An optimized method for measuring the volume between intraocular implants and anterior segment structures.

Abstract

This invention relates to the field of computer image processing technology, and particularly to an optimized method for measuring the volume between intraocular implants and anterior segment structures. The invention utilizes anterior segment OCT grayscale images, preprocesses the OCT images using specific image processing methods, generates a high-resolution OCT grayscale image mask using a specific mask generation method, and combines mathematical methods to automatically identify, 3D reconstruct, and measure the volume between the posterior surface of the intraocular implant and the anterior surface of the lens in the image. This invention fills the gap in the fully automated measurement of the volume between intraocular implants and anterior segment structures based on anterior segment OCT grayscale image recognition, proposes a new image preprocessing workflow, optimizes the speed and accuracy of Mask RCNN for high-resolution anterior segment OCT grayscale image recognition, and simultaneously enables 3D reconstruction and automatic measurement of the spatial volume between the posterior surface of the intraocular implant and the anterior surface of the lens.

Description

Technical Field

[0001] This invention relates to the field of computer image processing technology, and in particular to an optimized method for measuring the volume between an intraocular implant and anterior segment structures. Background Technology

[0002] Anterior segment OCT measurement is a technique that uses optical interferometry for two-dimensional imaging and quantitative analysis based on the different optical scattering properties of ocular tissue structures. Anterior segment OCT is widely used in clinical practice due to its advantages such as non-contact operation, high resolution, high repeatability, and fast image acquisition. Currently, there are methods for partial anterior segment 3D reconstruction based on anterior segment OCT image recognition. However, in practice, due to differences in the anterior segment structure of the patient, poor fixation during scanning, and optical distortion, artifacts and noise appear in the scanned images. Therefore, current anterior segment 3D reconstruction based on anterior segment OCT is mostly limited to natural human structures with high signal intensity, such as the cornea, iris, and lens. Intraocular implants have relatively low signal intensity, making them difficult to distinguish from image artifacts and noise. Furthermore, the position and refractive power of intraocular implants vary greatly among different patients. General computer vision contour extraction methods cannot meet the needs of planar recognition of intraocular implants, let alone accurately identify and reconstruct the 3D structure of intraocular implants. Currently, there is no suitable method for measuring the volume between intraocular implants and anterior segment structures.

[0003] Semantic segmentation refers to the technique of pixel-level classification of objects in a two-dimensional image. Building upon semantic segmentation, a segmentation method that can distinguish different instances of the same type of object is called instance segmentation, which requires object detection as a prerequisite. Currently, there are many neural network architectures capable of both object detection and instance segmentation. The classic network is Mask R-CNN, which has good generalization ability and high final accuracy. The standard Mask R-CNN implementation uses ResNet101 as the backbone feature extraction network and is tested using the COCO dataset. Therefore, the image size entering the network must be an integer multiple of 2 to the power of 6, and the image must have 3 channels (RGB). The high-precision front-end OCT image is used as a grayscale image, with only one grayscale channel. Its high resolution means that copying two more channels to combine into a 3-channel image for the neural network would significantly increase the memory load and time cost. Furthermore, when training the network with a large training set, even if only grayscale data is input... Figure 1The high resolution of the OCT image in the first channel still places a significant burden on the GPU memory, while most pixels offer little benefit to the network's predictions. Directly reducing the size of the original image before feeding it into Mask R-CNN reduces the accuracy of the extracted pixel coordinates. Conventional methods have high hardware requirements and may even fail to converge under certain parameters, making it difficult for ordinary graphics cards to train Mask R-CNN using high-resolution grayscale image datasets. The 3D point cloud reconstruction surface can be fitted using the least squares method, but due to the irregular shape of intraocular implants, it does not conform to the classic quadratic surface equation, resulting in the loss of some feature data when using the quadratic surface equation for fitting. Using a back propagation (BP) neural network for fitting can produce less data loss, but overall, fitting still leads to the loss of information from the original data points. The fitting result needs to satisfy a certain functional relationship, leading to overly idealized results or overfitting.

[0004] In medical research, the need to investigate the spatial relationship between intraocular implants and anterior segment structures, and to acquire anterior segment information using an anterior segment OCT instrument to perform three-dimensional reconstruction of the intraocular implant to explore its volume changes between the implant and the anterior segment structure under different conditions, remains. Designing an optimized method suitable for measuring the volume between intraocular implants and anterior segment structures would facilitate clinicians in more intuitive follow-up and observation of the morphology and position of intraocular implants, significantly improving the efficiency of postoperative assessment of intraocular implants during clinical diagnosis and treatment.

