An AI feedback closed-loop-based child fundus photograph recognition and optic nerve development automatic grading method, system, terminal and storage medium

Abstract

The application relates to the technical field of intelligent image analysis, and discloses a child fundus photograph recognition and optic nerve development automatic grading method and system based on AI feedback closed loop, a terminal and a storage medium. The system comprises an image acquisition module, an image preprocessing module, a feature extraction module, a fusion diagnosis module, a report warning module and a feedback optimization module, the image acquisition module acquires a child fundus image; the image preprocessing module pre-processes the fundus image to obtain a target fundus image; the feature extraction module segments the target fundus image, and generates an image feature vector according to a segmentation result; the fusion diagnosis module inputs child clinical data and the image feature vector after multi-modal feature fusion into a deep learning model, and outputs an optic nerve development continuity score and an automatic grading result; and the report warning module generates a structured diagnosis report according to the automatic grading result and pushes an early warning information. The application realizes automatic evaluation of child optic nerve development.

Description

Technical Field

[0001] This invention relates to the field of intelligent image analysis technology, and in particular to a method, system, terminal, and computer-readable storage medium for automatic grading of pediatric fundus photo recognition and optic nerve development based on AI feedback closed loop. Background Technology

[0002] Children's eye health is a key public health concern worldwide. The developmental status of the optic nerve is an important indicator for assessing children's visual function and predicting the risk of eye diseases. Accurate assessment of the optic nerve is a core prerequisite for achieving early screening, timely intervention, and long-term management of childhood eye diseases. Currently, traditional assessments of optic nerve development mainly rely on the clinical experience and subjective judgment of ophthalmologists, as well as auxiliary examinations such as fundus endoscopy, optical coherence tomography (OCT), and visual field testing.

[0003] However, existing technologies have the following significant drawbacks: First, traditional examinations focus on disease diagnosis at a single point in time, lacking dynamic and continuous monitoring and graded assessment of children's optic nerve development, particularly lacking age-stratified baseline models for optic nerve development, making it difficult to determine whether a child's development is on a normal track. Second, the assessment of optic nerve development is highly dependent on the doctor's experience and subjective judgment; assessment results may vary significantly between different doctors, and even among the same doctor at different time points, leading to poor consistency and repeatability of diagnoses, especially in primary healthcare institutions. Third, current fundus screenings are often isolated events, lacking effective connections between diagnostic results and subsequent intervention plan development, follow-up management, and effectiveness evaluation, failing to form a closed loop. Furthermore, the vast amount of fundus image data accumulated by medical institutions has not been fully utilized, lacking systematic data annotation, feature extraction, and model training mechanisms, resulting in low data utilization.

[0004] Therefore, existing technologies still need to be improved and developed. Summary of the Invention

[0005] The main objective of this invention is to provide a method, system, terminal, and computer-readable storage medium for children's fundus image recognition and automatic grading of optic nerve development based on AI feedback closed loop, aiming to solve the problems of lack of dynamic monitoring and objective grading methods for children's optic nerve development, lack of closed loop for diagnosis and intervention, and low utilization rate of fundus image data in the prior art.

[0006] To achieve the above objectives, the present invention provides a child fundus image recognition and optic nerve development automatic grading system based on AI feedback closed loop. This system includes: an image acquisition module, an image preprocessing module, a feature extraction module, a fusion diagnosis module, a report and early warning module, and a feedback optimization module. The image acquisition module, image preprocessing module, feature extraction module, fusion diagnosis module, and report and early warning module are sequentially connected, and the feedback optimization module is connected to both the fusion diagnosis module and the report and early warning module. The image acquisition module is used to acquire fundus images of children; The image preprocessing module is used to preprocess the fundus image to obtain the target fundus image; The feature extraction module is used to segment the target fundus image to obtain a segmentation result, calculate fundus structure quantization parameters based on the segmentation result, and convert the fundus structure quantization parameters into an image feature vector. The fusion diagnostic module is used to acquire children's clinical data, fuse the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, input the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. The report and early warning module is used to generate a structured diagnostic report based on the automatic classification results and push early warning information through a push channel; The feedback optimization module is used to obtain the clinical gold standard data corresponding to the automatic grading results, use the clinical gold standard data as training data, and fine-tune the deep learning model online to obtain the updated deep learning model.

[0007] Furthermore, the image acquisition module includes: An image acquisition unit is used to acquire fundus images of children, perform quality assessment on the fundus images, and obtain quality assessment results; The re-acquisition control unit is used to re-acquire fundus images of the child when the quality assessment results are not up to standard, until fundus images that pass the quality assessment are acquired.

[0008] Furthermore, the image preprocessing module includes: An image denoising unit is used to denoise the fundus image to obtain a denoised fundus image. An enhancement and correction unit is used to perform contrast enhancement and color correction processing on the denoised fundus image to obtain the target fundus image.

[0009] Furthermore, the feature extraction module includes: An image segmentation unit is used to input the target fundus image into a deep learning segmentation network, the deep learning segmentation network segments the target fundus image and outputs segmentation results, wherein the segmentation results include optic disc segmentation results, optic cup segmentation results, retinal nerve fiber layer segmentation results and retinal vessel segmentation results; The feature encoding unit is used to calculate fundus structure quantization parameters based on the segmentation results and convert the fundus structure quantization parameters into image feature vectors. The fundus structure quantization parameters include cup-to-disc ratio parameters, retinal nerve fiber layer thickness parameters, and blood vessel morphology parameters.

[0010] Furthermore, the fusion diagnostic module includes: A multimodal fusion unit is used to acquire pediatric clinical data and use a graph neural network or Transformer model to fuse the image feature vector with the pediatric clinical data to generate fused features. A multi-task diagnostic unit is used to input the fused features into a multi-task learning deep learning model, which outputs a visual nerve development continuity score and an automatic grading result.

