Data processing method and system for eye trauma chamber angle retrogression detection of eye UBM image

Abstract

The invention provides a data processing method and system for eye trauma chamber angle receding detection of an eye UBM image, and is applied to the technical field of data processing. The method comprises the following steps: processing original eye UBM image data of a target patient based on a data enhancement strategy to generate target eye UBM image data; processing a room corner structure in the target eye UBM image data to generate annotation information; based on target eye UBM image data in the target data set and in combination with a YOLO-V8 deep learning algorithm, room corner structure features in the eye UBM images are processed, and an initial room corner retreating recognition model is generated; processing the initial room corner backing identification model based on a training set, a verification set and a test set in the target data set to generate a target room corner backing identification model; and processing the target eye UBM image data and the annotation information based on the target chamber angle retreating identification model to generate an eye chamber angle retreating identification result.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular to a data processing method and system for detecting angle recession of ocular trauma using UBM images. Background Art

[0002] In the field of ophthalmology, accurate diagnosis of angle recession is crucial for the prevention and treatment of eye diseases such as glaucoma. Currently, there are various methods for examining angle recession. Although gonioscopy is a commonly used method that can directly observe abnormalities such as the shape, width, and presence of angle recession, it has certain limitations, such as limited ability to resolve subtle structures, reliance on the doctor's subjective judgment, and susceptibility to interference from human factors. Optical coherence tomography (OCT), especially anterior segment OCT technology, can perform cross-sectional imaging of the angle, providing detailed information about the angle structure and tissue, but its observation of the overall structure of the angle is not comprehensive enough.

[0003] The ultrasound biomicroscope (UBM), a B-type high-frequency ultrasound diagnostic device capable of real-time observation and research of anatomical structures related to the living human eye, offers unique advantages. It provides noninvasive, high-resolution images of the ocular angle and ciliary body, clearly displaying structures related to angle morphology, such as iris cross-sections, iris root attachment locations, ciliary body morphology, and posterior chamber morphology. This eliminates the influence of light on angle examination results and facilitates real-time observation of the angle. However, traditional UBM-based methods for identifying angle recession often rely on manual interpretation by physicians, which is not only time-consuming and laborious, but also difficult to ensure diagnostic accuracy and consistency due to differences in physician experience.

[0004] It should be noted that the information disclosed in the above background technology section is only used to enhance the understanding of the background of the present disclosure, and therefore includes information that does not constitute prior art known to ordinary technicians in this field. Summary of the Invention

[0005] The present application aims to provide a data processing method and system for detecting angle recession in ocular trauma using UBM images. This method, at least to a certain extent, overcomes the challenges of the prior art. The method obtains relevant data from a target patient and a target dataset, then enriches the data by performing data enhancement operations such as brightness adjustment and the addition of Gaussian and salt-and-pepper noise. Next, image classification information is generated using preset horizontal box annotations. The backbone, neck, and head networks of the YOLO-V8 algorithm are used to collaboratively extract angle structural features to generate an initial model. Optimization iterations are then performed through splitting training, validation, and test sets to obtain a target model. The target eye's UBM image is then processed to generate annotation information for the angle structure. The image and annotation information are analyzed based on the target model to generate a probability score for angle recession. Comparative analysis reveals angle status and detailed judgment information to aid clinical diagnosis.

[0006] Other features and advantages of the present application will become apparent from the following detailed description, or may be learned in part by practice of the invention.

[0007] According to one aspect of the present application, a data processing method for detecting angle recession due to ocular trauma using ocular UBM images is provided, comprising: obtaining original ocular UBM image data and a target data set of a target patient; processing the original ocular UBM image data of the target patient based on a data enhancement strategy to generate target ocular UBM image data; processing the angle structure in the target ocular UBM image data to generate annotation information; processing the angle structure features in the ocular UBM image based on the target ocular UBM image data in the target data set in combination with the YOLO-V8 deep learning algorithm to generate an initial angle recession recognition model, wherein the target ocular UBM image data includes the original ocular UBM image and an enhanced ocular UBM image; processing the initial angle recession recognition model based on a training set, a validation set, and a test set in the target data set to generate a target angle recession recognition model; processing the target ocular UBM image data and the annotation information based on the target angle recession recognition model to generate an ocular angle recession recognition result.

[0008] Another aspect of the present application is a data processing device for detecting angle recession due to ocular trauma using an eye UBM image, characterized in that it includes: processing original eye UBM image data of a target patient based on a data enhancement strategy to generate target eye UBM image data; processing the angle structure in the target eye UBM image data to generate annotation information; processing the angle structure features in the eye UBM image based on the target eye UBM image data in a target data set in combination with the YOLO-V8 deep learning algorithm to generate an initial angle recession recognition model, wherein the target eye UBM image data includes the original eye UBM image and the enhanced eye UBM image; processing the initial angle recession recognition model based on the training set, validation set, and test set in the target data set to generate a target angle recession recognition model; processing the target eye UBM image data and the annotation information based on the target angle recession recognition model to generate an eye angle recession recognition result.

[0009] According to another aspect of the present application, a computer-readable storage medium is provided, on which a computer program is stored. When the computer program is executed by a second processor, the data processing method for detecting angle recession due to ocular trauma in ocular UBM images is implemented.

[0010] The present application provides a data processing method and system for detecting angle recession due to ocular trauma using UBM images of the eye. A server obtains relevant data of a target patient and a target dataset. The images in the dataset are then subjected to data enhancement operations such as brightness adjustment and the addition of Gaussian and salt-and-pepper noise to enrich the data. Subsequently, image classification information is generated using preset horizontal box annotations. The backbone, neck, and head networks of the YOLO-V8 algorithm are used to collaboratively extract angle structural features to generate an initial model. The target model is then obtained by optimizing and iterating through training, validation, and test sets. The angle structure of the target eye UBM image is then processed to generate annotation information. The image and annotation information are analyzed based on the target model to generate a probability score for angle recession. Comparative analysis reveals the angle status and detailed judgment information to assist in clinical diagnosis.

[0011] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] Figure 1 A flow chart showing a data processing method for detecting angle recession of ocular trauma using an ocular UBM image provided by an embodiment of the present application is shown;

[0013] Figure 2 A schematic structural diagram of a data processing device for detecting angle recession due to ocular trauma using UBM images is shown in an embodiment of the present application. DETAILED DESCRIPTION

[0014] The preferred embodiments of the present invention are described below with reference to the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0015] The following combination Figure 1 The data processing method for detecting angle recession of ocular trauma using an eye UBM image according to an exemplary embodiment of the present application is described below. Figure 1 As shown, the method is applied to the server and includes:

[0016] S101, obtaining original eye UBM image data and a target dataset of a target patient.

