Macular region retina layered structure automatic segmentation system based on OCT image

Abstract

The invention relates to the technical field of image processing, in particular to a macular region retina layered structure automatic segmentation system based on an OCT image. The system comprises a memory and a processor, and the processor executes computer programs stored in the memory to implement the following steps: acquiring an OCT image of a retina, and determining an ILM boundary and a complex region based on the gray level of pixel points; obtaining a BM boundary according to the gray scale change characteristics of each column of pixel points in the complex body region; determining a shielding area and a target normal area based on the position of the BM boundary, and further determining an RPE boundary; dividing an abnormal region according to the gray scale relative size of region pixel points between the RPE boundary and the BM boundary in the target normal region, and determining a raised boundary and a normal boundary; and repairing the lifted boundary, and obtaining a macular region retina layered structure segmentation result in combination with a repairing result and a normal boundary. According to the invention, the accuracy of the segmentation result of the macular region retina layered structure is improved.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to an automatic segmentation system for the retinal layered structure in the macular region based on OCT images. Background Technology

[0002] Optical coherence tomography (OCT) works on a principle similar to ultrasound, but it uses light waves instead of sound waves. It generates cross-sectional images by measuring the delay time and intensity of low-coherence light reflected from different retinal layers. Different retinal layers have different light reflection characteristics due to variations in their cellular structure, density, and composition. This manifests as different grayscale and texture patterns in OCT images. Segmentation is based on the differences in the microstructure of different retinal layers, resulting in distinguishable grayscale, texture, and boundary features in OCT images.

[0003] To automatically segment the retinal laminar structure in the macular region based on OCT images, when analyzing the retinal laminar structure in OCT images, lesions such as hemorrhage and scars may appear highly reflective and be incorrectly identified by the algorithm as layer boundaries (such as the RPE layer). Furthermore, macular edema and cystic edema of the inner / outer layers caused by central retinal vein occlusion can form low-reflectivity cavities, disrupting the normal laminar structure and thus affecting the accuracy of the automatic segmentation results. Summary of the Invention

[0004] To address the issue of low segmentation accuracy in existing methods for automatically segmenting the retinal layered structure, the present invention aims to provide an automatic segmentation system for the macular retinal layered structure based on OCT images. The specific technical solution adopted is as follows: This invention provides an automatic segmentation system for the retinal layer structure of the macular region based on OCT images. The system includes a memory and a processor. The processor executes a computer program stored in the memory to achieve the following steps: Obtain an OCT image of the target patient's retina, wherein the upper edge of the OCT image is the upper boundary of the retina; ILM boundary and complex region are determined based on the gray-scale distribution of pixels in OCT images; BM boundary is obtained based on the gray-scale change characteristics of each column of pixels in the complex region; candidate lesion connected regions are determined based on the location of the BM boundary; shielded regions and target normal regions are screened based on the shape characteristics of the candidate lesion connected regions; RPE boundary is determined based on the gray-scale change characteristics of each column of pixels in the target normal region and their relative position with the shielded region. The abnormal region is divided according to the relative grayscale values ​​of the pixels in the region to be analyzed; the region to be analyzed is the area between the RPE boundary and the BM boundary in the target normal region; the raised boundary and the normal boundary are determined according to the location of the abnormal region, the curvature at different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region. The raised boundary is repaired, and the retinal laminar structure segmentation result of the macular region is obtained based on the repair result and the normal boundary.

[0005] Preferably, the determination of the ILM boundary and complex region based on the grayscale distribution of pixels in the OCT image includes: In the OCT image, each pixel that is greater than the first grayscale threshold is obtained and recorded as the first pixel. The first first pixel in each column is taken as the marker pixel in the order from top to bottom. All marker pixels are filtered to obtain the ILM boundary. Below the ILM boundary, the pixels below the ILM boundary are divided based on the gray values ​​of the pixels using the Otsu thresholding method. The connected region with the largest area among the connected regions formed by pixels with a value greater than the threshold is taken as the complex region.

[0006] Preferably, obtaining the BM boundary based on the grayscale variation characteristics of each column of pixels in the composite region includes: The grayscale difference between two adjacent pixels in each column of pixels in the composite region is obtained respectively. The bottom pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the BM boundary. All pixels on the BM boundary constitute the BM boundary.

[0007] Preferably, determining the connected domain of candidate lesions based on the location of the BM boundary includes: The boundary of BM is translated upward by a preset first distance and downward by a preset second distance to obtain two line segments. The area between the two line segments is denoted as a strip region. Morphological closing operation is performed on the pixels in the strip region that are greater than the second grayscale threshold, and the resulting connected component is recorded as the first candidate connected component. Region growth is performed on each first candidate connected region to obtain a second candidate connected region; wherein, the condition for region growth is: before each growth, the average gray value of the pixels in the corresponding connected region is obtained; if the difference between the gray value of the pixel to be grown and the average gray value is less than a preset gray value difference threshold, then the pixel to be grown is merged with the current connected region. The second candidate connected component with an area greater than a preset area threshold is selected as the candidate lesion connected component.

