Eye fundus image detection method and device for diabetic retinopathy and intelligent equipment

Abstract

The invention discloses an eye fundus image detection method and device for diabetic retinopathy and intelligent equipment. The method comprises the steps of performing image preprocessing on an eye fundus image; differentiated enhancement processing is carried out on the standardized eye fundus images according to the pathological features; performing candidate area detection in each task image, determining a lesion unit of each task image, constructing a lesion unit set of each task image, and performing fusion processing on the lesion units in each lesion unit set; extracting feature parameters of each lesion unit in the global lesion unit set; performing category judgment on each lesion unit in the global lesion unit set through a preset judgment rule based on the feature parameters, and determining a pathological feature tag corresponding to each lesion unit; and mapping the pathological feature label to a corresponding position in the standardized fundus image, and generating a detection result distribution diagram of the diabetic retinopathy. According to the embodiment of the invention, the detection sensitivity of vascular lesions and exudative lesions can be improved.

Description

Technical Field

[0001] This invention relates to the field of image detection technology, and in particular to a method, apparatus, and intelligent device for detecting fundus images of diabetic retinopathy. Background Technology

[0002] Diabetic retinopathy (DR) is a common complication of diabetes and a significant cause of blindness in adults. Its main pathological features include microaneurysms, retinal hemorrhages, hard exudates, and cotton wool spots. Early detection and monitoring of these pathological features are crucial for protecting visual function and controlling disease progression.

[0003] Currently, the main imaging techniques for clinically assessing diabetic retinopathy include fluorescein fundus angiography (FFA) and optical coherence tomography angiography (OCTA). While FFA, as the traditional gold standard, can reveal hemodynamic abnormalities such as vascular leakage, non-perfusion areas, and neovascularization, it is an invasive procedure and time-consuming, making it unsuitable for routine screening. OCTA, as a non-invasive method, can visualize the vascular structure of the fundus; however, because OCTA is based on blood flow signals, its sensitivity for detecting static exudative lesions caused by vascular leakage (such as hard exudates and cotton wool spots) is relatively low. Summary of the Invention

[0004] This application discloses a method, apparatus, and intelligent device for detecting fundus images of diabetic retinopathy, which can highlight the imaging manifestations of vascular lesions and exudative lesions and improve the detection sensitivity of vascular lesions and exudative lesions.

[0005] In a first aspect, embodiments of this application disclose a method for detecting fundus images of diabetic retinopathy, comprising: acquiring a color fundus image; performing image preprocessing on the color fundus image to obtain a standardized fundus image; performing differential enhancement processing on the standardized fundus image according to the pathological features of diabetic retinopathy to obtain multiple task images, the pathological features including vascular lesions and exudative lesions; performing candidate region detection in each task image to determine the lesion units of each task image; constructing a lesion unit set for each task image based on the lesion units; performing fusion processing on the lesion units in each lesion unit set to generate a global lesion unit set; extracting feature parameters of each lesion unit in the global lesion unit set, the feature parameters including at least one of morphological features, texture features, and color features; classifying each lesion unit in the global lesion unit set according to the feature parameters and a preset judgment rule to determine the pathological feature label corresponding to each lesion unit; mapping the pathological feature label to the corresponding position in the standardized fundus image to generate a detection result distribution map of diabetic retinopathy.

[0006] Secondly, this application discloses a fundus image detection device for diabetic retinopathy, comprising: an acquisition unit for acquiring a color fundus image and performing image preprocessing on the color fundus image to obtain a standardized fundus image; a task image processing unit for performing differential enhancement processing on the standardized fundus image according to the pathological features of diabetic retinopathy to obtain multiple task images, the pathological features including vascular lesions and exudative lesions; a fusion unit for performing candidate region detection in each task image to determine the lesion units in each task image, constructing a lesion unit set for each task image based on the lesion units, and performing fusion processing on the lesion units in each lesion unit set to generate a global lesion unit set; an extraction unit for extracting feature parameters of each lesion unit in the global lesion unit set, the feature parameters including at least one of morphological features, texture features, and color features; a determination unit for classifying each lesion unit in the global lesion unit set based on the feature parameters and using preset judgment rules to determine the pathological feature label corresponding to each lesion unit; and a generation unit for mapping the pathological feature labels to corresponding positions in the standardized fundus image to generate a detection result distribution map of diabetic retinopathy.

[0007] Thirdly, embodiments of this application disclose an intelligent device, including a processor and a memory, wherein the processor calls a computer program stored in the memory to implement the method disclosed in the first aspect above.

[0008] Fourthly, embodiments of this application disclose a computer-readable storage medium storing a computer program or computer instructions that, when executed by a processor, implement the method disclosed in the first aspect above.