[0011] Furthermore, the reporting and early warning module includes: The report generation unit is used to generate a structured diagnostic report based on the automatic grading results, wherein the structured diagnostic report includes personalized follow-up recommendations and intervention recommendations; The early warning push unit is used to push early warning information to parents and doctors through a push channel when the automatic grading result reaches the preset early warning conditions.

[0012] Furthermore, the feedback optimization module includes: The data collection unit is used to acquire the clinical gold standard data corresponding to the automatic grading results. The clinical gold standard data includes expert review diagnosis results, OCT examination data, visual field examination data, intraocular pressure measurement data, and intervention effect data. The model fine-tuning unit is used to use the clinical gold standard data as training data to fine-tune the deep learning model online through federated learning or incremental learning techniques to obtain an updated deep learning model.

[0013] Furthermore, to achieve the above objectives, the present invention also provides a method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop. This method is applied to the aforementioned automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop. The method includes the following steps: The image acquisition module acquires fundus images of the child; The image preprocessing module preprocesses the fundus image to obtain the target fundus image; The feature extraction module segments the target fundus image to obtain a segmentation result, calculates fundus structure quantization parameters based on the segmentation result, and converts the fundus structure quantization parameters into an image feature vector. The fusion diagnostic module acquires children's clinical data, fuses the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, inputs the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. The report and early warning module generates a structured diagnostic report based on the automatic classification results and pushes early warning information through a push channel; The feedback optimization module obtains the clinical gold standard data corresponding to the automatic grading result, uses the clinical gold standard data as training data, and fine-tunes the deep learning model online to obtain an updated deep learning model.

[0014] Furthermore, to achieve the above objectives, the present invention also provides a terminal, wherein the terminal includes: a memory, a processor, and an AI-based feedback closed-loop-based automatic grading program for pediatric fundus image recognition and optic nerve development, stored in the memory and executable on the processor. When the AI-based feedback closed-loop-based automatic grading program for pediatric fundus image recognition and optic nerve development is executed by the processor, it implements the steps of the AI-based feedback closed-loop-based automatic grading method for pediatric fundus image recognition and optic nerve development as described above.

[0015] Furthermore, to achieve the above objectives, the present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop, wherein when the program for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop is executed by a processor, it implements the steps of the method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described above.

[0016] The beneficial effects of this invention are as follows: The invention acquires fundus images through an image acquisition module and improves image quality through preprocessing. The feature extraction module transforms image information into structured image feature vectors through segmentation and quantization, enabling the quantitative representation of the microscopic structural features of the fundus images. The fusion diagnosis module performs multimodal fusion of image features and pediatric clinical data, simultaneously outputting a continuous score for optic nerve development and automatic grading results using a deep learning model, achieving an objective and multidimensional quantitative assessment of optic nerve development status. The report and early warning module automatically generates a structured diagnostic report based on the grading results and pushes early warning information, ensuring timely communication of diagnostic conclusions to clinical decision-makers and achieving a closed loop between diagnosis and intervention. The feedback optimization module fine-tunes the model online by collecting clinical gold standard data, forming a data-driven continuous evolution mechanism that continuously improves the model's diagnostic capabilities with the accumulation of clinical data. The entire system forms a self-consistent data flow link from image acquisition to model optimization, not only achieving efficient utilization of fundus images and clinical data but also enabling the deep learning model to continuously adapt to new clinical data distributions through an online fine-tuning mechanism, providing dynamic updates and objective quantitative technical support for the assessment of pediatric optic nerve development. Attached Figure Description

[0017] Figure 1 This is a structural diagram of a preferred embodiment of the present invention, a child fundus photo recognition and optic nerve development automatic grading system based on AI feedback closed loop; Figure 2 This is a flowchart of a preferred embodiment of the present invention, which is a method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop; Figure 3 This is a structural diagram of a preferred embodiment of the terminal of the present invention. Detailed Implementation

[0018] This application provides a method, system, terminal, and storage medium for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop. To make the objectives, technical solutions, and effects of this application clearer and more explicit, the following detailed description, with reference to the accompanying drawings and embodiments, further illustrates this application. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0019] It will be understood by those skilled in the art that, unless otherwise defined, all terms used herein (including technical and scientific terms) have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains. It should also be understood that terms such as those defined in general dictionaries should be understood to have the same meaning as in the context of the prior art, and should not be interpreted in an idealized or overly formal sense unless specifically defined as herein.

[0020] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.

[0021] The preferred embodiment of the present invention describes a child fundus image recognition and automatic grading system for optic nerve development based on AI feedback closed loop, such as... Figure 1 As shown, the AI-based feedback closed-loop system for children's fundus image recognition and automatic grading of optic nerve development includes: an image acquisition module, an image preprocessing module, a feature extraction module, a fusion diagnosis module, a report and early warning module, and a feedback optimization module. The image acquisition module, the image preprocessing module, the feature extraction module, the fusion diagnosis module, and the report and early warning module are connected in sequence, and the feedback optimization module is connected to the fusion diagnosis module and the report and early warning module respectively. The image acquisition module is used to acquire fundus images of children; The image preprocessing module is used to preprocess the fundus image to obtain the target fundus image; The feature extraction module is used to segment the target fundus image to obtain a segmentation result, calculate fundus structure quantization parameters based on the segmentation result, and convert the fundus structure quantization parameters into an image feature vector. The fusion diagnostic module is used to acquire children's clinical data, fuse the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, input the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. The report and early warning module is used to generate a structured diagnostic report based on the automatic classification results and push early warning information through a push channel; The feedback optimization module is used to obtain the clinical gold standard data corresponding to the automatic grading results, use the clinical gold standard data as training data, and fine-tune the deep learning model online to obtain the updated deep learning model.