[0008] Preferably, the step of filtering the shielded region and the target normal region based on the shape characteristics of the connected regions of the candidate lesion includes: For any candidate lesion connected region: calculate the first ratio of the length to the width of the minimum bounding rectangle of the candidate lesion connected region; obtain the fitting loss value when performing ellipse fitting on the candidate lesion connected region; obtain the elongation of the candidate lesion connected region based on the first ratio and the fitting loss value. Based on the elongation of the connected regions of all candidate lesions, the Otsu threshold method is used to select the connected regions of candidate lesions that are less than the threshold as the connected regions of the lesions; morphological expansion of the connected regions of the lesions is then performed to obtain the shielding region. Draw two perpendicular lines through the leftmost and rightmost pixels of the shielded area, respectively, and denote them as the left perpendicular line and the right perpendicular line. The closed area enclosed by the left perpendicular line and the BM boundary and ILM boundary is taken as the target normal area, and the closed area enclosed by the right perpendicular line and the BM boundary and ILM boundary is also taken as the target normal area.

[0009] Preferably, determining the RPE boundary based on the grayscale variation characteristics of each column of pixels within the target normal area and their relative position to the shielded area includes: The grayscale difference between two adjacent pixels in each column of pixels in the target normal area is obtained respectively. The topmost pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the RPE boundary. The RPE boundary is obtained based on the pixels on the RPE boundary.

[0010] Preferably, the step of dividing the abnormal region based on the relative grayscale values ​​of pixels within the region to be analyzed includes: Based on the grayscale values ​​of pixels within the region to be analyzed, the Otsu thresholding method is used to classify the pixels within the region to be analyzed, and pixels with values ​​less than the threshold are recorded as second pixels. Based on the grayscale values ​​of the second pixels, the Otsu thresholding method is used to record all pixels with values ​​less than the threshold as third pixels. Morphological closing operations are performed on the region formed by the third pixels to obtain the initial abnormal region. The least squares method is used to perform ellipse fitting on the initial anomaly region to obtain the fitting loss value of the initial anomaly region; Based on the fitting loss value of the initial abnormal region, the Otsu threshold method is used to identify all initial abnormal regions whose fitting loss value is less than the threshold as abnormal regions.

[0011] Preferably, determining the raised boundary and the normal boundary based on the location of the abnormal region, the curvature at different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region includes: For any anomalous region: The leftmost and rightmost pixels passing through the edge line of any abnormal region are recorded as feature points. The two feature points divide the edge line of any abnormal region into an upper half edge and a lower half edge. The upper edge is divided into multiple boundary segments by using the maximum curvature of the pixels in the upper edge as the dividing point; the edge line directly above each boundary segment is recorded as the suspected affected edge line corresponding to each boundary segment; wherein, the edge line directly above the boundary segment is obtained by edge detection algorithm; Obtain the first distance between each boundary segment and its corresponding suspected affected edge line; Calculate the variance of the curvature of all suspected affected edge lines respectively; obtain the curvature change of each suspected affected edge line based on the first distance and the variance; The lifted boundary and the normal boundary are determined based on the curvature change.

[0012] Preferably, determining the lifted boundary and the normal boundary based on the curvature change includes: determining the suspected affected edge line where the curvature change is greater than a preset curvature change threshold as the lifted boundary, and determining the suspected affected edge line where the curvature change is less than or equal to the preset curvature change threshold as the normal boundary.

[0013] Preferably, the raised boundary is repaired using a variational model that minimizes curvature energy.

[0014] The present invention has at least the following beneficial effects: This invention, by combining the distribution characteristics of the LM boundary, complex region, BM boundary, lesion region, shielding region, and normal region in retinal OCT images, determines the ILM boundary, complex region, BM boundary, candidate lesion connected domain, shielding region, target normal region, and RPE boundary. It effectively eliminates interference from highly reflective lesions such as hemorrhage and exudation on layer boundary identification, accurately restoring the true location and morphology of key retinal structures such as the RPE boundary. Abnormal regions are identified by analyzing the relative grayscale values ​​of pixels within the region between the RPE boundary and the BM boundary in the target normal region, and further... By combining the location of abnormal areas, the curvature of the edge lines, and the characteristics of the edge lines above them, the system intelligently identifies the raised boundaries and normal boundaries, effectively addressing the deformation or damage of the retinal laminar structure caused by lesions such as cystic edema. By repairing the raised boundaries and combining them with the normal boundaries, accurate and coherent segmentation results of the macular retinal laminar structure are finally obtained. Even in complex pathological images with severe edema and lesions, precise restoration of each layer of structure can be achieved, improving the accuracy of the segmentation results of the macular retinal laminar structure and providing reliable and quantifiable image evidence for clinical diagnosis and disease assessment. Attached Figure Description

