Macular hole detection method

Abstract

The application relates to a macular hole detection method, which comprises the following steps: firstly, obtaining an OCT image of a macular region, performing extension processing on the OCT image to obtain a plurality of extension images, and increasing the diversity of image samples; then, extracting multi-channel feature data of the original image and the extension images, determining suspected lesion features by using the features of the original image, cross- verifying and strengthening the suspected lesion features by means of the features of the extension images, and obtaining accurate key lesion features; finally, calculating macular hole parameters based on the key lesion features, obtaining a detection result, and realizing efficient and accurate macular hole detection. The application combines multi-image extension, multi-channel feature extraction and cross-verification and strengthening, and improves the precision and automation degree of macular hole detection.

Description

Technical Field

[0001] This application relates to the field of macular hole detection technology, and in particular to a method for macular hole detection. Background Technology

[0002] Macular hole is a common lesion in the macular region of the fundus, causing symptoms such as decreased central vision and distorted vision. Accurate detection of its key parameters is crucial for clinical diagnosis and treatment planning. Currently, clinical practice largely relies on optical coherence tomography (OCT) images for macular hole detection, using core parameters such as the minimum diameter of the hole and choroidal thickness to assess the severity of the lesion.

[0003] Existing detection methods fall into two categories: manual detection and semi-automatic algorithm-assisted detection. Manual detection relies on physician experience to locate feature points, resulting in low efficiency and strong subjectivity. Measurement errors between different physicians are significant, leading to low detection accuracy and efficiency, making it difficult to meet the needs of large-scale clinical screening. Semi-automatic algorithms are mostly based on single feature extraction models, but are often affected by image noise and blurred lesion boundaries, resulting in insufficient feature extraction accuracy, requiring professional physician correction. Therefore, there is an urgent need for a more accurate and automated macular hole detection solution to address the problems of low efficiency and insufficient accuracy in existing technologies. Summary of the Invention

[0004] Based on this, the purpose of this application is to provide a method for detecting macular holes, which improves the accuracy and automation of macular hole detection by combining multi-image expansion, multi-channel feature extraction and cross-validation enhancement.

[0005] The method for detecting macular holes according to embodiments of this application includes the following steps: Acquire OCT images of the macular region of the eye to be examined; The OCT image of the macular region is expanded to obtain several expanded images; Extract the multi-channel feature data of the macular region OCT image and each of the extended images; determine suspected lesion features based on the multi-channel feature data of the macular region OCT image; combine the multi-channel feature data of each of the extended images to perform feature cross-validation and enhancement on the suspected lesion features to obtain key lesion features; The macular hole parameters are calculated based on the key lesion features; the macular hole detection results are obtained based on the macular hole parameters.

[0006] In this embodiment, after acquiring the OCT image of the macular region of the eye to be detected, it is expanded to generate several expanded images, enriching the diversity of image samples and providing a more comprehensive data foundation for subsequent accurate analysis. Next, multi-channel feature data of the macular region OCT image and each expanded image are extracted. These multi-channel features encompass rich details of the image and can reflect the condition of the macular region from multiple dimensions. After identifying suspected lesion features based on the macular region OCT image, cross-validation and enhancement are performed using the expanded image features. This effectively eliminates interference information caused by image noise, blurred lesion boundaries, and other factors, strengthening the correct lesion features and thus accurately identifying key lesion features, greatly improving the accuracy and reliability of feature extraction. Finally, macular hole parameters are calculated based on the key lesion features to obtain the detection results. The entire process is highly automated, avoiding subjective errors caused by differences in physician experience in manual measurement, greatly improving detection efficiency, and enabling rapid processing of large numbers of samples to meet the needs of large-scale clinical screening. At the same time, precise feature extraction and parameter calculation provide highly accurate data for clinical diagnosis, helping doctors to more scientifically assess the degree of lesions and formulate more reasonable treatment plans, thereby improving the treatment effect and quality of life of patients. Overall, it has brought positive and significant changes to the clinical detection and treatment of macular holes.

[0007] To better understand and implement this application, the following detailed description is provided in conjunction with the accompanying drawings. Attached Figure Description

[0008] Figure 1 This is a schematic flowchart of the macular hole detection method according to an embodiment of this application; Figure 2 This is a schematic diagram illustrating the steps of expanding the OCT image of the macular region in an embodiment of this application; Figure 3 This is a schematic diagram illustrating the steps of grayscale enhancement processing of the OCT image of the macular region in an embodiment of this application. Detailed Implementation

[0009] To make the objectives, technical solutions, and advantages of this application clearer, the embodiments of this application will be described in further detail below with reference to the accompanying drawings. Wherein, when the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements.

[0010] It should be understood that the embodiments described below do not represent all embodiments consistent with this application. Rather, they are merely examples of apparatuses and methods consistent with some aspects of this application as detailed in the appended claims. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of this application.

[0011] The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to be limiting of the application. The singular forms “a,” “the,” and “the” used in this application are also intended to include the plural forms unless the context clearly indicates otherwise. Furthermore, in the description of this application, unless otherwise stated, “a plurality” means two or more. It should also be understood that the term “and / or” as used herein refers to and includes any or all possible combinations of one or more associated listed items, for example, A and / or B, which can represent: A alone, A and B together, and B alone; the character “ / ” generally indicates that the preceding and following objects are in an “or” relationship.

[0012] It should be understood that although the terms first, second, third, etc., may be used in this application to describe various information, this information should not be limited to these terms, and these terms are only used to distinguish similar objects, and are not necessarily used to describe a specific order or sequence, nor should they be construed as indicating or implying relative importance. Those skilled in the art can understand the specific meaning of the above terms in this application according to the specific circumstances. Depending on the context, the word "if" as used in this application can be interpreted as "when," "when," or "in response to determination."

[0013] This application provides a method for detecting macular holes, which improves the accuracy and automation of macular hole detection by combining multi-image expansion, multi-channel feature extraction, and cross-validation enhancement.

[0014] Please refer to Figure 1 The method for detecting macular holes according to embodiments of this application includes the following steps: S101: Acquire OCT images of the macular region of the eye to be detected; S102: Perform expansion processing on the OCT image of the macular region to obtain several expanded images; S103: Extract the multi-channel feature data of the macular region OCT image and each of the extended images; determine suspected lesion features based on the multi-channel feature data of the macular region OCT image, and perform feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features; S104: Calculate the macular hole parameters based on the key lesion features; obtain the macular hole detection results based on the macular hole parameters.

