A Data Augmentation-Based Method for Lesion Segmentation and Quantitative Analysis in Fundus Images

Abstract

This invention discloses a data-enhanced method for lesion segmentation and quantitative analysis of fundus images, belonging to the field of fundus technology. The method includes the following steps: S1, image preprocessing; S2, optic disc and macula localization; S3, optic disc segmentation; S4, identification of myopic fundus feature changes; S5, quantification of leopard spot region features. This invention addresses the problem that existing methods cannot accurately segment and quantitatively analyze fundus image lesions, thus failing to provide high-quality diagnostic results and accurate treatment plans, reducing diagnostic accuracy and leading to poor treatment outcomes for users. This invention utilizes advanced artificial intelligence technology to perform in-depth analysis and processing of fundus images, accurately identifying and quantitatively assessing myopia-related fundus changes such as leopard spots, diffuse choroidal retinal atrophy, and patchy choroidal retinal atrophy. By generating intuitive heatmaps and quantitative reports, it provides doctors with comprehensive and objective diagnostic evidence, assisting them in developing more precise treatment plans.

Description

Technical Field

[0001] This invention relates to the field of fundus technology, specifically to a method for lesion segmentation and quantitative analysis of fundus images based on data augmentation. Background Technology

[0002] Fundus images are increasingly used in clinical ophthalmology, primarily for diagnosing and monitoring the progression of eye diseases. However, due to the complexity of eye diseases, lesions in fundus images are often unclear and variable, making accurate identification of lesions difficult.

[0003] Chinese patent CN103705207A discloses a fundus imaging system classified as a medical device for ophthalmic examination. It comprises an electro-optical fundus endoscope, a camera, and a computer. The camera is integrated into the electro-optical fundus endoscope. In use, the camera's transmission line is connected to the computer, allowing the fundus image to be transmitted and displayed on the monitor screen. It is suitable for electronic medical record systems and online diagnosis and treatment. However, this patent has the following drawbacks:

[0004] Existing technologies cannot accurately segment and quantify lesions in fundus images, cannot provide high-quality diagnostic results, reduce diagnostic accuracy, and cannot provide accurate treatment plans, resulting in poor treatment outcomes for users. Summary of the Invention

[0005] The purpose of this invention is to provide a data-enhanced method for lesion segmentation and quantitative analysis of fundus images. Utilizing advanced artificial intelligence technology, it performs in-depth analysis and processing of fundus images, accurately identifying and quantifying myopia-related fundus changes such as leopard spots, diffuse choroidal retinal atrophy, and patchy choroidal retinal atrophy. By generating intuitive heatmaps and quantitative reports, it provides doctors with comprehensive and objective diagnostic evidence, helping them to formulate more precise treatment plans and solving the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] A method for lesion segmentation and quantitative analysis of fundus images based on data augmentation, comprising the following steps:

[0008] S1. Image preprocessing:

[0009] Based on retinal camera to capture images of the user's fundus;

[0010] The system performs Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement on user fundus photos.

[0011] The preprocessed fundus image of the user was determined;

[0012] S2, Optic disc and macular localization:

[0013] Localization of the optic disc and macula based on deep learning object detection;

[0014] Train a model for optic disc and macula detection;

[0015] Output target detection bounding boxes based on optic disc and macula detection models;

[0016] The center point of the detection frame is used as the center point of the optic disc and the macula, respectively.

[0017] S3, Display Split:

[0018] The origin is obtained by positioning the center point of the viewing disc using the viewing disc itself.

[0019] Based on the computer vision attention mechanism, polar coordinate transformation is performed on the user's fundus image;

[0020] On a polar coordinate image, the edge detection operator is used to obtain the view disk edge in polar coordinates;

[0021] Perform inverse polar coordinate transformation on the user's fundus image to obtain the optic disc edge in a rectangular coordinate system, and then perform fine segmentation of the optic disc;

[0022] S4. Recognition of changes in fundus features in myopia:

[0023] The deep learning semantic segmentation network TransUnet was used to establish recognition models for leopard spot regions, patchy atrophy regions, and diffuse atrophy regions, respectively.

[0024] The samples from the training set are input into a deep learning semantic segmentation network model to segment leopard spot regions, patchy atrophy regions, and diffuse atrophy regions.

[0025] The segmentation accuracy of the deep learning semantic segmentation network model is evaluated using a test set, and the accuracy, sensitivity, and specificity are calculated respectively.

[0026] S5. Leopard Spot Region Feature Quantification:

[0027] Based on the results of ROI and optic disc segmentation, the density of leopard spots within 4.5 mm with the center of the optic disc as the origin, and the density of leopard spots within 1 mm, 3 mm, and 6 mm with the center of the macula as the origin, as well as the density of leopard spots in the corresponding areas of the upper and lower nasotemporal region and the upper, lower, upper temporal region and the lower temporal region.

[0028] The calculation formula is as follows:

[0029]

[0030] Where S' is the area of ​​leopard spots within the region, S is the fundus area of ​​the corresponding region, and ρ is a dimensionless value representing the proportion of leopard spots within the region.

[0031] Preferably, in step S1, the user's fundus photograph undergoes Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement processing, and the following operations are performed:

[0032] Obtain photos of the user's fundus;

[0033] Based on image cropping and magnification techniques, the Region of Interest (ROI) is extracted from the user's fundus photo.

[0034] Extract the regions containing important information from the user's fundus photos, remove irrelevant background from the user's fundus photos, and focus the analysis on the key parts of the user's fundus photos;

[0035] Based on advanced image denoising algorithms, noise reduction processing is performed on users' fundus photos;

[0036] The value of a point in the user's fundus photo is replaced by the median value of all points in a neighborhood of that point. Pixels with large differences in grayscale values ​​from surrounding pixels are replaced with values ​​that are close to those of their surrounding pixels. This process eliminates isolated noise points and removes noise from the user's fundus photo.

[0037] Based on histogram equalization and Gamma correction techniques, the user's fundus photos are normalized.

