Ophthalmology cornea health assessment system and method based on machine learning and multispectral imaging

Abstract

The invention relates to the technical field of ophthalmology informatization, in particular to an ophthalmology cornea health assessment system and method based on machine learning and multispectral imaging, and the system comprises a multispectral imaging module and a data processing module which is in communication connection with the multispectral imaging module; executing image preprocessing, feature extraction and image fusion based on the cornea multiband image data; the deep learning module is in communication connection with the data processing module and receives the fused multispectral image data sent by the data processing module; analyzing the fused multispectral image data by using a pre-trained convolutional neural network model to generate a cornea health assessment result; the user interface module is in communication connection with the deep learning module to receive the cornea health assessment result sent by the deep learning module; the cornea health assessment result and related suggestions are displayed to the user, processing parameters can be automatically adjusted according to different patients and different lesion types, and the accuracy and robustness of feature extraction are greatly improved.

Description

Technical Field

[0001] The present invention relates to the technical field of ophthalmic informatization, in particular to an ophthalmic corneal health assessment system and method based on machine learning and multispectral imaging. Background Art

[0002] The field of ophthalmic diagnosis has long faced challenges in the accuracy and efficiency of corneal health assessment. Traditional corneal examination methods, such as slit lamp microscopy, although widely used, have limitations in early lesion detection and objective quantitative assessment. With the development of medical imaging technology, advanced devices such as confocal microscopes and optical coherence tomography have been introduced into clinical practice, greatly improving the fineness of corneal structure observation. However, these technologies often require complex operations and professional interpretations, making it difficult to achieve rapid and widespread screening applications.

[0003] In recent years, multispectral imaging technology has shown great potential in the field of biomedicine. By collecting tissue images at different wavelengths, this technology can provide rich spectral information, which helps to reveal tissue characteristics that are difficult to observe with the naked eye. Some researchers have tried to apply multispectral imaging to ophthalmic examinations, but most are limited to the observation of specific parts such as the retina or iris, and there is still insufficient for the comprehensive assessment of the cornea. In addition, existing multispectral imaging systems are often bulky and complex to operate, making it difficult to promote their use in primary medical institutions.

[0004] At the same time, artificial intelligence technology, especially deep learning algorithms, has made breakthrough progress in the field of medical image analysis. Some research teams have begun to explore the application of machine learning to the automatic analysis of corneal images, but most of the work still stays at the identification of single lesion types or the analysis based on a single imaging modality, making it difficult to achieve a comprehensive and accurate assessment of corneal health status.

[0005] The main problems of the existing technical solutions are as follows: First, there is a lack of an imaging method that can quickly and non-invasively collect multi-dimensional optical characteristics of the cornea; Second, existing image analysis algorithms are difficult to effectively integrate and utilize multispectral information, and cannot fully utilize the advantages of multispectral imaging; Third, the training data of machine learning models often come from a single source, making it difficult to cope with complex and changeable clinical situations; Fourth, most systems lack interpretability and continuous learning ability, making it difficult to gain the trust of doctors and maintain high performance in the long term.

[0006] In view of the above problems, there is an urgent need for a corneal health assessment system that can comprehensively utilize multispectral imaging and advanced machine learning technology to achieve a rapid, accurate, and comprehensive assessment of corneal health status. Summary of the Invention

[0007] The present invention aims to solve the deficiencies of existing corneal health assessment methods in terms of accuracy, efficiency, and comprehensiveness, and provides an ophthalmic corneal health assessment system and method based on machine learning and multispectral imaging. By innovatively combining multispectral imaging technology and deep learning algorithms, the system achieves a comprehensive, accurate, and efficient assessment of corneal health status.

[0008] The present invention proposes an ophthalmic corneal health assessment system based on machine learning and multispectral imaging, including:

[0009] A multispectral imaging module, used for:

[0010] Collecting images of the cornea at different wavelengths;

[0011] Obtaining the RGB image and trichromatic light images of the cornea;

[0012] A data processing module, communicatively connected to the multispectral imaging module, used for:

[0013] Receiving the corneal multi-band image data sent by the multispectral imaging module;

[0014] Based on the corneal multi-band image data, performing image preprocessing, feature extraction, and image fusion;

[0015] A deep learning module, communicatively connected to the data processing module, used for:

[0016] Receiving the fused multispectral image data sent by the data processing module;

[0017] Using a pre-trained convolutional neural network model to analyze the fused multispectral image data and generate a corneal health assessment result;

[0018] A user interface module, communicatively connected to the deep learning module, used for:

[0019] Receiving the corneal health assessment result sent by the deep learning module;

[0020] Showing the corneal health assessment result and related suggestions to the user.