[0017] In one embodiment, the target patient's basic information is Mr. Li, a 45-year-old male who sought medical attention for an eye injury. He has a five-year history of hypertension and his blood pressure is usually well controlled. At a hospital, a doctor used professional UBM equipment to perform an eye examination on Mr. Li. According to clinical operating procedures, the acquisition process ensures the appropriate contact angle and pressure between the device probe and the eye to obtain a clear image. The original UBM image of the eye obtained shows that some structures in the angle area are blurred. The preliminary judgment is that there may be signs of angle recession, but it is difficult to accurately determine.

[0018] The target dataset contains raw ocular UBM images from different patients, assuming there are patients A, B, and C. Patient A's image data: Patient A, a 32-year-old female, sought medical attention after sustaining an eye injury while playing sports. The acquired raw UBM image clearly shows the angle structure, with uniform angle width, a normal connection between the ciliary body and the iris root, and a clear trabecular meshwork texture. The expert labeled the image "normal angle," used a horizontal box to delineate the angle range, and saved it in YOLO format. Patient B's image data: Patient B, a 58-year-old male with a family history of glaucoma, had recently experienced eye discomfort. His raw UBM image showed a widening of the angle in some areas, with slight signs of separation between the ciliary body and the iris root. After careful assessment, the expert labeled the image "angle recession (suspected early stage)," similarly used a horizontal box to delineate the angle's location and range, and converted it into YOLO format. Patient C's image data: Patient C, a 25-year-old female, suffered an eye injury after an accidental fall. The original UBM images were of average quality, with incomplete visualization of the angle structures. After image processing and enhancement, significant tearing was observed at the angle, with a clear separation between the ciliary body and the iris root. Experts labeled the images as "angle recession (severe)" and, after annotation and format conversion, included them in the dataset. The original UBM images from these different patients were combined to form the original dataset.

[0019] S102 : Processing the original eye UBM image data of the target patient based on the data enhancement strategy to generate target eye UBM image data.

[0020] In one embodiment, the target eye UBM image data in the target data set is brightness-adjusted to generate eye UBM images under different lighting conditions. In Li's original eye UBM image, the RGB value of a certain pixel is (80, 90, 100). When brightening, 70 is added to the value of each channel, and the pixel becomes (150, 160, 170). Then, the np.clip function is used to limit the pixel value to the range of 0-255 to ensure normal image display. When dimming, 70 is subtracted, and the pixel becomes (10, 20, 30), and the np.clip function is also used for clipping. After such processing, eye UBM images under different lighting conditions are generated, simulating the image effects collected in strong light and weak light environments, so that the model can learn the characteristic differences of the chamber angle structure under different lighting conditions.

[0021] The target ocular UBM image data was processed with Gaussian noise and salt-and-pepper noise, respectively, to generate a noisy ocular UBM image. Assuming that the image size of Li's eye UBM is 512×512, the pixel values ​​of the image were first normalized from 0–255 to the range of 0–1. Then, the np.random.normal function was used to generate a Gaussian-distributed random noise matrix of the same size, with a mean of 0 and a standard deviation of 0.3. For example, if the value corresponding to a pixel in the noise matrix is ​​0.15 and the normalized value of that pixel is 0.4, the sum of the two equals 0.55. The np.clip function was then used to clamp the pixel values ​​to the range of 0–1. Finally, the processed image was denormalized to the range of 0–255. This resulted in an image with Gaussian noise added, simulating the noise interference experienced during actual acquisition and enabling the model to learn to identify the anterior chamber angle structure in the presence of noise. For Li's ocular UBM image, the salt-and-pepper noise ratio was set to 0.3. The number of pixels to which noise needs to be added is calculated to be approximately 512 × 512 × 0.3 ≈ 78,643. Using the randint() function in the Python standard library, we randomly determine pixel positions in a loop. For example, if the randomly determined pixel position is (200, 300), we use the randint() function again to generate 0 or 1. If the generated random number is 0, the pixel at that position is set to black (0); if it is 1, it is set to white (255). In this way, approximately 78,643 pixels are processed one by one to complete the addition of salt and pepper noise, making the image present more complex features and enhancing the adaptability of the model.

[0022] The UBM images of the eye under different lighting conditions and the noise-treated UBM images were processed to generate target UBM image data. Images of Li after brightness adjustment (brightening and dimming) and images after adding Gaussian noise and salt and pepper noise were combined. These images exhibit diverse characteristics in terms of brightness and noise, and together constitute the target UBM image data. For example, some images were brightened and then Gaussian noise was added, while others were dimmed and then salt and pepper noise was added. This rich data allows the model to learn the characteristics of the anterior chamber angle structure under different conditions, helping to improve the model's accuracy and stability in identifying anterior chamber angle recession, enabling it to make more accurate judgments when faced with a variety of actual collected UBM images of the eye.

[0023] S103: Process the angle structure in the target eye UBM image data to generate annotation information.

[0024] In one embodiment, image processing and analysis techniques are used to deeply analyze the angle structure in the target eye's UBM image data. Edge detection and feature extraction algorithms are used to obtain the location information of the angle, such as the coordinate range of the angle in the image. Simultaneously, morphological features of the angle are extracted, such as the angle, the connection between the ciliary body and the iris root, and the texture of the trabecular meshwork. Based on the morphological standards of a normal angle, an abnormal angle is determined and an angle recession abnormality indicator is generated. Taking the UBM image of patient Li as an example, an edge detection operator, namely the Sobel operator, is used in the image. This operator combines Gaussian smoothing and differential derivation to calculate the approximate gradient of the image grayscale function. This operator contains convolution kernels in both horizontal and vertical directions, respectively, for detecting horizontal and vertical edges. In Python, the OpenCV library is used to implement Sobel operator edge detection to determine the boundary coordinates of the angle region. Assuming the coordinates of the upper left corner of the angle region are (x1, y1) and the coordinates of the lower right corner are (x2, y2), this represents the location information of the angle structure. Further analysis of the anterior chamber angle revealed a significantly larger angle between the ciliary body and the iris root compared to a normal anterior chamber angle, and the trabecular meshwork texture was blurred. These data represent morphological features. Based on the morphological knowledge of normal anterior chamber angles, the anterior chamber angle was determined to be abnormal, and an angle recession abnormality indicator was generated, indicating that this angle may be receding.

[0025] The angle structure position information, morphological feature data, and angle recession abnormality indicators obtained in the previous step are comprehensively processed. To accurately determine the angle status, classification rules and thresholds are developed based on extensive clinical research and medical knowledge. For example, after studying numerous UBM images of normal and angle-recessed eyes, it was found that when the angle between the ciliary body and the iris root exceeds a specific angle (45 degrees), and when the trabecular meshwork texture exhibits abnormalities such as blurring or fractures, the angle is highly likely to be receding. These key features and values ​​are set as the judgment criteria. When information related to Mr. Li's angle image is obtained, it is compared with these established criteria. In Mr. Li's angle image, the angle between the ciliary body and the iris root is measured to exceed 45 degrees, and the trabecular meshwork texture is abnormally blurred, meeting the pre-defined angle recession judgment rules. Therefore, the angle is determined to be "angle recession," and this is the generated classification information.