[0038] Eliminate differences in brightness and contrast between user fundus photos to identify consistent user fundus images;

[0039] Based on image enhancement technology, the user's fundus image is enhanced.

[0040] The contrast is enhanced and sharpened on the user's fundus images to highlight the detailed features in the images and make the lesion areas more obvious.

[0041] Preferably, the user's fundus image undergoes contrast enhancement processing, including:

[0042] Extract the adjusted pixel grayscale values ​​and the unadjusted pixel grayscale values ​​corresponding to noise points in the user's fundus image;

[0043] Extract the pixel grayscale value corresponding to the pixel adjacent to the noise point in the user's fundus image;

[0044] The pixel grayscale adjustment coefficient for each noise point is obtained by using the adjusted pixel grayscale value and the unadjusted pixel grayscale value corresponding to each noise point in the user's fundus image, as well as the pixel grayscale values ​​of the adjacent pixels of the noise point in the user's fundus image.

[0045] The pixel grayscale adjustment coefficient corresponding to each noise point is obtained by the following formula:

[0046]

[0047] Where L represents the pixel grayscale adjustment coefficient corresponding to each noise point; n represents the total number of pixels adjacent to the noise point in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted and unadjusted pixel grayscale values ​​corresponding to noise points; H z represents the median grayscale value of n pixels adjacent to the noise point; g represents the adjustment factor, which is obtained by the following formula:

[0048]

[0049] Where g represents the adjustment factor; n represents the total number of pixels adjacent to noise points in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted pixel grayscale value and the original pixel grayscale value corresponding to the noise point;

[0050] The user's fundus image is contrast-enhanced by using the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image.

[0051] Preferably, the user's fundus image is subjected to contrast enhancement processing using pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image, including:

[0052] Extract the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image;

[0053] Pixels in the user's fundus image whose difference between the average grayscale value of the pixels and the average grayscale value of the adjusted pixels corresponding to the noise points is lower than a preset grayscale difference threshold are extracted as target pixels.

[0054] The contrast adjustment coefficient is obtained using the pixel grayscale value corresponding to the target pixel.

[0055] The contrast adjustment coefficient is obtained by the following formula:

[0056]

[0057] Where k represents the contrast adjustment coefficient; m represents the number of target pixels; H mi H represents the grayscale value of the i-th target pixel; sp H represents the average adjusted pixel grayscale value corresponding to all noise points; sz H represents the median adjusted pixel grayscale value corresponding to a noisy point; min H represents the minimum grayscale value of the pixel corresponding to the target pixel. sm This represents the maximum adjusted pixel grayscale value corresponding to the noise point;

[0058] The target contrast is obtained by using the contrast adjustment coefficient and the pixel grayscale adjustment coefficient corresponding to all noise points contained in the user's fundus image;

[0059] The target contrast is obtained by the following formula:

[0060]

[0061] Among them, D m Indicates the target contrast; D represents the original grayscale value of the user's fundus image; a represents the total number of noise points contained in the user's fundus image; L i L represents the pixel grayscale adjustment coefficient corresponding to the i-th noise point; z This represents the median value of the pixel grayscale adjustment coefficient corresponding to n noise points; k represents the contrast adjustment coefficient.

[0062] The user's fundus image is adjusted to enhance contrast according to the target contrast.

[0063] Preferably, in step S2, after training the optic disc and macula detection model, the following operations are performed:

[0064] Based on the user's requirements for optic disc and macula localization in fundus images, annotation samples of the optic disc and macula regions are collected to determine annotation-based optic disc and macula samples.

[0065] The network model of Single-Step Detection (SSD) for object detection was selected, and the backbone network in the network structure was selected as ResNet-50.

[0066] The network model for single-step target detection (SSD) is trained based on labeled visual disc and macular samples, and the network model parameters are continuously optimized.

[0067] The detection models for the optic disc and macula were determined.

[0068] Preferably, in step S2, the target detection box is output based on the optic disc and macula detection model, and the following operations are performed:

[0069] Obtain optic disc and macula detection models;

[0070] The preprocessed fundus images of the user are input into the optic disc and macular detection model;

[0071] Target detection is performed on preprocessed user fundus images based on optic disc and macula detection models, enabling accurate detection of target regions at different scales.

[0072] The target detection box is output based on the optic disc and macula detection model, and the center point of each region is extracted as the center point of the optic disc and macula.

[0073] Preferably, in step S3, the viewing disc is finely segmented, and the following operations are performed:

[0074] Polar coordinate transformation: Converting the user's fundus image into polar coordinates with the center of the optic disc as the origin;

[0075] Edge detection: using edge detection operators to identify the edges of the visual disk;

[0076] Edge delineation: The identified edges are converted back to Cartesian coordinates for delineation, and the visual disk area is accurately segmented by delineating the edges.

[0077] Preferably, in step S4, the leopard-spot area, patchy atrophy area, and diffuse atrophy area are segmented, and the following operations are performed:

[0078] Based on the need to identify changes in fundus characteristics in myopia, the samples were labeled by two doctors;

[0079] One doctor provides the initial annotations, and the other doctor reviews and corrects them;

[0080] Among them, the labeled samples include leopard spot areas, patchy atrophy areas, and diffuse atrophy areas;

[0081] The labeled samples were divided into training and test sets in a 7:3 ratio;

[0082] During training, samples from the training set are input into the deep learning semantic segmentation network model;

[0083] Among them, the deep learning semantic segmentation network model uses the Transformer structure in the encoder to extract high-level features from the input user fundus image;

[0084] During the decoding stage, upsampling is used to restore the extracted high-level features to the resolution of the input image;

[0085] The confidence probability map of the deep learning semantic segmentation network model output and the extracted high-level features;

[0086] Threshold segmentation is performed based on confidence probability maps to further segment leopard spot regions, patchy atrophy regions, and diffuse atrophy regions.

[0087] Preferably, in step S4, the calculation of accuracy, sensitivity, and specificity involves the following operations:

[0088] Using image pixels as the smallest unit of analysis, the accuracy, sensitivity, and specificity of leopard spot regions, patchy atrophy regions, and diffuse atrophy regions were calculated separately.