[0021] Preferably, the multispectral imaging module includes:

[0022] A light source unit, used for emitting light of different wavelengths to irradiate the cornea;

[0023] An image acquisition unit, used for acquiring corneal images under the irradiation of light of different wavelengths;

[0024] A wavelength selection unit, communicatively connected to the light source unit, used for controlling the light source unit to emit light of different wavelengths in a preset order.

[0025] Preferably, the data processing module includes:

[0026] A preprocessing unit for denoising, contrast enhancement, and image segmentation of the corneal multi-band image data;

[0027] A feature extraction unit, communicatively connected to the preprocessing unit, for extracting feature parameters such as specular reflectance, optical density, and light transmittance of the cornea from the preprocessed image;

[0028] An image fusion unit, communicatively connected to the feature extraction unit, for performing multi-band superposition fusion on images of different bands and their extracted features.

[0029] Preferably, the deep learning module includes:

[0030] A model training unit for training a convolutional neural network model using a labeled corneal health sample data set;

[0031] A model application unit, communicatively connected to the model training unit, for analyzing the input fused multi-spectral image using the trained convolutional neural network model;

[0032] A result generation unit, communicatively connected to the model application unit, for generating a corneal health assessment grade and a detailed parameter analysis result based on the output of the convolutional neural network.

[0033] Preferably, it further includes:

[0034] A database module, communicatively connected to the deep learning module and the user interface module, for:

[0035] Storing the user's corneal image data, personal information, and historical assessment results;

[0036] Providing a training and validation data set for the deep learning module;

[0037] A real-time monitoring module, communicatively connected to the multi-spectral imaging module and the deep learning module, for:

[0038] Regularly collecting the user's corneal multi-spectral images;

[0039] Analyzing the dynamic changes in the corneal health status through the deep learning module.

[0040] Preferably, the user interface module includes:

[0041] A data input unit for receiving the user's personal information and relevant medical history;

[0042] A result display unit for graphically displaying the corneal health assessment results and historical change trends;

[0043] A suggestion generation unit, communicatively connected to the result display unit, for generating personalized health suggestions according to the evaluation results.

[0044] Preferably, it further includes:

[0045] A cloud computing module, communicatively connected to the deep learning module, for:

[0046] Storing the trained convolutional neural network model parameters;

[0047] Performing computationally intensive deep learning tasks;

[0048] Regularly updating and optimizing the deep learning model.

[0049] Preferably, the data processing module further includes:

[0050] A multi-modal data fusion unit, for fusing corneal image data with information such as the user's age, gender, medical history, etc. to generate a comprehensive feature vector;

[0051] A data augmentation unit, communicatively connected to the multi-modal data fusion unit, for augmenting the training data set by means of image rotation, scaling, flipping, etc.

[0052] Preferably, the deep learning module further includes:

[0053] A model interpretation unit, for generating a visual interpretation of the decision-making process of the convolutional neural network to help doctors understand the basis of the evaluation results;

[0054] A model optimization unit, communicatively connected to the model interpretation unit, for continuously optimizing the performance of the convolutional neural network model based on the newly added labeled data and the model interpretation results.

[0055] An ophthalmic corneal health assessment method based on machine learning and multi-spectral imaging, applied to the described system, includes the following steps:

[0056] S1. Collect images of the cornea in different bands through a multi-spectral imaging device to obtain the RGB image and the three-primary-color light images of the cornea;

[0057] S2. Preprocess the collected multi-band corneal images, including noise reduction, contrast enhancement, and image segmentation;

[0058] S3. Extract feature parameters such as the specular reflectance, optical density, and light transmittance of the cornea from the preprocessed images;

[0059] S4. Perform multi-band superposition fusion on the images in different bands and their extracted features to generate a fused multi-spectral image;

[0060] S5. Input the fused multi - spectral image into the pre - trained convolutional neural network model to obtain the corneal health assessment result;

[0061] S6. Generate a corneal health assessment grade and a detailed parameter analysis report based on the assessment result;

[0062] S7. Display the assessment result and the analysis report to the user through the user interface and provide personalized health advice;

[0063] S8. Regularly collect the corneal multi - spectral images of the user and analyze the dynamic changes of the corneal health status through the deep learning model;

[0064] S9. Use the newly collected data to continuously optimize the convolutional neural network model and improve the accuracy and adaptability of the assessment.