[0022] In the above-described solution, this invention upgrades traditional static fundus screening into a dynamic and intelligent optic nerve development management system. The system employs a three-layer cloud-edge collaborative architecture: the terminal layer includes fundus camera equipment and supporting edge computing units, deploying image acquisition and quality assessment functions; the cloud layer deploys AI analysis models, databases, and report generation and model optimization functions; the application layer provides a doctor's workstation web interface or client application, as well as a mobile application for parents. All modules interact via encrypted communication protocols. During system operation, the image acquisition module first integrates an AI quality assessment mechanism at the acquisition end to ensure that the input image quality meets standards; the image preprocessing module performs denoising, enhancement, and color correction on qualified images to improve image quality; the feature extraction module performs computationally intensive image segmentation and feature extraction at the edge, reducing cloud computing pressure and protecting the privacy of original data; the fusion diagnosis module uses a deep learning model to perform multimodal fusion analysis, outputting quantitative developmental scores and grading results, realizing a shift from subjective experience-based judgment to objective data-driven analysis; the report and early warning module transforms the analysis results into actionable clinical decision support information, pushing early warnings through multiple channels; and the feedback optimization module achieves continuous model evolution through federated learning and incremental learning. Through the collaborative work of the above modules, this invention achieves objectivity, precision, dynamism, and intelligence in the assessment of children's optic nerve development, effectively solving problems such as missing developmental dimensions, strong subjectivity, closed-loop breaks, low data utilization, and insufficient privacy protection in existing technologies.

[0023] Furthermore, the image acquisition module includes: An image acquisition unit is used to acquire fundus images of children, perform quality assessment on the fundus images, and obtain quality assessment results; The re-acquisition control unit is used to re-acquire fundus images of the child when the quality assessment results are not up to standard, until fundus images that pass the quality assessment are acquired.

[0024] The purpose of limiting the image acquisition module is to ensure that the fundus images input to the subsequent analysis module have sufficient quality, so as to avoid judgment errors caused by image quality problems.

[0025] In this embodiment, the image acquisition unit employs a high-resolution (≥10 megapixels) non-contact fundus camera (such as the Topcon TRC-NW400 or Canon CR-2 PLUS AF). This device supports non-mydriatic imaging, features autofocus and automatic exposure, and is adaptable to the cooperation characteristics of children during examinations. For primary healthcare institutions or community screening scenarios, a portable handheld fundus scope (such as the Volk Pictor Plus or Optomed Smartscope) can be used, connecting to a tablet computer via Wireless Fidelity (Wi-Fi) or Bluetooth for mobile screening. During acquisition, the child is guided to fix their head position and gaze at the fixation target within the device according to standardized operating procedures. Images of multiple visual field regions centered on the optic disc and the macula are acquired separately. The device supports output in multiple image formats, including Digital Imaging and Communications in Medicine (DICOM), Joint Photographic Experts Group (JPEG), and Portable Network Graphics (PNG).

[0026] The real-time image quality assessment unit employs a lightweight convolutional neural network (CNN) model (such as MobileNetV2 or SqueezeNet architecture). This model is deployed in the acquisition device or edge computing unit and performs quality scoring immediately after image acquisition. Evaluation metrics include: Sharpness score, used to determine image blurriness based on image sharpness and edge gradients, employing Laplacian variance or BRISQUE no-reference image quality assessment methods; Brightness score, used to assess whether overall brightness is appropriate, avoiding overexposure or underexposure, and using image histogram analysis to determine whether pixel value distribution is within a reasonable range (e.g., pixel values ​​in an 8-bit grayscale image should be between 50 and 240); Contrast score, used to assess whether local contrast is sufficient, using a contrast sensitivity function or local standard deviation analysis; Visual field integrity score, used with a lightweight semantic segmentation network to determine whether key anatomical structures such as the optic disc and macula are intact within the visual field, and whether there is eyelid or eyelash obstruction; and Artifact detection score, used to identify and label the presence of artifacts such as reflection spots, motion blur, and tear streaks.

[0027] When the overall quality score falls below a preset threshold (e.g., 0.85, normalized to the 0-1 range), the resampling control unit displays a prompt message on the screen or issues a prompt sound, informing the operator which specific indicator is not up to standard and suggesting adjustments. For example, it may prompt "Insufficient image clarity, please refocus" or "Incomplete field of view, please adjust the shooting angle." Through this mechanism, the present invention ensures that only qualified images enter the subsequent analysis process, effectively improving the reliability of the diagnosis.

[0028] It should be noted that the image quality assessment model can employ knowledge distillation technology to reduce computational load while maintaining assessment accuracy, enabling it to run smoothly on mobile devices. For multiple consecutive substandard images, the system can automatically enter a troubleshooting mode, prompting users to check the cleanliness of the device lens or the patient's cooperation. The image quality assessment model can be periodically fine-tuned using newly acquired image data to adapt to changes in different device types and acquisition environments.

[0029] Furthermore, the image preprocessing module includes: An image denoising unit is used to denoise the fundus image to obtain a denoised fundus image. An enhancement and correction unit is used to perform contrast enhancement and color correction processing on the denoised fundus image to obtain the target fundus image.

[0030] The purpose of the aforementioned image preprocessing module is to eliminate noise interference in fundus images, enhance image contrast and color consistency, and make key anatomical structures clearer and more identifiable.

[0031] In this embodiment, the image denoising unit first performs denoising processing on the color fundus image that has passed the quality assessment. Specifically, a nonlocal mean filtering algorithm is used to remove image noise. This algorithm utilizes the self-similarity in the image to effectively suppress noise while preserving edge details. Wavelet denoising or deep learning-based denoising networks (such as DnCNN) can also be used to obtain better denoising results. Taking nonlocal mean filtering as an example, this algorithm searches for similar pixel blocks in the image for each pixel point and uses the weighted average of these similar blocks to estimate the true value of the current pixel, thereby preserving key structural information such as optic disc boundaries and blood vessel edges to the greatest extent while denoising.

