histogram equalization to the normalized original image. This makes the potential leopard spot texture in low-contrast areas more prominent, thereby improving the model's recognition ability in weakly textured scenes. On the other hand, the high-frequency texture enhancement channel is based on frequency domain filtering: the three channels of the RGB image are transformed to the frequency domain, and a circular high-pass filter mask is constructed at the center of the spectrum to suppress low-frequency brightness and large-scale background components, retaining only high-frequency detail textures. Subsequently, an inverse Fourier transform and amplitude normalization are performed to obtain an enhanced image containing rich high-frequency structural information. This high-frequency image can highlight the fine, linear textures in the fundus, helping to strengthen the expression of the edges and fine structures of the leopard spots.

[0069] Ultimately, the standardized original image, the CLAHE image, and the high-frequency texture enhancement image are input into the segmentation network in a multi-channel format. This provides the model with three complementary features: original color structure, local contrast enhancement, and high-frequency texture enhancement. This enables the network to have a richer and more stable source of multi-scale information when learning leopard spot texture, providing high-quality input features for subsequent refined segmentation.

[0070] like Figure 4 After completing multi-channel image enhancement and obtaining the voting mask, this invention divides the wide-angle fundus image into local image patches of size 512×512 pixels with a stride of 64 pixels to obtain high-resolution samples for model training. Subsequently, all patches are divided into training, validation, and test sets in a 50%:20%:30% ratio: the training set is used for parameter learning of the segmentation model; the validation set is used for hyperparameter tuning and model selection; and the test set is used for independent performance evaluation of the final model. This figure illustrates the data construction process of the leopard spot segmentation model of this invention, including slicing operations and structured partitioning of the dataset, providing basic data support for subsequent model training and evaluation.

[0071] like Figure 5 Using sliced ​​multi-channel image patches as input, pixel-by-pixel prediction of leopard spot regions is achieved through a U-Net-based encoder-decoder structure.

[0072] First, the input wide-angle fundus image is normalized for brightness, contrast, and size, and then combined with its corresponding CLAHE image and denoised texture enhancement image to form a multi-channel image set. This multi-channel image is then sliced ​​using a sliding window method with dimensions of 512×512 pixels and a stride of 64 pixels to generate local image patches covering the entire fundus image. This slicing method can fully utilize local structural features while maintaining high spatial resolution, enabling the model to have higher recognition capabilities in areas with fine textures.

[0073] Subsequently, the patch is input into a segmentation network based on the U-Net architecture. The U-Net encoder extracts multi-scale semantic features through successive convolutions and downsampling operations, enabling the model to capture the texture patterns of leopard spots at different scales. The decoder recovers spatial details through successive upsampling and skip connections, allowing the model to accurately reconstruct the local boundaries of the leopard spots at the output layer. Because the input includes contrast information from the original image, CLAHE local enhancement information, and denoised high-frequency texture information, the model can synthesize multi-dimensional features for discrimination, thereby improving its ability to recognize weakly textured regions and fine structures.

[0074] During supervised training, this invention uses a highly consistent mask obtained by fusing pixel-level annotations from three annotators through majority voting as the training label. This strategy significantly reduces noise caused by differences in manual annotations, enabling the model to learn leopard spot structural features with more coherent boundaries and more stable morphology. Finally, U-Net outputs a corresponding leopard spot probability map for each input patch, providing local prediction results for the subsequent whole-image reconstruction module.

[0075] like Figure 6 The reconstruction module of this invention is used to stitch together and fuse the various prediction patches output by the model according to their spatial positions into a complete leopard spot prediction mask. The specific calculation process is as follows:

[0076] First, for the same source image, record the coordinates of the top-left corner of each prediction patch. and its corresponding prediction probability map ,in, For patch index, These are the coordinates within the patch. Based on the maximum coordinate value of all patches and the preset patch size. The original full image size is inferred to be and initialize the reconstructed image at that size. With weight accumulation matrix Let be a zero matrix, where, These are the pixel coordinates of the entire image;

[0077] Subsequently, this invention constructs a two-dimensional Gaussian weight kernel using a one-dimensional Gaussian function. Specifically, within the interval... Take evenly from the top sampling points ,calculate: ;

[0078] And on Perform maximum value normalization, and then use the outer product to obtain the two-dimensional Gaussian weight matrix:

[0079] ;

[0080] This weighting kernel assigns a larger weight to the center region of the patch and a smaller weight to the edge region, which is beneficial for achieving a smooth transition in overlapping patch areas.