[0065] The beneficial effects of the present invention are mainly reflected in the following aspects:

[0066] First of all, at the data acquisition level, the multi - spectral imaging module of this system can quickly and non - invasively obtain the images of the cornea in different bands, providing rich spectral information for subsequent analysis. This not only greatly increases the information content of the imaging, but also enables the system to capture subtle lesions that are difficult to observe with the naked eye.

[0067] Secondly, in terms of data processing, the improved bilateral filtering algorithm and the adaptive wavelet transform feature extraction method developed in the present invention can effectively extract the key features of the cornea from the multi - spectral image. The adaptability of these algorithms enables the system to automatically adjust the processing parameters according to different patients and different lesion types, greatly improving the accuracy and robustness of feature extraction.

[0068] In the core health assessment link, the deep learning module adopted in the present invention can not only efficiently process the fused multi - spectral data, but also, through innovative model architectures and training strategies, achieve accurate identification of various corneal lesions. In particular, the attention mechanism and multi - scale feature fusion technology introduced in this system enable the model to better capture the local and global features of the cornea, improving the detection ability for early and mild lesions.

[0069] At the practical application level, the system of the present invention has significant clinical value. It can not only provide accurate corneal health assessment results, but also display the assessment process in a visual way, enhancing the interpretability and credibility of the system. This is of great significance for assisting doctors in diagnosis and patients in understanding their own conditions. In addition, the real - time monitoring function of the system makes it possible to long - term track the corneal health status of patients, providing strong support for the formulation and adjustment of personalized treatment plans.

[0070] From a technological innovation perspective, the present invention demonstrates unique advantages in multiple aspects. For example, the introduction of the multi-modal data fusion unit enables the system to comprehensively consider image data and patient personal information, providing more comprehensive and personalized assessment results. The anomaly detection algorithm based on variational autoencoders offers an innovative solution for the early warning of corneal health status.

[0071] Finally, the system of the present invention has good scalability and adaptability. Through the cloud computing module and continuous optimization mechanism, the system can continuously learn new data, improve model performance, and adapt to changing clinical needs. This dynamic optimization ability ensures that the system can maintain a high level of performance during long-term use and can handle newly emerging types of corneal lesions.

[0072] In summary, the ophthalmic corneal health assessment system and method based on machine learning and multi-spectral imaging provided by the present invention, through the innovative integration of multiple advanced technologies, not only solve the problems existing in the existing corneal health assessment methods, but also achieve significant breakthroughs in accuracy, efficiency, comprehensiveness, and scalability. This system is expected to play an important role in clinical practice, bringing a revolutionary change to the field of ophthalmic diagnosis and health management. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] Figure 1 is the logical block diagram of the overall system of the present invention.

[0074] Figure 2 is the logical block diagram of the multi-spectral imaging module of the present invention.

[0075] Figure 3 is the logical block diagram of the data processing module of the present invention.

[0076] Figure 4 is the logical block diagram of the deep learning module of the present invention.

[0077] Figure 5 is the logical block diagram of the user interface module of the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0078] Referring to Figures 1-5 , the present invention provides an ophthalmic corneal health assessment system and method based on machine learning and multi-spectral imaging. The system realizes the accurate assessment of ophthalmic corneal health status through the innovative combination of multi-spectral imaging technology and deep learning algorithms. The following will describe the technical solutions of the present invention in detail.

[0079] The system of the present invention includes a multi-spectral imaging module 1, a data processing module 2, a deep learning module 3, and a user interface module 4. These modules work together to form a complete corneal health assessment process.

[0080] The multispectral imaging module 1 is used to collect images of the cornea at different wavelengths and obtain the RGB image and trichromatic light images of the cornea. Preferably, this module adopts high-precision spectral imaging technology and can collect multi-band images of the cornea at intervals of 10 nm within the wavelength range of 400 - 800 nm. This high-resolution multi-band acquisition can capture the subtle changes of the cornea under different spectra, providing a rich data basis for subsequent health assessment.