[0032] The enhancement and correction unit performs contrast enhancement and color correction on the denoised fundus image. For contrast enhancement, Adaptive Histogram Equalization (AHE) or Contrast Limited Adaptive Histogram Equalization (CLAHE) algorithms are used to enhance local contrast and highlight fine structures such as blood vessels and optic nerve fibers. The CLAHE algorithm divides the image into multiple small regions, performs histogram equalization on each region, and limits the contrast amplification factor to prevent excessive noise amplification, making it particularly suitable for enhancing blood vessels and optic disc edges in fundus images. Gamma correction or Retinex algorithms (such as Multi-Scale Retinex with Color Restoration (MSRCR)) can also be used to improve illumination unevenness. For color correction, color space conversion is performed (e.g., from Red Green Blue (RGB) space to Hue Saturation Value (HSV) space or CIELAB color space), and color constancy algorithms are applied to eliminate color deviations caused by different acquisition devices or lighting conditions, ensuring consistency of input to subsequent segmentation networks. For example, for images with color casts due to device differences, the enhancement correction unit calculates the color mean of the image in the LAB space and corrects it to the standard reference white balance value, ensuring consistency of optic disc and blood vessel colors across different devices. The preprocessed image is the target fundus image.

[0033] It should be noted that for scenarios with high processing speed requirements, a simplified preprocessing workflow can be adopted, performing only necessary median filtering for noise reduction and histogram equalization to achieve a balance between image quality and processing speed. Preprocessing parameters (such as filter kernel size and contrast enhancement intensity) can be adaptively adjusted based on the image quality assessment results. For images with good quality, the preprocessing intensity can be reduced to retain more original details.

[0034] Furthermore, the feature extraction module includes: An image segmentation unit is used to input the target fundus image into a deep learning segmentation network, the deep learning segmentation network segments the target fundus image and outputs segmentation results, wherein the segmentation results include optic disc segmentation results, optic cup segmentation results, retinal nerve fiber layer segmentation results and retinal vessel segmentation results; The feature encoding unit is used to calculate fundus structure quantization parameters based on the segmentation results and convert the fundus structure quantization parameters into image feature vectors. The fundus structure quantization parameters include cup-to-disc ratio parameters, retinal nerve fiber layer thickness parameters, and blood vessel morphology parameters.

[0035] The purpose of the aforementioned feature extraction module is to accurately extract biological features with clinical diagnostic value from the target fundus image and to uniformly vectorize multi-source heterogeneous features, thereby realizing the transformation from image pixels to clinical parameters.

[0036] In this embodiment, the image segmentation unit employs multiple pre-trained deep learning segmentation networks to perform pixel-level precise segmentation of the target fundus image. Specifically: Optic disc and optic cup segmentation: Pixel-level segmentation of the optic disc and optic cup is performed using U-Net, DeepLabV3+, or Mask R-CNN semantic segmentation networks. Taking U-Net as an example, this network employs an encoder-decoder structure. The encoder extracts image features layer by layer through convolutional and pooling layers, while the decoder restores spatial resolution through upsampling and skip connections, ultimately outputting a probability map of each pixel belonging to either the optic disc or optic cup region. The training data must include a large number of meticulously annotated fundus images (at least 5000 cases, covering different ages, sexes, and disease states) by professional ophthalmologists, with annotations including the optic disc boundary and the optic cup boundary. Based on the segmentation results, the feature encoding unit accurately calculates the optic disc area, optic cup area, and cup-to-disc ratio (CDR). Simultaneously, the optic disc edge sharpness (by analyzing the gradient change rate of the edge region) and optic disc morphological features (such as optic disc ellipticity and tilt angle) are calculated.

[0037] Retinal Nerve Fiber Layer (RNFL) Segmentation: RNFL segmentation is performed using a deep learning segmentation network based on Residual Network (ResNet) or Densely Connected Convolutional Network (DenseNet). When optical coherence tomography (OCT) data is unavailable, the system indirectly assesses the RNFL status by analyzing the texture features and vascular patterns of fundus images. When OCT equipment is available, multimodal fusion can be performed using OCT data to obtain a more accurate RNFL thickness map.

[0038] Retinal vessel segmentation: Pixel-level segmentation of retinal vessels was performed using UNet++ or Attention U-Net deep learning segmentation networks. UNet++ improves upon traditional U-Net through nested dense skip connections, enabling better capture of multi-scale vascular features. Based on segmentation, vessel diameter, color, and topological features were further utilized to distinguish between arteries and veins. Extracted vascular morphology parameters included: vessel diameter (average diameter of major vessels), tortuosity (ratio of the actual length of a vessel segment to the straight-line distance between its two ends), branching angle (angle between two branches at a bifurcation point), fractal dimension (a quantitative indicator of vascular network complexity), vessel density (proportion of vessel pixels to total pixels), and vessel course patterns (such as parallel course, crossing compression, and other abnormal patterns).

[0039] The feature encoding unit integrates all the extracted features, including optic disc features (area, CDR, edge sharpness, etc.), optic cup features (area, etc.), RNFL features (average thickness, sector thickness, defect area, etc.), and vascular features (diameter, tortuosity, fractal dimension, density, etc.). This unit first normalizes each feature, then uses Principal Component Analysis (PCA) to reduce the dimensionality of the high-dimensional feature space to a low-dimensional space, preserving key information while reducing redundancy; or it uses an autoencoder for unsupervised feature encoding, generating a high-dimensional image feature vector in a unified format (e.g., a 512-dimensional or 1024-dimensional floating-point vector). This feature vector is uploaded to the cloud-based AI analysis and grading module via an encrypted channel.

[0040] It should be noted that all of the above segmentation networks can run at edge computing devices (such as edge computing boxes equipped with graphics processing units (GPUs) or high-performance tablets) at real-time or near real-time speeds (processing time for a single image is less than 2 seconds). In network-free environments, the system can cache the original images and extracted feature vectors, and upload them in batches after the network recovers. Each segmentation network can be fine-tuned periodically using newly labeled data to improve segmentation accuracy.

[0041] Furthermore, the fusion diagnostic module includes: A multimodal fusion unit is used to acquire pediatric clinical data and use a graph neural network or Transformer model to fuse the image feature vector with the pediatric clinical data to generate fused features. A multi-task diagnostic unit is used to input the fused features into a multi-task learning deep learning model, which outputs a visual nerve development continuity score and an automatic grading result.