[0015] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0016] Figure 1 A flowchart illustrating the method performed by an automatic segmentation system for the macular retinal layer structure based on OCT images, as provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the target normal area provided in an embodiment of the present invention; Figure 3 This is a structural block diagram of an automatic segmentation device for the retinal layered structure in the macular region based on OCT images, provided in an embodiment of the present invention. Detailed Implementation

[0017] To further illustrate the technical means and effects adopted by the present invention to achieve the intended purpose, the following detailed description of an automatic segmentation system for the retinal layer structure in the macular region based on OCT images, in conjunction with the accompanying drawings and preferred embodiments, is provided below.

[0018] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0019] The following description, in conjunction with the accompanying drawings, details the specific scheme of an automatic segmentation system for the macular retinal layer structure based on OCT images provided by this invention.

[0020] Example of an automatic segmentation system for the macular retinal layer structure based on OCT images: The specific scenario addressed in this embodiment is as follows: When segmenting the retinal laminar structure in the macular region of patients with diabetic macular edema and central retinal vein occlusion with cystic edema lesions in OCT images, in order to prevent the influence of edema on the segmentation results, the raised boundaries are selected based on the distribution characteristics of fixed structures in the retina and repaired to improve the accuracy of subsequent macular retinal laminar structure segmentation results.

[0021] This embodiment proposes an automatic segmentation system for the retinal laminar structure of the macular region based on OCT images. The system includes a memory and a processor. The processor executes a computer program stored in the memory, such as... Figure 1As shown, the method performed by the automatic segmentation system for the macular retinal layer structure based on OCT images in this embodiment includes the following steps: Step S1: Obtain an OCT image of the target patient's retina, wherein the upper edge of the OCT image is the upper boundary of the retina.

[0022] When automatically segmenting the retinal laminar structure of the macular region in OCT images of the target patient's retina, the target patient is first placed in a standardized dark room environment. Retinal angiography combined with optical coherence tomography (Spectralis HRA-OCT imaging system) is used to acquire images of the target patient's retina. During OCT image acquisition, the target patient is strictly kept in a stable fixation state. First, both eyes are fully dilated. Then, the patient's position is adjusted so that the chin is stably placed on a chin rest. The camera lens is perpendicular to the corneal surface of the target patient, and a beam of light is used to vertically scan the strip-shaped area in the retina. The OCT image acquisition ensures complete coverage of the 6mm × 6mm area of ​​the macular region, guaranteeing comprehensive capture of the three-dimensional distribution characteristics of cystic edema and its morphological changes to the retinal laminar structure. The upper edge of the acquired retinal OCT image is the upper boundary of the retina. Then, a small (3 × 3) Gaussian kernel is used to filter the acquired retinal OCT image to suppress the inherent speckle noise in the OCT image. It should be noted that the OCT images of the target patient's retina mentioned later are all filtered images.

[0023] The target patients are those with diabetic macular edema and central retinal vein occlusion with cystoid edema lesions.

[0024] Thus, this embodiment has acquired an OCT image of the target patient's retina, with the upper edge of the OCT image representing the upper boundary of the retina.

[0025] Step S2: Determine the ILM boundary and complex region based on the grayscale distribution of pixels in the OCT image; obtain the BM boundary based on the grayscale change characteristics of each column of pixels in the complex region; determine the candidate lesion connected region based on the location of the BM boundary; filter the shielded region and the target normal region based on the shape characteristics of the candidate lesion connected region; determine the RPE boundary based on the grayscale change characteristics of each column of pixels in the target normal region and its relative position with the shielded region.

[0026] Since the ILM boundary in an OCT image is generally the top horizontal highlight line, each pixel in the OCT image with a value greater than the first grayscale threshold is designated as the first pixel. The first pixel in each column is then designated as the marker pixel from top to bottom. Median filtering is applied to all marker pixels to smooth out any spurs. The line segment obtained after smoothing is then used as the boundary of the Internal Limiting Membrane (ILM).

