Retinal detachment area identification system based on wide-angle fundus photograph images

Abstract

The present application relates to the field of image processing, and is used to solve the problem that traditional methods are difficult to combine various manifestations of retinal detachment, resulting in single evaluation means and misjudgment in auxiliary identification of pathological images, and specifically relates to a retinal detachment area identification system based on wide-angle fundus photograph images, which performs imaging parameter correction, artifact removal and hierarchical brightness compensation based on blood vessel areas on wide-angle fundus images to solve the image quality degradation problem caused by edge distortion, light attenuation and artifact interference, on the basis of which, the planar boundary features of the detachment area, the blood vessel curvature change features and the pseudo 3D protrusion height features are synchronously extracted, the planar range of the detachment area, the blood vessel deformation and the three-dimensional height are comprehensively evaluated, finally, the three features are mapped and analyzed in terms of coincidence degree, and multi-gear risk quantitative evaluation and macular area approximation early warning are performed, thereby providing reliable auxiliary basis for clinical diagnosis.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to a retinal detachment region identification system based on wide-angle fundus photography images. Background Technology

[0002] Retinal detachment is a common blinding emergency in ophthalmology. Its pathological feature is the separation of the neuroepithelial layer and the pigment epithelium of the retina, which leads to ischemia and hypoxia of photoreceptor cells. If not intervened in time, it can cause irreversible visual impairment. Early and accurate identification of the extent, boundaries and progression risk of the detachment area is of decisive significance for clinical surgical planning and prognosis assessment.

[0003] In recent years, wide-angle fundus photography technology has made great strides. Ultra-wide-angle imaging devices can acquire digital images covering a large area of ​​the retina in one go without pupil dilation or with small pupils. At the same time, deep learning image segmentation technology is increasingly widely used in the field of ophthalmic imaging. Some studies have attempted to use semantic segmentation networks to automatically label lesion areas in fundus images, initially achieving pixel-level classification without regard to the region.

[0004] However, traditional techniques still have significant shortcomings in clinical applications. Conventional methods mostly treat retinal detachment as a routine segmentation task, relying solely on two-dimensional image features of a single modality for prediction. Due to the inherent characteristics of wide-angle fundus photography images, such as severe edge distortion, significant peripheral illumination attenuation, and prominent eyelid and eyelash artifact interference, the segmentation accuracy of such methods at the detachment boundary drops sharply, and the false positive rate remains high. Meanwhile, conventional methods also neglect the essential pathological manifestations of retinal detachment in imaging, namely the changes in the course of retinal vessels and the three-dimensional morphological changes caused by retinal bulging in the detachment area. Relying solely on two-dimensional grayscale information makes it difficult to distinguish true retinal detachment from other low-reflection lesions, resulting in incorrect identification results when identifying retinal detachment. This fails to provide effective assistance for treatment and may even interfere with treatment decisions. Summary of the Invention

[0005] This invention addresses image quality degradation caused by edge distortion, illumination attenuation, and artifact interference by using a distortion correction module to correct imaging parameters, remove artifacts, and perform graded brightness compensation based on vascular regions in wide-angle fundus images. Building upon this, a multi-dimensional feature extraction module simultaneously extracts planar boundary features of the detachment area, vascular curvature change features, and pseudo-3D elevation height features estimated based on shadow and texture gradients. This achieves a joint characterization of the three pathological manifestations of the detachment area's planar extent, vascular deformation, and three-dimensional height, overcoming the limitations of single two-dimensional grayscale segmentation. Finally, coordinate mapping and overlap analysis are performed on the three features to construct a high-confidence predicted detachment range from the intersection area. Multi-level risk quantification assessment and macular region approach warning are also conducted, providing reliable auxiliary evidence for clinical diagnosis. This addresses the aforementioned technical deficiencies and proposes a retinal detachment area identification system based on wide-angle fundus photography.