[0089] Among them, the accuracy, sensitivity and specificity of the leopard spot region were 0.9766, 0.9551 and 0.9801, respectively;

[0090] Among them, the accuracy, sensitivity and specificity of patchy atrophy areas were 0.9993, 0.9795 and 0.9995, respectively;

[0091] The accuracy, sensitivity, and specificity of the diffuse atrophy area were 0.997, 0.978, and 0.998, respectively.

[0092] Preferably, in step S5, after the density of the leopard spots is calculated and quantified, a quantitative report of the user's fundus is generated. The quantitative report of the user's fundus includes detailed information on the area, shape, and severity of the lesion area. Based on the quantitative report of the user's fundus, doctors are assisted in developing personalized treatment plans for the user and providing personalized treatment.

[0093] Compared with the prior art, the beneficial effects of the present invention are:

[0094] This invention utilizes advanced artificial intelligence technology to perform in-depth analysis and processing of fundus images, accurately identifying and quantifying myopia-related fundus changes such as leopard spots, diffuse choroidal retinal atrophy, and patchy choroidal retinal atrophy. By generating intuitive heat maps and quantitative reports, it provides doctors with comprehensive and objective diagnostic evidence, helping them to formulate more precise treatment plans. Attached Figure Description

[0095] Figure 1 This is a flowchart of the fundus image lesion segmentation and quantitative analysis method of the present invention;

[0096] Figure 2 This is a flowchart of the lesion extraction process for the fundus image lesion segmentation and quantitative analysis method of the present invention.

[0097] Figure 3 This is a schematic diagram of the pathological myopia characteristic indicators of the fundus image lesion segmentation and quantitative analysis method of the present invention. Detailed Implementation

[0098] 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.

[0099] To address the current limitations of precise segmentation and quantitative analysis of lesions in fundus images, which result in low-quality diagnostic results, reduced diagnostic accuracy, and inability to provide accurate treatment plans, leading to poor treatment outcomes for users, please refer to [link to relevant documentation]. Figures 1-3 This embodiment provides the following technical solution:

[0100] A method for lesion segmentation and quantitative analysis of fundus images based on data augmentation, comprising the following steps:

[0101] S1. Image preprocessing:

[0102] Based on retinal camera to capture images of the user's fundus;

[0103] The system performs Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement on user fundus photos.

[0104] It can eliminate non-fundus structures in the user's fundus photos, improve the clarity of internal feature edges, and reduce differences between images. Such preprocessing lays a solid foundation for subsequent advanced image analysis.

[0105] The preprocessed fundus image of the user was determined;

[0106] It should be noted that by connecting a retinal camera to a cloud server, fundus images can be acquired instantly and online diagnoses can be performed, providing convenient and efficient telemedicine services for a wide range of patients. Regardless of their location, as long as they have internet access, patients can enjoy professional fundus examination services, greatly alleviating the problem of strained medical resources and improving the accessibility and convenience of medical services. Specifically, this includes:

[0107] Real-time image transmission: Through real-time connection with retinal cameras and cloud servers, the patient's fundus images can be acquired instantly and uploaded to the cloud server for analysis and processing. In this way, no matter where the patient is, as long as there is network access, fundus examination can be performed in real time and professional diagnostic results can be obtained.

[0108] Online diagnosis: It can perform real-time analysis and diagnosis of uploaded fundus images and return the diagnosis results to doctors and patients. Doctors can view the diagnosis results through the online platform and make further diagnosis and treatment suggestions. Patients can also obtain the diagnosis results through the online platform and understand their eye health status in a timely manner.

[0109] Telemedicine services: Through telemedicine services, patients can receive professional fundus examinations without going to the hospital. This is especially suitable for remote areas and areas with scarce medical resources. Telemedicine services not only improve the accessibility and convenience of medical services, but also reduce patients' travel time and costs, and improve patients' medical experience.

[0110] Electronic medical record management: Supports storing diagnostic results in the electronic medical record system, making it convenient for doctors to review and compare changes in the patient's fundus at any time. The application of electronic medical records improves the continuity and accuracy of medical services and provides strong support for long-term eye disease management and follow-up.

[0111] In this embodiment, the user's fundus photograph undergoes Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement processing, and the following operations are performed:

[0112] Obtain photos of the user's fundus;

[0113] Based on image cropping and magnification techniques, the Region of Interest (ROI) is extracted from the user's fundus photo.

[0114] Extract the regions containing important information from the user's fundus photos, remove irrelevant background from the user's fundus photos, and focus the analysis on the key parts of the user's fundus photos;

[0115] Based on advanced image denoising algorithms, noise reduction processing is performed on users' fundus photos;

[0116] The value of a point in the user's fundus image is replaced by the median value of all points in its neighborhood. Pixels with large differences in grayscale values ​​from surrounding pixels are replaced with values ​​close to those of their neighbors. This eliminates isolated noise points, removes noise from the user's fundus image, improves the clarity and contrast of the image, and ensures that details in the lesion area can be accurately captured.

[0117] Based on histogram equalization and Gamma correction techniques, the user's fundus photos are normalized.

[0118] Eliminate differences in brightness and contrast between user fundus photos to identify consistent user fundus images;

[0119] Based on image enhancement technology, the user's fundus image is enhanced.

[0120] The contrast enhancement and sharpening process is applied to the user's fundus images to highlight the detailed features in the images, making the lesion areas more obvious and providing a reliable foundation for subsequent automated identification and diagnosis.

[0121] Specifically, contrast enhancement processing is performed on the user's fundus images, including:

[0122] Extract the adjusted pixel grayscale values ​​and the unadjusted pixel grayscale values ​​corresponding to noise points in the user's fundus image;

[0123] Extract the pixel grayscale value corresponding to the pixel adjacent to the noise point in the user's fundus image;

[0124] The pixel grayscale adjustment coefficient for each noise point is obtained by using the adjusted pixel grayscale value and the unadjusted pixel grayscale value corresponding to each noise point in the user's fundus image, as well as the pixel grayscale values ​​of the adjacent pixels of the noise point in the user's fundus image.