[0009] Fifthly, embodiments of this application disclose a computer program product, which includes computer program code, such that when the computer program code is run by a processor, the above-described method is executed.

[0010] As described above, this application provides a method, apparatus, and intelligent device for detecting fundus images of diabetic retinopathy. This method uses color fundus images as input, eliminating the need for invasive examinations such as fluorescein angiography. It is suitable for routine follow-up and large-scale screening of diabetic patients. Based on the pathological characteristics of diabetic retinopathy, the standardized fundus images are differentially enhanced, highlighting the imaging manifestations of vascular and exudative lesions and improving their detection sensitivity. Secondly, by separately detecting candidate regions, a set of lesion units for each task image is constructed, dividing complex lesions into independent units for processing, avoiding the omission of small lesions and improving the comprehensiveness and sensitivity of the detection. Then, based on the feature parameters of each lesion unit in the global lesion unit set, category determination is performed, improving the objectivity of lesion identification. Finally, the determined pathological feature labels are mapped to corresponding positions in the standardized fundus images to generate a detection result distribution map, enhancing the intuitiveness and interpretability of the detection results. Attached Figure Description

[0011] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0012] Figure 1 This is a schematic flowchart of a fundus image detection method for diabetic retinopathy disclosed in an embodiment of this application; Figure 2 This is a schematic flowchart of another fundus image detection method for diabetic retinopathy disclosed in the embodiments of this application; Figure 3 This is a schematic diagram of the structure of a fundus image detection device for diabetic retinopathy disclosed in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of a smart device disclosed in an embodiment of this application. Detailed Implementation

[0013] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings.

[0014] This application discloses a method, apparatus, and intelligent device for detecting fundus images of diabetic retinopathy, which can highlight the imaging manifestations of vascular lesions and exudative lesions, and improve the detection sensitivity of vascular lesions and exudative lesions. These will be described in detail below.

[0015] To better understand the embodiments of this application, the relevant technologies are described below.

[0016] Diabetic retinopathy (DR) is a common complication of diabetes and a significant cause of blindness in adults. Its main pathological features include microaneurysms, retinal hemorrhages, hard exudates, and cotton wool spots. Early detection and monitoring of these pathological features are crucial for protecting visual function and controlling disease progression.

[0017] Currently, the main imaging techniques for clinically assessing diabetic retinopathy include fluorescein fundus angiography (FFA) and optical coherence tomography angiography (OCTA).

[0018] In addition to the aforementioned imaging techniques, recent studies have also attempted to use large-scale neural network models based on deep learning to automatically analyze fundus images in order to identify features related to diabetic retinopathy. These methods typically require a large amount of labeled data for training, resulting in a massive number of model parameters and high computational resource consumption during training and inference. Furthermore, fundus images vary significantly depending on the device and the individual, limiting the generalization ability of deep learning models and making them prone to misclassification or missed diagnoses.

[0019] To overcome the aforementioned technical challenges, in this embodiment, fundus images are acquired, and image preprocessing is performed on the fundus images to obtain standardized fundus images. Differential enhancement processing is then applied to the standardized fundus images based on the pathological characteristics of diabetic retinopathy to obtain multiple task images. The pathological characteristics include vascular lesions and exudative lesions. Candidate region detection is performed in each task image to determine the lesion units in each task image. A lesion unit set for each task image is constructed based on the lesion units. The lesion units in each lesion unit set are fused to generate a global lesion unit set. Feature parameters of each lesion unit in the global lesion unit set are extracted. The feature parameters include at least one of morphological features, texture features, and color features. Based on the feature parameters, a preset judgment rule is used to classify each lesion unit in the global lesion unit set, determining the pathological feature label corresponding to each lesion unit. The pathological feature labels are mapped to the corresponding positions in the standardized fundus images to generate a detection result distribution map of diabetic retinopathy. As can be seen, this application, by preprocessing fundus images, ensures the consistency, stability, and reliability of subsequent processing. It is also applicable to fundus images from various sources, enhancing the generalization ability of this application and enabling large-scale screening. Differential enhancement processing is applied to the standardized fundus images obtained after preprocessing according to different pathological features, resulting in multiple task images. Each task image is optimized for detecting specific pathological features, improving the sensitivity and specificity of subsequent detection. Candidate region detection is performed separately to determine the lesion units in each task image, maximizing the discovery of possible pathological features. The lesion units in each lesion unit set are fused to generate a global lesion unit set, resolving the issue of the same lesion appearing in different task images. To address the issue of repeated detections, multiple detection results are merged into a global lesion unit set. Based on feature parameters and using preset judgment rules, each lesion unit in the global lesion unit set is classified, and the pathological feature label corresponding to each lesion unit is determined. According to the preset judgment rules, the extracted feature parameters are used to determine the pathological feature label of each lesion unit, which can avoid subjective bias and fatigue error caused by manual interpretation and ensure the consistency and repeatability of detection results. The pathological feature labels are mapped to the corresponding positions in standardized fundus images to generate a detection result distribution map of diabetic retinopathy. Finally, the determined pathological feature labels are mapped to the corresponding positions in standardized fundus images to generate a detection result distribution map, which can enhance the intuitiveness and interpretability of detection results.

