Retinal hemorrhage area identification method and system

Abstract

An embodiment of the present invention provides a method and system for identifying retinal hemorrhage areas, belonging to the field of image processing technology. The retinal hemorrhage area identification method includes: collecting image information of the user's eyes; performing optic disc and blood vessel identification in the image information based on an optic disc identification model and a blood vessel identification model, respectively, to obtain optic disc and blood vessel identification results; based on the optic disc and blood vessel identification results, performing optic disc and blood vessel peeling in the image information, respectively, to obtain peeled image information; preprocessing the peeled image information, and performing hemorrhage area identification based on the preprocessed image information to obtain an identification result. The solution of the present invention performs retinal hemorrhage identification in images after optic disc and retinal peeling, ensuring identification accuracy.

Description

Technical Field

[0001] The present invention relates to the technical field of image processing, and in particular to a retinal hemorrhage area recognition method and a retinal hemorrhage area recognition system. Background Art

[0002] When studying the identification and segmentation of hemorrhage areas in retinal fundus images, traditional image processing and deep learning methods typically preprocess the data using traditional methods, such as brightness balancing, contrast enhancement, denoising, and normalization. The model is then used directly to identify and segment the preprocessed data. However, retinal hemorrhage areas are different from other fundus structures, such as blood vessels, optic discs, and macula. Other fundus structures are present as visible targets in retinal fundus images, while hemorrhage is not a symptom in a normal retina. Furthermore, their shape features are irregular, with varying sizes and random distribution. Therefore, identification is inevitably affected by visible targets such as blood vessels and optic discs. Previous researchers have rarely considered the impact of other fundus structures on hemorrhage areas when conducting research on the identification and segmentation of hemorrhage areas.

[0003] In order to address the problem that the existing schemes have inaccurate retinal hemorrhage identification results because they do not consider the impact of other fundus tissue structures on the hemorrhage area, it is necessary to create a new scheme for retinal hemorrhage area identification. Summary of the Invention

[0004] The purpose of the embodiments of the present invention is to provide a method and system for identifying retinal hemorrhage areas, so as to at least solve the problem that the existing solutions do not consider the influence of other fundus tissue structures on the hemorrhage area, resulting in inaccurate retinal hemorrhage identification results.

[0005] To achieve the above objectives, the present invention provides, in a first aspect, a method for identifying a retinal hemorrhage area, the method comprising: acquiring image information of a user's eye; performing optic disc and blood vessel identification in the image information based on an optic disc identification model and a blood vessel identification model, respectively, to obtain optic disc and blood vessel identification results; performing optic disc and blood vessel stripping in the image information based on the optic disc and blood vessel identification results, respectively, to obtain image information after stripping; preprocessing the stripped image information, and performing hemorrhage area identification based on the preprocessed image information to obtain an identification result.

[0006] Optionally, the optic disc recognition model is a UNet++ model; and the blood vessel recognition model is a DC-UNet model.

[0007] Optionally, after the optic disc and blood vessel stripping is completed, the method further includes: performing pixel filling on the optic disc and blood vessel stripping area, including: mapping pixel values ​​of the contour adjacent area of ​​the area to be filled to a three-dimensional coordinate space; clustering all pixel values ​​based on the distribution of pixels of the contour adjacent area of ​​the area to be filled in the three-dimensional coordinate space using the density of pixels in the three-dimensional space coordinates to obtain clustered pixel values; taking values ​​of filling pixels based on the clustered pixel values; and performing pixel filling on the stripping area based on the taken values ​​of the filling pixels; wherein,

[0008] The calculation formula for filling pixels is:

[0009]

[0010] Among them, M represents the total number of all classified pixels; m represents the number of all pixels in a cluster; P is the filling pixel; and N is the number of types.