[0027] As one implementation method, the first grayscale threshold can be determined in the following way: Specifically, historically acquired retinal OCT images are collected, and then the grayscale distribution characteristics of pixels on the ILM boundary in these OCT images are statistically analyzed. Based on these distribution characteristics, the first grayscale threshold is set to effectively extract the grayscale value of the ILM boundary. This method of determining the first grayscale threshold based on the statistical analysis results of the grayscale value of the ILM boundary in historically acquired retinal OCT images can better extract pixels on the ILM boundary and improve the accuracy of subsequent lesion area screening results.

[0028] Searching below the defined ILM boundary can eliminate interference from the vitreous portion above the retina; to obtain the retinal laminar structure in the macular region, it is also necessary to determine the BM boundary in the image. Since there is a bright band in the OCT image formed by the RPE and BM boundary, the complex region is first determined by combining this feature, and then the BM boundary is obtained.

[0029] Specifically, below the ILM boundary, based on the grayscale values ​​of pixels, the Otsu thresholding method is used to divide the pixels below the ILM boundary, and the connected region with the largest area among the connected regions formed by pixels with values ​​greater than the threshold is taken as the complex region. The Otsu thresholding method is existing technology and will not be described in detail here.

[0030] The BM is the basement membrane, one of the rigid structures of the eyeball, so it will not move due to retinal edema or exudation; and the lower edge of the RPE-BM complex located at the BM will form a very sharp gray-scale decrease trend due to the strong reflection of the BM layer.

[0031] Based on the above characteristics, the grayscale difference between two adjacent pixels in each column of the composite region is calculated. When calculating the grayscale difference between two adjacent pixels, the absolute value of the difference is used as the grayscale difference between those two pixels. The bottom pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the BM boundary. In this way, multiple pixels on the BM boundary can be obtained.

[0032] Since the pixels on the obtained BM boundary may be discrete, if the pixels on the BM boundary obtained in adjacent columns are not adjacent, then a straight line is used to connect the two pixels to obtain the complete BM boundary.

[0033] The image obtained by the above operation shows the BM boundary above the RPE (retinal pigment epithelium) layer in the retina. Normally, the RPE layer can be directly obtained using the RPE-BM layer. However, since hemorrhages and exudates in the macular region are usually located near or above the RPE layer, this can interfere with the acquisition of the RPE layer. Therefore, the BM boundary is shifted upward by a preset first distance and downward by a preset second distance to obtain two line segments. The area between these two line segments is denoted as a band-shaped region; the preset first distance is 150 micrometers, and the preset second distance is 50 micrometers.

[0034] Considering that hemorrhagic exudative lesions (hemoglobin, lipids) have a strong reflectivity to OCT, resulting in a large grayscale value in the lesion area, morphological closing operations are performed on pixels in the band region that are larger than the second grayscale threshold. This causes the discrete pixels to cluster together, and the resulting connected component is recorded as the first candidate connected component. The first candidate connected component is more likely to be the region where the lesion is located.

[0035] One specific implementation method is to analyze and determine the second gray-scale threshold based on historically acquired retinal OCT images of lesion areas. Specifically, historically acquired retinal OCT images of lesion areas are collected, and the gray-scale distribution characteristics of the lesion areas in these images are statistically analyzed. Based on these distribution characteristics, a second gray-scale threshold is set to a gray-scale value that can effectively distinguish lesion areas from normal areas, for example, 240. This method of determining the second gray-scale threshold based on the statistical analysis results of the gray-scale values ​​of lesion areas in historically acquired retinal OCT images of lesion areas allows the second gray-scale threshold to better distinguish lesion areas and improve the accuracy of subsequent lesion area screening results. Another implementation method is to directly set the second gray-scale threshold to an empirical value of 238.

[0036] The reflection at the edge of the lesion may be slightly weaker, which may prevent threshold segmentation from completely capturing the potential connected regions of the lesion. Therefore, it is necessary to perform region growing on each first candidate connected region to obtain second candidate connected regions. The condition for region growing is as follows: before each growing operation, the average gray value of the pixels in the current connected region is obtained. If the difference between the gray value of the pixel to be grown and the average gray value is less than a preset gray difference threshold, the pixel to be grown is merged with the current connected region. In this embodiment, the preset gray difference threshold is 5. In specific applications, the implementer can set it according to the specific situation. The region growing result of the first candidate connected region is recorded as the second candidate connected region.

[0037] Considering that if the area of ​​the second candidate connected region is too small, it is highly likely to be noise, the second candidate connected region with an area greater than a preset area threshold is selected as the candidate lesion connected region. In this embodiment, the preset area threshold is 100 square micrometers. In specific applications, the implementer can set it according to the specific circumstances.

[0038] Since the RPE region in the image is a narrow strip, while the hemorrhage and exudation lesion appears as a cluster in the image, the shape features of the connected regions of the candidate lesion will be analyzed.