[0020] Please see Figure 1 , Figure 1 This is a schematic flowchart of a fundus image detection method for diabetic retinopathy disclosed in an embodiment of this application. Figure 1 As shown, the fundus image detection method for diabetic retinopathy may include the following steps.

[0021] 101. Obtain fundus images, perform image preprocessing on the fundus images to obtain standardized fundus images.

[0022] Fundus images can be acquired using a fundus camera, a smartphone fundus accessory, or other ocular imaging devices. Fundus images can be color digital images, with common color spaces including RGB, HSV, or Lab mode. Image resolution can be selected according to accuracy requirements, such as no less than 2896×1944 pixels, to ensure that subsequent processing can capture details of tiny lesions such as microaneurysms.

[0023] Image preprocessing of fundus images may include: uniformly scaling or cropping the fundus images to a fixed size (e.g., 1024×1024 pixels) to ensure consistency of input for subsequent algorithms. During cropping, registration is performed using the optic disc or macula center as a reference to eliminate translation and rotation differences caused by different shooting angles. It should be noted that an initial pixel coordinate system can be defined with the upper left corner of the fundus image as the origin (0,0), extending the X-axis to the right and the Y-axis downwards.

[0024] Furthermore, to reduce color deviations caused by different devices and shooting conditions, color transfer or correction methods based on standard color charts can be used. For example, the image can be converted to the Lab color space, and the mean and standard deviation of its a and b channels can be adjusted to the range of a preset standard reference image to ensure consistent color representation.

[0025] 102. Based on the pathological characteristics of diabetic retinopathy, standardized fundus images were subjected to differential enhancement processing to obtain multiple task images. The pathological characteristics included vascular lesions and exudative lesions.

[0026] Vascular lesions include microaneurysms and hemorrhages, which appear as dark spots or punctate structures with a reddish hue. Exudative lesions include hard exudates and cotton wool spots, which appear as bright spots with specific colors (such as the yellow of hard exudates) and textures (such as the fibrous texture of cotton wool spots).

[0027] Step 102 may include the following steps.

[0028] 1021: Extract the green channel component from the standardized fundus image to obtain the first intermediate image; process the first intermediate image through morphological opening operation to obtain the second intermediate image; use a multi-scale matched filter to perform speckle enhancement processing on the second intermediate image to generate the microaneurysm detection task image.

[0029] Since microaneurysms exhibit the highest contrast with the background in the green spectral band, extracting the green channel component of a standardized fundus image can initially highlight the lesion. Morphological opening operations are then performed on the first intermediate image using disk-shaped structuring elements (e.g., with a radius of 3-5 pixels). This process suppresses elongated blood vessel structures and noise points in the image while preserving approximately circular microaneurysm candidate points, resulting in a second intermediate image. A multi-scale matched filter is then used to perform speckle enhancement on the second intermediate image. The multi-scale matched filter can be a multi-scale Laplacian of Gaussian (LoG) filter bank. The scale parameters can be set according to the common size range of microaneurysms (e.g., diameter 2-10 pixels). After taking the absolute value of the filtered response image at each scale, the maximum response value of each pixel across all scales is taken to generate the microaneurysm detection task image.

[0030] 1022: Extract the saturation component in the HSV color space and the red channel component in the RGB color space of the standardized fundus image, and fuse the saturation component and the red channel component to obtain the third intermediate image; use morphological bottom cap operation to perform dark area enhancement processing on the third intermediate image to generate the hemorrhage detection task image.

[0031] Bleeding points typically possess both high saturation and high red component color features. These two components can be normalized to the [0, 1] interval, and then their pixel-level product can be calculated to obtain a third intermediate image. This ensures that only pixels with both high saturation and high red value have higher values ​​in the third intermediate image, thus specifically enhancing the bleeding points. Subsequently, a morphological bottom-hat operation is used to enhance the dark areas of the third intermediate image. The bottom-hat operation is defined as the difference between the original image and the result of the morphological closing operation. The bottom-hat operation effectively extracts areas darker than the surrounding background, thus highlighting bleeding point candidate regions and generating images for the bleeding point detection task.