[0011] Optionally, all pixel values ​​are clustered using the density of pixels in three-dimensional space coordinates. The clustering algorithm used is an improved DBSCAN algorithm. The clustering process includes: presetting an initial radius and a minimum number of samples in three-dimensional space coordinates; randomly selecting an unvisited pixel point, and based on the initial radius, finding all pixel points that can be reached based on the initial radius and counting the number of points; comparing the counted number of points with the minimum number of samples. If the counted number of points is greater than the minimum number of samples, the pixel point selected this time is used as the core point and a new cluster is created; otherwise, the pixel point selected this time is used as noise; and diffusion clustering is performed based on the determined core point until all pixel points are traversed.

[0012] Optionally, the preprocessing of the stripped image information includes: preprocessing the stripped image information based on different preset preprocessing models, respectively, to obtain the preprocessed image information under each preprocessing model; comparing the preprocessed image information under each preprocessing model, comparing the segmentation effects based on preset standards, and identifying the preprocessed image information under the preprocessing model with the best segmentation effect as the preprocessing result.

[0013] Optionally, the different preprocessing models include: a single brightness balance model, a single contrast enhancement model, a single denoising model, a brightness balance and contrast enhancement combined model, a denoising and brightness balance combined model, a denoising and contrast enhancement combined model, and a denoising plus brightness balance plus contrast enhancement combined model.

[0014] Optionally, the bleeding area identification based on the preprocessed image information includes: after the preprocessing is completed, using the preprocessed image information as an input parameter, training in a pre-built bleeding recognition model to obtain a training result; judging whether there is a bleeding area in the preprocessed image information based on the training result; if so, calibrating the bleeding area.

[0015] Optionally, the method further includes: performing bleeding recognition model training, including: collecting a public data set of eye images in a public database, the public data set including: IDRiD data set, DRIVE data set, STARE data set, CHASE-DB1 data set; collecting historical image data specified by the user; using the public data set and the historical image data as training sets, performing model training in a preset neural network, and obtaining a bleeding recognition model.

[0016] A second aspect of the present invention provides a retinal hemorrhage area identification system, comprising: an acquisition unit for acquiring image information of a user's eyes; a processing unit for performing optic disc and blood vessel identification in the image information based on an optic disc identification model and a blood vessel identification model, respectively, to obtain optic disc and blood vessel identification results; a stripping unit for performing optic disc and blood vessel stripping in the image information based on the optic disc and blood vessel identification results, respectively, to obtain stripped image information; and an identification unit for preprocessing the stripped image information, identifying the hemorrhage area based on the preprocessed image information, and obtaining an identification result.

[0017] On the other hand, the present invention provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, enable the computer to execute the above-mentioned method for identifying a retinal hemorrhage area.

[0018] Through the above technical solution, the present invention first identifies the optic disc and blood vessels in the captured eye image. Because these two tissue structures most interfere with retinal hemorrhage detection, these two areas are then separated based on the identification results to prevent them from affecting subsequent retinal hemorrhage identification. After the optic disc and retina are separated, retinal hemorrhage identification is performed in the remaining images to ensure identification accuracy.

[0019] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] The accompanying drawings are used to provide a further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following detailed description, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention. In the accompanying drawings:

[0021] Figure 1 is a flowchart of the steps of a method for identifying a retinal hemorrhage area provided by one embodiment of the present invention;

[0022] Figure 2 This is a system structure diagram of a retinal hemorrhage area identification system provided by one embodiment of the present invention. DETAILED DESCRIPTION

[0023] The following describes the specific embodiments of the present invention in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present invention and are not intended to limit the present invention.

[0024] When studying the identification and segmentation of hemorrhage areas in retinal fundus images, traditional image processing and deep learning methods typically preprocess the data using traditional methods, such as brightness balancing, contrast enhancement, denoising, and normalization. The model is then used directly to identify and segment the preprocessed data. However, retinal hemorrhage areas are different from other fundus structures, such as blood vessels, optic discs, and macula. Other fundus structures are present as visible targets in retinal fundus images, while hemorrhage is not a symptom in a normal retina. Furthermore, their shape features are irregular, with varying sizes and random distribution. Therefore, identification is inevitably affected by visible targets such as blood vessels and optic discs. Previous researchers have rarely considered the impact of other fundus structures on hemorrhage areas when conducting research on the identification and segmentation of hemorrhage areas.