[0015] In this embodiment, after acquiring the OCT image of the macular region of the eye to be detected, it is expanded to generate several expanded images, enriching the diversity of image samples and providing a more comprehensive data foundation for subsequent accurate analysis. Next, multi-channel feature data of the macular region OCT image and each expanded image are extracted. These multi-channel features encompass rich details of the image and can reflect the condition of the macular region from multiple dimensions. After identifying suspected lesion features based on the macular region OCT image, cross-validation and enhancement are performed using the expanded image features. This effectively eliminates interference information caused by image noise, blurred lesion boundaries, and other factors, strengthening the correct lesion features and thus accurately identifying key lesion features, greatly improving the accuracy and reliability of feature extraction. Finally, macular hole parameters are calculated based on the key lesion features to obtain the detection results. The entire process is highly automated, avoiding subjective errors caused by differences in physician experience in manual measurement, greatly improving detection efficiency, and enabling rapid processing of large numbers of samples to meet the needs of large-scale clinical screening. At the same time, precise feature extraction and parameter calculation provide highly accurate data for clinical diagnosis, helping doctors to more scientifically assess the degree of lesions and formulate more reasonable treatment plans, thereby improving the treatment effect and quality of life of patients. Overall, it has brought positive and significant changes to the clinical detection and treatment of macular holes.

[0016] The macular hole detection method of this application uses a computer as the execution entity. The following provides a detailed description of each step.

[0017] For step S101, acquire an OCT image of the macular region of the eye to be detected.

[0018] Among them, OCT images, or Optical Coherence Tomography (OCT) images, are a technique that uses the principle of light interference to perform high-resolution imaging of biological tissues. In the field of ophthalmology, they are widely used for imaging structures such as the retina. They can clearly show the fine structure of the macular region and provide important evidence for the diagnosis of macular holes.

[0019] This step uses specialized OCT imaging equipment, such as a spectral-domain optical coherence tomography (SD-OCT) scanner, to scan the patient's eye. This device generates high-resolution OCT images of the macular region by emitting and receiving light signals, utilizing the principle of light interference. These images contain detailed structural information about the macular region, such as the morphology and thickness of the various layers of the retina, and form the basis for subsequent detection and analysis.

[0020] In step S102, the OCT image of the macular region is expanded to obtain several expanded images.

[0021] Image processing algorithms are used to expand the original macular region OCT image. In one embodiment, the expansion process includes image enhancement, image segmentation, and image rotation. Image enhancement involves processing the original OCT image using specific algorithms to improve image quality. For example, the dual-branch lightweight grayscale image enhancement algorithm (IEA) is specifically designed for single-channel grayscale images, enabling pixel-level precision enhancement and suitable for scenarios with high detail preservation requirements, such as medical imaging. This algorithm can suppress background interference, highlight lesion areas, and make the macular hole features in the image clearer, facilitating subsequent analysis. Image segmentation uses image segmentation algorithms, such as Mask R-CNN, to accurately segment the lesion area in the macular region OCT image, thereby generating segmented images containing only the specific region of the hole. For example, masking the left and right sides of the macular hole separately yields a segmented image containing only the left side of the hole (SIOL) and a segmented image containing only the right side of the hole (SIOR), increasing the data volume to three times the original size and providing more focused image data for subsequent feature analysis. Image rotation involves rotating the original OCT image at different angles. Rotation can simulate viewing the macular region from different perspectives, increasing the diversity of image samples. For example, rotating an image by a certain angle may make previously less obvious lesion features more prominent from the new perspective, helping to discover lesion information from different angles, providing more references for subsequent feature extraction, and improving the comprehensiveness and accuracy of detection.

[0022] Please refer to Figure 2 In one embodiment, step S102, which involves expanding the OCT image of the macular region to obtain several expanded images, includes: Step S1021: Perform grayscale enhancement processing on the OCT image of the macular region to obtain an enhanced OCT image of the macular region.

[0023] Gray-scale enhancement refers to the process of adjusting the gray values ​​of an image using specific algorithms to improve its visual effect and enhance useful information. In medical images, gray-scale enhancement can highlight the features of lesion areas, making image details clearer and facilitating subsequent analysis and processing.

[0024] Step S1022: Masking is performed on the left and right regions of the hole in the enhanced macular region OCT image to obtain the left segmented image and the right segmented image of the hole.

[0025] Masking is an image processing technique that extracts or performs specific operations on a target region from an image by creating a mask with the same shape as the target region. In macular hole detection, masking can extract the left and right regions of the hole separately, forming independent segmented images. In one embodiment, an instance segmentation model is used to perform masking on the left and right regions of the hole in the enhanced macular region OCT image. An instance segmentation model is a computer vision model that can not only identify different categories of objects in an image but also perform precise pixel-level segmentation of each object instance. In macular hole detection, the instance segmentation model can accurately segment the left and right regions of the hole from the image, providing more accurate data for subsequent analysis.

[0026] This step extracts the left and right regions of the crack separately, creating independent segmented images of the left and right sides of the crack. These segmented images more effectively display the features on both sides of the crack, avoiding interference from information from other regions, and facilitating subsequent accurate analysis and parameter measurement of the crack's features.

[0027] Step S1023: Perform diagonal flipping on the enhanced macular region OCT image to obtain the corresponding flipped derivative image.

[0028] Diagonal flipping is an image transformation method that flips an image along its diagonal to generate a new image. This process can increase image diversity, enrich image feature information, and help improve the accuracy and robustness of subsequent detection.

[0029] Step S1024: Based on the enhanced macular region OCT image, the segmented image on the left side of the macula, the segmented image on the right side of the macula, and the flipped derived image, the plurality of extended images are obtained.

[0030] This embodiment significantly improves image quality by performing grayscale enhancement processing on OCT images of the macular region, highlighting the features of the lesion area and providing a clearer image foundation for subsequent processing. An instance segmentation model is used to mask the left and right sides of the tear in the enhanced image, resulting in a segmented image that more centrally displays the features on both sides of the tear, avoiding interference from information from other regions and facilitating accurate analysis of the features on both sides of the tear. Combining the enhanced image, the segmented image, and their inverted derivative images enriches the diversity of image samples, showcasing image features from different perspectives, enabling subsequent detection algorithms to capture lesion information more comprehensively.

[0031] Please refer to Figure 3In one embodiment, step S1021, which involves performing grayscale enhancement processing on the macular region OCT image to obtain an enhanced macular region OCT image, includes: Step S10211: Normalize the OCT image of the macular region to obtain a normalized single-channel grayscale image; obtain the maximum grayscale value of the normalized single-channel grayscale image and generate the corresponding normalized single-value feature.