[0005] Therefore, an optimized method is proposed for measuring the volume between intraocular implants and anterior segment structures. Summary of the Invention

[0006] The purpose of this invention is to provide an optimized method for measuring the volume between intraocular implants and anterior segment structures, thereby filling the gap in current clinical and scientific research on the spatial three-dimensional position of intraocular implants.

[0007] To achieve the above objectives, the technical solution of the present invention is as follows:

[0008] An optimized method for measuring the volume between intraocular implants and anterior segment structures is proposed. This method is implemented through a planar image preprocessing module, a planar image anterior segment recognition module, a planar data processing module, a three-dimensional reconstruction module, and a volume measurement module. The optimized method includes the following steps:

[0009] High-resolution front-segment OCT grayscale images were acquired; image preprocessing methods were used to filter artifacts from the front-segment OCT grayscale images; computer vision methods were used to identify the artifact-filtered front-segment OCT grayscale images, determine the regions of interest (ROIs), and adjust the size of the front-segment OCT grayscale images according to the requirements of Mask RCNN. Mask RCNN was trained on a self-annotated OCT image dataset. RCNN identifies regions of interest (ROIs) and, combined with preprocessed binary images, extracts the coordinates of pixels on the posterior surface of the intraocular implant and the anterior surface of the lens. It averages different pixels at the same X-coordinate value in a single image using a pixel averaging method at the same location, integrating data from multiple anterior segment OCT grayscale images. It reconstructs the 3D point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens using coordinate transformation, and generates surface coordinate matrices for the posterior surface and the anterior surface of the lens using linear interpolation, then visualizes the data. It obtains the height difference between each sampling position of the two surfaces by subtracting the corresponding positions of the height coordinate matrix of the posterior surface of the intraocular implant from the height coordinate matrix of the anterior surface of the lens, and measures the volume of the irregular cylinder between the two surfaces using a differential method. Finally, it summarizes and organizes the data and generates corresponding images and / or reports.

[0010] Optionally, the image preprocessing of the preceding OCT grayscale image based on the image processing method to filter artifacts in the preceding OCT grayscale image includes: zero-filling the periphery of the preceding OCT grayscale image, performing convolution based on the Sobel vertical edge detection operator to enhance the edge signal in the vertical direction, and then performing binarization processing, followed by zero-filling the periphery and then performing convolution using the horizontal isolated noise filtering operator, and then performing binarization processing to filter horizontally isolated noise in the preceding OCT grayscale image.

[0011] Optionally, the step of identifying the anterior segment OCT grayscale image after artifact filtering using computer vision methods, determining the region of interest (ROI), and adjusting the size of the anterior segment OCT grayscale image according to the requirements of Mask R-CNN includes: using a pre-cropped image of the posterior corneal surface at the same resolution as a template; locating the perpendicular bisector of the image to the center of the posterior corneal surface using a template matching method based on computer vision; and cropping a rectangular region as the region of interest by taking one-Nth of the image height value to the left and right of this central axis; scaling the image size of the region of interest according to the requirements of Mask R-CNN to obtain a reduced ROI image; and using this ROI image as the input to Mask R-CNN.

[0012] Optionally, the Mask RCNN requires that the length and width of the image be integer multiples of the power of 2 to the sixth power.

[0013] Optionally, the Mask R-CNN trained on a self-labeled OCT image dataset identifies the Region of Interest (ROI), and extracts the coordinates of pixels on the posterior surface of the intraocular implant and the anterior surface of the lens by combining the preprocessed binary image. This includes: manually masking the scaled ROI image using Labelme software, labeling several images as a training set, setting the number of channels for Mask R-CNN, using ResNet50 as the backbone feature extraction network of Mask R-CNN, and inputting the scaled ROI image into the Mask R-CNN dataset. RCNN is used for training to generate small masks for the anterior surface of the lens and the posterior surface of the intraocular implant. The original image mask is scaled to the original ROI image size using bilinear interpolation to obtain an enlarged mask. Then, zero-padding is performed on both sides to restore it to the size of the anterior segment OCT grayscale image to obtain the original image mask. A preprocessed binary image is generated. The original image mask is compared with the preprocessed binary image to obtain the white pixels in the mask area of ​​the original image mask in the preprocessed binary image. The coordinates of the white pixels are extracted as the coordinates of the pixels on the posterior surface of the intraocular implant and the anterior surface of the lens.