[0025] To address the problem of inaccurate retinal hemorrhage identification results caused by existing solutions not considering the influence of other fundus tissue structures on the hemorrhage area, the present invention proposes a new retinal hemorrhage area identification method. The present invention first identifies the optic disc and blood vessels in the collected eye image. Because these two tissue structures have the greatest interference with retinal hemorrhage identification, these two areas are separated based on the identification results to prevent their influence on subsequent retinal hemorrhage identification. After the optic disc and retina are separated, retinal hemorrhage identification is performed in the remaining image to ensure identification accuracy.

[0026] Figure 1 This is a flow chart of a method for identifying retinal hemorrhage areas provided by one embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a method for identifying a retinal hemorrhage area, the method comprising:

[0027] Step S10: collecting image information of the user's eyes.

[0028] Specifically, the present invention is based on the user's eye image and retinal hemorrhage recognition after image processing. Therefore, it is necessary to accurately capture the user's eye image to ensure subsequent recognition accuracy.

[0029] Step S20: performing optic disc and blood vessel recognition in the image information based on the optic disc recognition model and the blood vessel recognition model respectively to obtain optic disc and blood vessel recognition results.

[0030] Specifically, retinal hemorrhages have irregular shapes and sizes, appear bright or dark red, and are randomly distributed. Therefore, they are inherently susceptible to interference. The optic disc and ocular blood vessels share similar characteristics with retinal hemorrhages. Existing methods that directly identify hemorrhage areas in eye images often misidentify the optic disc or blood vessels as retinal hemorrhages. For this reason, the present invention requires optic disc and blood vessel dissection before retinal hemorrhage removal. Accurately dissecting these two areas requires ensuring accurate identification of the optic disc and blood vessels.

[0031] Preferably, the optic disc recognition model is a UNet++ model; and the blood vessel recognition model is a DC-UNet model.

[0032] Step S30: Based on the optic disc and blood vessel recognition results, the optic disc and blood vessels are stripped in the image information to obtain the stripped image information.

[0033] Specifically, the average similarity of vascular and optic disc separation across the entire retinal fundus image is calculated by summing and averaging the similarities between the padded pixels in the vascular and optic disc regions and their neighboring pixels, ensuring that the similarity for the entire image is between [0, 1]. A similarity closer to 1 indicates a higher similarity, and the closer the pixel padded value is to the surrounding pixels, the better the effect. Conversely, a greater difference in the pixel padded value from the neighboring pixels indicates a worse effect. The target formula for vascular separation across the entire image is as follows:

[0034]

[0035] The objective formula for disc stripping of the entire image is as follows:

[0036]

[0037] in, represents the average pixel similarity of the blood vessel stripping and filling of the entire image, and It represents the average similarity of the pixels filled in the optic disc peeling of the whole image, N represents the representative value of the neighboring pixels of the filling pixel, V represents the filling pixels in the vascular peeling area, and O represents the filling pixels in the optic disc peeling area. It represents the similarity between the filling pixel of the ith point in the vascular stripping area and its neighboring pixels, and It represents the similarity between the filling pixel of the jth point in the optic disc peeling area and its neighboring pixels. n represents the number of peeled blood vessel pixels, and m represents the number of peeled optic disc pixels.

[0038] The similarity between a single fill pixel and its neighboring pixels is calculated using a weighted average method. This method uses the average, median, and mode values ​​of the neighboring pixels to calculate the similarity with the fill pixel. The weighted average is then used to calculate the overall similarity between the fill pixel and its neighboring pixels. The target formula for vascular stripping is as follows:

[0039]

[0040] The target formula for optic disc dissection is as follows:

[0041]

[0042] Among them, S represents the similarity between the filled pixel and the adjacent pixels, represents the filled pixels in the vascular stripping area, represents the fill pixels of the optic disc detachment area, and represents the average value of neighboring pixels, represents the median of neighboring pixels, represents the mode of neighboring pixels. α and β are weight coefficients of the similarity between different representative values ​​and the filled pixel.