[0026] Annotation confidence reflects the reliability of the diagnosis result. It is not randomly derived but calculated based on a comprehensive set of factors. For one thing, the degree of feature matching is important. In Li's case, various features of the anterior chamber angle were carefully compared with the standard features of angle recession. In addition to the angle and trabecular meshwork texture, other features such as the depth of the anterior chamber angle and the morphology of the iris were also considered. The higher the degree of match between these features and the standard features, the higher the confidence score. For example, the calculated similarity between multiple features of Li's anterior chamber angle and the standard features of angle recession reached a high level. Taking these factors into account, a deep learning-based confidence assessment model, which has learned the feature relationships between a large number of normal and abnormal anterior chamber angle images during training, ultimately determined that the annotation confidence for Li's anterior chamber angle was 0.85, indicating an 85% confidence level for this judgment. This allows doctors to better balance the accuracy and reliability of the diagnosis when considering the judgment results, providing more valuable evidence for subsequent diagnosis and treatment.

[0027] Based on the classification information of normal or receding anterior chamber angles and the annotation confidence information, the angle structure-related features in the target eye's UBM image data are screened. Features that are important for determining the anterior chamber angle status are retained, while redundant or interfering features are removed. The screened features are then integrated and annotated to generate the annotation information screening results. For the anterior chamber angle image of Mr. Li, who was diagnosed with "angle receding" with a confidence level of 0.85, the most critical features for determining angle receding were screened, such as the separation distance between the ciliary body and the iris root and the abnormal texture characteristics of the trabecular meshwork. These features are integrated and annotated in a specific format, such as recording the specific value of the separation distance and the type of abnormal texture, to form the annotation information screening results, paving the way for subsequent standardized processing.

[0028] The results of the annotation information screening were systematically organized and standardized according to unified standards and formats. This ensured that the annotation information completely and accurately characterized the state of the angle structure and related detailed information in the target eye's UBM image, including the angle's location, morphological characteristics, whether it was receding, and the confidence level of the judgment, to generate final annotation information for model training. The annotation information screening results for Li's angle image were organized and standardized according to a prescribed format, such as the YOLO format. The angle's location information (x1, y1, x2, y2), morphological feature descriptions (distance between the ciliary body and iris root, abnormal trabecular meshwork texture, etc.), classification information (angle receding), and annotation confidence level (0.85) were integrated to form standardized annotation information. This way, in subsequent training based on the YOLO-V8 model, the model can accurately read this annotation information, learn the characteristic patterns of angle receding, and improve recognition accuracy and reliability.

[0029] S104 , based on the target eye UBM image data in the target data set and combined with the YOLO-V8 deep learning algorithm, the chamber angle structure features in the eye UBM image are processed to generate an initial chamber angle recession recognition model.

[0030] In one embodiment, data augmentation is performed on the target ocular UBM image data in the target dataset to generate data-enhanced ocular UBM images. This includes, but is not limited to, adjusting brightness and adding noise to obtain different images, thereby expanding the number of images in the dataset and reducing the number of images that cannot be successfully processed due to insufficient original data. Assume that the target dataset contains an original ocular UBM image from patient A, which clearly shows the normal anterior chamber angle structure. According to the data augmentation strategy, the image is brightness-adjusted. In the RGB color space, the red, green, and blue channel values ​​of each pixel are added by 70 to brighten the image. For example, the original RGB value of a pixel in the image is (50, 60, 70), which becomes (120, 130, 140) after brightening. The np.clip function is then used to limit the pixel value to the range of 0-255 to prevent overflow. Similarly, the image can be dimmed by subtracting 70. Specifically, in the field of image processing, the RGB color space is a commonly used color representation. In this space, the color of each pixel is determined by the values ​​of the red (R), green (G), and blue (B) channels, each of which typically ranges from 0 to 255 (represented in 8 bits). Increasing the value of each channel increases the brightness of the pixel, thereby brightening the image. However, since pixel values ​​cannot exceed the specified range, they must be clipped after adjustment to ensure they remain between 0 and 255. Each pixel in the read image array is processed. Specifically, the red, green, and blue channel values ​​of each pixel are added by 70. This step utilizes tools specifically designed for array processing, efficiently increasing the channel values ​​of all pixels in the image uniformly, thereby increasing the overall image brightness.

[0031] The np.clip function in the NumPy library is used to clamp the pixel values ​​after brightness enhancement to the range of 0–255. This function automatically checks pixel values, adjusting pixel values ​​less than 0 to 0 and pixel values ​​greater than 255 to 255. This ensures pixel value validity and prevents image display errors caused by out-of-range values. The image, after brightness adjustment and pixel value clipping, is saved to a designated storage location. When saving, use appropriate image saving tools to ensure the correct image format for subsequent reading and use. The brightened image simulates UBM images acquired under different lighting conditions, enriching the data diversity. This helps the model learn the characteristics of normal anterior chamber angle structures under different lighting conditions. In real-world applications, the model can more accurately identify anterior chamber angle structures when exposed to images under various lighting conditions, improving its generalization and adaptability.

[0032] The principle behind image dimming is the opposite of brightening. By subtracting a fixed positive number, 70, from the red, green, and blue channel values ​​of each pixel, the image brightness is reduced, thus achieving image dimming. Similarly, to ensure pixel values ​​are within a reasonable range, post-processing clipping is required. Similar to the image brightening process, the original UBM image of patient A is read using the OpenCV library's read function and converted into a multidimensional array for subsequent processing. Within the read image array, each pixel is iterated over and the red, green, and blue channel values ​​are subtracted by 70, reducing the overall image brightness. The np.clip function is then used to process the reduced pixel values, ensuring that all pixel values ​​are within the normal range of 0-255. This prevents image display errors caused by pixel values ​​less than 0. The dimmed images further enrich the dataset, simulating UBM images under low-light conditions. This enables the model to learn the characteristics of normal anterior chamber angle structures under low-light conditions, enhancing its adaptability to diverse imaging conditions and ultimately improving its ability to accurately identify anterior chamber angle structures in complex environments.