[0014] Optionally, the three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens are reconstructed based on the coordinate transformation method, wherein the formula for the coordinate transformation is:

[0015]

[0016] In the formula: θ is the rotation angle, x and y are the plane coordinates before rotation, and x′ and y′ are the plane coordinates after rotation.

[0017] Optionally, the height difference between the two surfaces is obtained by subtracting the corresponding positions of the height coordinate matrix of the posterior surface of the intraocular implant from the height coordinate matrix of the anterior surface of the lens. The volume of the irregular cylinder between the two surfaces is then measured using a differential method, wherein the formula for the differential method is:

[0018]

[0019] In the formula: d is the diameter of the base area, ZL is the height of the sampling position i on the anterior surface of the lens, ZI is the height of the intraocular implant at the sampling position i, i is the sampling position, and ILV is the volume between the two curved surfaces.

[0020] The beneficial effects of this invention are as follows:

[0021] This invention fully considers the problem of low efficiency of Mask RCNN in processing high-resolution front-segment OCT grayscale images and optimizes the recognition results. Its beneficial effects are specifically reflected in the following aspects:

[0022] (1) This invention requires no human intervention, is a fully automatic measurement process, eliminates human interference, and has strong repeatability of results;

[0023] (2) The measurement process of the present invention has low requirements for computers; ordinary graphics cards can be used and trained without the need for a deep learning workstation.

[0024] (3) The present invention designs the configuration and mask generation mode of Mask RCNN, which improves the efficiency of Mask RCNN in recognizing high-resolution front-segment OCT grayscale images and generating high-resolution masks.

[0025] (4) The present invention uses two convolution and binarization methods to filter noise and artifacts in the original anterior segment OCT grayscale image, and fully extracts the contour features of the intraocular implant and its own ocular structure.

[0026] (5) This invention fills the gap in the prior art regarding the fully automated measurement of the volume between intraocular implants and anterior segment structures based on anterior segment OCT grayscale image recognition. The optimization method of this invention proposes a new image preprocessing process to optimize the speed and accuracy of Mask RCNN for high-resolution anterior segment OCT grayscale image recognition.

[0027] (6) The present invention uses the pixel averaging method at the same position to further eliminate the influence of noise on the measurement results, uses the linear interpolation method to sample the three-dimensional point cloud data at equal intervals to generate a surface matrix, and uses the differential method to calculate the volume between the intraocular implant and the anterior segment structure, which has the advantages of high accuracy and fast speed. Attached Figure Description

[0028] Figure 1 In this embodiment of the invention, a high-resolution anterior segment OCT grayscale image is obtained, wherein 1-the boundary between the artifact light stripe and the posterior surface of the intraocular implant, and 2-noise near the posterior surface of the intraocular implant.

[0029] Figure 2 The above is the result and mask of the anterior segment OCT grayscale image after Mask RCNN recognition in the embodiment of the present invention, wherein ICL is an intraocular implant and LENS is the anterior surface of the lens;

[0030] Figure 3 This is the image after convolution using the Sobel vertical edge detection operator in this embodiment of the invention;

[0031] Figure 4 This is the image after the first binarization in this embodiment of the invention;

[0032] Figure 5 This is the image after convolution using the horizontal isolated noise filtering operator in this embodiment of the invention;

[0033] Figure 6 This is the image after the second binarization in this embodiment of the invention;

[0034] Figure 7 This is a flowchart of step S102 in an embodiment of the present invention;

[0035] Figure 8 This is an image in this embodiment of the invention, in which the recognition results are marked on the original anterior segment OCT grayscale image, wherein 1-positioning frame at the top of the posterior corneal surface, 2-rotation axis positioning line (Z-axis), 3-recognition result of the posterior surface of the intraocular implant, and 4-recognition result of the anterior surface of the lens;

[0036] Figure 9 This is an image representing the result of restoring the point cloud data of 18 anterior segment OCT grayscale images of the lens anterior surface in three-dimensional space in an embodiment of the present invention;

[0037] Figure 10 This is a screenshot of the coordinates of pixels identified from a single front-segment OCT grayscale image in an embodiment of the present invention.