[0043] When calculating the similarity between a single filled pixel and its neighboring pixels, the weighted average method is also used. Since the retinal fundus image data is a 24-bit RGB three-channel image, namely the Red channel, the Green channel, and the Blue channel. Therefore, the value of the pixel point can be regarded as a three-dimensional coordinate point, so the Euclidean distance between the two pixels can be calculated, and the straight-line distance between the two pixels can be determined based on the Euclidean distance between the pixels to judge the similarity. On the other hand, the pixel point can also be regarded as a vector starting from the coordinate origin or a fixed point, and the cosine value of the angle between the two pixel mapping vectors is calculated based on the cosine similarity to judge the similarity between the two pixels. The target formula can be expressed as:

[0044]

[0045] in, 𝜆 is the similarity between two pixels calculated using Euclidean distance, with a value range of [0, 1]. A larger value indicates a higher similarity in Euclidean distance and a better filling effect. cos(𝜃) is the cosine similarity, with a value range of [0, 1]. The closer the value is to 1, the higher the similarity. 𝜆 is the balance coefficient between Euclidean distance and cosine similarity.

[0046] When calculating the Euclidean distance and cosine similarity based on the three channel values ​​of the pixel, since the grayscale features of the retinal fundus image data in the three channels are different, the sensitivity to the three channels is also different.

[0047] According to the single-channel output of the retinal fundus image, the grayscale features of the retinal fundus image are more obvious in the green channel, while the blue channel is the weakest. Therefore, the target formula of the Euclidean distance similarity can be expressed as:

[0048]

[0049] Where σ is the sensitivity of the Euclidean distance. When the Euclidean distance is equal, the higher the sensitivity, the higher the similarity. Conversely, the lower the sensitivity, the lower the similarity. For example, if the sensitivity σ is set to 20, when the Euclidean distance is 20, the similarity is 0.9048. When the sensitivity σ is 10 and the Euclidean distance is also 20, the similarity is 0.8187. d is the Euclidean distance between two pixels, and its calculation formula is:

[0050]

[0051] Where a, b, and c are the weight coefficients of the three channels, Ri>Gi>Dian are the three-channel pixel values ​​of the filling pixel, and R, G, and 3 are the three-channel pixel values ​​of the adjacent representative pixels.

[0052] The formula for calculating the cosine similarity of two pixels is:

[0053]

[0054] Specifically, after the optic disc and blood vessels are stripped, the method further includes: performing pixel filling in the optic disc and blood vessels stripping area, including: mapping pixel values ​​of an area adjacent to the outline of the area to be filled to a three-dimensional coordinate space; clustering all pixel values ​​based on the distribution of pixels in the area adjacent to the outline of the area to be filled in the three-dimensional coordinate space using the density of pixels in the three-dimensional space coordinates to obtain clustered pixel values; selecting filling pixels based on the clustered pixel values ​​according to the cluster center values; and performing pixel filling in the stripping area based on the selected filling pixels; wherein the calculation formula for the filling pixels is:

[0055]

[0056] Among them, M represents the total number of all classified pixels; m represents the total number of pixels in a class; P represents the filling pixels; and N represents the number of types.

[0057] Preferably, the clustering algorithm used to cluster all pixel values ​​using the density of pixels in three-dimensional space coordinates is an improved DBSCAN algorithm, and the clustering process includes: presetting an initial radius and a minimum number of samples in three-dimensional space coordinates; randomly selecting an unvisited pixel point, finding all pixel points that can be reached based on the preset initial radius based on the radius, and counting the number; comparing the statistical number with the minimum number of samples, if the statistical number is greater than the minimum number of samples, then the point is regarded as a core point and a new cluster is created; otherwise, the point is regarded as noise; and diffusion clustering is performed based on the determined core point until all points are traversed.