[0032] Normalization, in particular, scales the pixel values ​​of an image to a specific range (e.g., 0 to 1) to eliminate differences between different images caused by factors such as brightness and contrast, making subsequent processing more stable and reliable. For OCT images of the macular region, normalization can highlight details and reduce noise interference. The maximum grayscale value of the normalized single-channel grayscale image is obtained and normalized into a single-value feature. This single-value feature represents the brightest pixel value in the image, reflecting the overall brightness information and providing an important reference for subsequent feature fusion.

[0033] Step S10212: Perform multi-stage convolution, pooling, and discarding operations on the normalized single-channel grayscale image to extract spatial layering features.

[0034] Multi-stage convolution operations extract local features from the image by sliding convolution kernels of different sizes across the image. As the convolutional layers deepen, the extracted features gradually transition from basic texture features to regional structural features and global high-level features. Pooling operations (such as max pooling) are used to reduce the dimensionality of the feature map, reducing computational cost while retaining important feature information. Dropout is a regularization technique that randomly discards a portion of neurons to prevent overfitting and improve the model's generalization ability. Through these operations, multi-level spatial features of the macular region OCT image can be extracted. In this embodiment, the spatially layered features include basic texture features, regional structural features, and global high-level features, providing rich information for subsequent feature fusion.

[0035] Step S10213: Perform two fully connected operations on the normalized single-valued features, map them to a preset dimension feature space, and then perform upsampling to obtain threshold features that are aligned with the size of the spatial hierarchical features.

[0036] Two fully connected operations map the normalized single-valued features to a feature space of a preset dimension, enabling effective fusion with other spatial features. The upsampling operation maps the low-dimensional features back to a high-dimensional space, aligning their dimensions with the spatially hierarchical features for subsequent fusion operations. In this way, the threshold features can be fused with the spatially hierarchical features in the same dimension, improving the feature fusion effect.

[0037] Step S10214: Fuse the spatial layering features and the threshold features to obtain initial fused features; optimize the initial fused features by using a preset convolutional layer to remove redundant information and enhance lesion-related features to obtain optimized fused features.

[0038] Spatial hierarchical features and threshold features are fused element-wise to fully utilize the information from both and establish a correlation mechanism where a specific threshold corresponds to enhancement of a specific spatial region. The initial fused features are then optimized using pre-defined convolutional layers. These layers learn the complex relationships between features, remove redundant information, and enhance features relevant to the lesion. This makes the optimized fused features more focused on the lesion area of ​​the macular hole, improving the accuracy of subsequent detection.

[0039] Step S10215: Perform two-stage upsampling and feature complementation processing on the optimized fusion features to obtain an enhanced macular region OCT image.

[0040] The two-stage upsampling operation gradually restores the size of the optimized fusion features to the original image size. After the first stage of upsampling, low-level texture features are stitched together. These low-level texture features contain detailed information about the image, and stitching them together can supplement the missing details in the optimized fusion features, making the features richer. After the second stage of upsampling, the grayscale information of the original normalized single-channel grayscale image is fused. This preserves the grayscale distribution information of the original image, making the enhanced image closer to the real image while highlighting the features of the lesion area.

[0041] This embodiment normalizes the OCT images of the macular region, eliminating differences between images and providing a stable foundation for subsequent processing. Through multi-stage convolution, pooling, and discarding operations, multi-level spatial features are extracted, comprehensively reflecting the structural information of the image. Normalized single-value feature mapping is fused with spatial features to establish a correlation between the threshold and the spatial region, enhancing the expressive power of the features. By optimizing the fused features, redundant information is removed, and lesion-related features are strengthened. Finally, through two-stage upsampling and feature complementation processing, the image size is restored, and details and grayscale information are supplemented, resulting in an enhanced OCT image of the macular region. This image quality is significantly improved, and lesion features are more prominent, providing a more accurate and clearer image foundation for subsequent hole detection and parameter measurement.

[0042] For step S103, extract the OCT image of the macular region and the multi-channel feature data of each of the extended images; determine the suspected lesion features based on the multi-channel feature data of the OCT image of the macular region, and perform feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features.

[0043] Multichannel feature data refers to various types of information extracted from an image, which can describe the image's features from different dimensions (such as color, texture, and edges). In macular hole detection, multichannel feature data can more comprehensively reflect the lesion status of the macular region, helping to accurately identify lesion characteristics.

[0044] Suspected lesion features are preliminary characteristics that may represent macular hole lesions, determined based on multi-channel feature data from OCT images of the macular region. However, due to potential interference from noise, artifacts, and other factors in the images, these suspected lesion features may not be entirely accurate.

[0045] Key lesion features are accurate and reliable lesion features obtained through feature cross-validation and enhancement of suspected lesion features. These features accurately reflect the actual lesion condition of the macular hole and form the basis for subsequent calculations of macular hole parameters.

[0046] This step employs advanced feature extraction algorithms, such as deep learning-based convolutional neural networks (CNNs), to extract multi-channel feature data from the original macular region OCT image and each extended image. These feature data can include image texture features (such as features extracted from gray-level co-occurrence matrix and wavelet transform) and edge features (such as features extracted from the Canny edge detection operator). Multi-channel feature data can describe the image features from multiple dimensions, more comprehensively reflecting the lesion status of the macular region. Furthermore, based on the multi-channel feature data of the macular region OCT image, specific algorithmic models (such as classification algorithms and clustering algorithms) are used to initially determine suspected lesion features. These features may be regions in the image that differ from normal structures and possess certain lesion characteristics; however, due to potential interference from noise, artifacts, and other factors, their accuracy needs further verification. Finally, the suspected lesion features are cross-validated using the multi-channel feature data from each extended image. By comparing suspected lesion features in the macular region OCT image with corresponding features in the extended image, the credibility of the suspected lesion feature is high if similar feature changes appear in the same or similar locations across multiple extended images. Conversely, if the features at that location in the extended image differ significantly from those in the macular region OCT image, the suspected lesion feature may be caused by noise or artifacts and needs to be excluded. This cross-validation method reinforces correct lesion features, eliminates interfering information, and ultimately yields accurate key lesion features.

[0047] In one embodiment, step S103 involves extracting the multi-channel feature data of the macular region OCT image and each of the extended images; determining suspected lesion features based on the multi-channel feature data of the macular region OCT image; and performing feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features. Step S1031: Input the OCT image of the macular region and each of the extended images into a preset lesion feature extraction model.