[0038] Figure 11 This is a screenshot of the average value of the Y values ​​of pixels with the same X value identified from a single front-segment OCT grayscale image in an embodiment of the present invention.

[0039] Figure 12 This is a partial screenshot showing the result of generating the three-dimensional coordinates of each pixel using a coordinate transformation matrix in an embodiment of the present invention.

[0040] Figure 13 This is a visualization result of the anterior surface matrix of the lens generated using interpolation in an embodiment of the present invention.

[0041] Figure 14 This is a visualization result of the surface matrix of the intraocular implant generated using interpolation in an embodiment of the present invention.

[0042] Figure 15 This is a topographic map of the posterior surface matrix of the intraocular implant and the anterior surface matrix of the lens in an embodiment of the present invention, wherein, Figure 15-1 This is a topographic map of the surface matrix behind an intraocular implant. Figure 15-2 This is a topographic map of the anterior surface matrix of the lens. Figure 15-1 and Figure 15-2 The central triangle represents the central position. Figure 15-1 and Figure 15-2 The middle circle represents the highest position;

[0043] Figure 16This is a topographic map showing the difference between the two matrices and the calculated volume result in an embodiment of the present invention;

[0044] Figure 17 This is a flowchart of an optimized method for measuring the volume between an intraocular implant and an anterior segment structure in an embodiment of the present invention. Detailed Implementation

[0045] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. 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.

[0046] The present invention relates to an optimized method for measuring the volume between an intraocular implant and the anterior segment structure. The method utilizes anterior segment OCT to scan the anterior segment of the subject (the anterior segment OCT scanning probe rotates 170 degrees around the scanning axis, and a scan is performed every 10 degrees) to generate 18 anterior segment OCT grayscale images. The OCT images are preprocessed using a specific image processing method, and a high-resolution OCT grayscale image mask is generated using a specific mask generation method. The method also combines mathematical methods to automatically identify, reconstruct, and measure the volume between the posterior surface of the intraocular implant and the anterior surface of the lens in the images.

[0047] The technical solution of this invention is described in detail below:

[0048] An optimized method for measuring the volume between intraocular implants and anterior segment structures is proposed. This optimization method is implemented through a planar image preprocessing module, a planar image anterior segment recognition module, a planar data processing module, a three-dimensional reconstruction module, and a volume measurement module.

[0049] Furthermore, the planar image preprocessing module of the present invention is used to filter artifacts in high-resolution anterior segment OCT grayscale images, enhance the signals of intraocular implants and the anterior surface of the lens, determine the region of interest (ROI) for the anterior segment recognition module of the planar image, and improve recognition efficiency.

[0050] Furthermore, the anterior segment recognition module for planar images involved in this invention is used to identify regions of interest (ROIs) using Mask RCNN trained on a self-labeled OCT image dataset, and to extract the coordinates of pixels on the posterior surface of intraocular implants and the anterior surface of the lens by combining the preprocessed binary image.

[0051] Furthermore, the planar data processing module of the present invention is used to average different pixels at the same X coordinate value in a single image and to integrate data from multiple previous OCT grayscale images.

[0052] Furthermore, the three-dimensional reconstruction module of the present invention is used to realize the reconstruction of three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens based on the coordinate transformation method, and to generate the surface coordinate matrix of the posterior surface of the intraocular implant and the anterior surface of the lens through linear interpolation, and to visualize the data.

[0053] Furthermore, the volume measurement module of the present invention is used to obtain the height difference between each sampling position between the two surfaces by subtracting the corresponding positions of the height coordinate matrix of the posterior surface of the intraocular implant from the height coordinate matrix of the anterior surface of the lens, and to measure the volume of the irregular cylinder between the two surfaces based on the differential method; the data is summarized and organized and corresponding images and / or reports are generated.

[0054] Furthermore, the development environment setup involved in this invention includes: using Python 3.6, Tensorflow 2.4.0, Keras 2.4.3, CUDA 11.4, CPU: Intel Core i7-6700K, graphics card: NVIDIA GeForce RTX 3080Ti to build a deep learning platform and write process code.

[0055] For details, please refer to the appendix. Figure 17 The optimization method involved in this invention includes the following steps:

[0056] S101. Obtain a high-resolution front-segment OCT grayscale image.