[0058] Specifically, the core idea of ​​the DBSCAN algorithm is to classify pixels in high-density areas into a class where pixels are close to each other, while pixels in low-density areas are isolated and treated as noise. Considering that there may be other pixels to be filled around the pixel to be filled, and these other pixels cannot be counted as surrounding pixels, the number of pixels surrounding the pixel to be filled is uncertain, and the number of clusters is also uncertain. A major advantage of DBSCAN is that the number of clusters does not need to be specified in advance and it can handle noise and outliers. Using DBSCAN to find cluster centers and ignore the influence of noise is very suitable.

[0059] The algorithm steps of DBSCAN are as follows: first set a radius eps and the minimum number of samples MinPts. Then randomly select an unvisited point, find all points that can be reached by the radius based on the radius eps, and if these points exceed MinPts, then consider the point as the core point and create a new cluster. If the number is less than MinPts, then mark it as noise. Finally, use the core point to diffuse until all points are traversed. The main purpose of filling is to fill the area to be filled with surrounding pixels so that it has a high similarity with the pixels in the surrounding background area. The pixels in the background area are evenly distributed and the similarity between pixels is high, while the area to be filled has many pixels of different shades. Filling is to restore the white area to pixels similar to the surrounding pixels. Use DBSCAN to cluster the two areas, set the DBSCAN radius to 30, and the minimum number of samples Pts to 5.

[0060] In one possible embodiment, the distribution of each pixel in the three-dimensional spatial coordinate system shows that the distribution of the background area is very regular, and the difference in the value range on the three coordinate axes of X, Y, and Z does not exceed 12. The value range of the red channel of the X-axis, i.e., the pixel is [190, 191, ..., 202], the value range of the green channel of the Y-axis, i.e., the pixel is [92, 93, ..., 102], and the value range of the blue channel of the Z-axis, i.e., the pixel is [24, 25, ..., 34]. This shows that the pixel density distribution of the background area is very close. Compared with the background area, the pixel distribution of the to-be-filled area is very different. The pixel distribution of the spatial coordinate system of the to-be-filled area shows that there is a lot of noise (noise is marked as black dots and represented by -1), and the difference between the noise and the classified pixels is very large. Therefore, in order to eliminate the influence of noise, this section uses the classified pixels to calculate the filling pixel values ​​when selecting the filling pixels, thereby eliminating the influence of noise on the selection of filling pixels. Of course, the clustering result shown in the figure only has one cluster, but in reality there may be two or more clusters. Overall, the idea of ​​density-based clustering in DBSCAN is similar to the idea of ​​pixel filling. DBSCAN has good noise handling capabilities and does not require manual setting of the number of clusters. Pixel filling also needs to eliminate the influence of noise and try to make the pixels as similar as possible to the majority of pixels.

[0061] Step S40: pre-processing the stripped image information, and after the pre-processing is completed, identifying the bleeding area based on the pre-processed image information to obtain an identification result.

[0062] Specifically, based on different preset preprocessing models, the stripped image information is preprocessed respectively to obtain the preprocessed image information under each preprocessing model; the preprocessed image information under the preprocessing model is compared, and the segmentation effect is compared based on the preset standard, and the preprocessed image information under the preprocessing model with the best segmentation effect is identified as the preprocessing result.

[0063] Furthermore, the different preprocessing models include: a single brightness balance model, a single contrast enhancement model, a single denoising model, a brightness balance and contrast enhancement combination model, a denoising and brightness balance combination model, a denoising and contrast enhancement combination model, and a denoising plus brightness balance plus contrast enhancement combination model.

[0064] 1) Regarding brightness balance, for retinal fundus images with large brightness differences and uneven brightness, the effect of deep learning retinal image segmentation is not satisfactory. In order to solve this problem, a brightness balance method based on local areas is used in the preprocessing stage to make the background of the entire retinal fundus image dataset consistent. The algorithm principle of this method is: the dataset has N images, the resolution of each image is X*Y, and the grayscale level is . Then the average brightness of the data set is:

[0065]

[0066] Where p(j, k) is the brightness value of the pixel at coordinates (j, k) in the image. This method uses the local area brightness equalization method to balance the brightness of the image. Set the local area size to m*n local area blocks to block a retinal fundus image. The average brightness of the sub-block is:

[0067]

[0068] It can be concluded that the scaling factor of the average value of the sub-block brightness and the brightness of the entire data set is:

[0069]

[0070] When alpha is less than 1, it indicates that the brightness of the local sub-block is higher than the average brightness of the dataset, and the local sub-block needs to be adjusted to reduce the brightness. When alpha is greater than 1, it indicates that the brightness of the local sub-block is lower than the average brightness of the dataset, and the local sub-block needs to be adjusted to increase the brightness. Through this method, the sub-block scaling factor of the entire image is calculated, and then the image brightness is adjusted according to the scaling factor. The smaller the size of the local sub-block, the better the image brightness adjustment effect, but the time complexity will also increase accordingly.

[0071] 2) As for contrast enhancement, there is a great similarity between the blood vessels and hemorrhage areas in the retinal fundus image and the retinal fundus background. In order to improve the visual effect, the details of the hemorrhage area are made clearer and more conspicuous, which is more conducive to the deep learning network to extract features. The contrast enhancement method is used to increase the obvious degree of the brightness difference between the hemorrhage area and the background in the image. The solution of the present invention adopts the contrast-limited adaptive histogram equalization (CLAHE) method to enhance the contrast of the retinal fundus image. This method divides the image into many small sub-images, performs histogram equalization on the sub-images, and then uses the contrast limitation method to prevent excessive enhancement of the image contrast, so that the image looks more natural. CLAHE not only does not sacrifice accuracy, but also minimizes latency.

[0072] 3) As for denoising, due to the influence of internal and external environments, there is more or less noise in the retinal fundus image, which affects the quality of the retinal fundus image and causes certain interference for the subsequent deep learning algorithm to identify and segment the hemorrhage area. Therefore, in order to minimize the impact of other factors on the experiment, it is necessary to reduce the noise in the image to improve the image quality and make the image clear and readable. In the scheme of the present invention, two methods, median filtering and Gaussian filtering, are used to process the retinal fundus image. The principle of median filtering is to select a pixel point and the pixel values ​​of the adjacent pixels around the pixel point, and then sort these pixel values, and finally use the median of the sequence as the pixel value of the current pixel. This method makes the surrounding pixel values ​​close to the true value to achieve the purpose of eliminating isolated noise points. It can be expressed by the formula:

[0073]

[0074] in represents the original image, g(x, y) represents the filtered image, and W is a two-dimensional template, i.e., the size of the neighborhood sliding window, which is usually 3*3, but can also be different shapes, such as circular or linear. The median filter process will slide all pixels through the neighborhood sliding window. Set the neighborhood sliding window size to 3*3, and sort all the pixel values ​​in its neighborhood to obtain the sequence [3, 34, 34, 36, 44, 46, 48, 85, 88]. Since 44 is the median of the sequence, the original value will be replaced by 44, and 44 will be used as the new pixel value of the center point, and the noise will be processed.

[0075] Median filtering effectively suppresses salt and pepper noise in images. The principle of Gaussian filtering is to scan each pixel through a Gaussian convolution kernel and replace the value of the central pixel with the weighted average of the pixels in the neighborhood determined by the convolution kernel. The characteristic of Gaussian filtering is that the weight of the center point of the Gaussian convolution kernel is greater than the weight of points farther away from the center. The Gaussian filtering formula is as follows:

[0076]

[0077] Among them, 90 represents the value of the (very)th pixel of the filtered image, represents the value of the 0th pixel of the original image, N is the filter radius, represents the Gaussian filter kernel, and the formula is as follows:

[0078]

[0079] Here, b is the standard deviation of the Gaussian filter, and m and n represent the offset of the filter kernel in the horizontal and vertical directions of the image, respectively. Gaussian filtering not only suppresses noise but also preserves image detail. However, as the standard deviation of the Gaussian filter increases, the filtering effect becomes stronger, and image detail becomes less noticeable. Therefore, choosing an appropriate standard deviation is key to implementing Gaussian filtering.