[0033] Adding Gaussian noise is a key method for data augmentation of ocular UBM images. Its purpose is to simulate real-world image noise caused by factors such as sensor noise and signal transmission, thereby increasing data diversity and improving the model's robustness to interference. Since Gaussian noise is added to the image, and subsequent computations involve operations such as probability distribution, to facilitate computation and ensure data consistency, the image's pixel values ​​must first be normalized from the original range of 0-255 to 0-1. This changes the value range of each pixel in the image and prepares for the subsequent addition of noise. The np.random.normal function is used to generate a random noise matrix of the same size as the original image. For a 512×512 image, each element in the noise matrix generated by this function follows a Gaussian distribution with a mean of 0 and a standard deviation of 0.3. The Gaussian distribution is a common probability distribution with a specific shape and characteristics. These characteristics make the generated noise more consistent with real-world noise distribution patterns. For example, in the generated noise matrix, the value corresponding to a pixel in the original image is 0.1. The generated Gaussian noise matrix is ​​superimposed on the normalized image. Specifically, each element in the noise matrix is ​​added to the element at the same position in the image matrix. For example, if the normalized value of a pixel is 0.2, and the corresponding noise matrix element is 0.1, the new value of that pixel becomes 0.3. In this way, Gaussian noise is incorporated into the image, simulating the effect of an image affected by noise.

[0034] After adding noise, some pixel values ​​may exceed the range of 0-1. To ensure that pixel values ​​remain within a reasonable range, the np.clip function is used to clip the pixel values. This function adjusts pixel values ​​outside the 0-1 range, setting values ​​less than 0 to 0 and values ​​greater than 1 to 1, thereby ensuring the validity and rationality of the image data. Since the image was previously normalized, in order to restore the image to a format suitable for subsequent processing and use, the cropped pixel values ​​need to be denormalized from the 0-1 range to the 0-255 range. This step is achieved through a mathematical transformation method that is the opposite of normalization, returning the image pixel values ​​to their original value range, ultimately resulting in the UBM image of the eye after adding Gaussian noise.

[0035] In the process of data enhancement of eye UBM images, adding salt and pepper noise is an important means to expand data diversity and simulate image transmission interference. In order to simulate the random interference of the image during transmission, the researchers set the ratio of adding salt and pepper noise to 0.3. This means that in the processed image, 30% of the pixels will be randomly replaced with black (0) or white (255), thereby increasing the diversity and complexity of the image and enabling the model to learn more different image features during training. Taking a 512×512 eye UBM image as an example, based on the determined noise ratio of 0.3, the number of pixels that need to be added with noise is calculated. The formula shows that: 512×512×0.3≈78643. This number determines the total number of pixels that need to be randomly changed in subsequent operations.

[0036] Using the randint() function in the Python standard library's random function, we determine the pixel positions and colors to be added to the noise pixel by pixel through a loop. Each time through the loop, we first use the randint() function to randomly generate an integer within the width and height of the image, respectively, to determine the horizontal and vertical coordinates of the pixel to be added to the noise. For example, the randomly determined pixel position is (100, 200). Then, we use the randint() function again to generate one of the two random numbers, 0 or 1. If the generated random number is 0, the pixel at that position is set to black (pixel value 0); if the generated random number is 1, the pixel is set to white (pixel value 255). For example, at the pixel position (100, 200) determined above, when the generated random number is 1, the pixel becomes white (255). In this way, the approximately 78,643 calculated pixels are processed one by one.

[0037] After adding salt and pepper noise, the image is processed using other data augmentation methods, such as brightness adjustment and Gaussian noise. Brightening or dimming the image simulates imaging effects under different lighting conditions, while adding Gaussian noise simulates real-world noise caused by factors such as sensor noise and signal transmission. These processing steps ultimately yield various enhanced UBM images of the eye. These enhanced images exhibit diverse characteristics in terms of brightness and noise, significantly enriching the original dataset and providing richer, more representative data for subsequent training of the YOLO-V8-based deep learning model, helping to improve the model's generalization and recognition accuracy.

[0038] The ratio of adding salt and pepper noise is determined to be 0.3, and the number of pixels that need to add noise to the image is calculated (512×512×0.3≈78643). Through a loop, the randint() function in the Python standard library random is used each time to randomly generate horizontal and vertical coordinates within the width and height of the image to determine the pixel position. The randint() function is then used to generate 0 or 1. If it is 0, the pixel is set to black (0), and if it is 1, it is set to white (255). For example, if the randomly determined pixel position is (100,200) and the random number generated is 1, the pixel becomes white (255). After brightness adjustment, adding Gaussian noise and salt and pepper noise, a variety of data-enhanced UBM images of the eye are obtained.

[0039] The augmented UBM images and the target eye's UBM images are annotated using preset horizontal boxes to generate image classification information. This provides accurately labeled samples for model training, enabling the model to learn the characteristic differences between normal and receding angle structures. Labelimg annotation software is used to annotate the augmented images and the original target eye's UBM images. Consider an augmented image. An ophthalmologist observes that the angle structure is clear, with the ciliary body and iris root connected properly, consistent with normal angle characteristics. In Labelimg, the ophthalmologist uses the preset horizontal box tool to precisely select the angle region and then annotates the image as "normal angle." Another image, showing an abnormal angle structure with clear signs of separation between the ciliary body and iris root, is annotated as "recessed angle." After annotation, the software generates image classification information, recording the angle category and horizontal box coordinates. This information serves as a crucial basis for model training.

[0040] Based on the YOLO-V8 deep learning algorithm, the backbone network, neck network, and head network collaborate to extract features from the augmented eye UBM image and the target eye UBM image, generating angle-of-angle structural features. Taking a 640×640×3 eye UBM image as an example, the backbone network extracts features: the image first enters the ConvModule module, where convolution operations slide a specific convolution kernel across the image, performing a weighted summation of local regions to extract basic features such as edges and textures. Batch normalization normalizes the input to the intermediate layers, making the data more stable and facilitating model learning. The activation function introduces nonlinearity to enhance the model's expressiveness. After processing by this module, the image's feature map size is reduced to 320×320×64, with an increased number of channels, allowing for the initial extraction of basic features. Next, the image enters the CSPLayer_2Conv module, which splits the input features into two parts. One part is processed through two convolutional layers for further feature extraction, while the other part is directly concatenated. Finally, the two parts are concatenated along the channel dimension to output richer features. As the stages progress, the feature map size gradually decreases after passing through multiple similar modules. For example, after passing through four CSPLayer_2Conv modules, the size becomes 20×20×1024. Finally, the image enters the SPPF module, which fuses features through max pooling operations at different scales, expanding the receptive field and outputting a feature map of uniform size, completing the feature extraction of the backbone network.