[0048] The lesion feature extraction model employs a deep learning architecture based on dual-stream feature fusion, capable of simultaneously processing multi-scale, multi-view information from both the original OCT image and the extended image. The model's input layer is designed to be compatible with extended images of different sizes (such as rotated, scaled, and segmented mask images), and an adaptive spatial pyramid pooling layer ensures uniformity of feature map size, guaranteeing the stability of subsequent feature extraction. For example, the model can integrate inputs from OCT images of the macular region, segmented images of the left and right sides of the lesion, and their flipped derivative images, utilizing multiple input channels to collaboratively mine the implicit correlation features of the lesion region.

[0049] Step S1032: Extract the multi-channel feature data of the macular region OCT image and each of the extended images using the lesion feature extraction model; determine suspected lesion features based on the multi-channel feature data of the macular region OCT image; and perform feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features.

[0050] The model employs a multi-branch feature extraction module. The backbone network extracts basic texture and structural features, while auxiliary branches capture deformation information of lesion boundaries through deformable convolutions. In one embodiment, multi-channel feature data can include dimensions such as gray-level gradient, texture roughness, edge sharpness, and local binary pattern (LBP) to form a high-dimensional feature vector. For example, in the crack region, gray-level gradient features can highlight the steep changes in the lesion edge, and texture roughness features can quantify the non-uniformity of the lesion region. These multi-channel features together construct a three-dimensional feature representation of the lesion region.

[0051] Furthermore, based on the multi-channel features of the macular region OCT images, an attention mechanism is used to focus on high-probability lesion areas (such as low-grayscale, high-texture-variety areas) to generate initial suspected lesion features. Subsequently, multi-channel features of the extended images are used for cross-validation: for example, features from the segmented image to the left of the lesion can verify the consistency of lesions in the left region of the macular region OCT image, and features from the flipped derived image can test the stability of lesion features under different viewpoints. Through feature similarity matching and inconsistency detection, misjudged features caused by noise or artifacts are eliminated, and lesion features that are stable under multiple viewpoints are strengthened, ultimately forming key lesion features with high confidence.

[0052] This embodiment achieves end-to-end collaborative feature analysis from the original image to the expanded image through a multi-input, multi-channel feature fusion mechanism in the lesion feature extraction model. The expanded image generated by grayscale enhancement and instance segmentation provides the model with rich multi-view, multi-scale information, enabling multi-channel feature extraction to cover the comprehensive feature representation of the lesion region. The feature cross-validation and enhancement process, through feature comparison between multiple images, effectively suppresses the interference of noise and artifacts in a single image on feature extraction, while simultaneously enhancing the stable features of the lesion region under multiple views, significantly improving the accuracy and robustness of key lesion features. Finally, the macular hole parameters (such as minimum diameter and choroidal thickness) calculated based on high-confidence key lesion features are more accurate.

[0053] In one embodiment, the lesion feature extraction model includes a grayscale adaptation module, a channel attention module, a lightweight backbone network layer, and a fusion detection layer connected in sequence. Step S1032, which involves extracting multi-channel feature data of the macular region OCT image and each of the extended images using the lesion feature extraction model; determining suspected lesion features based on the multi-channel feature data of the macular region OCT image; and performing feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features, includes: Step S10321: The input macular region OCT image and single-channel grayscale data of each extended image are obtained through the grayscale adaptation module; for any single-channel grayscale data, the single-channel grayscale features are mapped to three-channel features, and the local texture features of the macular region and the structural features of the lesion-related region are extracted simultaneously.

[0054] The grayscale adaptation module is a feature transformation unit designed specifically for single-channel medical images. It maps single-channel grayscale data into three-channel features through learnable convolutional kernels. While preserving the original grayscale information, it expands the feature dimensions to adapt to subsequent multi-channel processing. At the same time, it extracts the microtexture features of the macular region (such as the fine structure between retinal layers) and the structural features of the lesion-related region (such as the continuity of the tear edge) through local filter banks.

[0055] Step S10322: Based on the three-channel features, the channel attention module generates adaptive channel weights to enhance the three-channel features, the local texture features of the macular region, and the structural features of the lesion-related region, thereby obtaining the corresponding multi-channel feature data.

[0056] The channel attention module aggregates the spatial information of the three-channel features through global average pooling and generates adaptive channel weights through a fully connected layer.

[0057] In this step, the channel attention module receives three-channel features and synchronously extracted texture / structure features. It compresses spatial information through global average pooling and generates adaptive channel weights through a two-layer fully connected network. These weights can dynamically adjust the contribution of different channels—for example, assigning higher weights to high-contrast channels in the pore region and lower weights to channels with uniform background. Finally, feature weighting enhancement is achieved through channel-by-channel multiplication, highlighting lesion-related features and suppressing background noise.

[0058] Step S10323: Extract and fuse lesion features from the multi-channel feature data of the macular region OCT image through the lightweight backbone network layer to obtain suspected lesion features.

[0059] The lightweight backbone network layer is optimized based on the YOLO architecture. Through spatial pyramid pooling and feature fusion structure, it significantly reduces the number of parameters and computational cost while maintaining feature extraction capabilities, making it suitable for the need to quickly process multiple extended images in clinical scenarios.

[0060] This step involves lightweighting the backbone network layer to perform multi-level feature extraction and fusion on the multi-channel features of the enhanced original image. Through depthwise separable convolution, lesion features from local texture to global structure are gradually extracted to generate initial suspected lesion features. Step S10324: The fusion detection layer performs feature cross-validation and enhancement operations on the suspected lesion features based on the multi-channel feature data of each of the extended images to obtain key lesion features. The fusion detection layer integrates multi-view feature comparison and consistency verification functions. It cross-validates suspected lesion features through confidence comparison and fusion, eliminates artifact interference in single images, and enhances lesion features that are stable across multiple views.

[0061] This step of the fusion detection layer utilizes the multi-channel features of each extended image (such as segmented images of the left and right sides of the lesion and flipped derived images) to perform multi-view cross-validation of suspected lesion features through feature similarity calculation and inconsistency detection. For example, it verifies the consistency of the features on the left side of the lesion in the original image and the segmented image on the left side, eliminates misjudged features caused by image artifacts, and finally strengthens the stable lesion features from multiple perspectives through feature fusion to form key lesion features with high confidence.

[0062] This embodiment utilizes a grayscale adaptation module to achieve feature expansion from single-channel to three-channel and simultaneous extraction of low-level features, providing richer feature dimensions. The channel attention module, through adaptive weight allocation, intelligently enhances lesion-related features, suppresses background noise, and improves the effectiveness and reliability of features. A lightweight backbone network layer completes multi-level feature extraction and fusion at extremely low computational cost, ensuring rapid processing capabilities in clinical scenarios. The fusion detection layer, through feature cross-validation and enhancement from multiple expanded images, effectively eliminates artifact interference in single images, strengthens lesion features stable across multiple perspectives, and significantly improves the accuracy and robustness of key lesion features. Ultimately, the macular hole parameters (such as minimum diameter and choroidal thickness) calculated based on high-confidence key lesion features are more accurate, and the detection results more closely match the actual clinical lesion severity, providing reliable data support for doctors to develop personalized treatment plans.