[0032] It should be noted that the closing operation can use a disk-shaped structuring element (radius = 15 pixels) to fill in all dark details in the image smaller than this structuring element and fit the local background brightness of the image. The dark region is the pixel region in the third intermediate image whose gray value is lower than the local background gray value fitted by the morphological closing operation.

[0033] 1023: Extract the blue channel component of the standardized fundus image in the RGB color space to obtain the fourth intermediate image; process the fourth intermediate image using Contrast Limited Adaptive Histogram Equalization (CLAHE) to obtain the fifth intermediate image; perform bright area enhancement processing on the fifth intermediate image using morphological top-hat operation to generate the hard exudation detection task image.

[0034] The hard exudate is bright yellow, appearing as a distinct dark area in the blue channel, contrasting sharply with the background.

[0035] The fourth intermediate image is processed using CLAHE, which significantly improves the local contrast and makes the hard exudation areas more prominent, resulting in the fifth intermediate image. Then, a morphological top-hat operation is used to enhance the bright areas of the fifth intermediate image. The top-hat operation is defined as the difference between the original image and the result of the morphological opening operation. The top-hat operation effectively extracts small areas in the image that are brighter than the surrounding background, thereby further enhancing the hard exudation candidate points and generating the initial task image for hard exudation.

[0036] Furthermore, after generating the hard exudation detection task image, the boundary sharpness of each candidate region in the hard exudation detection task image can be calculated. Candidate regions are then filtered based on boundary sharpness, removing artifact regions with blurred boundaries. Boundary sharpness can be quantified by calculating the average gradient magnitude at the same location in the original normalized image corresponding to that region. Specifically, the Sobel operator can be used to calculate the image gradient, and the average gradient magnitude of all pixels within the candidate region can be calculated. Based on boundary sharpness, candidate regions are filtered by setting a gradient threshold (e.g., the average gradient magnitude must be greater than a preset value T1), removing artifact regions with blurred boundaries (such as uniform bright spot noise), and retaining hard exudation candidates with clear boundaries to generate the final hard exudation detection task image.

[0037] It should be noted that the opening operation can use a disk-shaped structuring element (radius = 3 pixels) to eliminate all bright details in the image smaller than this structuring element and fit the local background brightness of the image. The bright areas are the pixel regions in the fifth intermediate image whose gray values ​​are higher than the local background gray values ​​fitted by the morphological opening operation.

[0038] 1024: Extract the green channel component of the standardized fundus image in the RGB color space to obtain the sixth intermediate image; process the sixth intermediate image with Contrast Limited Adaptive Histogram Equalization (CLAHE) to obtain the seventh intermediate image; perform weak edge smoothing and region enhancement processing on the seventh intermediate image using morphological closing operation to generate the cotton lint detection task image.

[0039] In the green channel, the contrast between the cotton-like texture and the background is quite obvious in the cotton-like spots. The sixth intermediate image was processed using CLAHE to obtain the seventh intermediate image. Morphological closing operations were then applied to the seventh intermediate image for weak edge smoothing and region enhancement. This process can fuse the weak edges adjacent to the cotton-like spots, fill in their small gaps, and smooth their contours, while suppressing dark noise smaller than the structuring elements. This results in an enhanced cotton-like spot image with better connectivity and a more complete region, generating the cotton-like spot detection task map.

[0040] Furthermore, after generating the enhanced image of cotton lint spots, the mean value of the boundary gradient magnitude or the local entropy value of each candidate region in the enhanced image of cotton lint spots can be calculated; the candidate regions are then filtered based on the mean value of the boundary gradient magnitude or the local entropy value to remove artifact regions with sharp boundaries or coarse internal textures, so as to generate the final cotton lint spot detection task image.

[0041] 103. Perform candidate region detection in each task image to determine the lesion units in each task image. Construct a lesion unit set for each task image based on the lesion units. Perform fusion processing on the lesion units in each lesion unit set to generate a global lesion unit set.

[0042] In this embodiment, candidate region detection is performed in each task image to determine the lesion units in each task image. This can include: performing binarization processing on the task image using an adaptive threshold segmentation algorithm to obtain a binary mask of the candidate region. Specifically, for each task image, Otsu's method or an adaptive thresholding algorithm (such as a Gaussian weighted average based on local neighborhoods) is used for image binarization to separate the enhanced foreground (pathological features) from the background, resulting in a binary mask. Connectivity analysis is performed on the binary mask to determine each independent connected region as a lesion unit, and the position and contour information of each lesion unit are recorded. The position information is the coordinates of the minimum bounding rectangle of the lesion unit; the contour information is the sequence of pixel coordinates constituting the boundary of the lesion unit.