[0125] The pixel grayscale adjustment coefficient corresponding to each noise point is obtained by the following formula:

[0126]

[0127] Where L represents the pixel grayscale adjustment coefficient corresponding to each noise point; n represents the total number of pixels adjacent to the noise point in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted and unadjusted pixel grayscale values ​​corresponding to noise points; H z represents the median grayscale value of n pixels adjacent to the noise point; g represents the adjustment factor, which is obtained by the following formula:

[0128]

[0129] Where g represents the adjustment factor; n represents the total number of pixels adjacent to noise points in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted pixel grayscale value and the original pixel grayscale value corresponding to the noise point;

[0130] The user's fundus image is contrast-enhanced by using the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image.

[0131] The technical effects of the above solution are as follows: This solution first focuses on noise points in the user's fundus image. By extracting the pixel grayscale values ​​of these noise points before and after adjustment, as well as the pixel grayscale values ​​of their adjacent pixels, it provides basic data for subsequent processing. This targeted processing method helps to more accurately identify and process noise in the image, avoiding unnecessary interference to the overall image. By calculating the pixel grayscale adjustment degree coefficient L corresponding to each noise point, this solution can dynamically determine the degree of grayscale adjustment for each noise point based on the difference in grayscale values ​​between the noise point and its adjacent pixels. This dynamic adjustment mechanism makes contrast enhancement processing more flexible and precise, and can better adapt to the characteristics of different fundus images.

[0132] The introduction of the adjustment factor g further enhances the flexibility and adaptability of this scheme. By comprehensively considering the gray value differences between noise points and their neighboring pixels, as well as the total number of neighboring pixels n, the adjustment factor g can more comprehensively reflect the local features of the fundus image, thereby guiding the calculation of the pixel gray-level adjustment coefficient, making the contrast enhancement processing more reasonable and effective. Through contrast enhancement processing, this scheme can significantly improve the quality of the user's fundus image. On the one hand, by reducing the interference of noise points, the image becomes clearer and more accurate; on the other hand, by enhancing the contrast of the image, the fundus structures become more prominent and easier to identify, which is helpful for subsequent medical diagnosis and analysis.

[0133] In summary, this technical solution achieves more refined and effective contrast enhancement of user fundus images through targeted noise point processing, dynamic adjustment of pixel grayscale, introduction of an adjustment factor g, and improvement of image quality. This is of great significance for improving the accuracy of fundus image recognition and diagnostic efficiency.

[0134] Specifically, the user's fundus image is subjected to contrast enhancement processing using the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image, including:

[0135] Extract the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image;

[0136] Pixels in the user's fundus image whose difference between the average grayscale value of the pixels and the average grayscale value of the adjusted pixels corresponding to the noise points is lower than a preset grayscale difference threshold are extracted as target pixels.

[0137] The contrast adjustment coefficient is obtained using the pixel grayscale value corresponding to the target pixel.

[0138] The contrast adjustment coefficient is obtained by the following formula:

[0139]

[0140] Where k represents the contrast adjustment coefficient; m represents the number of target pixels; H mi H represents the grayscale value of the i-th target pixel; sp H represents the average adjusted pixel grayscale value corresponding to all noise points; sz H represents the median adjusted pixel grayscale value corresponding to a noisy point; min H represents the minimum grayscale value of the pixel corresponding to the target pixel. sm This represents the maximum adjusted pixel grayscale value corresponding to the noise point;

[0141] The target contrast is obtained by using the contrast adjustment coefficient and the pixel grayscale adjustment coefficient corresponding to all noise points contained in the user's fundus image;

[0142] The target contrast is obtained by the following formula:

[0143]

[0144] Among them, D m Indicates the target contrast; D represents the original grayscale value of the user's fundus image; a represents the total number of noise points contained in the user's fundus image; L i L represents the pixel grayscale adjustment coefficient corresponding to the i-th noise point; z This represents the median value of the pixel grayscale adjustment coefficient corresponding to n noise points; k represents the contrast adjustment coefficient.

[0145] The user's fundus image is adjusted to enhance contrast according to the target contrast.

[0146] The technical effect of the above solution is as follows: By extracting the pixel grayscale adjustment coefficients corresponding to all noise points in the user's fundus image, this solution can perform fine-grained contrast adjustment for each noise point. This fine-grained processing helps reduce the impact of noise on image quality while preserving important details and features in the fundus image. By setting a grayscale difference threshold, pixels with grayscale values ​​close to the adjusted target noise point are selected as target pixels. This more accurately reflects the grayscale characteristics of local areas in the fundus image, providing a reliable basis for subsequent calculation of contrast adjustment coefficients.

[0147] The calculation of the contrast adjustment coefficient k comprehensively considers multiple factors, including the grayscale value of the target pixel, the average, median, maximum, and minimum grayscale values ​​of the noise points after adjustment. This dynamic adjustment mechanism makes contrast enhancement processing more flexible and adaptable, automatically adjusting the contrast according to the specific conditions of the fundus image to achieve the best visual effect. The target contrast is calculated by combining the original grayscale values ​​of the user's fundus image, the pixel grayscale adjustment coefficient of the noise points, and the contrast adjustment coefficient. This calculation method not only considers the overall grayscale characteristics of the fundus image but also takes into account the contrast requirements of local areas, making the adjusted image more natural and realistic. After the above contrast enhancement processing, the quality of the user's fundus image is significantly improved. Details in the image are clearer, and the contrast is more vivid, helping doctors to better observe and diagnose fundus lesions. At the same time, this scheme also has a certain degree of versatility and scalability, and can be applied to different types of fundus image processing and diagnostic tasks.

[0148] In summary, this technical solution achieves effective contrast enhancement of user fundus images through steps such as refined contrast adjustment, target pixel selection, dynamic contrast adjustment coefficient calculation, and target contrast calculation. This processing method not only improves image quality but also provides strong support for the accurate diagnosis of fundus lesions.