[0041] Neck network feature fusion: The high-level features (e.g., 20×20) output by the backbone network are upsampled (e.g., bilinear interpolation) to the same size as the mid-level features (e.g., 40×40), concatenated with the features processed by the ConvModule, and further fused via CSPLayer_2Conv, resulting in a fused feature output (e.g., 80×80). This process is repeated, with the fused features further upsampled and concatenated with lower-level features (e.g., 80×80), completing the top-down transfer of semantic information. Simultaneously, the top-down fused features are downsampled (e.g., 3×3 convolution with a stride of 2), concatenated with high-level features, and further processed via CSPLayer_2Conv, transferring detailed information from the lower layers to the higher layers, ultimately outputting a multi-scale fused feature. Head network detection and prediction (primarily focusing on feature extraction): The multi-scale feature maps (e.g., 80×80, 40×40, and 20×20) output by the neck are fed into the head network. Although the head network is primarily used for detection and prediction, it also further processes features during this process. For example, the convolutional layers in the classification and regression branches perform convolution operations on the input feature maps to extract more detailed features related to the room angle category and location. These features together constitute the room angle structural features.

[0042] The angle structural features and image classification information generated in the previous steps are fed into the YOLO-V8 deep learning algorithm. The algorithm learns from this data and adjusts the model parameters, enabling the model to establish a correlation between angle structural features and image classification. For example, during training, the model continuously attempts to analyze and judge the angle structural features, compares them with the image classification information (normal angle or angle recession), and calculates a loss value (for example, using the BCE loss function to measure classification accuracy, and using the DFL and CIoULoss loss function to measure the accuracy of bounding box regression). Through the backpropagation algorithm, the parameters of the backbone network, neck network, and head network are continuously adjusted, ensuring that the model's predictions are increasingly close to the actual image classification information. After a certain number of training cycles, an initial angle recession recognition model is generated that can initially identify angle recession in ocular UBM images.

[0043] S105 , processing the initial angle recession recognition model based on the training set, the validation set, and the test set in the target data set to generate a target angle recession recognition model.

[0044] In one embodiment, in deep learning model training, in order to accurately evaluate the performance of the model and avoid overfitting, the dataset needs to be divided into different subsets. The training set is used to allow the model to learn the features and patterns in the data; the validation set is used to evaluate the performance of the model during training, adjust hyperparameters, and prevent the model from overfitting on the training set; the test set is used to finally evaluate the generalization ability of the model and test the performance of the model on unseen data. The dataset is divided according to the ratio of 7:2:1. For example, assume that the target dataset has a total of 1,000 UBM images of the eye, including 600 images of normal anterior chamber angle and 400 images of anterior chamber angle recession. This distribution of different types of images reflects the possible situations in actual clinical data. The dataset was divided into two groups according to a 7:2:1 ratio: 420 images (600×0.7) of the normal anterior chamber angle images were assigned to the training set, which served as the primary data source for the model to learn normal anterior chamber angle features; 120 images (600×0.2) were assigned to the validation set, used to monitor the model's recognition performance for normal anterior chamber angle images during training; and 60 images (600×0.1) were assigned to the test set, used to ultimately evaluate the model's generalization ability to normal anterior chamber angle images. For angle-recessed images, 280 images (400×0.7) were assigned to the training set, 80 images (400×0.2) to the validation set, and 40 images (400×0.1) to the test set, similarly serving the roles of learning, validation, and testing, respectively.

[0045] Directly randomly partitioning the entire dataset may lead to correlation between the data in the training, validation, and test sets, resulting in overfitting during model evaluation and failing to accurately reflect the model's true performance. To prevent this, the original dataset is first divided into normal angle images and angle recession images in a 7:2:1 ratio. This is because normal angle images and angle recession images have significant feature differences. If they are mixed together in a random partition, some similar images will appear in multiple subsets, causing bias in model training and evaluation. For example, if a group of similar angle recession images are concentrated in both the training and validation sets due to random partitioning, the model may overlearn the features of these similar images and perform poorly when encountering images with truly different angle recession. By partitioning separately, each subset ensures that it contains images with different characteristics, reducing data correlation. After partitioning, data augmentation is performed on each of the six folders (training, validation, and test folders for normal angles and training, validation, and test folders for angle recession).

[0046] Data augmentation is an important means to improve the generalization ability of the model. Data augmentation operations such as brightness adjustment, adding Gaussian noise and salt and pepper noise were performed on the images in these six folders. Brightness adjustment can simulate UBM images obtained under different lighting conditions. By increasing or decreasing the RGB values ​​of image pixels by a certain value, the image can be brightened or dimmed, so that the model can learn the characteristics of the anterior chamber angle structure under different lighting conditions. Adding Gaussian noise is to simulate the image noise caused by factors such as sensor noise and signal transmission in the real world. A Gaussian distribution random noise matrix with the same size as the image, a mean of 0 and a standard deviation of 0.3 is generated and added to the original image, so that the model can learn to recognize the anterior chamber angle structure in the presence of noise interference. Adding salt and pepper noise simulates the random interference of the image during transmission, randomly replacing some pixels in the image with black (0) or white (255), further increasing the diversity of the data. Through these data augmentation operations, rich training, validation and test data are obtained. These expanded data allow the model to be exposed to more different forms of images and learn a wider range of features, thereby enhancing the model's generalization ability and enabling it to more accurately identify angle recession when faced with various actual situations.

[0047] For the original ocular UBM images and their enhanced images from the training set, target hyperparameters were set and then fed into an initial angle recession recognition model built using YOLO-V8. Hyperparameter settings included disabling caching (cache=False) to optimize data loading efficiency; setting the input image size to 640×640 pixels (imgsz=640) to balance feature resolution and computational cost; and disabling mosaic data augmentation (close_mosaic=0) to preserve the original scene feature distribution. The model's backbone, neck, and head networks worked together, updating parameters via backpropagation to generate preliminary training results. During training, the Adam optimizer (optimizer='Adam') was used instead of traditional SGD, with an initial learning rate of 0.01 and a batch size of 8 (batch=8). During model training, the backbone network extracted basic image features, such as edges and texture; the neck network integrated features from different layers; and the head network performed classification and regression predictions. Through the back-propagation algorithm, the network parameters are continuously adjusted according to the difference between the predicted results and the true labels, so that the model's predicted results gradually approach the actual situation, thereby generating preliminary training results. Through preliminary training, the model begins to learn the characteristics of the chamber angle structure, laying the foundation for subsequent optimization and evaluation.

[0048] The preliminary training results were validated on the validation set to generate validation evaluation metrics. These metrics include Precision, Recall, mAP0.5 (mean average precision @ 0.5IoU), and mAP0.5:0.95 (mean average precision @ 0.5-0.95IoU). For example, suppose there are 100 images in the validation set, and the model predicts 80 of them as angle recession, of which 70 actually have angle recession. Precision = 70 ÷ 80 = 0.875, and Recall = 70 ÷ (the number of images with actual angle recession, assuming 80) = 0.875. By calculating the average precision at different IoU thresholds, we obtain metrics such as mAP0.5 and mAP0.5:0.95, which comprehensively reflect the model's performance on the validation set. The validation set is used to evaluate the performance of the model on unseen data. By generating evaluation indicators, we can understand the strengths and weaknesses of the model and provide a basis for subsequent optimization.