[0043] In this embodiment, constructing a set of lesion units for each task image based on lesion units, and performing fusion processing on the lesion units in each lesion unit set to generate a global lesion unit set may include: calculating the spatial overlap between lesion units in each lesion unit set; if the spatial overlap is greater than a preset overlap threshold, then merging the lesion units between different lesion unit sets into a global lesion unit; and constructing a global lesion unit set based on the global lesion unit and the lesion units that have not been merged.

[0044] 104. Extract the feature parameters of each lesion unit in the global lesion unit set. The feature parameters include at least one of morphological features, texture features, and color features.

[0045] Extracting the morphological features of each lesion unit in the global lesion unit set can include: calculating the geometric morphological feature parameters of the lesion unit based on the contour information of each lesion unit recorded by 103. The geometric morphological feature parameters can include area, perimeter, roundness, aspect ratio and convexity.

[0046] Specifically, the total number of all internal pixels constituting the outline of the lesion unit can be counted to determine the area of ​​the lesion unit; the sum of the Euclidean distances between all adjacent pixels on the outline of the lesion unit can be calculated to determine the perimeter of the lesion unit; the ratio of the width to the height of the smallest bounding rectangle of the lesion unit can be calculated to determine the aspect ratio of the lesion unit; and the ratio of the area of ​​the lesion unit to the area of ​​its convex hull can be calculated to determine the convexity of the lesion unit. The smaller this value is, the greater the degree of concavity of the outline and the more irregular the shape.

[0047] Extracting the texture features of each lesion unit in the global lesion unit set can include: calculating the texture feature parameters of each lesion unit in the corresponding region on the green channel of the standardized fundus image. The texture feature parameters can include gray-level co-occurrence matrix features.

[0048] Specifically, the gray-level co-occurrence matrix (GLCM) can be calculated, and the following parameters can be extracted based on the GLCM: contrast, which measures the sharpness of the image and the depth of texture grooves. The larger the value, the deeper the texture; energy, which reflects the uniformity and regularity of the image texture. The larger the value, the more uniform the texture; homogeneity, which measures the local consistency of the image texture; and correlation, which represents the consistency of the image texture direction.

[0049] Extracting the color features of each lesion unit in the global lesion unit set can include: calculating the color feature parameters of each lesion unit in the region corresponding to each lesion unit in the original RGB or Lab color space of the standardized fundus image, including color statistics, similarity to standard colors, and color distribution.

[0050] Specifically, the average value and standard deviation of all pixels within the lesion unit across each color channel (e.g., R, G, B, L, a, b) can be calculated to obtain color statistics. The Euclidean distance between the average color of the region and a predefined standard lesion color template in the Lab color space can be calculated, and the similarity to the standard color can be determined using the Euclidean distance. The color histogram or color moments (e.g., first-order and second-order color moments) of the region can be calculated to obtain the color distribution.

[0051] 105. Based on feature parameters, the category of each lesion unit in the global lesion unit set is determined by a preset judgment rule, and the pathological feature label corresponding to each lesion unit is determined.

[0052] For example, a preliminary judgment can be made based on morphological and color features.

[0053] Screening microaneurysm candidates: If a lesion unit has a roundness greater than 0.85, an area between [5, 50] pixels, and a color difference of less than 15 from standard red, it is marked as a microaneurysm candidate.

[0054] Screening for bleeding point candidates: If a lesion unit has a roundness of less than 0.7, an area between [50, 500] pixels, and a color difference of less than 12 from standard red, it is marked as a bleeding point candidate.

[0055] Screening for hard exudation candidates: If the average brightness of a lesion unit is greater than the preset brightness threshold and the color difference with standard yellow is less than 10, it is marked as a hard exudation candidate.

[0056] Screening for cotton wool spot candidates: If the homogeneity of a lesion unit is greater than 0.7, the contrast is less than the preset contrast threshold, and the color difference with standard white is less than 8, then it is marked as a cotton wool spot candidate.

[0057] For candidates that pass the initial screening, texture rules can be further applied for confirmation and differentiation.

[0058] Microaneurysm identification: Microaneurysm candidates have high energy values ​​(uniform texture) and moderate contrast values.

[0059] Confirming hard exudates: Hard exudate candidates have a boundary sharpness higher than the threshold and a low internal color standard deviation (uniform color).

[0060] Confirm cotton lint spots: The local entropy of the cotton lint spot candidate is lower than the threshold (uniform internal texture), and its average gradient magnitude is low (blurred boundary).

[0061] If a lesion does not meet all of the above category characteristics, or if its characteristic values ​​fall outside all typical ranges, it can be identified as an artifact and assigned a corresponding label.