[0006] The objective of this invention can be achieved through the following technical solution: a retinal detachment region identification system based on wide-angle fundus photography images, including an image input module, a distortion correction module, a multi-dimensional feature extraction module, a fusion positioning module, and a report output module; The image input module is used to acquire fundus images generated by the wide-angle fundus photography method and simultaneously record the image capture time; The distortion correction module acquires fundus images through the image input module, and performs imaging parameter correction, edge occlusion filtering, and color standard transformation on the fundus images according to preset distortion correction steps to obtain optimized fundus images. The multidimensional feature extraction module performs feature extraction based on the fundus optimized image. During feature extraction, it simultaneously performs feature extraction of detached regions, feature extraction of vascular curvature changes, and pseudo-3D perception features through the fundus optimized image, thereby obtaining detached region features, detached vascular features, and detached stereo features, and sends the extracted features to the fusion positioning module. The fusion localization module can perform fusion evaluation on the extracted features to obtain quantitative evaluation results of the deviated area and approximation evaluation results of the risk area. The report output module will output results independent of regional assessment results and risk area proximity assessment results.

[0007] In a preferred embodiment of the present invention, when the distortion correction module corrects the imaging parameters, it first obtains the imaging parameters of the wide-angle fundus photography device, generates corresponding image modification parameters based on the imaging parameters, and restores the image edge position through the image modification parameters. When the distortion correction module performs edge occlusion filtering, it uses a trained recognition algorithm to identify the eyelashes and reflective areas at the corners of the eyes at the image edges, delineates the areas to be disturbed, and removes the disturbed areas from the fundus image.

[0008] In a preferred embodiment of the present invention, the method for the distortion correction module to perform color standard transformation is as follows: Select a fundus image and divide it layer by layer from the geometric center into the central region, transition region and edge region; Identify blood vessels in the central region, transition region, and edge region, and label them as blood vessel image blocks; The image brightness of the blood vessel image block in the central region is acquired, and the image brightness is arithmetically averaged to obtain the average brightness of the central blood vessel. The image brightness of blood vessel image blocks in the edge region and transition region is acquired, and the image brightness of each region is arithmetically averaged to obtain the average brightness of edge blood vessels and the average brightness of transition blood vessels. The average brightness of peripheral blood vessels is compared with that of central blood vessels to obtain the edge brightness difference. This edge brightness difference is then applied to the edge area to increase the brightness of the image in the edge area. The average brightness of transitional blood vessels is compared with that of central blood vessels to obtain the transition brightness difference. This transition brightness difference is then applied to the transition area to increase the brightness of the image in the transition area.

[0009] In a preferred embodiment of the present invention, when the multidimensional feature extraction module extracts features from the detached region, it first calculates the gradient magnitude and gradient direction of the pixels in the optimized fundus image, then extracts candidate edge pixels through the edge detection operator, and constructs a candidate edge pixel set through the candidate edge pixels. Then, a preset typical gradient threshold is obtained, and the gradient magnitude of the candidate edge pixel set is compared with the typical gradient threshold. Candidate edge pixels that are less than the typical gradient threshold are removed. The continuous edges formed by the candidate edge pixel set are subjected to image detection, and continuous edges whose aspect ratio is within the set discrete cluster range are removed to obtain a high confidence edge pixel set. Finally, the average grayscale difference between the inner and outer sides of the high-confidence edge pixel set is calculated. If the average grayscale difference is lower than the set threshold, it is determined to be a low-confidence boundary. If the average grayscale difference is greater than the set threshold, it is determined to be a feature deviating from the region.

[0010] In a preferred embodiment of the present invention, the method for the multidimensional feature extraction module to extract features detached from blood vessels is as follows: The centerline of blood vessels is obtained by optimizing fundus images, and the curvature of the centerline of blood vessels is calculated at each position. The multidimensional feature extraction module also identifies vascular regions adjacent to normal areas in the optimized fundus image, and obtains a curvature threshold through polynomial regression fitting. It compares the region at each position of the vascular centerline with the curvature threshold to obtain the curvature difference, and records the positions with curvature differences greater than the set threshold as detached vascular features.

[0011] In a preferred embodiment of the present invention, the method by which the multidimensional feature extraction module obtains features detached from stereo features is as follows: Extract suspected shadow locations from the optimized fundus image, then obtain the light source direction from the wide-angle fundus photograph. Based on the light source direction and suspected shadow locations, calculate the bulge height and slope angle in reverse. Record locations with bulge heights greater than a set threshold as detached from stereo features.

[0012] In a preferred embodiment of the present invention, the fusion positioning module generates coordinates for the optimized fundus image, covers the entire range of the optimized fundus image through a coordinate system, performs coordinate transformation on the positions of detachment from regional features, detachment from vascular features, and detachment from three-dimensional features, and generates corresponding coordinate ranges based on the transformed coordinates. The coordinates that are repeated in the coordinate ranges corresponding to the three features are recorded as high-confidence coordinates, and the expected detachment range is constructed based on all high-confidence coordinates.