[0057] Further, see attached document. Figure 1 This invention reads 18 high-resolution anterior segment OCT grayscale images. In this embodiment, the resolution of a single anterior segment OCT grayscale image is 4002×1805.

[0058] S102. Based on image processing methods, perform image preprocessing on the front-segment OCT grayscale image to filter out artifacts in the front-segment OCT grayscale image; identify the front-segment OCT grayscale image after artifact filtering using computer vision methods, determine the region of interest (ROI), and adjust the size of the front-segment OCT grayscale image according to the requirements of Mask RCNN.

[0059] For details, please refer to the appendix. Figure 2-7The image preprocessing method for the preceding OCT grayscale image, which filters artifacts from the preceding OCT grayscale image, includes: zero-filling the periphery of the preceding OCT grayscale image, performing convolution based on the Sobel vertical edge detection operator to enhance the edge signal in the vertical direction, and then performing binarization processing. Subsequently, after zero-filling the periphery, convolution is performed using the horizontal isolated noise filtering operator, followed by binarization processing to filter out isolated noise points in the horizontal direction in the preceding OCT grayscale image.

[0060] In one embodiment, since the anterior segment OCT grayscale image often has an artifact band in the vertical direction, and this band is often connected to the posterior surface of the intraocular implant, an image processing method is needed to disconnect the potential artifact band from the posterior surface of the intraocular implant. This ensures that when extracting pixels from the posterior surface of the intraocular implant, pixels from the artifact band are not extracted simultaneously. The specific method is as follows:

[0061] After zero-padding the grayscale image (this operation aims to ensure that the image size remains unchanged before and after convolution), the Sobel vertical edge detection operator is used for convolution to enhance the edge signal in the vertical direction. Then, the image is binarized (the threshold for binarization is 55, that is, the brightness value of pixels with a brightness greater than this threshold is set to 255, and the brightness value of pixels with a brightness lower than this threshold is set to 0). After zero-padding the perimeter, the horizontal isolated noise filtering operator is used for convolution. The convolution kernel size is 5×5, which aims to reduce the brightness of isolated pixels in the horizontal direction (specifically, the pixel with few left and right adjacent pixels). Then, binarization is performed to filter out pixels with a brightness (grayscale value) lower than (255 / 50)×20, thus filtering out horizontal isolated noise in the previous OCT grayscale image.

[0062] The Sobel vertical edge detection operator involved in this invention is:

[0063]

[0064] The convolution kernel of the horizontal isolated noise filtering operator involved in this invention is:

[0065]

[0066] Specifically, the step of identifying the anterior segment OCT grayscale image after artifact filtering using computer vision methods, determining the region of interest (ROI), and adjusting the size of the anterior segment OCT grayscale image according to the requirements of Mask R-CNN includes: using a pre-cropped image of the posterior corneal surface at the same resolution as a template; locating the perpendicular bisector of the image to the center of the posterior corneal surface using a template matching method based on computer vision; and cropping a rectangular region as the region of interest by taking N parts of the image height to the left and right of this central axis; scaling the image size of the region of interest according to the requirements of Mask R-CNN to obtain a reduced ROI image; and using this ROI image as the input to Mask R-CNN; the requirements of Mask R-CNN are that the length and width of the image are both integer multiples of the power of 2.

[0067] In one embodiment, the present invention uses computer vision methods to identify each image and find the region of interest (ROI). The specific operations include:

[0068] Using a pre-extracted image of the posterior corneal surface at the same resolution as a template, the vertical axis of the image is located at the center of the posterior corneal surface using a template matching method based on computer vision (in this embodiment, the vertical axis position is w = 2001). With this axis as the center, half of the image height value is cropped to the left and right (902 in this embodiment) to extract a rectangular area (a square area in this embodiment) as the region of interest (ROI) (size is 1805×1805 in this embodiment). Based on the local mean method, the ROI is scaled down to a reduced ROI image of size 1024×1024 as the input of Mask RCNN.

[0069] The reason for cropping the region of interest is that the resolution of the anterior segment OCT grayscale image is too high, and directly inputting it into Mask R-CNN would cause a single image to occupy a large amount of video memory, which is not conducive to training Mask R-CNN with a large number of datasets. The signal intensity of the posterior corneal surface on the anterior segment OCT grayscale image is extremely high. The lens and intraocular implants are located directly below the posterior corneal surface. Removing the rest of the OCT image and keeping only the region we need can greatly improve the efficiency of training and prediction.