[0081] The data processing module 2 is communicatively connected to the multispectral imaging module 1 and is used to receive the multi-band image data of the cornea sent by the multispectral imaging module 1 and perform image preprocessing, feature extraction, and image fusion based on these data. In an embodiment of the present invention, the data processing module 2 adopts an improved bilateral filtering algorithm for image denoising. The mathematical expression of this algorithm is as follows:

[0082]

[0083] where, I filtered (x) is the filtered image, I(x) is the original image, Ω is the window centered on pixel x, f r and g s are the range kernel and the spatial kernel respectively, and W p is the normalization factor. By adjusting the parameters of f r and g s , the present invention realizes the effective removal of different types of noise while retaining the edge information of the cornea image. W p is the normalization factor, defined as:

[0084]

[0085] The present invention improves the traditional bilateral filtering and introduces an adaptive kernel function selection mechanism.

[0086] Specifically, the forms of f r and g s are dynamically adjusted according to the local characteristics of the image:

[0087]

[0088] where, σ r (x) and σ s (x) are parameters adaptively calculated according to the local image characteristics:

[0089] σ r (x) = k r ·std(I(x)),

[0090] σ s (x) = ks ·edge(x),

[0091] Here, std(I(x)) is the standard deviation of the pixel values within a local window centered at x, edge(x) is the edge intensity calculated using the Sobel operator, and k r and k s are adjustable coefficients.

[0092] The deep learning module 3 is communicatively connected to the data processing module 2, and is configured to receive the fused multi-spectral image data sent by the data processing module 2, and analyze these data using a pre-trained convolutional neural network model to generate a corneal health assessment result. The present invention adopts an improved ResNet structure as the basic architecture of the convolutional neural network, and improves the model's ability to identify minor corneal lesions by introducing an attention mechanism and multi-scale feature fusion.

[0093] The user interface module 4 is communicatively connected to the deep learning module 3, and is configured to receive the corneal health assessment result sent by the deep learning module 3, and display the assessment result and related suggestions to the user. This module adopts an intuitive graphical interface, enabling complex assessment results to be presented to the user in an easy-to-understand manner.

[0094] Furthermore, the multi-spectral imaging module 1 includes a light source unit 11, an image acquisition unit 12, and a wavelength selection unit 13. The light source unit 11 is configured to emit light of different wavelengths to irradiate the cornea. Preferably, the light source unit 11 adopts LED array technology, which can precisely control the wavelength and intensity of the emitted light. The image acquisition unit 12 is configured to acquire corneal images under illumination with different wavelengths of light. This unit adopts a high-sensitivity CMOS sensor, which can capture corneal images under low-light conditions. The wavelength selection unit 13 is communicatively connected to the light source unit 11, and is configured to control the light source unit 11 to emit light of different wavelengths in a preset order. Through this design, the present invention realizes precise imaging of the cornea at multiple specific wavelengths.

[0095] The data processing module 2 includes a preprocessing unit 21, a feature extraction unit 22, and an image fusion unit 23. The preprocessing unit 21 is configured to perform noise reduction, contrast enhancement, and image segmentation on the corneal multi-band image data. In addition to the aforementioned bilateral filtering algorithm, the preprocessing unit 21 also adopts adaptive histogram equalization technology to enhance the image contrast. The feature extraction unit 22 is communicatively connected to the preprocessing unit 21, and is configured to extract feature parameters such as specular reflectivity, optical density, and light transmittance of the cornea from the preprocessed image. The present invention has developed a multi-scale feature extraction algorithm based on wavelet transform, and its mathematical expression is as follows:

[0096]

[0097] Among them, F(a,b) is the wavelet transform coefficient, f(t) is the input image, ψ(t) is the wavelet basis function, and a and b are the scale and translation parameters respectively. By selecting appropriate wavelet basis functions and scale parameters, the present invention can effectively extract multi-scale features of the cornea.

[0098] The present invention uses an improved wavelet basis function called Corneal Feature Adaptive Wavelet (CFAW), and its form

[0099]

[0100] Among them, ω is the fundamental wave frequency, σ controls the width of the Gaussian envelope, α is the edge adaptation coefficient, and edge(t) is the edge intensity of the image at position t.