[0063] In one embodiment, the grayscale adaptation module maps single-channel features to three-channel features using a 3×3 convolution kernel; The channel attention module compresses the spatial dimension of the three-channel features by performing global average pooling, and uses one-dimensional convolution to learn the inter-channel dependencies to generate adaptive channel weights. The lightweight backbone network layer includes a first lightweight module and a second lightweight module. The first lightweight module is a dual-branch lightweight feature fusion module, which extracts features through 1×1 convolution, block operation, lightweight bottleneck structure feature enhancement processing, and splicing integration. The second lightweight module is an adaptive residual enhancement lightweight feature integration module, which performs local feature extraction and residual feature enhancement through dual parallel branches, and then splices and integrates them to obtain suspected lesion features.

[0064] In this embodiment, the grayscale adaptation module uses a 3×3 convolutional kernel to map single-channel grayscale features to three-channel features. This process simultaneously extracts local texture features of the macular region and structural features of lesion-related areas. An adaptive filling strategy maintains the feature map size, avoiding information loss. The channel attention module, based on an improved ECA mechanism, dynamically adjusts the kernel coverage by adaptively adjusting the 1D convolutional kernel size to accurately capture local channel relationships. The first lightweight module maps the input channel number using 1×1 convolutions. Block operations divide the feature map into multiple sub-regions. Lightweight bottleneck structures (such as Ghost Bottleneck) enhance features through depthwise separable convolutions and low-cost feature generation. A concatenation and integration operation fuses multi-branch features to form a multi-level feature representation. The second lightweight module uses a dual-parallel branch design. Branch 1 uses a 1×1 convolution to reduce channels and then extracts local features through a 3×3 convolution. Branch 2 enhances feature representation through the Bottleneck residual module. The two branches are concatenated and integrated through a 1×1 convolution, combined with residual connections to alleviate gradient vanishing and improve high-level feature extraction capabilities.

[0065] In one embodiment, step S10324, which involves performing feature cross-validation and enhancement operations on the suspected lesion features based on the multi-channel feature data of each of the extended images through the fusion detection layer to obtain key lesion features, includes: Step S103241: The fusion detection layer is used to locate the lesion region in the global features of the suspected lesion features to obtain the first set of lesion feature parameters and corresponding confidence scores; the lesion feature parameters are extracted from the local segmentation features from the left and right segmentation images of the lesion in the suspected lesion features to obtain the second set of lesion feature parameters and corresponding confidence scores.

[0066] The fusion detection layer first performs multi-scale lesion region localization on the global features of suspected lesions in the original image. It integrates feature maps of different scales through a spatial pyramid pooling layer, and uses a group of convolutional kernels to perform multi-dimensional scanning of the global features, locating global lesion feature parameters such as the center region of the lesion and boundary contours (e.g., minimum diameter and maximum height of the lesion). A confidence score is then generated based on the feature matching degree and the model's prediction probability. For example, when the matching degree between a low-grayscale continuous region and a high-texture variation region in the global features exceeds a threshold, a high confidence score is assigned, and the region is identified as a stable lesion.

[0067] For the local segmentation features of the left and right segmented images of the crack, the fusion detection layer uses a dual-branch feature extractor to process them separately. The left branch focuses on the jagged structural features of the left edge of the crack, extracting parameters such as the left crack height and edge sharpness; the right branch analyzes the texture roughness and structural continuity of the right region, extracting parameters of the right crack. Each local segmentation feature parameter generates a confidence score through feature similarity calculation and artifact detection algorithms. For example, when the edge sharpness in the left segmentation feature exceeds a preset threshold and matches the global feature, a high confidence score is assigned.

[0068] Step S103242: From the first group of lesion feature parameters and the second group of lesion feature parameters, select the lesion feature parameter with the highest confidence among the same lesion feature parameters, and determine the lesion feature parameter with the highest confidence as the key lesion feature.

[0069] The fusion detection layer performs homologous feature parameter matching between the first set of global feature parameters and the second set of local segmentation feature parameters. Through confidence score ranking and consistency verification, the lesion feature parameter with the highest confidence is selected as the key lesion feature. For example, when the minimum diameter of the lesion exists in both the global features and the left / right segmentation features, the parameter with the highest confidence score among the three sets of parameters is automatically selected as the key lesion feature. This process achieves cross-dimensional parameter association through a feature parameter mapping table, ensuring parameter consistency of the same lesion feature under different perspectives. For example, when the confidence of both the global feature parameter and the left segmentation feature parameter is higher than that of the right, the global feature parameter is selected first; if the left / right segmentation feature parameter has a higher confidence and matches the global feature, the local segmentation parameter is selected. This ultimately forms a key lesion feature with multi-view verification and high confidence, providing a reliable basis for subsequent parameter calculations. In one embodiment, before selecting the lesion feature parameter with the highest confidence among the same lesion feature parameters, a sharpening operation is performed on the confidence scores corresponding to the maximum height of the left and right lesions in the second set of lesion features, raising them to a preset confidence threshold. By employing a synergistic mechanism of confidence sharpening and pre-set confidence thresholds, the accuracy and robustness of screening key lesion features of macular hole were significantly improved.

[0070] This embodiment achieves multi-dimensional lesion feature extraction and confidence assessment from the overall to the local level by co-localizing global and local segmentation features of the fusion detection layer. Global feature localization ensures the macroscopic accuracy of the lesion area, while local segmentation feature extraction enhances the reliability of detailed features such as the edge of the macular hole. Through the highest confidence screening mechanism of homologous feature parameters, artifact interference and misjudged features in a single image are effectively eliminated, strengthening the lesion features that are stable across multiple perspectives. The resulting key lesion features possess high confidence and multi-view verification characteristics, significantly improving the accuracy of macular hole parameter calculation and the reliability of detection results.

[0071] In one embodiment, the lesion feature extraction model is trained through the following steps: Step S201: Obtain several macular region OCT image samples and several extended images of each of the macular region OCT image samples.

[0072] This step collects OCT images clinically diagnosed as macular holes as raw samples and generates expanded images for each sample using image preprocessing techniques. For example, the raw images are segmented to the left and right sides of the hole (symmetrically cut along the central axis of the hole), generating left and right segmented images to highlight the local features of the hole edge, respectively. Simultaneously, derivative images are generated through flipping operations (horizontal / vertical flipping) to simulate the lesion's appearance from different shooting angles. The generation of expanded images broadens the distribution of training data and enhances the model's adaptability to multi-view lesion features.