[0062] 106. Map the pathological feature labels to the corresponding locations in the standardized fundus image to generate a distribution map of the detection results for diabetic retinopathy.

[0063] Please refer to the figure. Figure 2 This is a schematic flowchart of another fundus image detection method for diabetic retinopathy disclosed in an embodiment of this application. Figure 2 As shown, this application may also include the following steps.

[0064] 201. Obtain the coordinates of the optic disc center and the fovea of ​​the macula from the standardized fundus image, and establish a distribution reference coordinate system based on the coordinates of the optic disc center and the fovea of ​​the macula.

[0065] Points in the initial pixel coordinate system are uniformly mapped to a distributed reference coordinate system with the fovea of ​​the macula as the origin and the line connecting the optic disc and the macula as the reference axis. This satisfies the requirement that the location of any lesion can be accurately described by its absolute distance and direction relative to the fovea of ​​the macula, thereby enabling quantitative spatial analysis between images taken from different patients at different time points.

[0066] 202. Calculate the Euclidean distance between the centroid coordinates and the foveal coordinates of each lesion unit in the global lesion unit set in the distributed reference coordinate system, and determine the lesion region to which each lesion unit belongs based on the Euclidean distance.

[0067] According to internationally accepted classification standards for diabetic retinopathy (such as the ETDRS standard), a series of concentric circle regions are constructed with the fovea as the center, and each lesion unit is divided into the corresponding region based on the Euclidean distance between the centroid coordinates of each lesion unit and the fovea coordinates.

[0068] Central region: A circular region with radius R1 centered on the fovea. Inner ring region: A ring-shaped region with inner radius R1 and outer radius R2 centered on the fovea. Outer ring region: A ring-shaped region with inner radius R2 and outer radius R3 centered on the fovea. For a given lesion unit, its lesion region is determined by the distance D between its centroid and the fovea. If the Euclidean distance is less than or equal to R1, the lesion is located in the central region. If the Euclidean distance is less than or equal to R2 and greater than R1, the lesion is located in the inner ring region. If the Euclidean distance is less than or equal to R3 and greater than R2, the lesion is located in the outer ring region.

[0069] 203. Output statistical results based on the lesion area and pathological feature labels. The statistical results may include the feature parameters of the lesion units corresponding to different pathological features within each lesion area. As shown in Table 1. Table 1 is an example of lesion area statistics.

[0070]

[0071] exist Figure 1 The described fundus image detection method for diabetic retinopathy uses color fundus images as input, eliminating the need for invasive examinations such as fluorescein angiography. It is suitable for routine follow-up and large-scale screening of diabetic patients. Differential enhancement processing is applied to standardized fundus images based on the pathological characteristics of diabetic retinopathy, highlighting the imaging manifestations of vascular and exudative lesions and improving their detection sensitivity. Secondly, by separately detecting candidate regions, a lesion unit set is constructed for each task image, dividing complex lesions into independent units for processing, avoiding the omission of small lesions and improving the comprehensiveness and sensitivity of the detection. Then, category determination is performed based on the feature parameters of each lesion unit in the global lesion unit set, improving the objectivity of lesion identification. Finally, the determined pathological feature labels are mapped to corresponding positions in the standardized fundus image to generate a detection result distribution map, enhancing the intuitiveness and interpretability of the detection results.

[0072] It should be understood that the same or corresponding information in the different embodiments described above can be referenced in relation to each other.

[0073] Please see Figure 3 , Figure 3 This is a schematic diagram of the structure of a fundus image detection device for diabetic retinopathy disclosed in an embodiment of this application. Figure 3 As shown, the fundus image detection device for diabetic retinopathy may include an acquisition unit 301, a task image processing unit 302, a fusion unit 303, an extraction unit 304, a determination unit 305, and a generation unit 306.

[0074] The acquisition unit 301 is used to acquire a fundus color image, perform image preprocessing on the fundus color image, and obtain a standardized fundus image. The task image processing unit 302 is used to perform differential enhancement processing on standardized fundus images according to the pathological features of diabetic retinopathy to obtain multiple task images. The pathological features include vascular lesions and exudative lesions. The fusion unit 303 is used to perform candidate region detection in each task image, determine the lesion units in each task image, construct a lesion unit set for each task image based on the lesion units, perform fusion processing on the lesion units in each lesion unit set, and generate a global lesion unit set. Extraction unit 304 is used to extract feature parameters of each lesion unit in the global lesion unit set. The feature parameters include at least one of morphological features, texture features and color features. The determination unit 305 is used to classify each lesion unit in the global lesion unit set based on feature parameters and a preset determination rule, and to determine the pathological feature label corresponding to each lesion unit. The generation unit 306 is used to map pathological feature labels to corresponding positions in a standardized fundus image to generate a distribution map of detection results for diabetic retinopathy.