[0101] The image fusion unit 23 is communicatively connected to the feature extraction unit 22 and is used to perform multi-band superposition fusion on images of different bands and the features extracted therefrom. The present invention adopts an improved weighted average fusion algorithm, and its fusion weights are determined by the saliency and information entropy of the features, thereby ensuring the high quality and rich information of the fused image.

[0102] Through the collaborative work of the above modules, the system of the present invention can comprehensively and accurately evaluate the health status of the cornea, providing a powerful tool for ophthalmic diagnosis. The system can not only detect common corneal diseases such as keratitis and corneal ulcer, but also identify early corneal lesions, providing the possibility for timely treatment. In addition, the system of the present invention has good scalability and adaptability, and can improve its evaluation accuracy by continuously learning new data, providing technical support for personalized ophthalmic diagnosis and treatment.

[0103] In a preferred embodiment of the present invention, as described in claim 4, the deep learning module 3 includes a model training unit 31, a model application unit 32, and a result generation unit 33. This structural design enables the system of the present invention to continuously learn and improve, thereby improving the accuracy and reliability of corneal health assessment.

[0104] The model training unit 31 is used to train a convolutional neural network model using a labeled corneal health sample data set. The present invention adopts an innovative transfer learning strategy. First, the model is pre-trained on a large-scale natural image data set, and then fine-tuned on a specific corneal image data set. This method can effectively solve the problem that medical image data sets are usually small in scale and improve the generalization ability of the model. Preferably, the present invention uses an improved ResNet-101 architecture as the basic network and introduces an attention mechanism to enhance the model's perception ability of key features. The loss function of the model adopts focal loss, and its expression is as follows:

[0105] FL(p t ) = -α t (1 - p t ) γ log(p t ),

[0106] where p t is the probability predicted by the model, α t is the class weight factor, and γ is the modulation factor. By adjusting α t and γ, the system of the present invention can better handle the class imbalance problem and improve the recognition ability for rare corneal diseases.

[0107] The model application unit 32 is communicatively connected to the model training unit 31 and is used to analyze the input fused multi - spectral image by applying the trained convolutional neural network model. In actual application, the model application unit 32 adopts a sliding window technique to analyze the entire corneal image block by block, so as to accurately locate the possible lesion areas.

[0108] The result generation unit 33 is communicatively connected to the model application unit 32 and is used to generate the corneal health assessment level and the detailed parameter analysis result based on the output of the convolutional neural network. The present invention has developed a multi - dimensional scoring system, which not only gives the overall health level, but also provides a number of specific indicators including transparency, curvature, thickness, etc. Such a comprehensive assessment result provides strong support for ophthalmologists to formulate treatment plans.

[0109] The system of the present invention further includes a database module 5 and a real - time monitoring module 6. The database module 5 is communicatively connected to the deep learning module 3 and the user interface module 4 and is used to store the corneal image data, personal information and historical assessment results of users, and at the same time provide training and validation data sets for the deep learning module 3. Preferably, the database module 5 adopts a distributed storage architecture to ensure the efficient management and fast access of large - scale data.

[0110] The real - time monitoring module 6 is communicatively connected to the multi - spectral imaging module 1 and the deep learning module 3 and is used to regularly collect the corneal multi - spectral images of users and analyze the dynamic changes of the corneal health status through the deep learning module 3. A feature of the present invention is that the real - time monitoring module 6 adopts an adaptive sampling strategy. For users with stable health status, the system will reduce the sampling frequency to reduce unnecessary examinations; while for users detected with potential risks, the system will increase the sampling frequency to achieve closer monitoring. Such an intelligent monitoring method not only improves the efficiency of the system, but also ensures timely attention to high - risk users.

[0111] In another embodiment of the present invention, as described in claim 6, the user interface module 4 includes a data input unit 41, a result display unit 42, and a suggestion generation unit 43. The data input unit 41 is used to receive the user's personal information and relevant medical history. To improve the efficiency and accuracy of information collection, the data input unit 41 adopts an intelligent question-and-answer system that can automatically generate relevant follow-up questions based on the user's initial input.

[0112] The result display unit 42 is used to display the corneal health assessment results and historical change trends in a graphical manner. The present invention has developed an intuitive visualization interface that uses heat map technology to display the health status of each region of the cornea and shows the historical changes of key indicators through line charts. This design makes complex medical data easy to understand and helps patients better grasp their own health conditions.