[0049] In object detection tasks, mAP0.5 and mAP0.5:0.95 are important metrics for evaluating the performance of angle recession recognition models in eye UBM images. mAP0.5 measures the average precision of all categories when the IoU (Intersection over Union, used to measure the degree of overlap between the predicted box and the true box) threshold is 0.5, reflecting the overall performance of the model under moderate overlap. mAP0.5:0.95 calculates the average precision of all categories within the IoU threshold range of 0.5 to 0.95 (step size 0.05), focusing more on evaluating the robustness of the model under different overlap levels. For example, when a model performs poorly on the mAP0.5:0.95 metric, it means that the model's ability to accurately identify angle recession areas and locate their boundaries under different overlap levels needs to be improved. What may happen is that the angle recession area predicted by the model and the actual angle recession area do not overlap enough in many cases, making it impossible to accurately determine the range and degree of angle recession. This may lead to misdiagnosis or missed diagnosis in actual clinical diagnosis, affecting the accurate judgment of the patient's condition.

[0050] Optimize the model structure by adjusting the number of convolution kernels: Convolution kernels play a key role in extracting image features in convolutional neural networks. Increasing the number of convolution kernels in certain layers allows the model to learn richer and more detailed features. For example, if a layer in the backbone network originally has 64 convolution kernels and the model's localization accuracy is insufficient, increase the number of kernels to 128. This allows the model to extract features of the anterior chamber angle structure from different angles and scales when processing images, potentially capturing more subtle features related to angle recession, such as subtle changes in tissue texture at the angle, thereby improving the model's ability to localize areas of angle recession. Modify the network layer connections: In addition to adjusting the number of convolution kernels, the network layer connections can also be optimized. In the neck network, change the path or method of top-down and bottom-up feature fusion. Weighted fusion assigns different weights to different features based on their importance, making the model more rational when fusing multi-scale features, better integrating information from different levels, and improving the recognition of angle recession features.

[0051] Optimize model parameters, specifically by adjusting the learning rate. The learning rate determines the step size for parameter updates during model training. If the model converges too slowly or tends to get stuck in a local optimum during training, optimize the learning rate. For example, if the initial learning rate is set to 0.01 and the model's validation set loss decreases slowly during training and shows no significant improvement for a long time, consider lowering the learning rate, such as to 0.001. A smaller learning rate makes the model more cautious when updating parameters, avoiding skipping optimal solutions. This helps the model fit the data more accurately and improves the accuracy of angle recession recognition. Optimize other parameters: In addition to the learning rate, parameters such as the batch size and optimizer selection also affect model performance. The batch size determines the number of samples fed into the model during each training session. Setting the batch size too large can lead to insufficient memory; setting it too small can make the training process unstable and slow convergence. If the current batch size is 8 and you notice significant model fluctuations during training, you can try adjusting it to 16. Observe the model's performance on the training and validation sets to find an appropriate batch size for more stable and efficient model training. Furthermore, different optimizers can also affect the convergence speed and performance of the model. For example, switching from the traditional stochastic gradient descent (SGD) optimizer to the adaptive moment estimation (Adam) optimizer can adaptively adjust the learning rate based on the parameter gradient, helping to improve the model's accuracy in identifying angle recession.

[0052] After adjusting the model structure and parameters, the adjusted model is retrained and validated using the test set data. This process repeats the training, validation, and evaluation steps. After each training session, the validation set is used to evaluate model performance and observe improvements in various metrics. If model performance is still unsatisfactory, further metrics are analyzed, the model structure or parameters are adjusted, and training and validation are repeated. For example, after one adjustment, if the model's mAP0.5:0.95 performance improves but still falls short of the expected clinical application standard, the model needs to be re-examined, possibly with further adjustments to the number of convolution kernels or the learning rate, before continuing training. Through multiple iterations of optimization, the model continuously adapts to the characteristics of the data, gradually improving its generalization and accuracy. When the model achieves satisfactory results across all evaluation metrics, the target angle recession recognition model is generated. This model can more accurately identify angle recession at varying degrees of overlap, improving the accuracy and efficiency of ocular disease diagnosis.

[0053] S106 , processing the target eye UBM image data and the annotation information based on the target angle recession recognition model to generate an eye angle recession recognition result.

[0054] In one embodiment, after the UBM image of the patient Li's eye is input into the model, the backbone network (Backbone) starts working. Taking the backbone network of YOLO-V8 as an example, the image first enters the ConvModule module, which integrates the basic operations of convolution, batch normalization and activation function. The convolution operation slides on the image through a specific convolution kernel, and performs weighted summation on the local pixels in the angle area of ​​the chamber to extract basic features such as edges, textures and other information. The convolution kernel detects the edge contour of the tissue at the angle of the chamber and preliminarily outlines the shape of the angle of the chamber; the batch normalization operation normalizes the input data of the intermediate layer, which helps the model learn features faster and better; the activation function introduces nonlinear factors to enhance the expression ability of the model, enabling the model to learn more complex feature relationships. After processing by this module, the size of the feature map of the image changes, the number of channels increases, and the basic features are preliminarily extracted. The image then enters the CSPLayer_2Conv module, which splits the input features into two parts. One part is further extracted through two convolutional layers, while the other part is directly concatenated and finally concatenated along the channel dimension to output richer features. As the stages progress, through multiple similar modules, the feature map size is gradually reduced, continuously achieving multi-scale feature extraction, capturing characteristic information of the chamber angle from different levels and angles.

[0055] The features extracted by the backbone network then enter the neck network (Neck). The neck network further integrates feature information at different levels through a top-down and bottom-up feature fusion process. For example, the high-level features output by the backbone network (such as a feature map of size 20×20) are upsampled (such as bilinear interpolation) to the same size as the mid-level features (such as 40×40), and then spliced ​​with the features processed by ConvModule. They are further fused through CSPLayer_2Conv and the fused features (such as 80×80) are output. This process is repeated continuously, and the fused features are further upsampled and spliced ​​with the lower-level features. At the same time, the underlying detailed information is passed to the higher layers through downsampling, and finally the multi-scale fused features are output, allowing the model to obtain more comprehensive and representative chamber angle feature information.