[0073] Step S202: The key lesion features of each macular region OCT image sample are obtained by pixel-level feature point annotation of the key parameters of macular hole in each macular region OCT image sample using a feature annotation tool.

[0074] This step uses feature annotation tools such as LabelMe to perform pixel-level annotation on each original OCT image, marking the location of key parameters of the macular hole (such as the hole's edge contour and the choroid-retinal junction). The annotation process combines the grayscale distribution of the OCT image (low reflectance areas correspond to the hole cavity) with structural features (high texture variation areas correspond to the hole edge) to ensure the clinical accuracy of the labeled data. The annotated key lesion features serve as supervisory signals for model training, guiding the model to learn the mapping relationship between lesion features and image pixels.

[0075] Step S203: Based on the several macular region OCT image samples, several extended images corresponding to each macular region OCT image sample, and key lesion features, construct several sets of training samples; each set of training samples includes sample data and label data; wherein, the sample data is the macular region OCT image sample and several corresponding extended images; the label data is the corresponding key lesion features.

[0076] This step combines the original OCT image samples and their corresponding extended images into sample data, and uses the labeled key lesion features as label data to construct a paired "sample data-label data" training set. For example, a training sample set may contain the original image, the left-side segmented image, the right-side segmented image, and the flipped derived image as input, with the corresponding labels being parameters such as the minimum diameter and maximum height of the lesion annotated by the doctor. This multi-image-single-label sample design forces the model to extract common lesion features from images from different perspectives, improving the robustness of feature extraction.

[0077] Step S204: Input the several sets of training samples into the initial feature extraction model and output the predicted key lesion features corresponding to each set of training samples; calculate the loss value based on the predicted key lesion features and the corresponding label data; adjust the parameters of the feature extraction model based on the loss value; until the loss value is less than a preset threshold or the number of adjustments is greater than a preset number, the trained feature extraction model is obtained; use the feature extraction model as the lesion feature extraction model.

[0078] The initial feature extraction model is based on a deep learning neural network architecture, integrating a grayscale adaptation module, a channel attention module, and a lightweight backbone network layer, which is used to automatically extract and verify macular hole features from OCT images.

[0079] This step inputs training samples into the initial feature extraction model. The model uses a grayscale adaptation module to map single-channel OCT images into three-channel features. A channel attention module weights the features by channel to highlight lesion-related channels. A lightweight backbone network layer extracts multi-level features and generates predicted key lesion features. The loss value (e.g., mean squared error loss) is calculated by comparing the predicted values ​​with the label values. The model parameters (e.g., convolutional kernel weights, attention mechanism weights) are adjusted using the backpropagation algorithm. Training is repeated iteratively until the loss value is less than a preset threshold (e.g., 0.01) or the number of iterations exceeds a preset number (e.g., 500 times). At this point, training is complete, and a lesion feature extraction model that can accurately extract macular hole features is obtained.

[0080] This embodiment constructs a training framework of "multi-view extended images - pixel-level annotation labels," combining grayscale adaptation, channel attention, and a lightweight backbone network layer model architecture to achieve high-precision training and optimization of the macular hole feature extraction model. The introduction of extended images simulates the lesion manifestations under different shooting angles in clinical practice, forcing the model to extract common features from multi-view data, significantly improving the model's generalization ability for key features such as the jagged structure of the hole edge and low-reflection signal areas. Pixel-level annotation labels provide the model with accurate supervision signals, and combined with loss function minimization optimization, the model's prediction results highly match actual clinical parameters, reducing the risk of false positives and false negatives. The finally trained lesion feature extraction model can automatically and quickly extract key parameters of the macular hole from OCT images, providing doctors with objective and accurate lesion assessment basis, promoting the leap from manual interpretation to automation and intelligence in macular hole detection, and significantly improving the efficiency of clinical diagnosis and the scientific nature of treatment decisions.

[0081] For step S104, the macular hole parameters are calculated based on the key lesion features; and the macular hole detection results are obtained based on the macular hole parameters.

[0082] Based on the key lesion characteristics obtained, a pre-defined computational model is used to calculate relevant parameters of macular holes, such as minimum diameter at the hole (MDH), maximum diameter at the hole base (MDBH), choroidal thickness (TC), maximum height on the left side of the hole (MHLH), and maximum height on the right side of the hole (MHRH). These parameters accurately reflect the morphology and severity of macular holes. Then, by comparing these parameters with normal ranges, the detection results of macular holes are obtained, determining the presence and extent of lesions, and providing a basis for clinical diagnosis and treatment.

[0083] In one embodiment, the key lesion features include the coordinates of the minimum diameter endpoint of the tear, the coordinates of the base endpoint of the tear, the coordinates of the thickness endpoint of the choroid, the coordinates of the maximum height endpoint of the left tear, and the coordinates of the maximum height endpoint of the right tear. The endpoint coordinates of the minimum diameter of the macular hole are the pixel coordinates of the two endpoints corresponding to the minimum diameter of the macular hole in the OCT image. They are used to locate the spatial position of the narrowest part of the hole and are the key geometric parameters for calculating the actual minimum diameter of the hole.

[0084] The endpoint coordinates of the tear base are the pixel coordinates of the two endpoints of the widest region of the tear base (i.e., the junction between the tear and the retinal pigment epithelium). They are used to determine the maximum width of the tear base and reflect the degree of lateral expansion of the tear.

[0085] Choroid thickness endpoint coordinates are the pixel coordinates of the two endpoints of the choroid (the vascular layer located between the retina and sclera) at the thinnest or thickest point in the tear region. They are used to calculate the actual thickness of the choroid and to help assess the compression or traction effect of the tear on the choroid.

[0086] The maximum height endpoint coordinates of the left / right crack are the pixel coordinates corresponding to the highest point from the bottom to the top of the crack on the left or right edge of the crack. They are used to calculate the actual height of the left and right sides of the crack respectively, reflecting the longitudinal asymmetry of the crack.

[0087] Step S104, which involves calculating macular hole parameters based on the key lesion features and obtaining macular hole detection results based on the macular hole parameters, includes: Step S1041: Obtain the image length, image width, horizontal pixel count, and vertical pixel count of the OCT image of the macular region, and establish a mapping relationship between image pixels and actual size.