[0081] During the weighted write phase, for each patch, its internal coordinates are traversed. Predicting probability Multiply by Gaussian weights Accumulate to the corresponding position in the whole image Simultaneously, the corresponding weights are accumulated:

[0082] ;

[0083] ;

[0084] After all patches are written, pixel-by-pixel weight normalization is performed on the entire image. To avoid division by zero, when a pixel... Setting it to 1, we obtain the final leopard spot probability map:

[0085] ;

[0086] Normalized probability graph Then through the preset threshold (e.g., 0.5) is binarized when Pixels are identified as leopard spots if they are not, otherwise they are identified as background, resulting in a binary segmentation mask for the leopard spots in the entire wide-angle fundus image. This Gaussian-weighted reconstruction process can achieve continuous and smooth fusion in overlapping areas of multiple patches, effectively reducing boundary breaks and block artifacts caused by the sliding window.

[0087] Figure 7This figure demonstrates the segmentation performance of the pre-trained leopard spot segmentation model on the test set. The three columns in the figure, from left to right, are: the original wide-angle fundus image, the mask generated by majority voting from three annotators, and the predicted mask output by the model of this invention. The voting mask shown in the second figure, as a highly consistent annotation result, serves as the core supervision information for subsequent model training; the predicted mask of the pre-trained model shown in the third figure demonstrates the initial segmentation performance achievable by the model when relying only on a limited number of labeled samples. The performance metrics obtained by the pre-trained model on the test set are listed below the figure: Dice coefficient of 0.7002, accuracy of 0.9102, sensitivity of 0.6415, specificity of 0.9669, and AUC of 0.9401. The results show that the model can provide coarse-grained segmentation results with high availability and low structural bias, providing stable initial predictions for the pseudo-supervised optimization stage, ensuring that the subsequent model can further improve segmentation accuracy and robustness on larger-scale data.

[0088] like Figure 8 As shown, the pseudo-supervised optimization module of this invention is used to further expand the scale of training data under the condition of a limited number of manually labeled samples, thereby improving the performance and generalization ability of the leopard spot segmentation model. The specific process of this module is as follows: First, the UNet model trained in the first stage is applied to unlabeled wide-angle fundus images to automatically generate corresponding leopard spot pseudo-labels. Since pseudo-labels may have slight offsets, omissions, or mislabeling in local areas, this invention performs manual quality control correction on the pseudo-labels after acquisition to ensure that the overall regional morphology, boundary coherence, and texture distribution of the pseudo-labels are consistent with the real leopard spot features. The quality of the pseudo-labels after manual quality control correction is significantly better than the direct output of the model and can be used as reliable new training samples, together with the original manually labeled samples, for the second stage of UNet retraining. By adding the participation of high-quality pseudo-labels, the model can learn richer leopard spot structural patterns and a wider range of image distribution features, effectively improving robustness under different imaging conditions, different texture densities, and different image qualities.

[0089] In summary, this invention proposes an automatic leopard spot segmentation method for wide-angle fundus images. Through the synergistic integration of key technologies such as multi-channel texture enhancement, deep learning segmentation, and pseudo-supervised optimization, it achieves high-precision recognition of leopard spot regions under conditions of large field of view and complex texture.

[0090] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A deep learning-based method for segmenting leopard spots in wide-angle fundus images, characterized in that: The method includes: S100. The acquired wide-angle fundus images are labeled to construct data for supervised training of the subsequent segmentation model. S200: The multi-channel texture enhancement module is used to preprocess and enhance the texture of the wide-angle fundus image to generate a standardized image, a CLAHE enhanced image, and a high-frequency texture enhanced image. S300. Using the UNet-based leopard spot segmentation module, the enhanced image is input into the segmentation model to obtain local leopard spot prediction results; the prediction results are weighted and reconstructed to generate a leopard spot probability mask for the entire fundus image, and binarization segmentation is completed. S400: The pseudo-supervised optimization module generates leopard spot pseudo-labels on unlabeled images based on the initially trained model, and further improves the model performance through pseudo-label optimization and retraining. The S500, based on the collaborative work of three core modules—a multi-channel texture enhancement module, a leopard spot segmentation module, and a pseudo-supervised optimization module—constructs a dedicated deep learning model for leopard spot segmentation in wide-angle fundus images.

2. The leopard spot segmentation method for wide-angle fundus images based on deep learning as described in claim 1, characterized in that, The step S100 involves constructing labeled data from the acquired wide-angle fundus images, including: Wide-angle fundus images were acquired, and three professionally trained human annotators independently annotated the leopard-spot regions of the same batch of wide-angle fundus images pixel by pixel; the annotation tool used was Labelme. After obtaining three independent annotation masks, a majority voting strategy is used to perform pixel-level fusion of the three. This strategy compares the annotation results of the three annotators at each pixel position: if at least two annotators determine that the pixel is a leopard spot, then the pixel is marked as a leopard spot in the fused mask; otherwise, it is marked as background.