[0056] The features fused by the neck network enter the head network. The classification branch in the head network is responsible for calculating the probability of angle recession based on these fused features. The classification branch performs convolution operations on the feature map through convolutional layers to further extract fine-grained features related to angle recession. For example, convolution operations can more accurately identify features closely related to angle recession, such as texture changes and angle abnormalities at the angle. Based on these features, the BCE (Binary Cross Entropy) loss function is used to measure the difference between the predicted result and the true label (the feature pattern in the annotation information). The model parameters are continuously adjusted to ensure that the predicted result is as close to the true label as possible. During this process, the model assigns a score indicating the probability of angle recession based on the degree of feature matching. Suppose the model outputs a score of 0.8 for the probability of angle recession. This score is based on the model's analysis of the UBM image of Mr. Li's eye and the degree of match with the angle recession feature pattern in the annotation information. This score reflects the model's prediction of angle recession in the current image, meaning that the model believes there is an 80% probability of angle recession in this image.

[0057] Taking Mr. Li's angle likelihood score as an example, assume the target angle recession identification model sets an angle recession threshold range of 0.6-0.9, and Mr. Li's angle likelihood score is 0.8. When performing a comparison, the model automatically compares this score with the threshold range. Since 0.8 falls between 0.6 and 0.9, the model's prediction is within a reasonable range, but relatively close to the lower limit. This indicates that the model is not sensitive enough to the characteristics of angle recession. While it can determine the possibility of angle recession, it may not fully consider some subtle feature changes. Even if the score is within the threshold range, if it approaches the boundary, it is necessary to fine-tune the model to improve its accuracy. Adjusting the model parameters enhances the model's ability to recognize angle recession characteristics, enabling it to make more accurate judgments in various complex situations.

[0058] The model calculates a decision bias correction vector based on the difference between the score and the threshold. The model determines which parameters need to be adjusted, as well as the magnitude and direction of the adjustments, based on factors such as the degree and direction of the score's deviation from the threshold center and the model's structural characteristics. For example, if the score approaches the lower threshold, the model determines that sensitivity to angle recession features needs to be increased, and thus adjusts relevant parameters to enhance their response to these features. For example, adjusting the weights of the convolutional layers in the classification branch might include information about weight adjustments for multiple convolutional layers. For example, for a convolutional layer sensitive to angle recession, its weight might be increased by 0.05. This increase in weight means that the convolutional layer will respond more strongly to the relevant features when extracting features, enabling the model to more sensitively capture angle recession. In addition to the magnitude of the weight adjustment, the vector may also include information about the direction of the adjustment, such as whether to increase or decrease the weight, as well as the strategy for adjusting the weights of different channels. This information is used to ensure that the model's subsequent predictions more accurately reflect the true state of the angle. By generating and applying this decision bias correction vector, the model can optimize its own parameters. When processing new ocular UBM images subsequently, the adjusted model can more accurately extract chamber angle features and more reasonably calculate the probability score of chamber angle recession, thereby improving the overall recognition accuracy and providing more reliable support for clinical diagnosis.

[0059] The judgment bias correction vector is parsed and transformed to generate an angle recession identification result, clarifying whether the angle is receding and providing detailed information. In Li's case, when parsing the judgment bias correction vector, the model first reads the parameter adjustment portion of the vector. For example, if the vector indicates that the weight of a convolutional layer needs to be increased, the model interprets this as a need to strengthen the capture of specific angle features. These features may be related to the presence, location, or degree of angle recession. Based on this adjustment information, the model re-examines the previously extracted angle structural features. It focuses on feature signals that are enhanced after the weight adjustment, such as the manifestation of specific texture changes, angle abnormalities, and other features in the image. By re-analyzing and integrating these features, the model begins to determine the state of the angle.

[0060] After parsing the vector, the model determines whether the anterior chamber angle has receded based on certain criteria. This criterion is based on the characteristic patterns of normal anterior chamber angles and anterior chamber angle recedes, learned during previous training. If, after reanalyzing the features, it is found that the anterior chamber angle features closely match the characteristic patterns of anterior chamber angle recedes, reaching the judgment threshold set by the model, the model will determine that the anterior chamber angle has receded. For example, in the anterior chamber angle image of Mr. Li, the model, after adjusting the weights, found that certain key features of the anterior chamber angle, such as the connection between the ciliary body and the root of the iris, and the anterior chamber angle, were similar to the features of the samples with anterior chamber angle recedes in the training set, and the degree of similarity exceeded the set threshold. Therefore, the model determined that Mr. Li's anterior chamber angle had receded.

[0061] Determining the degree of angle recession also relies on the model's feature analysis. The model comprehensively considers changes in multiple features to assess the degree of recession. For example, features such as the magnitude of angle change, the distance between the ciliary body and the iris root, and the degree of damage to the trabecular meshwork are all included in the assessment. In Li's case, the model found that the angle was somewhat enlarged compared to the normal range, and the distance between the ciliary body and the iris root reached the standard for moderate recession. The trabecular meshwork also showed some blurring. Combining these features, the model determined that the angle recession was moderate. Determining the specific location of angle recession requires the model to carefully analyze the features of different regions of the angle in the image. The model monitors changes in features in different regions and locates the location of recession by comparing the differences in features between normal and receding angles. In Li's case, when the model analyzed the image, it found that the feature changes in the upper area of ​​the chamber angle were the most obvious. For example, the separation between the ciliary body and the root of the iris in the upper area was more prominent, and the texture abnormality of the trabecular meshwork in the upper area was more significant. Therefore, it was judged that the chamber angle recession was mainly concentrated in the upper area of ​​the chamber angle.

[0062] The server focuses on identifying angle recession in ocular UBM images, using a deep learning algorithm to build a model to improve diagnostic efficiency and accuracy. First, the target patient's original ocular UBM image data, clinical background information, and target dataset are obtained. Next, data enhancement is performed on the images in the target dataset, including brightness adjustment and the addition of Gaussian and salt-and-pepper noise to increase data diversity. Image classification information is then generated based on preset horizontal bounding box annotations. The backbone, neck, and head networks of the YOLO-V8 deep learning algorithm are then used to collaboratively extract angle structural features, thereby generating an initial angle recession recognition model. The initial model is then iteratively optimized by partitioning into training, validation, and test sets, resulting in the target angle recession recognition model. Next, the angle structure in the target ocular UBM image data is processed to generate annotations. Finally, the target model analyzes the image data and annotations to generate an angle recession probability score. Through comparison and analytical transformation, the presence of angle recession and detailed judgment information are determined, providing strong support for clinical diagnosis.