[0042] The purpose of the aforementioned fusion diagnostic module is to deeply integrate multi-source heterogeneous data (image features and clinical data), and use a multi-task deep learning model to achieve accurate, quantitative assessment and automatic grading of the optic nerve development status, thus upgrading from single-modal analysis to multi-modal comprehensive diagnosis.

[0043] In this embodiment, the multimodal fusion unit first acquires the child's clinical structured data. Specifically, the following data is acquired through the Hospital Information System (HIS) interface or manual input: age (accurate to the month), gender, ethnicity, birth history (whether premature, birth weight, etc.), family history (parental myopia, family history of fundus diseases, etc.), refractive error (spherical, cylindrical, axial length), axial length (acquired via A-scan or optical biometry), corneal curvature (acquired via corneal topography), intraocular pressure (acquired via a non-contact tonometer), past medical history (whether previously diagnosed with eye diseases, whether systemic diseases affect the eyes, etc.), and eye habits (average daily outdoor activity time, average daily electronic product usage time, reading and writing posture, etc.). The above data is standardized, for example, age is encoded by month, refractive error is converted to equivalent spherical power, categorical variables are one-hot encoded, and continuous variables are Z-score standardized, ultimately forming a clinical structured feature vector.

[0044] The multimodal fusion unit deeply fuses image feature vectors with clinical structured feature vectors. Fusion strategies can include early fusion (directly concatenating the two feature vectors to form a joint feature vector), late fusion (training models for each modality separately and then weighting or voting on the prediction results), or hybrid fusion. This embodiment preferably adopts a hybrid fusion strategy based on an attention mechanism: first, image features and clinical features are processed separately through two independent fully connected branches; then, a cross-attention module is introduced to capture the correlation between image features and clinical features; finally, a gated fusion layer adaptively determines the contribution weight of each modality to the final prediction. Specifically, a graph neural network (GNN) or a Transformer model is used to handle the complex correlations of heterogeneous data. Taking the Transformer model as an example, image feature vectors and clinical feature vectors are used as two types of tokens, respectively. The multi-head self-attention mechanism of the Transformer learns the deep correlation between the two types of features and outputs the fused feature vector.

[0045] The multi-task diagnostic unit inputs the fused multimodal features into a pre-trained multi-task learning deep learning model. This model employs a hybrid architecture: a convolutional neural network (such as EfficientNet-B3) is used as the backbone network for image feature extraction, followed by a Transformer encoder or Long Short-Term Memory (LSTM) network to process temporal features (such as historical axial length growth curves and refractive error trends). A Squeeze-and-Excitation Network (SE-Net) or a Convolutional Block Attention Module (CBAM) is introduced to automatically focus on the channels and spatial regions most important for diagnosis. The model uses a multi-task learning framework, sharing backbone network parameters and simultaneously outputting results for three tasks: a continuous score for optic nerve development (a continuous value from 0 to 100, reflecting the overall state of optic nerve development), automatic grading results (discrete classification outputs from level 0 to level 3), and specific disease risk predictions (such as the probability of myopia onset within the next two years and the prediction of myopia progression rate).

[0046] Specifically, the multi-task diagnostic unit constructed baseline curves and standard deviation ranges for normal optic nerve development at different age groups (3 to 15 years old, stratified by six-month intervals) based on large-scale multicenter data of healthy children. This baseline data includes multicenter data from at least 5000 healthy children, stratified by age and sex, and covers percentile curves (e.g., 5%, 25%, 50%, 75%, 95%) for key indicators such as CDR value, RNFL thickness, and vascular density, constructed using quantile regression or Gaussian process regression methods. For the input child data, the deviation (Z-score) of each optic nerve development indicator from the baseline mean for the corresponding age group is calculated. For example, if a 10-year-old boy has a CDR value of 0.55, while the baseline mean for that age group is 0.40 ± 0.08, then the deviation Z = (0.55 - 0.40) / 0.08 = 1.875. Taking into account the deviation of multiple indicators, genetic background (family history), eye habits, and other factors, a comprehensive risk assessment is performed using Bayesian networks or decision trees, automatically classifying the risk level. The grading criteria are as follows: Grade 0 indicates normal development (all indicators are within the normal baseline range, and the absolute Z-score is less than 1), with routine follow-up recommended (re-examination every 12 months); Grade 1 indicates mild abnormality or risk (some indicators slightly deviate from baseline, 1 ≤ absolute Z-score less than 2), with regular follow-up observation recommended (re-examination every 6 months); Grade 2 indicates moderate abnormality or risk (multiple indicators deviate from baseline, 2 ≤ absolute Z-score less than 3), with a shortened follow-up period recommended (re-examination every 3 months) and initial intervention; Grade 3 indicates severe abnormality or risk (indicators significantly deviate from baseline, and the absolute Z-score is ≥ 3), with immediate referral to an ophthalmologist recommended. The grading criteria may be dynamically adjusted based on the latest clinical guidelines and expert consensus.

[0047] The multi-task diagnostic unit also integrates Explainable Artificial Intelligence (XAI) technology. It uses Grad-weighted Class Activation Mapping (Grad-CAM), Local Interpretable Model-agnostic Explanations (LIME), or Shapley Additive Explanations (SHAP) to generate heatmaps overlaid on the original fundus image. This highlights the spatial regions that the model prioritizes when making grading decisions, such as areas of thinning RNFL, abnormal optic disc margins, or areas of abnormally dense blood vessels. Simultaneously, SHAP is used to analyze the contribution of different features to the diagnostic results and generate a ranking of feature importance. It can also generate explanatory reports based on counterfactual interpretations, such as "how the diagnostic result would change if a certain indicator improved to the normal range." These interpretable outputs greatly enhance doctors' trust in the AI ​​diagnostic results.