[0149] S2, Optic disc and macular localization:

[0150] Localization of the optic disc and macula based on deep learning object detection;

[0151] Based on the user's requirements for optic disc and macula localization in fundus images, annotation samples of the optic disc and macula regions are collected to determine annotation-based optic disc and macula samples.

[0152] The network model for single-step detection of objects (SSD) was selected. The backbone network in the network structure is the ResNet-50 network, which can quickly and efficiently identify and locate the optic disc and macula in the user's fundus image.

[0153] The network model for single-step detection (SSD) of the selected target detection was trained based on labeled optic disc and macular samples. The network model parameters were continuously optimized. The model was trained on a large-scale fundus image dataset to continuously optimize the model parameters and improve the accuracy and robustness of detection.

[0154] The optic disc and macula detection models were determined;

[0155] Obtain optic disc and macula detection models;

[0156] The preprocessed fundus images of the user are input into the optic disc and macular detection model;

[0157] Target detection is performed on preprocessed user fundus images based on optic disc and macular detection models, enabling accurate detection of target regions at different scales. Multi-scale feature maps are used for target detection to ensure accurate detection of target regions at different scales, thereby improving the accuracy and stability of detection.

[0158] Based on the visual disc and macula detection model, target detection boxes are output, and the center point of each region is extracted as the center point of the visual disc and macula.

[0159] It should be noted that the optic disc and macula are important reference points in fundus photographs. To accurately identify these areas, deep learning object detection technology is used, especially the SSD (SingleShotDetection) network model. Through training, this model can efficiently locate the position of the optic disc and macula in fundus photographs, providing key reference information for subsequent fundus analysis.

[0160] S3, Display Split:

[0161] The origin is obtained by positioning the center point of the viewing disc using the viewing disc itself.

[0162] Based on the computer vision attention mechanism, polar coordinate transformation is performed on the user's fundus image;

[0163] On a polar coordinate image, the edge detection operator is used to obtain the view disk edge in polar coordinates;

[0164] Perform inverse polar coordinate transformation on the user's fundus image to obtain the optic disc edge in a rectangular coordinate system, and then perform fine segmentation of the optic disc;

[0165] In this embodiment, the viewing disc is finely segmented, and the following operations are performed:

[0166] Polar coordinate transformation: Using the center of the optic disc as the origin, the user's fundus image is converted into polar coordinates, simplifying the identification and segmentation of the optic disc edge;

[0167] Edge detection: Edge detection operators are used to identify the edges of the visual disk, such as Canny edge detection, which improves the accuracy of edge extraction.

[0168] Edge delineation: The identified edges are converted back to Cartesian coordinates for delineation, and the visual disk area is accurately segmented by delineating the edges.

[0169] It should be noted that after determining the location of the optic disc, the attention mechanism in computer vision is further used to achieve accurate segmentation of the optic disc. By converting the fundus image to polar coordinates and using an edge detection operator to identify the edge of the optic disc, and then converting it back to Cartesian coordinates for delineation, the accurate extraction of the optic disc region is achieved. This has an important auxiliary role in the subsequent analysis of myopic fundus changes.

[0170] S4. Recognition of changes in fundus features in myopia:

[0171] We used the deep learning semantic segmentation network TransUnet to establish identification models for leopard spot regions, patchy atrophy areas, and diffuse atrophy areas, respectively. We utilized the TransUnet structure to enhance the global feature expression capability and ensure the effective extraction of large-area lesions.

[0172] Based on the need to identify changes in fundus characteristics in myopia, the samples were labeled by two doctors;

[0173] One doctor provides the initial annotations, and the other doctor reviews and corrects them;

[0174] Among them, the labeled samples include leopard spot areas, patchy atrophy areas, and diffuse atrophy areas;

[0175] The number of labeled samples were 1160, 2700, and 2780, respectively, and they were divided into training and test sets in a 7:3 ratio.

[0176] During training, samples from the training set are input into the deep learning semantic segmentation network model;

[0177] Among them, the deep learning semantic segmentation network model uses the Transformer structure in the encoder to extract high-level features from the input user fundus image;

[0178] During the decoding stage, upsampling is used to restore the extracted high-level features to the resolution of the input image;

[0179] The confidence probability map of the deep learning semantic segmentation network model output and the extracted high-level features;

[0180] Threshold segmentation is performed based on confidence probability maps to further segment leopard spot regions, patchy atrophy regions, and diffuse atrophy regions;

[0181] The segmentation accuracy of the deep learning semantic segmentation network model was evaluated using a test set. Image pixels were used as the smallest unit of analysis to calculate the accuracy, sensitivity, and specificity of the leopard spot region, patchy atrophy region, and diffuse atrophy region, respectively.

[0182] Specifically, pathological myopia is an eye disease characterized by progressive increase in refractive power, continuous elongation of the axial length, and progressive damage to the contents of the eye and the retinal and choroidal tissues, leading to visual dysfunction. It occurs at an early age and progresses rapidly, and may cause serious complications such as macular atrophy, retinal pigment epithelium atrophy, choroidal neovascularization, and subretinal hemorrhage. It is one of the leading causes of blindness in my country.

[0183] Leopard-like fundus is a fundus lesion characterized by increased visibility of large choroidal vessels around the fovea and arch vessels, accompanied by thinning of the retinal epithelium and choroidal capillaries. Multiple factors are associated with the occurrence of leopard-like fundus lesions, including age, myopia, early age-related macular degeneration, glaucoma, and body mass index. The characteristic description includes increased visibility of large choroidal vessels around the fovea and arch vessels, and thinning of the retinal epithelium and choroidal capillaries. Factors leading to leopard-like fundus lesions include aging, myopia progression, early age-related macular degeneration, glaucoma, and body mass index. Fundus image analysis identifies the leopard-like lesion area and performs quantitative assessment, providing detailed diagnostic information.