[0113] The suggestion generation unit 43 is communicatively connected to the result display unit 42 and is used to generate personalized health suggestions based on the assessment results. The system of the present invention integrates an expert knowledge base and can provide targeted lifestyle and treatment suggestions according to the assessment results and the user's personal situation. For example, for users detected with mild dry eye symptoms, the system may suggest increasing indoor humidity, reducing screen usage time, and recommending suitable artificial tear products.

[0114] The system of the present invention further includes a cloud computing module 7, which is communicatively connected to the deep learning module 3. The cloud computing module 7 is used to store the trained convolutional neural network model parameters, execute computationally intensive deep learning tasks, and regularly update and optimize the deep learning model. By introducing cloud computing technology, the system of the present invention realizes flexible scheduling of computing resources and continuous optimization of the model. In particular, the cloud computing module 7 adopts federated learning technology, which can improve the model performance by using data from multiple medical institutions while protecting user privacy. This distributed learning method not only improves the generalization ability of the model but also solves the legal and ethical issues of medical data sharing.

[0115] In summary, the ophthalmic corneal health assessment system based on machine learning and multispectral imaging of the present invention realizes the full-process automation from data collection, processing, analysis to result display through the collaborative work of multiple functional modules. The design of the system fully considers the needs of medical practice, can provide high-precision health assessments, and ensure a good user experience and data security. This innovative technical solution brings new possibilities to the field of ophthalmic diagnosis and health management and is expected to play an important role in preventive medicine and personalized medicine.

[0116] In another preferred embodiment of the present invention, the data processing module 2 further includes a multi-modal data fusion unit 24 and a data augmentation unit 25. The introduction of these two units further improves the data processing ability and model training effect of the system of the present invention.

[0117] The multi-modal data fusion unit 24 is used to fuse corneal image data with information such as the user's age, gender, medical history, etc., to generate a comprehensive feature vector. The present invention adopts an innovative attention mechanism to achieve effective fusion of multi-modal data. Specifically, the system first performs independent feature extraction on the data of each modality, then calculates the correlation between the features of different modalities through the self-attention mechanism, and finally fuses the weighted features. This method can adaptively adjust the importance of different modality data, thereby improving the expression ability of the fused features. The mathematical expression of the fusion process is as follows:

[0118]

[0119] where F is the final fused feature, F i is the feature of the i-th modality, α i is the corresponding attention weight, and N is the number of modalities. The attention weight α i is calculated through the softmax function:

[0120]

[0121] where W i is a learnable parameter matrix.

[0122] The data augmentation unit 25 is communicatively connected to the multi-modal data fusion unit 24 and is used to expand the training data set by means of image rotation, scaling, flipping, etc. Considering the particularity of medical images, the present invention has developed a specific augmentation strategy for ophthalmic images. For example, the system will simulate different degrees of corneal opacity effects, or add simulated corneal scars to enhance the model's recognition ability of various pathological states. In addition, the present invention also introduces a data augmentation method based on the generative adversarial network (GAN), which can generate high-quality synthetic corneal images, further expanding the diversity of the training data.

[0123] The deep learning module 3 further includes a model interpretation unit 34 and a model optimization unit 35. The addition of these two units enables the system of the present invention to not only provide accurate evaluation results, but also explain the decision-making process and continuously optimize the model performance.

[0124] The model interpretation unit 34 is used to generate a visual interpretation of the decision-making process of the convolutional neural network, helping doctors understand the basis for the evaluation results. The present invention adopts an improved Grad-CAM++ algorithm to generate class activation maps, visually showing the image regions that the model focuses on when making decisions. In addition, the system also introduces SHAP (SHapley Additive exPlanations) value analysis to quantify the contribution degrees of different features to the final decision. This multi-dimensional interpretation mechanism greatly improves the interpretability and credibility of the model, helping doctors better understand and apply the AI-assisted diagnosis results.

[0125] The model optimization unit 35 is communicatively connected to the model interpretation unit 34 and is used to continuously optimize the performance of the convolutional neural network model based on the newly added labeled data and the model interpretation results. The present invention develops a dynamic weight adjustment mechanism that automatically identifies the weaknesses of the model according to the model interpretation results and gives higher weights to the corresponding samples or features in subsequent training. This method can specifically improve the performance of the model, especially when dealing with rare cases or boundary situations. Preferably, the model optimization process adopts an online learning strategy, which can absorb new clinical data in real time and continuously improve the accuracy and robustness of the model.