[0075] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the above-described fundus image detection device for diabetic retinopathy, acquisition unit 301, task image processing unit 302, fusion unit 303, extraction unit 304, determination unit 305, and generation unit 306 can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0076] This application proposes a detection device that can automatically identify and quantify the core pathological features of diabetic retinopathy based on conventional color fundus images. It can accurately extract and analyze the number, area, spatial distribution, and morphological texture features of lesions such as microaneurysms, hemorrhages, hard exudates, and cotton wool spots in multiple dimensions, which can improve screening efficiency and diagnostic consistency. At the same time, it can provide reliable quantifiable data support for remote diagnosis and treatment via the Internet, and reduce the risk of vision impairment caused by missed or misdiagnosed early lesions.

[0077] In the several embodiments provided in this application, the coupling between the units can be electrical, mechanical or other forms of coupling.

[0078] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0079] Please refer to Figure 4 , Figure 4 This is a schematic diagram of the structure of a smart device disclosed in an embodiment of this application. Figure 4 As shown, the smart device may include a processor 401 and a memory 402. The memory 402 may store one or more computer programs. The one or more computer programs are configured to perform the methods described in the foregoing method embodiments. The memory 402 may be independent or integrated with the processor 401.

[0080] Processor 401 may include one or more processing cores. Processor 401 can connect to various parts of the terminal device using various interfaces and lines. It can perform various functions and process data of the terminal device by running or executing instructions, programs, code sets, or instruction sets stored in memory 402, and by calling data stored in memory 402. Optionally, processor 401 may be implemented using at least one hardware form of digital signal processing (DSP), field-programmable gate array (FPGA), or programmable logic array (PLA). Processor 401 may integrate one or more of the following: central processing unit (CPU), graphics processing unit (GPU), and modem. The CPU mainly handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem is used for wireless communication. It is understood that the modem may also not be integrated into processor 401 and may be implemented separately through a communication chip.

[0081] The memory 402 may include random access memory (RAM) or read-only memory (ROM). The memory 402 can be used to store instructions, programs, code, code sets, or instruction sets. The memory 402 may include a program storage area and a data storage area. The program storage area may store instructions for implementing an operating system, instructions for implementing at least one function (such as touch functionality, sound playback functionality, image playback functionality, etc.), and instructions for implementing the various method embodiments described above. The data storage area may also store data created by the terminal device during use (such as phonebook data, audio and video data, chat log data, etc.).

[0082] When the computer program instructions stored in memory 402 are executed, the processor 401 can be used to perform various operations performed by the terminal device in the above method embodiments. Specific implementations of these operations can be found in the preceding embodiments and will not be repeated here.

[0083] This application also discloses a computer-readable storage medium storing computer program code, which can be called by a processor to execute various operations in the above method embodiments. Specific implementations of each of the above operations can be found in the preceding embodiments and will not be repeated here.

[0084] Computer-readable storage media can be electronic storage devices such as flash memory, electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), hard disk, or ROM. Optionally, computer-readable storage media can include non-transitory computer-readable storage medium. The computer-readable storage medium has storage space for program code that performs any of the method steps described above. This computer program code can be read from or written to one or more computer program products. The computer program code can be compressed, for example, in a suitable form.

[0085] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A method for detecting fundus images in diabetic retinopathy, characterized in that, include: Acquire fundus images, perform image preprocessing on the fundus images to obtain standardized fundus images; Based on the pathological characteristics of diabetic retinopathy, the standardized fundus images are subjected to differential enhancement processing to obtain multiple task images. The pathological characteristics include vascular lesions and exudative lesions. Candidate region detection is performed in each of the task images to determine the lesion units in each task image. A lesion unit set for each task image is constructed based on the lesion units. The lesion units in each lesion unit set are fused to generate a global lesion unit set. Extract the feature parameters of each lesion unit in the global lesion unit set, wherein the feature parameters include at least one of morphological features, texture features and color features; Based on the feature parameters, the category of each lesion unit in the global lesion unit set is determined by a preset judgment rule, and the pathological feature label corresponding to each lesion unit is determined. The pathological feature labels are mapped to the corresponding positions in the standardized fundus image to generate a distribution map of the detection results for diabetic retinopathy.