[0070] In addition, since Mask R-CNN requires that the length and width of the image must be an integer multiple of 2 to the power of 6, the size of the input image in this embodiment is 1024×1024. 1024 is 16 times 64, which meets the requirement. Through the above method, ROI images can be pre-selected for the neural network. When they are input into Mask R-CNN, they are all valid data and no zero padding is required.

[0071] S103. Based on the self-labeled OCT image dataset, Mask RCNN is used to identify the region of interest (ROI). Combined with the preprocessed binary image, the coordinates of the pixels on the posterior surface of the intraocular implant and the anterior surface of the lens are extracted.

[0072] Specifically, the Mask R-CNN trained on a self-labeled OCT image dataset identifies Regions of Interest (ROIs). Combined with preprocessed binary images, it extracts the coordinates of pixels on the posterior surface of the intraocular implant and the anterior surface of the lens. This process includes: manually annotating scaled ROI images using Labelme software, using several annotated images as a training set, setting the number of Mask R-CNN channels, employing ResNet50 as the backbone feature extraction network of Mask R-CNN, and inputting the scaled ROI images into Mask R-CNN for training to generate small masks for the anterior surface of the lens and the posterior surface of the intraocular implant. Bilinear interpolation is used to scale the original image mask to the original ROI image size to obtain a magnified mask. Left and right zero-padding is then applied to restore the mask to the size of the anterior segment OCT grayscale image to obtain the original image mask. A preprocessed binary image is generated. The original image mask is compared with the preprocessed binary image to obtain white pixels within the masked area of ​​the original image mask in the preprocessed binary image. The coordinates of the white pixels are extracted as the coordinates of pixels on the posterior surface of the intraocular implant and the anterior surface of the lens.

[0073] In one embodiment, reference is made to the appendix. Figure 8 This invention uses Labelme software to manually annotate the scaled-down ROI images, labeling a total of 100 images as the training set. The Mask R-CNN channel number is set to 1, and ResNet50 is used as the backbone feature extraction network. This neural network is trained with 575 EPOCHs to accurately generate masks for the posterior surface of intraocular implants and the anterior surface of the lens. Key configurations are as follows:

[0074] class EyeFrontConfig(mrcnn.config.Config):

[0075] NAME = "EyeFront"

[0076] GPU_COUNT = 1

[0077] IMAGES_PER_GPU = 1

[0078] BACKBONE = "resnet50"

[0079] NUM_CLASSES=3

[0080] MEAN_PIXEL=np.array([123.7])

[0081] IMAGE_CHANNEL_COUNT = 1

[0082] RPN_ANCHOR_SCALES=(32,64,-128,-256,512)

[0083] IMAGE_MIN_DIM = 1024

[0084] IMAGE_MAX_DIM = 1024

[0085] TRAIN_ROIS_PER_IMAGE = 100

[0086] MAX_GT_INSTANCES = 10

[0087] Furthermore, the scaled ROI image dataset is fed into Mask RCNN for training, enabling it to accurately generate masks for the anterior surface of the lens and the posterior surface of intraocular implants.

[0088] In this embodiment, the model is used to predict the scaled-down ROI image in the above figure to generate a mask for the posterior surface of the eye implant and a small mask for the anterior surface of the lens. The mask image is a binary image, with the part outside the mask that does not belong to the object having a value of 0, and the part inside the mask that belongs to the object having a value of 1.

[0089] Furthermore, the mask image is enlarged back to the original ROI image size (i.e., 1805×1805) using bilinear interpolation, and then zero-padding is applied to restore it to the size of the front OCT grayscale image (4002×1805). Through the above operations, the efficiency of Mask RCNN in generating high-resolution front OCT grayscale image masks can be improved.

[0090] Furthermore, the mask generated in the above steps further narrows the range of pixels to be extracted, which can prevent noise outside the mask from affecting the pixel extraction results.

[0091] Furthermore, the original image mask is compared with the preprocessed binary image to identify the white pixels (brightness value equal to 255) within the mask area of ​​the original image mask in the preprocessed binary image. The relative coordinates (relative to the OCT scanning rotation axis) of these white pixels and the angle of the image are then saved in a file. Through the above image preprocessing and recognition methods, the coordinates of the pixels we need to identify in the high-resolution front-segment OCT grayscale image can be efficiently read.