[0080] 4) Regarding data augmentation, due to the limited size of the dataset, to prevent the network from over-learning the characteristics of the training data during training, the solution of the present invention performs data augmentation on the dataset. By expanding the dataset, the data diversity is increased, thereby improving the generalization ability of the model, which may lead to poor performance on the test data. The data augmentation methods used mainly include rotation, flipping, and cropping.

[0081] Based on three image processing methods—brightness balance, contrast enhancement, and denoising—the present invention combines these three methods to design seven data preprocessing methods to identify the most suitable preprocessing method for segmenting hemorrhage areas in retinal fundus images. These methods include a single brightness balance model, a single contrast enhancement model, a single denoising model, a combined brightness balance and contrast enhancement model, a combined denoising and brightness balance model, a combined denoising and contrast enhancement model, and a combined denoising plus brightness balance and contrast enhancement model. Experiments were conducted on the seven preprocessed datasets to select the one with the best segmentation results for subsequent experiments.

[0082] Preferably, the method further includes training a bleeding recognition model, including: collecting a public data set of eye images from a public database, the public data set including: an IDRiD data set, a DRIVE data set, a STARE data set, and a CHASE-DB1 data set; collecting historical image data specified by a user; using the public data set and the historical image data as training sets, performing model training in a preset neural network, and obtaining a bleeding recognition model.

[0083] Figure 2 This is a system structure diagram of a retinal hemorrhage area identification system provided by one embodiment of the present invention. Figure 2As shown, an embodiment of the present invention provides a retinal hemorrhage area recognition system, which includes: an acquisition unit for acquiring image information of a user's eyes; a processing unit for performing optic disc and blood vessel recognition in the image information based on an optic disc recognition model and a blood vessel recognition model, respectively, to obtain optic disc and blood vessel recognition results; a stripping unit for performing optic disc and blood vessel stripping in the image information based on the optic disc and blood vessel recognition results, respectively, to obtain stripped image information; and an identification unit for preprocessing the stripped image information, and identifying the hemorrhage area based on the preprocessed image information to obtain an identification result.

[0084] An embodiment of the present invention further provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, enables the computer to execute the above-mentioned retinal hemorrhage area identification method.

[0085] Those skilled in the art will appreciate that all or part of the steps in the methods of the aforementioned embodiments can be accomplished by instructing the relevant hardware through a program. The program is stored in a storage medium and includes a number of instructions for causing a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a mobile hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0086] The above describes in detail the optional embodiments of the present invention in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above embodiments. Within the technical concept of the embodiments of the present invention, a variety of simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the scope of protection of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner unless there is any contradiction. In order to avoid unnecessary repetition, the embodiments of the present invention will no longer describe the various possible combinations separately.