[0184] Diffuse choroidal atrophy is a relatively early-stage fundus lesion in myopic atrophic macular degeneration. It manifests as a yellowish-white lesion with indistinct borders in the posterior pole of the eyeball, with blurred choroidal vessels in the atrophic area. As the disease progresses, the atrophic lesions can be widely distributed throughout the posterior pole of the eyeball, making it one of the fundus lesions leading to myopic visual impairment. Diffuse atrophy is an important critical feature for diagnosing pathological myopia. Early intervention and treatment of this disease are significant in delaying the progression of pathological myopia and reducing visual impairment. Its characteristic description is a yellowish-white lesion with indistinct borders in the posterior pole of the eyeball, blurred choroidal vessels in the atrophic area, and the lesion can be widely distributed throughout the posterior pole of the eyeball. Diffuse atrophy is a relatively early-stage lesion in myopic atrophic macular degeneration. As the disease progresses, the atrophic lesions gradually enlarge, leading to visual impairment. Fundus image analysis can identify diffuse atrophic areas and perform quantitative assessments to provide detailed diagnostic information.

[0185] Among them, patchy choroidal retinal atrophy often presents as well-defined grayish-white lesions, ranging in size from one to several choroidal lobules, and commonly appears in the macular region or around the optic disc. The anatomical structure of patchy atrophy is essentially the loss of retinal pigment epithelium and most of the choroid. It is characterized by well-defined grayish-white lesions, usually appearing in the macular region or around the optic disc, with the lesion size ranging from one to several choroidal lobules. Patchy atrophy is caused by the loss of retinal pigment epithelium and most of the choroid, and is common in patients with high myopia. By analyzing fundus images, patchy atrophy areas can be identified and quantitatively assessed to provide detailed diagnostic information.

[0186] It should be noted that the calculation results for the leopard spot area, patchy atrophy area, and diffuse atrophy area are shown in Table 1:

[0187] Table 1: Calculation results of leopard spot area, patchy atrophy area and diffuse atrophy area

[0188]

[0189] Among them, the accuracy, sensitivity and specificity of the leopard spot region were 0.97678, 0.95506 and 0.98006, respectively;

[0190] Among them, the accuracy, sensitivity and specificity of patchy atrophy areas were 0.99929, 0.97951 and 0.99951, respectively;

[0191] The accuracy, sensitivity, and specificity of the diffuse atrophy area were 0.99627, 0.97734, and 0.99764, respectively.

[0192] It should be noted that after the density of leopard spots is calculated and quantified, a quantitative report of the user's fundus is generated. This report includes detailed information on the area, shape, and severity of the lesion. Based on this report, doctors are assisted in developing personalized treatment plans for the user.

[0193] It should be noted that by employing state-of-the-art deep learning algorithms, the fundus images are meticulously analyzed and processed. This allows for the automatic identification of lesion areas within the fundus images and the quantitative assessment of these areas, including measurements of lesion area, shape, and severity. Through this analysis, detailed lesion information is provided, helping doctors better understand the patient's eye condition. Intuitive heatmaps are generated, displaying the distribution and severity of lesion areas. These heatmaps clearly show the specific location and extent of the lesions, helping doctors quickly locate the affected areas for more accurate diagnosis and treatment, improving the precision and effectiveness of treatment. This significantly improves the accuracy and efficiency of fundus image analysis, enabling the processing of large volumes of fundus images in a short time, providing high-quality diagnostic results, reducing the workload of doctors, and enhancing diagnostic accuracy.

[0194] Specifically, by applying AI-based quantitative fundus technology, primary care ophthalmologists can more accurately detect and diagnose fundus lesions, improving their ability to prevent and control eye diseases. Simultaneously, by leveraging an AI platform to build an eye health big data center, the integration, analysis, and sharing of eye disease information can be achieved, providing strong data support for primary healthcare institutions and promoting the intelligent and precise prevention and treatment of eye diseases. This includes:

[0195] Primary care screening technology: It can provide advanced eye disease screening technology to primary healthcare institutions. Through regular fundus examinations, it can help primary care doctors to detect and diagnose fundus lesions in a timely manner, thereby improving the quality and level of primary healthcare services.

[0196] Enhanced prevention and control capabilities: By applying artificial intelligence technology, it is possible to automatically identify and quantify fundus lesions, provide detailed lesion information, help primary care physicians improve their ability to prevent and control eye diseases, and formulate more scientific prevention and control measures;

[0197] Big data support: By building an eye health big data center with the help of AI platform, a large amount of eye disease information can be integrated and analyzed, providing comprehensive data support for primary medical institutions. The results of big data analysis can not only help doctors better understand patients' conditions, but also provide scientific basis for the formulation of public health policies.

[0198] Intelligent prevention and control: Through intelligent eye disease screening and prevention and control systems, primary healthcare institutions can achieve early detection, early diagnosis and early treatment of eye diseases, improve the efficiency and quality of prevention and control work, and reduce the incidence and blindness rate of eye diseases.

[0199] Meanwhile, addressing the needs of chronic disease screening systems for children and adolescents, as well as in remote areas, regular screening and monitoring can promptly detect and intervene in the development of eye diseases such as myopia, reducing the risk of vision loss. Furthermore, by constructing a comprehensive screening system, long-term tracking and management of patients with chronic diseases can be achieved, providing strong support for developing personalized prevention and treatment plans; specifically including:

[0200] Regular screening and monitoring: Enables regular fundus examinations and monitoring for children, adolescents, and patients in remote areas, allowing for timely detection and diagnosis of fundus lesions, especially myopia-related eye diseases, and providing timely intervention measures.

[0201] Personalized prevention and treatment plans: Through detailed lesion analysis and diagnostic reports, doctors can be provided with comprehensive diagnostic information, assisting them in developing personalized prevention and treatment plans, improving prevention and treatment effectiveness, and reducing the incidence and blindness rate of eye diseases.

[0202] Chronic disease management: Supports long-term tracking and management of patients with chronic diseases, enabling timely understanding of changes in the patient's condition through regular check-ups and follow-ups, adjusting treatment plans, and providing continuous medical services.