2. The fundus image detection method as described in claim 1, characterized in that, The method further includes: Obtain the optic disc center coordinates and macular fovea coordinates of the standardized fundus image, and establish a distribution reference coordinate system based on the optic disc center coordinates and the macular fovea coordinates; In the distributed reference coordinate system, calculate the Euclidean distance between the centroid coordinates of each lesion unit in the global lesion unit set and the foveal coordinates of the macula, and determine the lesion region to which each lesion unit belongs based on the Euclidean distance; Statistical results are output based on the lesion area and the pathological feature label. The statistical results include the feature parameters of the lesion units corresponding to different pathological features in each lesion area.

3. The fundus image detection method as described in claim 1, characterized in that, The vascular lesions include microaneurysms. The standardized fundus images are differentially enhanced based on the pathological characteristics of diabetic retinopathy to obtain multiple task images, including: Extract the green channel component from the standardized fundus image to obtain the first intermediate image; The first intermediate image is processed by morphological opening operation to obtain the second intermediate image; A multi-scale matched filter is used to perform speckle enhancement processing on the second intermediate image to generate a microaneurysm detection task image.

4. The fundus image detection method as described in claim 1, characterized in that, The vascular lesions include hemorrhage points. The standardized fundus images are differentially enhanced based on the pathological characteristics of diabetic retinopathy to obtain multiple task images, including: The saturation component in the HSV color space and the red channel component in the RGB color space of the standardized fundus image are extracted, and the saturation component and the red channel component are fused to obtain a third intermediate image. Morphological bottom-hat operations are used to enhance the dark areas of the third intermediate image to generate a bleed point detection task image.

5. The fundus image detection method as described in claim 1, characterized in that, The exudative lesions include hard exudates. The standardized fundus images are differentially enhanced based on the pathological characteristics of diabetic retinopathy to obtain multiple task images, including: Extract the blue channel component of the standardized fundus image in the RGB color space to obtain the fourth intermediate image; The fourth intermediate image is processed using contrast-limited adaptive histogram equalization to obtain the fifth intermediate image; The bright areas of the fifth intermediate image are enhanced using morphological top-hat operations to generate a hard exudation detection task image.

6. The fundus image detection method as described in claim 1, characterized in that, The exudative lesions include cotton wool spots. The standardized fundus images are differentially enhanced based on the pathological characteristics of diabetic retinopathy to obtain multiple task images, including: The green channel component of the standardized fundus image in the RGB color space is extracted to obtain the sixth intermediate image; The sixth intermediate image is processed using contrast-limited adaptive histogram equalization to obtain the seventh intermediate image; Morphological closing operations are used to perform weak edge smoothing and region enhancement on the seventh intermediate image to generate a cotton lint detection task image.

7. The fundus image detection method as described in claim 1, characterized in that, The step of performing candidate region detection in each of the task images to determine the lesion units in each task image includes: An adaptive threshold segmentation algorithm is used to binarize the task image to obtain a binary mask of the candidate region. Connectivity analysis is performed on the binary mask to determine each independent connected region as a lesion unit, and the position and contour information of each lesion unit are recorded.

8. The fundus image detection method as described in claim 1, characterized in that, The step of constructing a lesion unit set for each task image based on the lesion units, and performing fusion processing on the lesion units in each lesion unit set to generate a global lesion unit set includes: Calculate the spatial overlap between lesion units in each of the lesion unit sets; If the spatial overlap is greater than a preset overlap threshold, then the lesion units between different sets of lesion units are merged into a global lesion unit. The global lesion unit set is composed of the global lesion units and the lesion units that have not merged.

9. A fundus image detection device for diabetic retinopathy, characterized in that, include: The acquisition unit is used to acquire a fundus color image and perform image preprocessing on the fundus color image to obtain a standardized fundus image. The task image processing unit is used to perform differential enhancement processing on the standardized fundus image according to the pathological features of diabetic retinopathy to obtain multiple task images, wherein the pathological features include vascular lesions and exudative lesions. The fusion unit is used to perform candidate region detection in each of the task images, determine the lesion units in each of the task images, construct a lesion unit set for each of the task images based on the lesion units, perform fusion processing on the lesion units in each of the lesion unit sets, and generate a global lesion unit set. An extraction unit is used to extract feature parameters of each lesion unit in the global lesion unit set, wherein the feature parameters include at least one of morphological features, texture features, and color features; The determination unit is used to classify each lesion unit in the global lesion unit set based on the feature parameters and a preset determination rule, and to determine the pathological feature label corresponding to each lesion unit. The generation unit is used to map the pathological feature labels to the corresponding positions in the standardized fundus image to generate a distribution map of the detection results of diabetic retinopathy.

10. A smart device, characterized in that, It includes a processor and a memory, wherein the processor invokes a computer program stored in the memory to implement the method as described in any one of claims 1-8.

Images