[0013] In a preferred embodiment of the present invention, the fusion positioning module performs multi-level threshold judgment based on the expected departure range, and determines the risk level to which the expected departure range belongs as the quantitative assessment result of the departure area. The fusion positioning module also marks the boundary of the expected escape range and calculates the distance between the boundary of the expected escape range and the macula. If the distance is less than a set threshold, a high-risk approach result is generated; otherwise, no response is made.

[0014] Compared with the prior art, the beneficial effects of the present invention are: 1. In this invention, the imaging parameters are corrected by a distortion correction module, and the eyelashes and reflective areas at the corners of the eyes at the image edges are identified and removed by a trained recognition algorithm. At the same time, a layer-by-layer partitioned brightness compensation strategy based on the blood vessel region is adopted to perform step-by-step brightness correction on the central region, transition region and edge region. This effectively solves the image quality degradation problem caused by edge distortion, illumination attenuation and artifact interference in wide-angle fundus images, and provides high-quality fundus optimized images for subsequent multi-dimensional feature extraction.

[0015] 2. In this invention, while extracting the planar boundary features of the detachment area, the curvature change features of retinal vessels and the pseudo-3D elevation height features estimated based on shadow and texture gradient are extracted simultaneously. This enables the joint identification of three pathological manifestations of the detachment area: planar range, vascular deformation, and three-dimensional height. It makes full use of the multiple manifestations of retinal detachment in imaging, effectively making up for the limitations of a single two-dimensional grayscale segmentation method in distinguishing detachment from other low-reflection lesions of the fundus, and significantly reducing the risk of misjudgment.

[0016] 3. This invention also performs coordinate mapping and overlap analysis on the detachment features, detachment vascular features, and detachment three-dimensional features. A high-confidence predicted detachment range is constructed from the intersection of the coordinate ranges of the three features. At the same time, a multi-level risk quantification assessment is performed on this range, and the distance between the boundary of the predicted detachment range and the macula is automatically calculated to achieve a high-risk approach warning. The detachment area is confirmed from multiple angles in a collaborative manner, which significantly improves the accuracy of the identification results. Attached Figure Description

[0017] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings.

[0018] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a system flowchart of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0020] Example 1: Please refer to Figure 1 - Figure 2 As shown, the retinal detachment region identification system based on wide-angle fundus photography images includes an image input module, a distortion correction module, a multi-dimensional feature extraction module, a fusion localization module, and a report output module that are connected in sequence via communication. The image input module is used to acquire fundus images generated by the wide-angle fundus photography device. While acquiring the fundus image, the image input module simultaneously records the capture timestamp of the image to facilitate subsequent longitudinal follow-up and comparative analysis. The image input module sends the acquired fundus image and its associated time information to the distortion correction module.

[0021] The distortion correction module receives fundus images from the image input module and preprocesses the fundus images according to the following sub-steps to generate optimized fundus images: (a) Imaging parameter correction The distortion correction module first reads the factory calibration parameters of the wide-angle fundus camera, including but not limited to the lens distortion coefficient, imaging target size, pixel physical spacing and lens projection model.

[0022] Based on the imaging parameters and the inverse projection transformation model of the imaging device, the distortion correction module generates corresponding image modification parameters, namely a remapping table that maps each pixel from distorted coordinates to theoretically distortion-free coordinates. Subsequently, the distortion correction module restores the image edge positions pixel by pixel according to the remapping table, so that the peripheral anatomical structures of the retina are restored to the correct spatial position relationship.

[0023] (ii) Edge-covering filtration After the imaging parameters are corrected, the distortion correction module further identifies and removes artifacts in the image edge region. It identifies and outputs the category label and bounding box coordinates of each disturbed region through a pre-deployed target detection network. The training dataset of this network contains a large number of wide-angle fundus images labeled with eyelash shadows, eyelid edges and reflective areas at the corners of the eyes. The distortion correction module delineates the precise contour of the interfered area based on the recognition results and removes the interfered area, thereby effectively eliminating artifact interference from the fundus image and preventing the subsequent feature extraction module from misidentifying artifacts as lesion areas.