[0088] The mapping relationship between image pixels and actual size is established by the imaging parameters of the OCT device (such as image length, width, and number of horizontal / vertical pixels) and the conversion ratio between pixel coordinates and actual physical size (such as micrometers). For example, each pixel corresponds to 5μm, which is used to convert pixel coordinates into clinically usable actual size parameters.

[0089] Step S1042: Based on the endpoint coordinates of the minimum diameter of the pore, the endpoint coordinates of the pore base, the endpoint coordinates of the choroid thickness, the endpoint coordinates of the maximum height of the left pore, the endpoint coordinates of the maximum height of the right pore, and the mapping relationship, the minimum diameter of the pore, the maximum diameter of the pore base, the choroid thickness, the maximum height of the left pore, and the maximum height of the right pore are calculated respectively.

[0090] Step S1043: Obtain the macular hole parameters based on the minimum diameter of the hole, the maximum diameter of the hole base, the choroid thickness, the maximum height of the left hole, and the maximum height of the right hole.

[0091] Step S1044: Obtain the macular hole detection result based on the macular hole parameters.

[0092] This embodiment achieves automated and high-precision calculation of key parameters of macular hole through precise pixel coordinate positioning and pixel-to-actual-size mapping. First, the mapping relationship established based on the imaging parameters of the OCT device ensures the accuracy of the conversion from pixel coordinates to actual size, avoiding parameter deviations caused by differences in device resolution. Second, by calculating the minimum diameter of the hole, the base diameter, the choroidal thickness, and the left and right side heights respectively, the lateral expansion, longitudinal asymmetry, and impact on the choroid of the hole are comprehensively quantified, providing multi-dimensional lesion assessment evidence for clinical practice. Finally, the automated calculation process (from coordinate extraction to parameter generation) significantly shortens the diagnostic time (from 5-15 minutes of manual measurement to seconds), while reducing human operational errors (such as measurement deviations caused by parallax), improving the objectivity and consistency of diagnostic results.

[0093] In one embodiment, the key lesion features further include confidence scores corresponding to the minimum diameter endpoint coordinates of the lesion, the base endpoint coordinates of the lesion, the thickness endpoint coordinates of the choroid, the maximum height endpoint coordinates of the left lesion, and the maximum height endpoint coordinates of the right lesion.

[0094] When the lesion feature extraction model outputs key coordinates such as the coordinates of the minimum diameter endpoint of the lesion and the coordinates of the base endpoint of the lesion, it simultaneously generates a confidence score for each coordinate. For example, the model might output the coordinates of the minimum diameter endpoint of the lesion as (150, 200) and (180, 250), with a confidence score of 0.92, indicating that the model has a 92% confidence in predicting the accuracy of those coordinates. The confidence score is generated through algorithms such as multi-scale feature fusion and attention mechanism weighting within the model, integrating local image texture, global structure, and contextual information to ensure the reliability of the score.

[0095] Step S1044, the step of obtaining the macular hole detection result based on the macular hole parameters, includes: Step S10441: The minimum diameter of the macular hole, the maximum diameter of the hole base, the choroidal thickness, the maximum height of the left hole, and the maximum height of the right hole in the macular hole parameters are associated with the confidence scores of the corresponding endpoint coordinates, and the anatomical location information of the macular corresponding to each parameter is marked.

[0096] The calculated macular hole parameters (such as the minimum hole diameter of 180 μm) are bound to the confidence scores (such as 0.92) of the corresponding endpoint coordinates, and the anatomical location information of the parameters is labeled.

[0097] Step S10442: Determine the reliability level of each macular hole parameter based on the confidence score; generate macular hole detection results based on each macular hole parameter, the corresponding confidence score, the corresponding macular anatomical location information, and the corresponding reliability level.

[0098] The reliability levels of parameters are determined based on their confidence scores. For example, thresholds are set: confidence scores ≥ 0.85 are considered "high reliability," 0.7-0.85 are "medium reliability," and < 0.7 are "low reliability." Each parameter (e.g., minimum pore diameter 180 μm), its confidence score (0.92), and its reliability level ("high reliability") are integrated into a single detection result entry. The final detection results are output in report form, containing complete information on all parameters.

[0099] This embodiment significantly improves the objectivity and clinical usability of macular hole detection results by introducing confidence scores and anatomical location annotations. First, the confidence score quantifies the predictive reliability of each parameter, enabling physicians to quickly identify high-confidence parameters (e.g., parameters with a confidence score > 0.85 can be directly used for diagnosis), while allowing for manual review or supplementary examination of low-confidence parameters (e.g., parameters with a confidence score < 0.7) to avoid misdiagnosis. Second, anatomical location information correlates parameters with the spatial structure of the retina, helping physicians understand the spatial distribution of lesions (e.g., whether the hole deviates from the fovea, whether changes in choroidal thickness are localized), providing crucial information for treatment decisions (e.g., surgical approach selection). Finally, the classification of reliability levels allows for a tiered presentation of detection results, highlighting core parameters (e.g., high-reliability hole diameter) and downplaying auxiliary parameters (e.g., moderate-reliability choroidal thickness), making the report more suitable for clinical reading habits.

[0100] The above are merely embodiments of this application and are not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for detecting macular holes, characterized in that, Includes the following steps: Acquire OCT images of the macular region of the eye to be examined; The OCT image of the macular region is expanded to obtain several expanded images; Extract the multi-channel feature data of the macular region OCT image and each of the extended images; determine suspected lesion features based on the multi-channel feature data of the macular region OCT image; combine the multi-channel feature data of each of the extended images to perform feature cross-validation and enhancement on the suspected lesion features to obtain key lesion features; The macular hole parameters are calculated based on the key lesion features; the macular hole detection results are obtained based on the macular hole parameters.

2. The method for detecting macular holes according to claim 1, characterized in that, The steps of expanding the OCT image of the macular region to obtain several expanded images include: The macular region OCT image is subjected to grayscale enhancement processing to obtain an enhanced macular region OCT image; Masking is performed on the left and right regions of the hole in the enhanced macular region OCT image to obtain a segmented image of the left side of the hole and a segmented image of the right side of the hole. The enhanced macular region OCT image is diagonally flipped to obtain the corresponding flipped derivative image; The several extended images are obtained based on the enhanced macular region OCT image, the segmented image of the left side of the hole, the segmented image of the right side of the hole, and the flipped derived image.