[0092] S104. The pixel averaging method based on the same location averages different pixels at the same X coordinate value in a single image and integrates the data from multiple previous OCT grayscale images.

[0093] In one embodiment, reference is made to the appendix. Figure 10-11 In each image, on the same plane, there may be multiple pixels with the same X coordinate value. Therefore, it is necessary to average these pixels. The pixel averaging method at the same location is used to average different pixels with the same X coordinate value in a single image to ensure that only one height value (Z value) corresponds to one X coordinate value. This step can reduce the error in the 3D reconstruction effect caused by too many pixels to a certain extent. After averaging, the results of the 18 previous OCT grayscale images are merged and stored.

[0094] S105. Based on the coordinate transformation method, restore the three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens, and generate the surface coordinate matrix of the posterior surface of the intraocular implant and the anterior surface of the lens through linear interpolation, and visualize the data.

[0095] Among them, reference appendix Figure 9 Appendix Figure 12-15 This invention uses a coordinate transformation matrix to restore discrete points on different planes to their corresponding positions in three-dimensional space, generating two sets of three-dimensional point cloud data (one set is the posterior surface of the intraocular implant, and the other set is the anterior surface of the lens). Linear interpolation is used to perform equidistant sampling of the surface to generate a coordinate matrix with a size of 200×200. Each set of surfaces generates a three-dimensional array (with a shape of [3, 200, 200], which stores the X, Y, and Z values ​​of each point in the matrix, and the data is visualized.

[0096] Specifically, the method of reconstructing the three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens based on coordinate transformation, wherein the formula for coordinate transformation is:

[0097]

[0098] In the formula: θ is the rotation angle, x and y are the plane coordinates before rotation, and x′ and y′ are the plane coordinates after rotation.

[0099] In addition, the z-coordinate used in this invention represents height and is not transformed.

[0100] S106. By subtracting the corresponding positions of the height coordinate matrix of the posterior surface of the intraocular implant from the height coordinate matrix of the anterior surface of the lens, the height difference between each sampling position between the two surfaces is obtained. The volume of the irregular cylinder within a 2.5mm radius between the two surfaces is measured based on the differential method.

[0101] For details, please refer to the appendix. Figure 16 The height difference between the two surfaces is obtained by subtracting the height coordinate matrix of the posterior surface of the intraocular implant from the height coordinate matrix of the anterior surface of the lens at corresponding positions. The volume (ILV) of the irregular cylinder between the two surfaces is then measured using a differential method, where the formula for the differential method is:

[0102]

[0103] In the formula: d is the diameter of the base area, ZL is the height of the sampling position i on the anterior surface of the lens, ZI is the height of the intraocular implant at the sampling position i, i is the sampling position, and ILV is the volume between the two curved surfaces.

[0104] The volume (ILV) involved in this invention is: the measurement of the volume of an irregular cylinder within a radius of 2.5 mm between two curved surfaces based on the differential method, that is, d is the diameter of the base area of ​​5 mm, specifically the radius 2.5 × 2 = 5 mm.

[0105] S107. Summarize and organize the data and generate corresponding images and / or reports.

[0106] Furthermore, the present invention relates to summarizing and organizing data and generating corresponding images and / or reports, namely, the generated images and / or reports include, but are not limited to: generating a topographic map of the posterior surface of the intraocular implant, a topographic map of the anterior surface of the lens, a topographic map of the height difference between the two, a three-dimensional image of the posterior surface of the intraocular implant, a three-dimensional image of the anterior surface of the lens, a three-dimensional image of the height difference, volume data between the posterior surface of the intraocular implant and the anterior surface of the lens, and a height difference data report, etc.

[0107] In summary, this invention discloses an optimized method for measuring the volume between intraocular implants and anterior segment structures. It enables fully automated 3D reconstruction and measurement of the spatial volume between the posterior surface of the intraocular implant and the anterior surface of the lens. This method addresses the problems of high memory usage and long training time caused by conventional methods that input high-resolution grayscale images into Mask R-CNN by using template matching to identify the region of interest in the center of the posterior corneal surface, scaling the image, and applying a mask. This method effectively reduces the memory usage of Mask R-CNN and eliminates noise and longitudinal artifacts in high-resolution OCT images through double convolution and binarization. It achieves precise pixel extraction between the posterior surface of the intraocular implant and the anterior surface of the lens, as well as fully automated 3D reconstruction of the anterior segment space, thus completing the quantitative measurement of the volume between the posterior surface of the intraocular implant and the anterior surface of the lens.