[0087] In addition, the various embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. A method for identifying retinal hemorrhage areas, characterized in that: The method comprises: Collect image information of the user's eyes; Recognize the optic disc and the blood vessels in the image information based on the optic disc recognition model and the blood vessel recognition model, respectively, to obtain optic disc and blood vessel recognition results; Based on the optic disc and blood vessel identification results, the optic disc and blood vessel are stripped in the image information to obtain the stripped image information; wherein, The average similarity of the vascular and optic disc separation of the entire retinal fundus image is calculated by summing and averaging the similarities between the filling pixels of the vascular area and the optic disc area and their adjacent pixels; The similarity between a single filling pixel and its neighboring pixels is calculated using a weighted average method, that is, the average value, median value, and mode value of the neighboring pixels are used to calculate the similarity with the filling pixel, and then the weighted average is taken as the comprehensive similarity between the filling pixel and the neighboring pixels; The similarity between a single filled pixel and its neighboring pixels is also calculated using a weighted average method. The pixel value is used as a three-dimensional coordinate point, and the Euclidean distance between the two pixels is calculated. The straight-line distance between the two pixels is determined based on the Euclidean distance between the pixels to determine the similarity. On the other hand, the pixel point is regarded as a vector starting from the coordinate origin or a fixed point. The cosine value of the angle between the two pixel mapping vectors is calculated based on the cosine similarity to determine the similarity between the two pixels. After the optic disc and blood vessel stripping is completed, the method further includes: pixel filling of the optic disc and blood vessel stripping area, including: mapping the pixel values ​​of the contour adjacent area of ​​the area to be filled to a three-dimensional coordinate space; clustering all pixel values ​​based on the distribution of the pixels of the contour adjacent area of ​​the area to be filled in the three-dimensional coordinate space using the density of the pixels in the three-dimensional space coordinates to obtain clustered pixel values; taking values ​​of the filling pixels based on the clustered pixel values; and filling the stripping area pixels based on the taken values ​​of the filling pixels; wherein, The calculation formula for filling pixels is: Where M represents the total number of pixels classified; m represents the number of pixels in a cluster; P represents the fill pixel; N represents the number of types; All pixel values ​​are clustered using the density of pixels in three-dimensional space coordinates. The clustering algorithm used is the improved DBSCAN algorithm. The clustering process includes: presetting an initial radius and a minimum number of samples in three-dimensional space coordinates; randomly selecting an unvisited pixel point, finding all pixels that can be reached based on the initial radius based on the initial radius and counting the number of points; comparing the number of counted points with the minimum number of samples. If the number of counted points is greater than the minimum number of samples, the selected pixel point is used as the core point and a new cluster is created; otherwise, the selected pixel point is used as noise; and diffusion clustering is performed based on the determined core point until all pixels are traversed. The stripped image information is preprocessed, and the bleeding area is identified based on the preprocessed image information to obtain an identification result.

2. The method according to claim 1, characterized in that The optic disc recognition model is a UNet++ model; the blood vessel recognition model is a DC-UNet model.

3. The method according to claim 1, characterized in that The preprocessing of the stripped image information includes: Based on different preset preprocessing models, the peeled image information is preprocessed respectively to obtain preprocessed image information under each preprocessing model; The preprocessed image information under each preprocessing model is compared, the segmentation effect is compared based on the preset standard, and the preprocessed image information under the preprocessing model with the best segmentation effect is identified as the preprocessing result.

4. The method according to claim 3, characterized in that The different pre-processing models include: Single brightness balance model, single contrast enhancement model, single denoising model, brightness balance and contrast enhancement combined model, denoising and brightness balance combined model, denoising and contrast enhancement combined model, denoising plus brightness balance plus contrast enhancement combined model.

5. The method according to claim 1, wherein The identifying of the bleeding area based on the pre-processed image information includes: After the preprocessing is completed, the preprocessed image information is used as an input parameter to perform training in a pre-built bleeding recognition model to obtain training results; Determining whether the pre-processed image information has a bleeding area based on the training result; If present, perform bleeding area demarcation.

6. The method according to claim 1, characterized in that The method further includes: performing bleeding recognition model training, including: Collect a public dataset of eye images from a public database, where the public dataset includes: IDRiD dataset, DRIVE dataset, STARE dataset, CHASE-DB1 dataset; Collect historical image data specified by the user; The public data set and the historical image data are used as training sets, and model training is performed in a preset neural network to obtain a bleeding recognition model.

7. A retinal hemorrhage area identification system, characterized in that: The system is applied to the retinal hemorrhage area identification method according to any one of claims 1 to 6, and the system comprises: A collection unit, used to collect image information of the user's eyes; a processing unit, wherein the user performs optic disc and blood vessel recognition in the image information based on the optic disc recognition model and the blood vessel recognition model respectively, and obtains optic disc and blood vessel recognition results; A stripping unit, configured to strip the optic disc and blood vessels from the image information based on the optic disc and blood vessel identification results, respectively, to obtain image information after stripping; The recognition unit is used to pre-process the stripped image information, recognize the bleeding area based on the pre-processed image information, and obtain a recognition result.

8. A computer-readable storage medium having instructions stored thereon, which, when executed on a computer, enables the computer to execute the retinal hemorrhage area identification method according to any one of claims 1 to 6.

Images