3. The leopard spot segmentation method for wide-angle fundus images based on deep learning as described in claim 1, characterized in that, The multi-channel texture enhancement module in S200 includes: S201. Perform preprocessing on the original image, including brightness correction, contrast unification, and size standardization. S202, Construct the CLAHE channel and the high-frequency texture enhancement channel respectively; The CLAHE channel enhances the local brightness difference of the fundus image by performing contrast-limited adaptive histogram equalization on the standardized original image, making the potential leopard spot texture in the low contrast area more visible, thereby improving the model's recognition ability in weak texture scenes. The high-frequency texture enhancement channel is implemented based on frequency domain filtering: the three channels of the RGB image are converted to the frequency domain respectively, and a circular high-pass filter mask is constructed at the center of the spectrum to suppress low-frequency brightness and large-scale background components, retaining only high-frequency detail textures; then, inverse Fourier transform and amplitude normalization are performed to obtain an enhanced image containing rich high-frequency structural information. S203. The unified original image, CLAHE image and high-frequency texture enhancement image are jointly input into the segmentation module in a multi-channel form to provide the model with three complementary features: original color structure, local contrast enhancement and high-frequency texture enhancement.

4. The leopard spot segmentation method for wide-angle fundus images based on deep learning as described in claim 1, characterized in that, The leopard spot segmentation module in S300 includes: S301. After completing multi-channel image enhancement and obtaining the supervised mask, the wide-angle fundus image is segmented into local image patches of size 512×512 pixels with a stride of 64 pixels to obtain high-resolution samples for model training. Subsequently, all local image patches are divided into training set, validation set, and test set in a ratio of 50%:20%:30%. The training set is used for parameter learning of the segmentation model; the validation set is used for hyperparameter tuning and model selection; and the test set is used for independent performance evaluation of the final model. S302. Using the sliced ​​multi-channel image patch as input, pixel-by-pixel prediction of the leopard spot region is achieved through a U-Net-based encoder-decoder structure model.

5. The leopard spot segmentation method for wide-angle fundus images based on deep learning as described in claim 4, characterized in that, The pixel-by-pixel prediction of the leopard spot region using the U-Net-based encoder-decoder structure model in S302 includes: The various predicted local image patches output by the model are stitched together according to their spatial location and fused into a complete leopard spot prediction mask: First, for the same source image, record the coordinates of the top-left corner of each prediction patch. and its corresponding prediction probability map ,in, For patch index, The coordinates are the internal coordinates of the patch; based on the maximum coordinate value of all patches and the preset patch size. The original full image size is inferred to be and initialize the reconstructed image at that size. With weight accumulation matrix Let be a zero matrix, where, These are the pixel coordinates of the entire image; Subsequently, a two-dimensional Gaussian weight kernel is constructed using a one-dimensional Gaussian function, specifically within the interval... Take evenly from the top sampling points ,calculate ; and on Perform maximum value normalization, and then use the outer product to obtain the two-dimensional Gaussian weight matrix: ; During the weighted write phase, for each patch, its internal coordinates are traversed. Predicting probability Multiply by Gaussian weights Accumulate to the corresponding position in the whole image Simultaneously, the corresponding weights are accumulated: ； ； After all patches are written, pixel-by-pixel weight normalization is performed on the entire image; to avoid division by zero, when a certain pixel... Setting it to 1, we obtain the final leopard spot probability map: ； Normalized probability graph Then through the preset threshold Binarization is performed when If a pixel is identified as a leopard spot, it is considered as background. This process yields a binary segmentation mask for the leopard spot in the entire wide-angle fundus image.

6. The leopard spot segmentation method for wide-angle fundus images based on deep learning as described in claim 1, characterized in that, The pseudo-supervised optimization module in S400 includes: After the initial training is completed, the trained UNet model is applied to unlabeled wide-angle fundus images to automatically generate corresponding leopard spot pseudo-labels, and these pseudo-labels are manually quality controlled. Specifically, operators manually inspect the pseudo-labels generated by the model and manually correct obviously unreasonable areas, including filling in missing areas, deleting obvious false positive areas, and adjusting discontinuous or large-deviation boundaries, so that the pseudo-labels can more accurately reflect the real leopard spot structure in the fundus image. The pseudo-labels that have undergone manual quality control are considered "expanded labeled samples" and are used together with the original manually labeled training data for model retraining.

Images