[0048] It should be noted that the fusion diagnostic module supports elastic computing resource expansion, automatically scaling up GPU computing nodes during high-concurrency periods to ensure response speed. Model inference latency is controlled within 2 seconds, and the total end-to-end latency (from image acquisition to graded result output) does not exceed 10 seconds. The model version of the multi-task diagnostic unit can be continuously updated through the feedback optimization module to ensure continuous improvement in diagnostic performance.

[0049] Furthermore, the reporting and early warning module includes: The report generation unit is used to generate a structured diagnostic report based on the automatic grading results, wherein the structured diagnostic report includes personalized follow-up recommendations and intervention recommendations; The early warning push unit is used to push early warning information to parents and doctors through a push channel when the automatic grading result reaches the preset early warning conditions.

[0050] The purpose of the aforementioned report and early warning module is to transform AI analysis results into diagnostic reports and early warning information that are easy for doctors and parents to understand and are actionable, thereby achieving effective transformation from diagnostic conclusions to clinical actions and ensuring that the grading results can be promptly communicated to the relevant responsible persons.

[0051] In this embodiment, the report generation unit automatically generates a structured, illustrated diagnostic report based on the grading results output by the fusion diagnostic module. The report includes: a patient basic information area displaying information such as the child's name, age, gender, examination date, and examination institution; a raw fundus image area displaying the acquired raw fundus images, supporting zooming and annotation; an AI assessment results area displaying the optic nerve development grade (e.g., "Grade 2—Moderate Abnormality"), developmental continuity score (e.g., "68 points out of 100"), and confidence score; and a key feature quantitative indicator area displaying the CDR value and age-related percentile (e.g., "0.55, 85th percentile for age-related age"), average RNFL thickness, and thickness values ​​for each sector in tabular form. The report includes percentiles, vascular density, and fractal dimension, color-coded as normal (green) or abnormal (red or yellow); a heatmap visualization area displays an overlay of Grad-CAM heatmaps, visually showing the AI-focused areas; a historical data comparison trend chart, used to show the changes in various indicators over time and mark the normal range if the child has historical examination records; and a personalized follow-up and intervention suggestion area, used to automatically generate suggestions based on the grading results, such as "re-examination recommended in 3 months," "increase daily outdoor activities to more than 2 hours," and "reduce close-range eye use time." The report supports exporting to a portable document format (PDF) for easy printing, archiving, or sharing.

[0052] The early warning push unit triggers different levels of early warning pushes based on the classification results. For levels 0 and 1, the system pushes routine follow-up reminders via the parent's mobile application, such as a notification saying, "Your child has completed a fundus examination, and the optic nerve development is normal. A follow-up examination is recommended in 12 months." For level 2, the system sends warnings through both mobile application pushes and SMS, such as a warning message saying, "Your child has moderate optic nerve development abnormalities. A follow-up examination at a professional ophthalmology institution is recommended within 3 months. Click to view the detailed report," and also pushes it to the linked attending physician's workstation. For level 3, the system immediately issues an emergency warning through multiple channels, including mobile application pushes, SMS, and telephone voice notifications, while automatically generating referral suggestions and pushing them to the partner ophthalmologist's end. All warning messages include a report viewing link and medical guidance.

[0053] It should be noted that the frequency and channels of alert notifications can be personalized based on parental preferences and doctor's recommendations. The system supports sharing reports to family groups for collaborative family monitoring. The alert notification module also supports integration with hospital information systems, enabling automatic archiving of alert information and automatic creation of follow-up tasks.

[0054] Furthermore, the feedback optimization module includes: The data collection unit is used to acquire the clinical gold standard data corresponding to the automatic grading results. The clinical gold standard data includes expert review diagnosis results, OCT examination data, visual field examination data, intraocular pressure measurement data, and intervention effect data. The model fine-tuning unit is used to use the clinical gold standard data as training data to fine-tune the deep learning model online through federated learning or incremental learning techniques to obtain an updated deep learning model.

[0055] The purpose of the aforementioned feedback optimization module is to enable the AI ​​model to continuously learn and evolve from clinical practice, ensuring its long-term effectiveness and accuracy.

[0056] In this embodiment, the data collection unit continuously collects various clinical validation data. Expert-reviewed diagnostic results refer to the final clinical diagnosis and grading given by a senior ophthalmologist (with over 10 years of clinical experience and a professional title of associate chief physician or above) after manually reviewing the AI ​​diagnostic report. OCT examination data includes objective quantitative parameters such as RNFL thickness maps and ganglion cell complex thickness. Visual field examination data includes visual field defect indicators such as mean deviation (MD) and pattern standard deviation (PSD). Intraocular pressure measurement data includes intraocular pressure values ​​measured by a non-contact tonometer. Intervention effect data includes changes in visual acuity, refractive error, and axial length growth recorded during subsequent follow-up. All data undergoes standardization and de-identification processing (removing direct identifiers such as name and ID number, retaining only age, gender, and other research-required information).

[0057] The model fine-tuning unit employs federated learning or incremental learning techniques to fine-tune the deep learning model online. Taking federated learning as an example, each medical institution stores a large amount of accumulated pediatric fundus images and diagnostic data on its local server. The model fine-tuning unit uses this data locally to train the deep learning model, calculates the gradient updates of the model weights, and then uploads only the encrypted gradient updates to the central federated learning server in the cloud. The original images and clinical data remain stored locally within the hospital and do not leave the hospital's network. The central server collects encrypted gradients from multiple medical institutions and uses a federated averaging (FedAvg) algorithm to perform a weighted average of all gradients (the weight factor is proportional to the number of local training samples at each institution) to generate new global model weights. Then, the central server distributes the aggregated global model parameters to each local device, and each local device updates its local model with the global parameters.

[0058] In terms of incremental learning, when the locally accumulated gold standard data reaches a preset threshold (such as weekly or every 500 newly labeled data examples), the model fine-tuning unit triggers the incremental training process. Incremental learning employs Elastic Weight Consolidation (EWC) or learning rate backoff strategies to avoid catastrophic forgetting (i.e., overwriting existing knowledge when learning new tasks, leading to a decline in the performance of old tasks) while absorbing new knowledge.