[0039] The following explanation uses a candidate lesion connected region as an example. Other candidate lesion connected regions can be processed using the method provided in this embodiment.

[0040] The RPE region has a large length-to-width ratio due to its narrow, elongated band-like structure, while the lesion's cluster-like structure makes the lesion more similar to a circle.

[0041] Therefore, for any candidate lesion connected region, the minimum bounding rectangle of the candidate lesion connected region is obtained, and the ratio of the length to the width of the minimum bounding rectangle of the candidate lesion connected region is calculated, and this ratio is recorded as the first ratio. The least squares method is used to perform ellipse fitting on the candidate lesion connected region, and the fitting loss value is obtained when performing ellipse fitting on the candidate lesion connected region; the elongation of the candidate lesion connected region is obtained according to the first ratio and the fitting loss value.

[0042] As a specific implementation, the elongation of the connected region of the candidate lesion can be determined as follows: the product of a first ratio and a fitting loss value is used as the elongation of the connected region of the candidate lesion. A larger first ratio indicates a longer long side and narrower width of the minimum bounding rectangle of the connected region of the candidate lesion, thus indicating a relatively slender connected region. A larger fitting loss value indicates a greater difference in shape between the connected region of the candidate lesion and an ellipse, and a less close resemblance to an ellipse. When both the first ratio and the fitting loss value are larger, the elongation of the connected region of the candidate lesion is greater.

[0043] The above method can be used to obtain the extension length of each candidate lesion connected region. Based on the extension length of all candidate lesion connected regions, the Otsu threshold method is used to select candidate lesion connected regions that are less than the threshold as lesion connected regions. The lesion connected regions are morphologically expanded outward by 2 pixels to obtain the shielding area.

[0044] Due to the influence of lesions, the boundaries of some RPE areas in the shielded area are obscured by the lesions. Therefore, it is necessary to obtain the boundary conditions based on the unaffected areas.

[0045] Specifically, two perpendicular lines are drawn through the leftmost and rightmost pixels of the shielded area, denoted as the left perpendicular line and the right perpendicular line; the closed area enclosed by the left perpendicular line and the BM boundary and ILM boundary is taken as the target normal area, and the closed area enclosed by the right perpendicular line and the BM boundary and ILM boundary is also taken as the target normal area. Figure 2 As shown in the figure, 1 is the ILM boundary, 2 is the BM boundary, 3 is the shielded area, 4 is the left vertical line, 5 is the right vertical line, and 6 and 7 are the normal target area.

[0046] The grayscale difference between any two adjacent pixels in each column of the target normal area is obtained. The topmost pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the RPE boundary. In the target normal area, if the pixels on the RPE boundary in two adjacent columns are not adjacent (i.e., not connected), these two pixels are connected by a straight line to obtain the RPE boundary.

[0047] Step S3: Divide the abnormal region according to the relative grayscale values ​​of the pixels in the region to be analyzed; the region to be analyzed is the area between the RPE boundary and the BM boundary in the target normal region; determine the raised boundary and the normal boundary based on the location of the abnormal region, the curvature of different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region.

[0048] Above the RPE boundary is the kernel layer, where edema may occur, potentially causing the kernel layer to be pushed up and resulting in inaccurate determination of the kernel layer thickness. Therefore, it is necessary to repair the edema above the RPE boundary.

[0049] Since the fluid cavity caused by edema contains serous fluid, which reflects almost no light signal, the grayscale value within the edema area is extremely low. Therefore, in this embodiment, the closed region between the RPE boundary and the BM boundary in the target normal region is designated as the region to be analyzed. Based on the grayscale value of the pixels within the region to be analyzed, the Otsu thresholding method is used to classify the pixels, and pixels with grayscale values ​​less than the threshold are designated as second pixels. Based on the grayscale value of the second pixels, the Otsu thresholding method is used to designate all pixels with grayscale values ​​less than the threshold as third pixels. That is, when selecting third pixels, the Otsu thresholding method is applied twice, selecting pixels with lower grayscale values ​​as third pixels. A morphological closing operation is performed on the region formed by the third pixels to fill in any small noise points that may exist inside the fluid cavity. The region after this morphological closing operation is designated as the initial abnormal region.

[0050] Considering that cystic edema in the retina typically presents as round or oval, while normal dark areas are usually irregular in shape, the least squares method was used to perform elliptical fitting on each initial abnormal region to obtain the fitting loss value for each initial abnormal region.

[0051] Based on the fitting loss value of the initial abnormal region, the Otsu threshold method is used to identify all initial abnormal regions whose fitting loss value is less than the threshold as abnormal regions.