[0024] (III) Color Standard Transformation Due to the ring light source illumination characteristics of wide-angle fundus photography equipment, the image often exhibits uneven illumination with a bright center and dark periphery. Therefore, the distortion correction module selects the fundus image processed by the aforementioned steps, takes the geometric center of the image as the origin, and divides the image layer by layer into the central region, transition region, and edge region based on the radial distance from the center point. Retinal vessels in the central region, transition region, and edge region are identified separately, and the identified vessel pixel sets are marked as vessel image blocks for each region. Collect the pixel brightness values ​​of all vascular image blocks in the central region, and calculate their arithmetic mean to obtain the average brightness of the central blood vessels. Similarly, collect the pixel brightness values ​​of vascular image blocks in the edge region and the transition region respectively, and calculate their arithmetic mean to obtain the average brightness of the edge blood vessels and the average brightness of the transition blood vessels. Calculate the edge brightness difference and use this difference as a brightness gain coefficient to increase the brightness of all pixels in the edge area pixel by pixel, so that the overall brightness of the edge area moves towards the center area. Calculate the transition brightness difference and use this difference as a brightness gain coefficient to increase the brightness of all pixels in the transition area pixel by pixel; After the above-mentioned graded brightness compensation, the optimized fundus image is obtained, and the distortion correction module sends the optimized fundus image to the multidimensional feature extraction module.

[0025] The multidimensional feature extraction module receives the optimized fundus image and performs deviated region boundary recognition, vascular curvature change analysis, and pseudo-3D stereo perception. The three recognitions run synchronously, generating deviated region features, deviated vascular features, and deviated stereo features respectively, and then sending them to the fusion localization module.

[0026] Specifically, the extraction method detached from regional features is as follows: Anisotropic diffusion filtering is applied to the fundus optimization image to remove noise interference while preserving edge information. Then, the gradient magnitude and gradient direction of each pixel in the image are calculated, and the edge detection operator (Gaussian filter standard deviation σ is set to 1.2, and high and low threshold ratio is set to 2:1) is used to extract candidate edge pixels. All candidate edge pixel sets are constructed into a candidate edge pixel set. Obtain a preset gradient threshold, compare the gradient magnitude of each pixel in the candidate edge pixel set with the typical gradient threshold, remove candidate edge pixels whose gradient magnitude is less than the typical gradient threshold, and retain only strong edge pixels. Neighborhood connectivity tracking is performed on the remaining candidate edge pixels to form a continuous edge chain. For each continuous edge, image detection is performed, and the aspect ratio of its bounding rectangle is calculated. If the aspect ratio is within the set discrete cluster range, the continuous edge is determined to be a blood vessel edge or a false edge caused by imaging artifacts and is removed. The edges that are not removed constitute a high-confidence edge pixel set. Calculate the average grayscale difference of pixels within a quantitative pixel width range on both sides of the region enclosed by the high-confidence edge pixel set. If the average grayscale difference is lower than the set threshold, it means that the grayscale on both sides of the boundary is close, and the boundary is determined to be a low-confidence boundary. If the average grayscale difference is greater than the set threshold, it means that the grayscale difference on both sides of the boundary is significant, and the boundary is determined to be a valid departure region boundary, and the output is the departure region feature. Specifically, the extraction method for features detached from blood vessels is as follows: The complete retinal vascular map is extracted and refined to obtain the vascular centerline with a single pixel width. The spatial coordinates of each pixel on each vascular centerline are recorded. Then, for each position on the vascular centerline, the local curvature value of the point is calculated using the three-point circle fitting method with a certain number of pixels before and after the position as the neighborhood range. The multidimensional feature extraction module defines an annular buffer zone of a certain width of pixels outside the boundary of the detached region. The blood vessel region within the buffer zone is identified as a blood vessel in the adjacent normal region. For the center line position of all blood vessels in the adjacent normal region, the distance between the pixel point and the center of the optic disc is used as the independent variable and the local curvature is used as the dependent variable. A third-order polynomial regression is used to fit the curvature baseline function in the normal state, and the curvature threshold is set at each position based on the baseline function. The obtained local curvature values ​​at each location are compared with the corresponding curvature threshold, the curvature difference is calculated, and the locations where the curvature difference is greater than the set threshold are recorded as abnormal points of vascular curvature and output as detached vascular features. Specifically, the extraction method detached from 3D features is as follows: The system acquires the light source direction information from a wide-angle fundus camera, detects regions with directional shadow features in the image, and marks pixel regions that simultaneously meet the conditions of low brightness and directional consistency as suspected shadow locations. Then, for each suspected shadow location, a local surface reflection equation is established in conjunction with the light source direction. By solving the reflection equation, the surface normal vector and local slope angle at each location are calculated in reverse. The bulge height is then estimated by continuous slope integral. Locations with bulge heights greater than a set threshold are recorded as bulge regions, generating detached stereo features.