[0203] Increase screening coverage: By promoting its application in remote areas and primary healthcare institutions, the coverage of eye disease screening can be significantly improved, ensuring that more patients can be diagnosed and treated in a timely manner, thereby improving their eye health.

[0204] Simultaneously, leveraging AI platforms enables the integration and sharing of eye disease information, providing comprehensive eye health management services to medical institutions at all levels. By constructing a tiered medical system, it facilitates the rational triage and effective treatment of eye disease patients, improving the efficiency of medical resource utilization and service quality. This contributes to enhancing the overall level of eye disease prevention and treatment in my country, bringing benefits to a vast number of eye disease patients. Specifically, this includes:

[0205] Information integration and sharing: Through the AI ​​platform, comprehensive integration and sharing of eye disease information can be achieved, providing detailed eye disease data support for medical institutions at all levels and promoting the optimal allocation of medical resources;

[0206] Tiered medical system: It can rationally divert patients to different levels of medical institutions for treatment according to the severity of their condition, ensuring that patients can receive the most appropriate medical services and improve treatment outcomes;

[0207] Resource utilization efficiency: Through the hierarchical medical system, the utilization efficiency of medical resources can be optimized, the waste of medical resources can be avoided, and the overall quality and efficiency of medical services can be improved.

[0208] Comprehensive management services: We provide comprehensive eye health management services, including prevention, screening, diagnosis, treatment and follow-up, to ensure that patients receive continuous and comprehensive medical services and improve patient satisfaction and treatment outcomes.

[0209] It should be noted that, to support model training and validation, a high-quality fundus image database was constructed based on the China High Myopia Research Consortium (CHARM). This database covers fundus images of different ages, genders, demographic characteristics, and various types of fundus cameras, and the data was labeled by multiple fundus specialists. This rich dataset provides a solid foundation for subsequent research and applications; specifically, it includes:

[0210] Data collection: Collect fundus images of different ages, genders, demographic characteristics, and types of fundus cameras to ensure data diversity and coverage;

[0211] Data annotation: Data annotation was performed by multiple ophthalmologists to ensure the accuracy and reliability of the annotation, providing high-quality data support for model training;

[0212] Database Management: Construct a fundus image database, manage and maintain the data, ensure the integrity and availability of the data, and provide a solid foundation for subsequent research and applications.

[0213] S5. Leopard Spot Region Feature Quantification:

[0214] Based on the results of ROI and optic disc segmentation, the density of leopard spots within 4.5 mm with the center of the optic disc as the origin, and the density of leopard spots within 1 mm, 3 mm, and 6 mm with the center of the macula as the origin, as well as the density of leopard spots in the corresponding areas of the upper and lower nasotemporal region and the upper, lower, upper temporal region and the lower temporal region.

[0215] The calculation formula is as follows:

[0216]

[0217] Where S' is the area of ​​leopard spots within the region, S is the fundus area of ​​the corresponding region, and ρ is a dimensionless value representing the proportion of leopard spots within the region.

[0218] Therefore, this system can automatically analyze, quantitatively assess, and assist in the diagnosis of fundus images, effectively overcoming the limitations of traditional qualitative analysis in fundus examinations. Through a human-machine collaborative operating system, the system can also provide ophthalmologists with lesion annotation and auxiliary treatment plans, contributing to the precise and comprehensive prevention and control of fundus diseases. The automatic analysis and assessment technology enables fully automated processing and analysis of fundus images, providing detailed lesion information to assist doctors in scientific diagnosis and treatment. The human-machine collaborative operating system, combined with artificial intelligence technology and doctors' professional knowledge, provides lesion annotation and auxiliary treatment plans, improving the accuracy and efficiency of diagnosis and promoting the development of fundus disease prevention and control towards precision and comprehensiveness.

[0219] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0220] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for lesion segmentation and quantitative analysis of fundus images based on data augmentation, characterized in that, Includes the following steps: S1. Image preprocessing: Based on retinal camera to capture images of the user's fundus; The system performs Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement on user fundus photos. Contrast enhancement processing is performed on the user's fundus image, including: Extract the adjusted pixel grayscale values ​​and the unadjusted pixel grayscale values ​​corresponding to noise points in the user's fundus image; Extract the pixel grayscale value corresponding to the pixel adjacent to the noise point in the user's fundus image; The pixel grayscale adjustment coefficient for each noise point is obtained by using the adjusted pixel grayscale value and the unadjusted pixel grayscale value corresponding to each noise point in the user's fundus image, as well as the pixel grayscale values ​​of the adjacent pixels of the noise point in the user's fundus image. The pixel grayscale adjustment coefficient corresponding to each noise point is obtained by the following formula: ； Where L represents the pixel grayscale adjustment coefficient corresponding to each noise point; n represents the total number of pixels adjacent to the noise point in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted and unadjusted pixel grayscale values ​​corresponding to noise points; H z represents the median grayscale value of n pixels adjacent to the noise point; g represents the adjustment factor, which is obtained by the following formula: ； Where g represents the adjustment factor; n represents the total number of pixels adjacent to noise points in the user's fundus image; H i H represents the pixel grayscale value of the i-th pixel adjacent to the noise point; q and H h This represents the adjusted pixel grayscale value and the original pixel grayscale value corresponding to the noise point; The user's fundus image is contrast-enhanced by using the pixel grayscale adjustment coefficients corresponding to all noise points contained in the user's fundus image. The preprocessed fundus image of the user was determined; S2, Optic disc and macular localization: Localization of the optic disc and macula based on deep learning object detection; Train a model for optic disc and macula detection; Output target detection bounding boxes based on optic disc and macula detection models; The center point of the detection frame is used as the center point of the optic disc and the macula, respectively. S3, Display Split: The origin is obtained by positioning the center point of the viewing disc using the viewing disc itself. Based on the computer vision attention mechanism, polar coordinate transformation is performed on the user's fundus image; On a polar coordinate image, the edge detection operator is used to obtain the view disk edge in polar coordinates; Perform inverse polar coordinate transformation on the user's fundus image to obtain the optic disc edge in a rectangular coordinate system, and then perform fine segmentation of the optic disc; S4. Recognition of changes in fundus features in myopia: The deep learning semantic segmentation network TransUnet was used to establish recognition models for leopard spot regions, patchy atrophy regions, and diffuse atrophy regions, respectively. The samples from the training set are input into a deep learning semantic segmentation network model to segment leopard spot regions, patchy atrophy regions, and diffuse atrophy regions. The segmentation accuracy of the deep learning semantic segmentation network model is evaluated using a test set, and the accuracy, sensitivity, and specificity are calculated respectively. S5. Leopard Spot Region Feature Quantification: Based on the results of ROI and optic disc segmentation, the density of leopard spots within 4.5 mm with the center of the optic disc as the origin, and the density of leopard spots within 1 mm, 3 mm, and 6 mm with the center of the macula as the origin, as well as the density of leopard spots in the corresponding areas of the upper and lower nasotemporal region and the upper, lower, upper temporal region and the lower temporal region. The calculation formula is as follows: ； Where S' is the area of ​​leopard spots within the region, S is the fundus area of ​​the corresponding region, and ρ is a dimensionless value representing the proportion of leopard spots within the region.

2. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 1, characterized in that, In step S1, the user's fundus photo undergoes Region of Interest (ROI) extraction, noise reduction, normalization, and enhancement processing, and the following operations are performed: Obtain photos of the user's fundus; Based on image cropping and magnification techniques, the Region of Interest (ROI) is extracted from the user's fundus photo. Extract the regions containing important information from the user's fundus photos, remove irrelevant background from the user's fundus photos, and focus the analysis on the key parts of the user's fundus photos; Based on advanced image denoising algorithms, noise reduction processing is performed on users' fundus photos; The value of a point in the user's fundus photo is replaced by the median value of all points in a neighborhood of that point. Pixels with large differences in grayscale values ​​from surrounding pixels are replaced with values ​​that are close to those of their surrounding pixels. This process eliminates isolated noise points and removes noise from the user's fundus photo. Based on histogram equalization and Gamma correction techniques, the user's fundus photos are normalized. Eliminate differences in brightness and contrast between user fundus photos to identify consistent user fundus images; Based on image enhancement technology, the user's fundus image is enhanced. The contrast is enhanced and sharpened on the user's fundus image to highlight the detailed features in the image and make the lesion area more obvious.

3. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 2, characterized in that, In step S2, the optic disc and macula detection models are trained, and the following operations are performed: Based on the user's need for optic disc and macula localization in fundus images, annotation samples of the optic disc and macula regions are collected to determine annotation-based optic disc and macula samples. The network model of Single-Step Detection (SSD) for object detection was selected, and the backbone network in the network structure was selected as ResNet-50. The network model for single-step target detection (SSD) is trained based on labeled visual disc and macular samples, and the network model parameters are continuously optimized. The detection models for the optic disc and macula were determined.

4. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 3, characterized in that, In step S2, a target detection bounding box is output based on the optic disc and macula detection model, and the following operations are performed: Obtain optic disc and macula detection models; The preprocessed fundus images of the user are input into the optic disc and macular detection model; Target detection is performed on preprocessed user fundus images based on optic disc and macula detection models, enabling accurate detection of target regions at different scales. The target detection box is output based on the optic disc and macula detection model, and the center point of each region is extracted as the center point of the optic disc and macula.

5. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 4, characterized in that, In step S3, the viewing disc is finely segmented, and the following operations are performed: Polar coordinate transformation: Converting the user's fundus image into polar coordinates with the center of the optic disc as the origin; Edge detection: Using edge detection operators to identify the edges of the visual disk; Edge delineation: The identified edges are converted back to Cartesian coordinates for delineation, and the visual disk area is accurately segmented by delineating the edges.

6. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 5, characterized in that, In step S4, the leopard-spot area, patchy atrophy area, and diffuse atrophy area are segmented, and the following operations are performed: Based on the need to identify changes in fundus characteristics in myopia, the samples were labeled by two doctors; One doctor provides the initial annotations, and the other doctor reviews and corrects them; Among them, the labeled samples include leopard spot areas, patchy atrophy areas, and diffuse atrophy areas; The labeled samples were divided into training and test sets in a 7:3 ratio; During training, samples from the training set are input into the deep learning semantic segmentation network model; Among them, the deep learning semantic segmentation network model uses the Transformer structure in the encoder to extract high-level features from the input user fundus image; During the decoding stage, upsampling is used to restore the extracted high-level features to the resolution of the input image; The confidence probability map of the deep learning semantic segmentation network model output and the extracted high-level features; Threshold segmentation is performed based on confidence probability maps to further segment leopard spot regions, patchy atrophy regions, and diffuse atrophy regions.

7. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 6, characterized in that, In step S4, the accuracy, sensitivity, and specificity are calculated, and the following operations are performed: Using image pixels as the smallest unit of analysis, the accuracy, sensitivity, and specificity of leopard spot regions, patchy atrophy regions, and diffuse atrophy regions were calculated separately. Among them, the accuracy, sensitivity and specificity of the leopard spot region were 0.9766, 0.9551 and 0.9801, respectively; Among them, the accuracy, sensitivity and specificity of patchy atrophy areas were 0.9993, 0.9795 and 0.9995, respectively; The accuracy, sensitivity, and specificity of the diffuse atrophy area were 0.997, 0.978, and 0.998, respectively.

8. The method for lesion segmentation and quantitative analysis of fundus images based on data augmentation according to claim 7, characterized in that, In step S5, after the density of the leopard spots is calculated and quantified, a quantitative report of the user's fundus is generated. The quantitative report of the user's fundus includes detailed information on the area, shape, and severity of the lesion area. Based on the quantitative report of the user's fundus, doctors are assisted in developing personalized treatment plans for the user.

Images