[0052] For any anomalous region: The leftmost and rightmost pixels along the edge of the abnormal region are designated as feature points. These two feature points divide the edge of the abnormal region into an upper and lower half. The curvature of each pixel in the upper half is obtained, and the point with the maximum curvature is used as the dividing point to divide the upper half into multiple boundary segments. The edge lines directly above each boundary segment are designated as the suspected affected edge lines for that segment. The edge lines directly above the boundary segments are detected using an edge detection algorithm, which is a current technology and will not be elaborated upon here.

[0053] Obtain the distance between each boundary segment and its corresponding suspected affected edge line, and record this distance as the first distance for each suspected affected edge line. Calculate the variance of the curvature of all suspected affected edge lines; obtain the curvature change of each suspected affected edge line based on the first distance and the variance.

[0054] In one specific implementation, for any suspected affected edge line, the sum of the first distance and the zero-prevention parameter corresponding to the suspected affected edge line is calculated. The normalized result of the ratio of the variance of all curvatures on the suspected affected edge line to the sum is taken as the curvature change of the suspected affected edge line. The closer the distance between the boundary segment and its corresponding suspected affected edge lines, and the larger the variance of all curvatures on the suspected affected edge lines, the higher the degree of curvature change of the suspected affected edge lines, i.e., the greater the curvature change. The zero-prevention parameter can be 0.001; in specific applications, the implementer can set it according to the specific situation. When normalizing the data, the maximum-minimum value normalization method is used. The maximum-minimum value normalization method is existing technology and will not be elaborated further here.

[0055] Using the above method, the curvature change of each suspected affected edge line can be obtained. Edge lines with curvature changes greater than a preset curvature change threshold are identified as raised boundaries, while edge lines with curvature changes less than or equal to the preset curvature change threshold are identified as normal boundaries.

[0056] As one specific implementation, a preset curvature change threshold can be determined by analyzing historically acquired retinal OCT images that have been confirmed to have raised boundaries. Specifically, retinal OCT images with confirmed raised boundaries are collected, and then the curvature change of the raised boundaries is calculated using the method provided in this embodiment. The distribution characteristics of the curvature change of the raised boundaries in these images are statistically analyzed. Based on these distribution characteristics, the preset curvature change threshold is set to the curvature change that can effectively distinguish between raised boundaries and normal boundaries, for example, it can be 0.6. This method of determining the preset curvature change threshold based on the statistical analysis results of the curvature change of raised boundaries in historically acquired retinal OCT images with confirmed raised boundaries enables the preset curvature change threshold to better distinguish between raised boundaries and normal boundaries, improving the accuracy of the screening results for raised boundaries and normal boundaries. As another implementation, the preset curvature change threshold can also be directly set to an empirical value of 0.6.

[0057] Step S4: Repair the raised boundary, and obtain the retinal layer structure segmentation result of the macular region based on the repair result and the normal boundary.

[0058] In this embodiment, the lifted boundaries were selected in step S3, and then the lifted boundaries were repaired using a variational model that minimizes curvature energy. The variational model that minimizes curvature energy is an existing model, and will not be described in detail here.

[0059] Based on the repaired raised boundary and the normal boundary, a neural network model can be used to segment the retinal laminar structure in the macular region of the target patient's retina in the OCT image. Specifically, the repaired raised boundary and the normal boundary are marked in the OCT image of the target patient's retina, and the marked image is input into a pre-trained neural network model. The neural network model outputs the segmentation result of the retinal laminar structure in the macular region, thus achieving the segmentation of the retinal laminar structure in the OCT image of the target patient's retina and obtaining the segmentation result. The training process of the neural network model is existing technology and will not be elaborated further here.

[0060] This embodiment accurately restores the true orientation of key boundaries such as the retinal pigment epithelium by intelligently identifying and eliminating interference from hemorrhage, exudation, and other factors in complex pathological images. Simultaneously, it can detect structural regions deformed or disappeared due to cystic edema and intelligently reconstruct the original morphology of these damaged layer boundaries based on the inherent smoothness and continuity of retinal tissue. The final output is clear and coherent boundary lines for all retinal layers, thus generating reliable and quantifiable layer thickness divisions for clinicians even in OCT images with severe edema and lesions. This achieves precise restoration and objective assessment of pathological retinal structures, providing crucial evidence for disease diagnosis and follow-up.