[0027] The fusion localization module receives the detachment region features, detachment vessel features, and detachment stereo features sent by the multidimensional feature extraction module. It establishes a unified Cartesian coordinate system with the optic disc center as the origin and the direction from the optic disc center to the fovea of ​​the macula as the positive x-axis. This coordinate system covers the entire pixel range of the fundus optimization image and converts the spatial location information contained in the detachment region features, detachment vessel features, and detachment stereo features into coordinate form, and generates corresponding coordinate ranges for each, namely the detachment region boundary coordinate set, the abnormal vessel segment coordinate set, and the raised region coordinate set. A pixel-by-pixel overlap analysis is performed on the coordinate ranges corresponding to the three features. Specifically, for each pixel coordinate (x, y) in the image, it is determined whether the coordinate is simultaneously covered by the detachment region feature, the detachment blood vessel feature, and the detachment stereo feature. Coordinates covered by all three features are recorded as high-confidence coordinates, while the remaining coordinates covered by only one or two features are recorded as low-confidence coordinates. The expected detachment range is constructed based on all high-confidence coordinates.

[0028] When conducting multi-level risk quantification assessment, the fusion positioning module makes multi-level threshold judgments based on the proportion of the expected detachment area to the total area of ​​the retina. Specifically, when the proportion of the expected detachment area to the total area of ​​the retina is less than the lowest threshold, it is judged as a low-risk level. When the ratio falls between the minimum and medium thresholds, it is classified as a medium-risk level. When the ratio falls between the medium and the highest thresholds, it is classified as a high-risk level. When the proportion exceeds the highest threshold, it is judged as an extremely high-risk level; The above-mentioned gear determination results are output as the quantitative assessment results of the departure area.

[0029] Example 2: Please refer to Figure 1 — Figure 2 As shown, the fusion positioning module can also perform macular region approach warning, specifically: First, the coordinates of the optic disc center are located. Then, based on the positional relationship between the optic disc and the macula, the coordinates of the fovea are automatically estimated. Next, the boundary of the expected deviance range is marked pixel by pixel, and the set of contour edge points of the expected deviance range is extracted. Then, the distance from each contour edge point to the coordinates of the fovea is calculated, and the minimum distance value is taken. The minimum distance value is compared with a set threshold: if it is less than the set threshold, a high-risk approximation result is generated; otherwise, no reaction is made. The fusion positioning module sends the quantitative assessment result of the deviance area and the high-risk approximation result to the report output module together.

[0030] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not describe all details exhaustively, nor do they limit the invention to specific implementations. Clearly, many modifications and variations can be made based on the content of this specification.

[0031] This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents. Thresholds, preset values, preset ranges, etc., are set for result comparison and analysis to determine whether something is good or bad. The magnitude of these values ​​is determined based on a combination of large-scale model analysis of sample data and human experience, and can also be appropriately adjusted based on seasonal or common-sense influencing factors.

Claims

1. A retinal detachment region identification system based on wide-angle fundus photography images, characterized in that, It includes an image input module, a distortion correction module, a multi-dimensional feature extraction module, a fusion localization module, and a report output module; The image input module is used to acquire fundus images generated by the wide-angle fundus photography method and simultaneously record the image capture time; The distortion correction module acquires fundus images through the image input module, and performs imaging parameter correction, edge occlusion filtering, and color standard transformation on the fundus images according to preset distortion correction steps to obtain optimized fundus images. The multidimensional feature extraction module performs feature extraction based on the fundus optimized image. During feature extraction, it simultaneously performs feature extraction of detached regions, feature extraction of vascular curvature changes, and pseudo-3D perception features through the fundus optimized image, thereby obtaining detached region features, detached vascular features, and detached stereo features, and sends the extracted features to the fusion positioning module. The fusion localization module can perform fusion evaluation on the extracted features to obtain quantitative evaluation results of the deviated area and approximation evaluation results of the risk area. The report output module will output results independent of regional assessment results and risk area proximity assessment results.

2. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, When the distortion correction module corrects the imaging parameters, it first obtains the imaging parameters of the wide-angle fundus camera and generates corresponding image modification parameters based on the imaging parameters. The image edge position is then restored using the image modification parameters. When the distortion correction module performs edge occlusion filtering, it uses a trained recognition algorithm to identify the eyelashes and reflective areas at the corners of the eyes at the image edges, delineates the areas to be disturbed, and removes the disturbed areas from the fundus image.

3. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, The method by which the distortion correction module performs color standard transformation is as follows: Select a fundus image and divide it layer by layer from the geometric center into the central region, transition region and edge region; Identify blood vessels in the central region, transition region, and edge region, and label them as blood vessel image blocks; The image brightness of the blood vessel image block in the central region is acquired, and the image brightness is arithmetically averaged to obtain the average brightness of the central blood vessel. The image brightness of blood vessel image blocks in the edge region and transition region is acquired, and the image brightness of each region is arithmetically averaged to obtain the average brightness of edge blood vessels and the average brightness of transition blood vessels. The average brightness of peripheral blood vessels is compared with that of central blood vessels to obtain the edge brightness difference. This edge brightness difference is then applied to the edge area to increase the brightness of the image in the edge area. The average brightness of transitional blood vessels is compared with that of central blood vessels to obtain the transition brightness difference. This transition brightness difference is then applied to the transition area to increase the brightness of the image in the transition area.

4. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, When extracting features from the detached region, the multidimensional feature extraction module first calculates the gradient magnitude and gradient direction of the pixels in the optimized fundus image, then extracts candidate edge pixels through the edge detection operator, and constructs a candidate edge pixel set through the candidate edge pixels. Then, a preset typical gradient threshold is obtained, and the gradient magnitude of the candidate edge pixel set is compared with the typical gradient threshold. Candidate edge pixels that are less than the typical gradient threshold are removed. The continuous edges formed by the candidate edge pixel set are subjected to image detection, and continuous edges whose aspect ratio is within the set discrete cluster range are removed to obtain a high confidence edge pixel set. Finally, the average grayscale difference between the inner and outer sides of the high-confidence edge pixel set is calculated. If the average grayscale difference is lower than the set threshold, it is determined to be a low-confidence boundary. If the average grayscale difference is greater than the set threshold, it is determined to be a feature deviating from the region.

5. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, The method for extracting features detached from blood vessels by the multidimensional feature extraction module is as follows: The centerline of blood vessels is acquired using optimized fundus images, and the curvature of the centerline of blood vessels is calculated at each location. The multidimensional feature extraction module also identifies vascular regions adjacent to normal areas in the optimized fundus image, and obtains a curvature threshold through polynomial regression fitting. It compares the region at each position of the vascular centerline with the curvature threshold to obtain the curvature difference, and records the positions with curvature differences greater than the set threshold as detached vascular features.

6. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, The method for the multi-dimensional feature extraction module to obtain features detached from stereo features is as follows: Extract suspected shadow locations from the optimized fundus image, then obtain the light source direction from the wide-angle fundus photograph. Based on the light source direction and suspected shadow locations, calculate the bulge height and slope angle in reverse. Record locations with bulge heights greater than a set threshold as detached from stereo features.

7. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 1, characterized in that, The fusion localization module generates coordinates for the optimized fundus image, covering the entire range of the optimized fundus image through a coordinate system. It also transforms the coordinates of the locations that deviate from regional features, vascular features, and stereoscopic features. Based on the transformed coordinates, it generates the corresponding coordinate ranges for each feature. The coordinates that are repeated in the coordinate ranges corresponding to the three features are recorded as high-confidence coordinates. The expected deviance range is constructed based on all the high-confidence coordinates.

8. The retinal detachment region identification system based on wide-angle fundus photography images according to claim 7, characterized in that, The fusion positioning module performs multi-level threshold judgments based on the expected departure range, and determines the risk level to which the expected departure range belongs as the quantitative assessment result of the departure area. The fusion positioning module also marks the boundary of the expected escape range and calculates the distance between the boundary of the expected escape range and the macula. If the distance is less than a set threshold, a high-risk approach result is generated; otherwise, no response is made.

Images