[0063] In one embodiment, Figure 2 As shown, the present application also provides a data processing device for detecting angle recession of ocular trauma using an eye UBM image, comprising:

[0064] An acquisition module 201 is used to acquire original eye UBM image data and a target data set of a target patient;

[0065] The processing module 202 is used to process the original eye UBM image data of the target patient based on the data enhancement strategy to generate target eye UBM image data; process the chamber angle structure in the target eye UBM image data to generate annotation information; process the chamber angle structure features in the eye UBM image based on the target eye UBM image data in the target data set in combination with the YOLO-V8 deep learning algorithm to generate an initial chamber angle recession recognition model, wherein the target eye UBM image data includes the original eye UBM image and the enhanced eye UBM image; process the initial chamber angle recession recognition model based on the training set, validation set, and test set in the target data set to generate a target chamber angle recession recognition model; and process the target eye UBM image data and annotation information based on the target chamber angle recession recognition model to generate an eye angle recession recognition result.

[0066] Each embodiment of this application is described in a related manner, and similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences from other embodiments. In particular, for the evaluation of the data processing method, electronic device, electronic device, and readable storage medium embodiment for detecting angle recession of ocular trauma using UBM images, since they are basically similar to the above-described data processing method embodiment for detecting angle recession of ocular trauma using UBM images, the description is relatively simple, and for related parts, refer to the partial description of the above-described data processing method embodiment for detecting angle recession of ocular trauma using UBM images.

Claims

1. A data processing method for detecting angle recession in ocular trauma using UBM images, characterized in that: include: Obtain the original eye UBM image data and target dataset of the target patient; The original eye UBM image data of the target patient is processed based on the data enhancement strategy to generate the target eye UBM image data; Processing the chamber angle structure in the target eye UBM image data to generate annotation information; Based on the target eye UBM image data in the target dataset and combined with the YOLO-V8 deep learning algorithm, the anterior chamber angle structural features in the eye UBM image are processed to generate an initial anterior chamber angle recession recognition model. The target eye UBM image data includes the original eye UBM image and the enhanced eye UBM image. The initial chamber angle recession recognition model is processed based on the training set, validation set, and test set in the target data set to generate a target chamber angle recession recognition model; The target eye UBM image data and annotation information are processed based on the target angle recession recognition model to generate the recognition result of the eye angle recession.

2. The method according to claim 1, wherein The original eye UBM image data of the target patient is processed based on the data enhancement strategy to generate the target eye UBM image data, including: Perform brightness adjustment on the target eye UBM image data in the target dataset to generate eye UBM images under different lighting conditions; Perform Gaussian noise and salt-and-pepper noise processing on the target eye UBM image data to generate a noise-processed eye UBM image; The eye UBM images under different lighting conditions and the eye UBM images after noise processing are processed to generate target eye UBM image data.

3. The method according to claim 1, wherein Process the angle structure in the target eye UBM image data to generate annotation information, including: Extract the position and morphological features of the angle structure in the target eye UBM image data and perform abnormality judgment processing to generate angle structure position information, morphological feature data and angle recession abnormality identification; Processing the angle structure position information, morphological feature data, and angle recession abnormality identification to generate classification information and annotation confidence information of whether the angle is normal or recession; Based on the classification information of normal or receding chamber angle and the annotation confidence information, the chamber angle structure-related features in the target eye UBM image data are screened and annotated to generate an annotation information screening result; The annotation information screening results are collated and standardized to generate annotation information, wherein the annotation information is used to characterize the state of the chamber angle structure and related detailed information in the UBM image of the target eye.

4. The method according to claim 1, wherein Based on the target eye UBM image data in the target dataset and combined with the YOLO-V8 deep learning algorithm, the anterior chamber angle structural features in the eye UBM image are processed to generate an initial anterior chamber angle recession recognition model, including: Performing data enhancement processing on the target eye UBM image data in the target data set to generate a data-enhanced eye UBM image; Annotate the data-enhanced eye UBM image and the target eye UBM image data based on the preset horizontal frame to generate image classification information; Based on the YOLO-V8 deep learning algorithm, the backbone network, neck network, and head network are used to perform feature extraction on the enhanced eye UBM image and the target eye UBM image data to generate the anterior chamber angle structure features. The structural features of the anterior chamber angle and image classification information are processed based on the YOLO-V8 deep learning algorithm to generate an initial anterior chamber angle recession recognition model.

5. The method according to claim 4, wherein The initial chamber angle recession recognition model is processed based on the training set, validation set, and test set in the target dataset to generate a target chamber angle recession recognition model, including: A training set was created from the target dataset in a 7:2:1 ratio. Target hyperparameters were set for the original eye UBM images and their enhanced images in the training set. The initial angle recession recognition model built based on YOLO-V8 was then fed into the model. The model's backbone, neck, and head networks worked together, and parameters were updated through backpropagation to generate preliminary training results. Verify the preliminary training results on the validation set and generate validation evaluation indicators; The initial angle recession recognition model was evaluated and optimized based on the validation evaluation indicators and test set. The performance of the model at different overlap levels was analyzed according to the indicators, and the model structure and parameter information were optimized. The adjusted model was combined with the test set data for retraining and verification. After iterative optimization, the target angle recession recognition model was generated.

6. The method according to claim 1, wherein The target eye angle recession recognition model is used to process the target eye UBM image data and annotation information to generate an eye angle recession recognition result, including: Based on the target angle recession recognition model, feature matching and analysis are performed on the target eye UBM image data and annotation information to generate a probability score for angle recession. Processing the classification and regression parameters within the target angle recession recognition model, comparing the angle recession probability score with the angle recession determination threshold range set by the model, and generating a corresponding determination deviation correction vector; The judgment deviation correction vector is analyzed and converted to generate an identification result of eye angle recession, clarifying whether the eye angle is recession and related detailed judgment information.

7. A data processing device for detecting angle recession in ocular trauma using UBM images, characterized in that: The device includes: An acquisition module is used to obtain the original eye UBM image data and target data set of the target patient; A processing module processes the original eye UBM image data of the target patient based on a data enhancement strategy to generate target eye UBM image data; Processing the chamber angle structure in the target eye UBM image data to generate annotation information; It is used to process the chamber angle structure features in the eye UBM image based on the target eye UBM image data in the target dataset combined with the YOLO-V8 deep learning algorithm to generate an initial chamber angle recession recognition model, where the target eye UBM image data includes the original eye UBM image and the enhanced eye UBM image; the initial chamber angle recession recognition model is processed based on the training set, validation set, and test set in the target dataset to generate a target chamber angle recession recognition model; the target eye UBM image data and annotation information are processed based on the target chamber angle recession recognition model to generate an eye chamber angle recession recognition result.

8. An electronic device, characterized in that: include: a first processor; and a memory for storing executable instructions of the first processor; The first processor is configured to execute the data processing method for detecting angle recession caused by ocular trauma in ocular UBM images according to any one of claims 1 to 6 by executing the executable instructions.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the second processor, the data processing method for detecting angle recession of ocular trauma using ocular UBM images according to any one of claims 1 to 6 is implemented.

Images