[0108] In this embodiment, only grayscale images of a specific resolution are selected to illustrate the experimental concept and features of the present invention. The purpose is to enable those skilled in the art to understand the content of the present invention and implement it accordingly. The scope of protection of the present invention is not limited to the above embodiment. The fully automated measurement process based on this invention includes, but is not limited to, measuring the volume between intraocular implants and other structures in the anterior segment of the eye. Therefore, all equivalent changes or modifications made based on the principles and experimental ideas disclosed in this invention are within the scope of protection of this invention.

Claims

1. An optimized method for measuring the volume between an intraocular implant and anterior segment structures in the human eye, characterized in that, This optimization method is implemented through a planar image preprocessing module, a planar image anterior segment recognition module, a planar data processing module, a 3D reconstruction module, and a volume measurement module. The optimization method includes the following steps: acquiring a high-resolution anterior segment OCT grayscale image; zero-filling the periphery of the anterior segment OCT grayscale image, performing convolution based on the Sobel vertical edge detection operator to enhance the vertical edge signal, followed by binarization; then, after periphery zero-filling, performing convolution using a horizontal isolated noise filtering operator, followed by binarization to filter horizontally isolated noise in the anterior segment OCT grayscale image; using a pre-cropped corneal posterior surface image of the same resolution as a template, locating the vertical axis of the image to the center of the corneal posterior surface using a computer vision-based template matching method, and using this central axis as the center, cropping N parts of the image height value to the left and right to extract a rectangular region as the region of interest (ROI); scaling the ROI image size according to the requirements of Mask RCNN to obtain a reduced ROI image, and using this ROI image as the Mask RCNN. The input to RCNN is as follows: Labelme software is used to manually annotate the scaled ROI images, and several annotated images are used as the training set. The number of channels in Mask RCNN is set, and ResNet50 is used as the backbone feature extraction network. The scaled ROI images are then input into Mask RCNN for training, generating small masks for the anterior surface of the lens and the posterior surface of the intraocular implant. Bilinear interpolation is used to scale the original image mask to the original ROI image size to obtain a magnified mask. Zero-padding is then applied to restore the mask to the size of the anterior segment OCT grayscale image, generating a preprocessed binary image. The original image mask is compared with the preprocessed binary image to obtain white pixels within the masked area of ​​the original image mask in the preprocessed binary image. The coordinates of these white pixels are extracted as the coordinates of pixels on the posterior surface of the intraocular implant and the anterior surface of the lens. A pixel averaging method based on the same location is used to average the coordinates of the pixels. Data from multiple anterior segment OCT grayscale images were integrated, focusing on different pixels at the same X-coordinate value from a single image. Three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens were reconstructed using coordinate transformation. Linear interpolation was used to generate surface coordinate matrices for the posterior and anterior surfaces of the intraocular implant, and the data was visualized. The height difference between each sampling position of the two surfaces was obtained by subtracting the corresponding positions of the height coordinate matrix of the posterior and anterior surfaces of the intraocular implant. The volume of the irregular cylinder between the two surfaces was measured using a differential method. The data were then summarized and processed to generate corresponding images and / or reports. The horizontal isolated noise filtering operator is a 5×5 convolution kernel with a center weight of 8 and decreasing weights around the perimeter.

2. The optimized method for measuring the volume between an intraocular implant and the anterior segment structures of the human eye according to claim 1, wherein, The Mask R-CNN requires that the length and width of the image be integer multiples of the power of 2 to the sixth power.

3. The optimized method for measuring the volume between an intraocular implant and anterior segment structures according to claim 1, characterized in that, The method described above reconstructs the three-dimensional point cloud data of the posterior surface of the intraocular implant and the anterior surface of the lens using coordinate transformation. The formula for the coordinate transformation is as follows: In the formula: θ is the rotation angle, x and y are the plane coordinates before rotation, and x′ and y′ are the plane coordinates after rotation.

Images