[0061] This embodiment, by combining the distribution characteristics of the LM boundary, complex region, BM boundary, lesion region, shielding region, and normal region in retinal OCT images, determines the ILM boundary, complex region, BM boundary, candidate lesion connected region, shielding region, target normal region, and RPE boundary. This effectively eliminates interference from highly reflective lesions such as hemorrhage and exudation on layer boundary identification, accurately restoring the true location and morphology of key retinal structures such as the RPE boundary. Abnormal regions are identified by analyzing the relative grayscale values ​​of pixels within the region between the RPE boundary and BM boundary in the target normal region, and further... By combining the location of abnormal areas, the curvature of the edge lines, and the characteristics of the edge lines above them, the system intelligently identifies the raised boundaries and normal boundaries, effectively addressing the deformation or damage of the retinal laminar structure caused by lesions such as cystic edema. By repairing the raised boundaries and combining them with the normal boundaries, accurate and coherent segmentation results of the macular retinal laminar structure are finally obtained. Even in complex pathological images with severe edema and lesions, precise restoration of each layer of structure can be achieved, improving the accuracy of the segmentation results of the macular retinal laminar structure and providing reliable and quantifiable image evidence for clinical diagnosis and disease assessment.

[0062] An embodiment of an automatic segmentation device for the macular retinal layer structure based on OCT images: See Figure 3 The diagram illustrates a structural block diagram of an automatic segmentation device for the retinal layered structure in the macular region based on OCT images, according to an embodiment of the present invention. The device may include an image acquisition module, a region filtering module, a boundary classification module, and a segmentation module.

[0063] The image acquisition module is used to acquire an OCT image of the target patient's retina, wherein the upper edge of the OCT image is the upper boundary of the retina. The region filtering module is used to determine the ILM boundary and complex region based on the gray-level distribution of pixels in the OCT image; obtain the BM boundary based on the gray-level change characteristics of each column of pixels in the complex region; determine the candidate lesion connected region based on the location of the BM boundary; filter the shielded region and the target normal region based on the shape characteristics of the candidate lesion connected region; and determine the RPE boundary based on the gray-level change characteristics of each column of pixels in the target normal region and its relative position with the shielded region. The boundary classification module is used to divide abnormal regions based on the relative grayscale values ​​of pixels within the region to be analyzed. The region to be analyzed is the area between the RPE boundary and the BM boundary in the target normal region. The raised boundary and the normal boundary are determined based on the location of the abnormal region, the curvature at different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region. The segmentation module is used to repair the raised boundary and obtain the segmentation result of the retinal layer structure in the macular region based on the repair result and the normal boundary.

[0064] It should be understood that Figure 3 The structural block diagram and modules of the automatic retinal laminar structure segmentation device based on OCT images shown can be implemented in various ways. For example, in some embodiments, the device and its modules can be implemented by hardware, software, or a combination of software and hardware. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by appropriate instructions, such as a microprocessor or dedicated hardware. Those skilled in the art will understand that the above-described methods and devices can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of this specification can be implemented not only by hardware circuits such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., but also by software, for example, executed by various types of processors, or by a combination of the above-described hardware circuits and software (e.g., firmware).

[0065] For more details about the above modules, please refer to other parts of this manual; they will not be repeated here.

[0066] It should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An automatic segmentation system for the retinal laminar structure in the macular region based on OCT images, characterized in that, The system includes a memory and a processor, the processor executing a computer program stored in the memory to perform the following steps: Obtain an OCT image of the target patient's retina, wherein the upper edge of the OCT image is the upper boundary of the retina; The ILM boundary and complex region are determined based on the grayscale distribution of pixels in the OCT image; the BM boundary is obtained based on the grayscale variation characteristics of each column of pixels in the complex region; the candidate lesion connected regions are determined based on the location of the BM boundary; and the shielded region and the target normal region are screened based on the shape characteristics of the candidate lesion connected regions. The RPE boundary is determined based on the grayscale variation characteristics of each column of pixels within the target normal area and their relative position with the shielded area. The abnormal region is divided according to the relative grayscale values ​​of the pixels in the region to be analyzed; the region to be analyzed is the area between the RPE boundary and the BM boundary in the target normal region; the raised boundary and the normal boundary are determined according to the location of the abnormal region, the curvature at different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region. The raised boundary is repaired, and the retinal laminar structure segmentation result of the macular region is obtained based on the repair result and the normal boundary.

2. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The determination of the ILM boundary and complex region based on the grayscale distribution of pixels in the OCT image includes: In the OCT image, each pixel that is greater than the first grayscale threshold is obtained and recorded as the first pixel. The first first pixel in each column is taken as the marker pixel in the order from top to bottom. All marker pixels are filtered to obtain the ILM boundary. Below the ILM boundary, the pixels below the ILM boundary are divided based on the gray values ​​of the pixels using the Otsu thresholding method. The connected region with the largest area among the connected regions formed by pixels with a value greater than the threshold is taken as the complex region.

3. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The step of obtaining the BM boundary based on the grayscale variation characteristics of each column of pixels in the composite region includes: The grayscale difference between two adjacent pixels in each column of pixels in the composite region is obtained respectively. The bottom pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the BM boundary. All pixels on the BM boundary constitute the BM boundary.

4. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The determination of candidate lesion connected regions based on the location of the BM boundary includes: The boundary of BM is translated upward by a preset first distance and downward by a preset second distance to obtain two line segments. The area between the two line segments is denoted as a strip region. Morphological closing operation is performed on the pixels in the strip region that are greater than the second grayscale threshold, and the resulting connected component is recorded as the first candidate connected component. Region growth is performed on each first candidate connected region to obtain a second candidate connected region; wherein, the condition for region growth is: before each growth, the average gray value of the pixels in the corresponding connected region is obtained; if the difference between the gray value of the pixel to be grown and the average gray value is less than a preset gray value difference threshold, then the pixel to be grown is merged with the current connected region. The second candidate connected component with an area greater than a preset area threshold is selected as the candidate lesion connected component.

5. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The step of filtering the shielded region and the target normal region based on the shape characteristics of the connected regions of the candidate lesions includes: For any candidate lesion connected region: calculate the first ratio of the length to the width of the minimum bounding rectangle of the candidate lesion connected region; obtain the fitting loss value when performing ellipse fitting on the candidate lesion connected region; obtain the elongation of the candidate lesion connected region based on the first ratio and the fitting loss value. Based on the elongation of the connected regions of all candidate lesions, the Otsu threshold method is used to select the connected regions of candidate lesions that are less than the threshold as the connected regions of the lesions; morphological expansion of the connected regions of the lesions is then performed to obtain the shielding region. Draw two perpendicular lines through the leftmost and rightmost pixels of the shielded area, respectively, and denote them as the left perpendicular line and the right perpendicular line. The closed area enclosed by the left perpendicular line and the BM boundary and ILM boundary is taken as the target normal area, and the closed area enclosed by the right perpendicular line and the BM boundary and ILM boundary is also taken as the target normal area.

6. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The determination of the RPE boundary based on the grayscale variation characteristics of each column of pixels within the target normal area and their relative position with the shielded area includes: The grayscale difference between two adjacent pixels in each column of pixels in the target normal area is obtained respectively. The topmost pixel of the two pixels corresponding to the maximum grayscale difference in each column is taken as the pixel on the RPE boundary. The RPE boundary is obtained based on the pixels on the RPE boundary.

7. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The process of dividing abnormal regions based on the relative grayscale values ​​of pixels within the region to be analyzed includes: Based on the grayscale values ​​of pixels within the region to be analyzed, the Otsu thresholding method is used to classify the pixels within the region to be analyzed, and pixels with values ​​less than the threshold are recorded as second pixels. Based on the grayscale values ​​of the second pixels, the Otsu thresholding method is used to record all pixels with values ​​less than the threshold as third pixels. Morphological closing operations are performed on the region formed by the third pixels to obtain the initial abnormal region. The least squares method is used to perform ellipse fitting on the initial anomaly region to obtain the fitting loss value of the initial anomaly region; Based on the fitting loss value of the initial abnormal region, the Otsu threshold method is used to identify all initial abnormal regions whose fitting loss value is less than the threshold as abnormal regions.

8. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The process of determining the raised boundary and the normal boundary based on the location of the abnormal region, the curvature at different positions on the edge line of the abnormal region, and the edge line directly above the abnormal region includes: For any anomalous region: The leftmost and rightmost pixels passing through the edge line of any abnormal region are recorded as feature points. The two feature points divide the edge line of any abnormal region into an upper half edge and a lower half edge. The upper edge is divided into multiple boundary segments by using the maximum curvature of the pixels in the upper edge as the dividing point; the edge line directly above each boundary segment is recorded as the suspected affected edge line corresponding to each boundary segment; wherein, the edge line directly above the boundary segment is obtained by edge detection algorithm; Obtain the first distance between each boundary segment and its corresponding suspected affected edge line; Calculate the variance of the curvature of all suspected affected edge lines respectively; obtain the curvature change of each suspected affected edge line based on the first distance and the variance; The lifted boundary and the normal boundary are determined based on the curvature change.

9. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 8, characterized in that, The step of determining the lifted boundary and the normal boundary based on the curvature change includes: determining the suspected affected edge line where the curvature change is greater than a preset curvature change threshold as the lifted boundary, and determining the suspected affected edge line where the curvature change is less than or equal to the preset curvature change threshold as the normal boundary.

10. The automatic segmentation system for the retinal laminar structure based on OCT images according to claim 1, characterized in that, The lifted boundary is repaired using a variational model that minimizes curvature energy.

Images