[0059] The model fine-tuning unit also employs an active learning strategy, continuously monitoring the confidence score output by the model. When the model's confidence score for a case's grading diagnosis falls below a preset threshold (e.g., the Softmax output probability is less than 0.75), the system automatically marks the case as a "high-value sample to be labeled" and pushes it to the expert review queue, prioritizing manual labeling by experts. The labeled data is then added to the training set for subsequent model optimization. This active learning strategy efficiently utilizes limited expert resources, specifically improving the model's performance on difficult cases.

[0060] The model fine-tuning unit also establishes a positive and negative feedback mechanism. When the expert review diagnosis result is consistent with the AI ​​diagnosis result, the sample is used as a positive sample for reinforcement learning, increasing its weight in training to strengthen the model's correct behavior. When the expert review diagnosis result is inconsistent with the AI ​​diagnosis result, the sample is used as a negative sample to trigger in-depth analysis, including checking whether the feature distribution of the sample is within the training set coverage, analyzing whether the model's region of interest (compared via Grad-CAM heatmap) is consistent with the expert's region of interest, and analyzing which specific features led to the misjudgment. The analysis results are used to guide model correction strategies, such as adjusting the network structure, increasing the training weights of relevant features, or supplementing with corresponding types of training data.

[0061] The model fine-tuning unit also performs version control for each model update, recording metadata such as version number, performance metrics, training data volume, and optimization strategies. A / B testing is performed before deploying a new version to ensure that its performance is no less than that of the older version. Through this multi-layered feedback optimization mechanism, the AI ​​model of this invention can continuously evolve, and its diagnostic accuracy continuously improves as clinical applications deepen.

[0062] It should be noted that the aggregation frequency of federated learning and the trigger threshold for incremental learning can be dynamically adjusted according to the data accumulation speed of each medical institution. The system supports A / B testing before and after model updates to ensure that the performance of the new model version is not lower than that of the old version. The model fine-tuning unit also supports continuous performance monitoring, monitoring the model's accuracy, recall, F1 score, AUC, false positive rate, and false negative rate in real time, and monitoring changes in the distribution of input data (data drift). When performance degradation or data drift is detected, an alert is automatically triggered and the model is recommended to be retrained or expert intervention is suggested.

[0063] Furthermore, such as Figure 2 As shown, based on the above-mentioned AI-based feedback closed-loop system for pediatric fundus image recognition and automatic grading of optic nerve development, this invention also provides a corresponding AI-based feedback closed-loop method for pediatric fundus image recognition and automatic grading of optic nerve development, which includes the following steps: S10, The image acquisition module acquires fundus images of the child; S20. The image preprocessing module preprocesses the fundus image to obtain the target fundus image; S30. The feature extraction module segments the target fundus image to obtain a segmentation result, calculates fundus structure quantization parameters based on the segmentation result, and converts the fundus structure quantization parameters into an image feature vector. S40. The fusion diagnosis module acquires children's clinical data, fuses the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, inputs the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. S50. The report warning module generates a structured diagnostic report based on the automatic classification results and pushes the warning information through the push channel. S60. The feedback optimization module obtains the clinical gold standard data corresponding to the automatic grading result, uses the clinical gold standard data as training data, and fine-tunes the deep learning model online to obtain the updated deep learning model.

[0064] Understandably, the AI-based feedback closed-loop method for pediatric fundus image recognition and automatic grading of optic nerve development provided by this invention acquires pediatric fundus images through an image acquisition module and performs quality assessment; an image preprocessing module performs denoising, enhancement, and color correction on the fundus images; a feature extraction module accurately extracts quantitative parameters of key structures such as the optic disc, optic cup, retinal nerve fiber layer, and blood vessels through a deep learning segmentation network and converts them into image feature vectors; a fusion diagnosis module performs multimodal fusion of image features and pediatric clinical data and simultaneously outputs a continuous score for optic nerve development, automatic grading results, and disease risk prediction through a multi-task deep learning model; a report and early warning module automatically generates a structured diagnostic report based on the grading results and intelligently pushes early warnings; and a feedback optimization module continuously collects clinical gold standard data and uses federated learning or incremental learning techniques to fine-tune the model online. This invention upgrades the traditional static and subjective fundus screening to a dynamic, objective, and intelligent optic nerve development management system. It addresses the lack of developmental dimensions by constructing age-stratified baseline developmental curves, overcomes the subjectivity of manual assessment through AI-based automatic grading and diagnosis, achieves effective connection between diagnosis and intervention through a biofeedback closed loop, optimizes medical resource allocation through a cloud-edge collaborative architecture, and ensures data security and privacy through federated learning technology. It significantly improves the accuracy, objectivity, and accessibility of optic nerve development assessment in children, providing technical support for early screening, timely intervention, and long-term management of pediatric eye diseases.

[0065] Furthermore, such as Figure 3 As shown, based on the above-mentioned method and system for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop, the present invention also provides a terminal, which includes a processor 10, a memory 20 and a display 30. Figure 3 Only some of the terminal components are shown; however, it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0066] In some embodiments, the memory 20 may be an internal storage unit of the terminal, such as a hard drive or memory. In other embodiments, the memory 20 may be an external storage device of the terminal, such as a plug-in hard drive, smart media card (SMC), secure digital card (SD), flash card, etc. Further, the memory 20 may include both internal and external storage devices. The memory 20 is used to store application software and various types of data installed on the terminal, such as program code installed on the terminal. The memory 20 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 20 stores a child fundus photo recognition and optic nerve development automatic grading program 40 based on AI feedback closed loop. This child fundus photo recognition and optic nerve development automatic grading program 40 based on AI feedback closed loop can be executed by the processor 10, thereby realizing the child fundus photo recognition and optic nerve development automatic grading method based on AI feedback closed loop in this application.

[0067] In some embodiments, the processor 10 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in the memory 20 or process data, such as executing the AI-based feedback closed-loop method for recognizing children's fundus photos and automatically grading optic nerve development.