[0126] Finally, the present invention also proposes an ophthalmic corneal health assessment method based on machine learning and multispectral imaging. The method includes the following steps:

[0127] S1. Collect images of the cornea in different bands through a multispectral imaging device to obtain the RGB image and trichromatic light images of the cornea;

[0128] S2. Preprocess the collected multi-band corneal images, including noise reduction, contrast enhancement, and image segmentation;

[0129] S3. Extract feature parameters of the cornea such as specular reflectivity, optical density, and light transmittance from the preprocessed images;

[0130] S4. Perform multi-band superposition fusion on the images in different bands and their extracted features to generate a fused multispectral image;

[0131] S5. Input the fused multispectral image into a pre-trained convolutional neural network model to obtain the corneal health assessment result;

[0132] S6. Generate a corneal health assessment grade and a detailed parameter analysis report based on the assessment result;

[0133] S7. Display the assessment result and the analysis report to the user through the user interface and provide personalized health suggestions;

[0134] S8. Regularly collect the corneal multi - spectral images of users and analyze the dynamic changes in corneal health status through a deep - learning model;

[0135] S9. Use the newly collected data to continuously optimize the convolutional neural network model and improve the accuracy and adaptability of the evaluation.

[0136] This method realizes the full - process automation from data collection to health assessment through systematic steps. In particular, in step S3, the present invention adopts a method based on Fourier analysis to extract the optical characteristic parameters of the cornea. For example, the specular reflectance can be calculated by the following formula:

[0137]

[0138] where \(F(\omega)\) is the Fourier transform of the image, \(f\) 1 and \(f\) 2 are the lower and upper limits representing the specular reflection frequency range.

[0139] In step S8, the present invention introduces an anomaly - detection algorithm based on the variational auto - encoder (VAE) to identify abnormal changes in corneal health status. By learning the latent - space representation of normal corneal images, VAE can effectively detect abnormal situations deviating from the normal pattern, thus realizing early warning of corneal health status.

[0140] The objective function of VAE is as follows:

[0141]

[0142] where \(x\) is the input corneal image, \(z\) is the latent variable, \(q\) φ (z|x) is the encoder network (parameters \(\varphi\)), \(p\) θ (x|z) is the decoder network (parameters \(\theta\)), \(p(z)\) is the prior distribution (usually assumed to be the standard normal distribution), and \(D\) KL is the KL divergence.

[0143] The present invention improves the standard VAE by introducing a corneal - feature - aware loss term:

[0144]

[0145] where is the reconstructed image, \(\lambda\) is the trade - off coefficient, and is the corneal - feature - aware loss:

[0146]

[0147] Here, \(F\) corneal is a pre - trained corneal - feature extraction network.

[0148] Through this improvement, the VAE can better preserve the key features of the cornea and improve the accuracy of anomaly detection. The anomaly score is calculated as follows:

[0149]

[0150] where α is the coefficient for balancing the reconstruction error and the KL divergence, and MSE is the mean squared error.

[0151] In summary, the ophthalmic corneal health assessment system and method based on machine learning and multispectral imaging proposed by the present invention achieve high-precision and high-efficiency corneal health assessment through innovative technical solutions. The modular design of the system and the systematic steps of the method not only improve the accuracy and reliability of the assessment, but also have good scalability and adaptability. This method combining multispectral imaging and deep learning brings new technological breakthroughs to the field of ophthalmic diagnosis, is expected to play an important role in clinical practice, and provides strong support for the early diagnosis and personalized treatment of ophthalmic diseases.

[0152] It should be noted that the above are only the preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention shall be included within the protection scope of the present invention.

Claims

1. An ophthalmic corneal health assessment system based on machine learning and multispectral imaging, characterized in that: include: Multispectral imaging module for: Collect images of the cornea in different wavelength bands; Acquire RGB images and three-primary-color images of the cornea; A data processing module is connected to the multi-spectral imaging module for: Receiving corneal multi-band image data sent by the multi-spectral imaging module; Based on the corneal multi-band image data, performing image preprocessing, feature extraction and image fusion; A deep learning module, which is in communication with the data processing module, is used to: Receiving the fused multispectral image data sent by the data processing module; Analyzing the fused multispectral image data using a pre-trained convolutional neural network model to generate a corneal health assessment result; A user interface module is communicatively connected to the deep learning module and is used to: Receiving a corneal health assessment result sent by the deep learning module; The corneal health assessment results and related recommendations are displayed to the user.