3. The method for detecting macular holes according to claim 2, characterized in that, The step of performing grayscale enhancement processing on the OCT image of the macular region to obtain an enhanced OCT image of the macular region includes: The OCT image of the macular region is normalized to obtain a normalized single-channel grayscale image; the maximum grayscale value of the normalized single-channel grayscale image is obtained, and the corresponding normalized single-value feature is generated. Perform multi-stage convolution, pooling, and discarding operations on the normalized single-channel grayscale image to extract spatial layering features; Two fully connected operations are performed on the normalized single-valued features, which are then mapped to a preset dimension feature space and upsampled to obtain threshold features that are aligned with the size of the spatial hierarchical features. The spatial layering features and the threshold features are fused to obtain an initial fused feature; the initial fused feature is then optimized by a preset convolutional layer to remove redundant information and enhance lesion-related features, resulting in an optimized fused feature. The optimized fusion features are subjected to two-stage upsampling and feature complementation processing to obtain an enhanced OCT image of the macular region.

4. The method for detecting macular holes according to claim 1, characterized in that, The steps of extracting multi-channel feature data from the OCT image of the macular region and each of the extended images; determining suspected lesion features based on the multi-channel feature data of the OCT image of the macular region; and performing feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features include: The OCT image of the macular region and each of the extended images are input into a preset lesion feature extraction model; The lesion feature extraction model extracts multi-channel feature data of the macular region OCT image and each of the extended images; based on the multi-channel feature data of the macular region OCT image, suspected lesion features are determined, and the suspected lesion features are cross-validated and enhanced by combining the multi-channel feature data of each of the extended images to obtain key lesion features.

5. The method for detecting macular holes according to claim 4, characterized in that, The lesion feature extraction model includes a grayscale adaptation module, a channel attention module, a lightweight backbone network layer, and a fusion detection layer connected in sequence. The steps of extracting multi-channel feature data of the macular region OCT image and each of the extended images using the lesion feature extraction model; determining suspected lesion features based on the multi-channel feature data of the macular region OCT image; and performing feature cross-validation and enhancement on the suspected lesion features by combining the multi-channel feature data of each of the extended images to obtain key lesion features include: The grayscale adaptation module acquires the input macular region OCT image and single-channel grayscale data of each extended image; for any single-channel grayscale data, the single-channel grayscale features are mapped to three-channel features, and the local texture features of the macular region and the structural features of the lesion-related region are extracted simultaneously. The channel attention module generates adaptive channel weights based on the three-channel features, and then weights and enhances the three-channel features, the local texture features of the macular region, and the structural features of the lesion-related region to obtain the corresponding multi-channel feature data. The lightweight backbone network layer is used to extract and fuse lesion features from the multi-channel feature data of the macular region OCT image to obtain suspected lesion features. The fusion detection layer performs feature cross-validation and enhancement operations on the suspected lesion features based on the multi-channel feature data of each of the extended images to obtain key lesion features.

6. The method for detecting macular holes according to claim 5, characterized in that, The steps of obtaining key lesion features by performing feature cross-validation and enhancement operations on the suspected lesion features based on the multi-channel feature data of each of the extended images through the fusion detection layer include: The fusion detection layer is used to locate the lesion region in the global features of the suspected lesion features, and obtains the first set of lesion feature parameters and corresponding confidence scores; the local segmentation features from the segmentation images on the left and right sides of the lesion in the suspected lesion features are used to extract lesion feature parameters, and obtain the second set of lesion feature parameters and corresponding confidence scores. From the first group of lesion feature parameters and the second group of lesion feature parameters, the lesion feature parameter with the highest confidence among the same lesion feature parameters is selected, and the lesion feature parameter with the highest confidence is determined as the key lesion feature.

7. The method for detecting macular holes according to claim 4, characterized in that, The lesion feature extraction model is trained through the following steps: Acquire several OCT image samples of the macular region and several extended images of each of the OCT image samples of the macular region; By using a feature annotation tool, the key parameters of macular hole in each of the OCT image samples of the macular region are annotated at the pixel level to obtain the key lesion features of each of the OCT image samples of the macular region. Based on the aforementioned macular region OCT image samples, several extended images corresponding to each macular region OCT image sample, and key lesion features, several sets of training samples are constructed; each set of training samples includes sample data and label data; wherein, the sample data consists of the macular region OCT image samples and several corresponding extended images; and the label data consists of the corresponding key lesion features. The initial feature extraction model is input into the several sets of training samples, and the predicted key lesion features corresponding to each set of training samples are output. The loss value is calculated based on the predicted key lesion features and the corresponding label data. The parameters of the feature extraction model are adjusted based on the loss value. The training is completed when the loss value is less than a preset threshold or the number of adjustments is greater than a preset number. The feature extraction model is then used as the lesion feature extraction model.

8. The method for detecting macular holes according to any one of claims 1 to 7, characterized in that, The key lesion features include the coordinates of the minimum diameter endpoint of the tear, the coordinates of the base endpoint of the tear, the coordinates of the choroidal thickness endpoint, the coordinates of the maximum height endpoint of the left tear, and the coordinates of the maximum height endpoint of the right tear. The macular hole parameters are calculated based on the key lesion characteristics; The steps for obtaining the macular hole detection results based on the macular hole parameters include: The image length, image width, horizontal pixel count, and vertical pixel count of the OCT image of the macular region are obtained, and a mapping relationship between image pixels and actual size is established. Based on the endpoint coordinates of the minimum diameter of the pore, the endpoint coordinates of the pore base, the endpoint coordinates of the choroid thickness, the endpoint coordinates of the maximum height of the left pore, the endpoint coordinates of the maximum height of the right pore, and the mapping relationship, the minimum diameter of the pore, the maximum diameter of the pore base, the choroid thickness, the maximum height of the left pore, and the maximum height of the right pore are calculated respectively. The macular hole parameters are obtained based on the minimum diameter of the hole, the maximum diameter of the hole base, the choroid thickness, the maximum height of the left hole, and the maximum height of the right hole. The macular hole detection results are obtained based on the macular hole parameters.

9. The method for detecting macular holes according to claim 8, characterized in that, The key lesion features also include the confidence scores corresponding to the minimum diameter endpoint coordinates of the tear, the base endpoint coordinates of the tear, the choroid thickness endpoint coordinates, the maximum height endpoint coordinates of the left tear, and the maximum height endpoint coordinates of the right tear. The step of obtaining the macular hole detection result based on the macular hole parameters includes: The minimum diameter of the macular hole, the maximum diameter of the hole base, the choroidal thickness, the maximum height of the left hole, and the maximum height of the right hole in the macular hole parameters are associated one by one with the confidence scores of the corresponding endpoint coordinates, and the anatomical location information of the macular corresponding to each parameter is marked. The reliability level of each macular hole parameter is determined based on the confidence score; the macular hole detection result is generated based on each macular hole parameter, the corresponding confidence score, the corresponding macular anatomical location information, and the corresponding reliability level.

Images