[0068] In some embodiments, the display 30 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. The display 30 is used to display information on the terminal and to display a visual user interface. The components of the terminal communicate with each other via a system bus.

[0069] The present invention also provides a computer-readable storage medium, wherein the computer-readable storage medium stores a program for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop, wherein when the program for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop is executed by a processor, the program implements the steps of the method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described above.

[0070] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or terminal that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or terminal. Unless otherwise specified, an element defined by the phrase "comprising one" does not exclude the presence of other identical elements in the process, method, article, or terminal that includes that element.

[0071] Of course, those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided by this invention can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0072] It should be understood that the application of the present invention is not limited to the examples above. Those skilled in the art can make improvements or modifications based on the above description, and all such improvements and modifications should fall within the protection scope of the appended claims.

Claims

1. A child fundus photograph recognition and optic nerve development automatic grading system based on AI feedback closed loop, characterized in that, The AI-based feedback closed-loop system for children's fundus image recognition and automatic grading of optic nerve development includes: an image acquisition module, an image preprocessing module, a feature extraction module, a fusion diagnosis module, a report and early warning module, and a feedback optimization module. The image acquisition module, the image preprocessing module, the feature extraction module, the fusion diagnosis module, and the report and early warning module are connected in sequence, and the feedback optimization module is connected to the fusion diagnosis module and the report and early warning module respectively. The image acquisition module is used to acquire fundus images of children; The image preprocessing module is used to preprocess the fundus image to obtain the target fundus image; The feature extraction module is used to segment the target fundus image to obtain a segmentation result, calculate fundus structure quantization parameters based on the segmentation result, and convert the fundus structure quantization parameters into an image feature vector. The fusion diagnostic module is used to acquire children's clinical data, fuse the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, input the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. The report and early warning module is used to generate a structured diagnostic report based on the automatic classification results and push early warning information through a push channel; The feedback optimization module is used to obtain the clinical gold standard data corresponding to the automatic grading results, use the clinical gold standard data as training data, and fine-tune the deep learning model online to obtain the updated deep learning model.

2. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The image acquisition module includes: An image acquisition unit is used to acquire fundus images of children, perform quality assessment on the fundus images, and obtain quality assessment results; The re-acquisition control unit is used to re-acquire fundus images of the child when the quality assessment results are not up to standard, until fundus images that pass the quality assessment are acquired.

3. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The image preprocessing module includes: An image denoising unit is used to denoise the fundus image to obtain a denoised fundus image. An enhancement and correction unit is used to perform contrast enhancement and color correction processing on the denoised fundus image to obtain the target fundus image.

4. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The feature extraction module includes: An image segmentation unit is used to input the target fundus image into a deep learning segmentation network, the deep learning segmentation network segments the target fundus image and outputs segmentation results, wherein the segmentation results include optic disc segmentation results, optic cup segmentation results, retinal nerve fiber layer segmentation results and retinal vessel segmentation results; The feature encoding unit is used to calculate fundus structure quantization parameters based on the segmentation results and convert the fundus structure quantization parameters into image feature vectors. The fundus structure quantization parameters include cup-to-disc ratio parameters, retinal nerve fiber layer thickness parameters, and blood vessel morphology parameters.

5. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The fusion diagnostic module includes: A multimodal fusion unit is used to acquire pediatric clinical data and use a graph neural network or Transformer model to fuse the image feature vector with the pediatric clinical data to generate fused features. A multi-task diagnostic unit is used to input the fused features into a multi-task learning deep learning model, which outputs a visual nerve development continuity score and an automatic grading result.

6. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The reporting and early warning module includes: The report generation unit is used to generate a structured diagnostic report based on the automatic grading results, wherein the structured diagnostic report includes personalized follow-up recommendations and intervention recommendations; The early warning push unit is used to push early warning information to parents and doctors through a push channel when the automatic grading result reaches the preset early warning conditions.

7. The automatic grading system for pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 1, characterized in that, The feedback optimization module includes: The data collection unit is used to acquire the clinical gold standard data corresponding to the automatic grading results. The clinical gold standard data includes expert review diagnosis results, OCT examination data, visual field examination data, intraocular pressure measurement data, and intervention effect data. The model fine-tuning unit is used to use the clinical gold standard data as training data to fine-tune the deep learning model online through federated learning or incremental learning techniques to obtain an updated deep learning model.

8. A method for automatic grading of pediatric fundus image recognition and optic nerve development based on an AI feedback closed-loop pediatric fundus image recognition and optic nerve development system according to any one of claims 1-7, characterized in that, The method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop includes the following steps: The image acquisition module acquires fundus images of the child; The image preprocessing module preprocesses the fundus image to obtain the target fundus image; The feature extraction module segments the target fundus image to obtain a segmentation result, calculates fundus structure quantization parameters based on the segmentation result, and converts the fundus structure quantization parameters into an image feature vector. The fusion diagnostic module acquires children's clinical data, fuses the image feature vector with the children's clinical data in a multimodal manner to generate fusion features, inputs the fusion features into a deep learning model, and the deep learning model outputs a continuous score for optic nerve development and an automatic grading result. The report and early warning module generates a structured diagnostic report based on the automatic classification results and pushes early warning information through a push channel; The feedback optimization module obtains the clinical gold standard data corresponding to the automatic grading result, uses the clinical gold standard data as training data, and fine-tunes the deep learning model online to obtain an updated deep learning model.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and an AI-based feedback closed-loop automatic grading program for pediatric fundus image recognition and optic nerve development, stored in the memory and executable on the processor. When the processor executes the AI-based feedback closed-loop automatic grading program for pediatric fundus image recognition and optic nerve development, it implements the steps of the AI-based feedback closed-loop automatic grading method for pediatric fundus image recognition and optic nerve development as described in claim 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop. When the program is executed by a processor, it implements the steps of the method for automatic grading of pediatric fundus image recognition and optic nerve development based on AI feedback closed loop as described in claim 8.

Images