2. The system according to claim 1, characterized in that The multispectral imaging module comprises: A light source unit, used for emitting light of different wavelengths to irradiate the cornea; An image acquisition unit, used for acquiring corneal images under illumination of light of different wavelengths; The wavelength selection unit is connected to the light source unit for controlling the light source unit to emit light of different wavelengths in a preset order.

3. The system according to claim 1, characterized in that The data processing module comprises: A preprocessing unit, used for performing noise reduction, contrast enhancement and image segmentation on the corneal multi-band image data; A feature extraction unit, which is in communication with the preprocessing unit and is used to extract characteristic parameters such as the specular reflectivity, optical density, and light transmittance of the cornea from the preprocessed image; The image fusion unit is connected to the feature extraction unit for performing multi-band superposition and fusion of images of different bands and features extracted therefrom.

4. The system according to claim 1, characterized in that The deep learning module includes: A model training unit, used for training a convolutional neural network model using a labeled corneal health sample data set; A model application unit, which is in communication connection with the model training unit and is used to apply the trained convolutional neural network model to analyze the input fused multispectral image; A result generating unit is communicatively connected to the model application unit, and is used to generate a corneal health assessment grade and detailed parameter analysis results based on the output of the convolutional neural network.

5. The system according to claim 1, characterized in that Also includes: A database module, in communication with the deep learning module and the user interface module, is used to: Stores the user's corneal image data, personal information, and historical assessment results; Provide training and validation datasets for deep learning modules; A real-time monitoring module is connected to the multispectral imaging module and the deep learning module for: Regularly collect multispectral images of the user's cornea; Analyze the dynamic changes of corneal health status through deep learning module.

6. The system according to claim 1, characterized in that The user interface module comprises: A data input unit for receiving the user's personal information and relevant medical history; A result display unit, used to graphically display corneal health assessment results and historical change trends; The suggestion generating unit is in communication with the result display unit and is used to generate personalized health suggestions according to the evaluation results.

7. The system according to claim 1, characterized in that Also includes: A cloud computing module is connected to the deep learning module for: Store the trained convolutional neural network model parameters; Perform computationally intensive deep learning tasks; Regularly update and optimize deep learning models.

8. The system according to claim 1, characterized in that The data processing module also includes: A multimodal data fusion unit, used to fuse corneal image data with the user's age, gender, medical history and other information to generate a comprehensive feature vector; The data enhancement unit is communicatively connected to the multimodal data fusion unit and is used to expand the training data set by image rotation, scaling, flipping, etc.

9. The system according to claim 1, characterized in that The deep learning module also includes: Model explanation unit, which is used to generate visual explanations of the decision-making process of the convolutional neural network to help doctors understand the basis of the evaluation results; A model optimization unit is communicatively connected to the model interpretation unit, and is used to continuously optimize the performance of the convolutional neural network model based on the newly added labeled data and the model interpretation results.

10. An ophthalmic corneal health assessment method based on machine learning and multispectral imaging, applied to the system according to any one of claims 1 to 9, characterized in that: The following steps are involved: S1. Collect corneal images in different wavelength bands through multispectral imaging equipment to obtain RGB images and three-primary color images of the cornea; S2. Preprocessing the collected multi-band corneal images, including noise reduction, contrast enhancement and image segmentation; S3. extracting characteristic parameters such as specular reflectivity, optical density, and light transmittance of the cornea from the preprocessed image; S4. Perform multi-band superposition and fusion of images of different bands and their extracted features to generate a fused multispectral image; S5. Input the fused multispectral image into the pre-trained convolutional neural network model to obtain the corneal health assessment result; S6. Generate a corneal health assessment grade and detailed parameter analysis report based on the assessment results; S7. present the evaluation results and analysis reports to the user through the user interface and provide personalized health advice; S8. Regularly collect multispectral images of the user's cornea and analyze the dynamic changes of corneal health status through deep learning models; S9. Use the newly collected data to continuously optimize the convolutional neural network model to improve the accuracy and adaptability of the evaluation.

Images