Cornea conus analysis system based on dynamic cornea texture

Abstract

The invention relates to a keratoconus analysis system based on dynamic cornea textures. The keratoconus analysis system comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit which are connected in sequence, the multi-time-point cornea dynamic image texture analysis method is provided, in order to ensure that the screened features have the optimal diagnostic capability, a feature set with the highest diagnostic value is screened through a recursive feature elimination method, and a foundation is laid for FFKC high-precision recognition. The machine learning classification system comprises a segmentation module, a model training module, a model testing module and a performance evaluation module, and is used for obtaining a machine learning classifier. The performance of the model is evaluated through an ROC curve and a confusion matrix, and high sensitivity and high specificity are ensured. A multi-modal recognition platform fusing dynamic textures and cornea biomechanical parameters is constructed, the platform sequentially comprises a data input module, a data processing module, a model module and an output module, and efficient FFKC recognition and auxiliary diagnosis are achieved.

Description

Technical Field

[0001] The present invention relates to the field of medical technology, and more specifically, to a keratoconus analysis system based on dynamic corneal texture. Background Art

[0002] Keratoconus (KC) is a progressive corneal ectasia disease characterized by progressive thinning and expansion of the cornea, leading to irregular astigmatism and significant vision loss, which seriously affects the patient's quality of life. If not treated in time, it may even lead to blindness. According to studies, the global prevalence of keratoconus is about 1 / 2000 to 1 / 1000, which is particularly common among young people and imposes a significant burden on society and the economy. With the rapid popularization of refractive surgery, the number of surgeries has increased sharply, and some patients have iatrogenic corneal ectasia complications after surgery. Failure to identify latent keratoconus in time is considered to be one of the main reasons for its occurrence.

[0003] Frustrated keratoconus (FFKC) is defined as the normal fellow eye of a patient who has been clinically diagnosed with keratoconus, with normal slit lamp examination and no abnormal corneal topography. According to the 2015 global expert consensus on keratoconus, "there is no true unilateral keratoconus", and although FFKC currently has no clinical symptoms and signs, it is at great risk of developing keratoconus. Studies have shown that although the cause of keratoconus has not yet been fully clarified, the current consensus is that its occurrence and development are closely related to changes in the regional biomechanical properties of the cornea. Even if FFKC has not yet found any abnormal signs clinically, subtle biomechanical changes may have begun to occur inside the cornea.

[0004] Early identification of FFKC is important for preventing disease progression and avoiding inappropriate refractive surgery. On the one hand, early intervention can delay or prevent disease progression through techniques such as corneal collagen cross-linking; on the other hand, accurate identification of FFKC can prevent high-risk patients from undergoing corneal refractive surgery that may induce corneal ectasia. However, due to the subtle changes in FFKC abnormalities, its detection and diagnosis remains a major challenge.

[0005] Existing technologies mainly rely on morphological and biomechanical evaluations. Among them, the main corneal morphological evaluation devices are Pentacam and optical coherence tomography (OCT). This type of method identifies abnormalities by measuring static parameters such as corneal curvature and thickness. Specifically, it includes corneal topography analysis: used to evaluate the anterior surface morphology and provide curvature distribution and asymmetry information; keratoconus percentage index (KISA%): comprehensive parameters such as corneal curvature asymmetry and central curvature value; wavefront aberration analysis: measuring high-order aberrations and evaluating optical quality changes; and OCT imaging: evaluating the integrity and regularity of each layer structure, especially the Bowman layer. These technologies mainly focus on the morphological characteristics of the cornea, and assist in diagnosis by identifying morphological abnormalities with diagnostic significance. However, due to the extremely small morphological changes in the early stage of FFKC, the sensitivity of existing morphological parameters to its detection is limited, and there is not enough sensitivity and specificity to capture early subtle changes.

[0006] It can be seen that the current FFKC diagnostic methods mainly rely on morphological analysis and simple biomechanical parameters, which have many limitations. Morphological methods only focus on the surface characteristics of the cornea and lack sensitivity to changes in the internal structure of the cornea. FFKC lesions are caused by internal collagen fiber disorder, which first manifests as changes in internal microlesions, while there are no obvious abnormalities in the existing morphological parameters. Although the traditional biomechanical parameters provided by the visual corneal biomechanical analyzer (Corvis ST, CVS) can evaluate the dynamic response of the cornea, they are mainly affected by central corneal thickness and intraocular pressure, and are mainly global parameters. They lack the ability to finely analyze local changes in the cornea, resulting in limited diagnostic accuracy.

[0007] In recent years, with the advancement of medical imaging technology, accurate assessment of corneal morphology and biomechanical properties has become possible. As a new type of corneal biomechanical assessment device, CVS stimulates corneal deformation through airflow pulses and records its dynamic response, providing a new method to quantify corneal biomechanical properties. At the same time, artificial intelligence (AI) technology is increasingly used in the field of medical diagnosis, providing new possibilities for the processing and analysis of complex medical data. Radiomics, as a technology for extracting high-dimensional quantitative features from medical images, has shown great potential in oncology and other medical fields. At present, radiomics still faces challenges in identifying early biomechanical abnormalities in FFKC, and lacks a stable and interpretable modeling method.

[0008] In this context, how to analyze FFKC based on corneal dynamic deformation data is the technical problem to be solved by this application. Summary of the invention

[0009] In view of the shortcomings of the existing technology, the present invention proposes a keratoconus analysis system based on dynamic corneal texture, which combines CVS, advanced imaging genomics analysis and machine learning algorithms to extract rich texture features from corneal deformation images at multiple time points, thereby enabling in-depth analysis of the internal structure and dynamic deformation behavior of the cornea, and facilitating the diagnosis of frustrated keratoconus.

[0010] On the one hand, the present invention provides a keratoconus analysis system based on dynamic corneal texture, and the technical solution is as follows: A keratoconus analysis system based on dynamic corneal texture comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit which are connected in sequence, wherein a machine learning classifier is arranged in the feature screening unit.

[0011] The image acquisition unit acquires the cornea's initial moment image, first applanation moment image and maximum deformation moment image through a visual corneal biomechanics analyzer.

[0012] The image preprocessing unit extracts the region of interest (ROI) from the image at each moment.

[0013] The feature extraction unit extracts a number of texture features from the region of interest of the image at each moment.

[0014] The feature screening unit uses a machine learning classifier to use a recursive feature elimination method to screen out a feature subset from the texture features of the image at each moment.

[0015] In a second aspect, the present invention provides a machine learning classification system, and the technical solution is as follows: The machine learning classification system includes a segmentation module, a model training module, a model testing module and a performance evaluation module which are connected in sequence.

[0016] A data set including normal eye samples and keratoconus eye samples is obtained, and a data segmentation module divides the data set into a training set and a test set.

[0017] The model training module builds and optimizes different types of machine learning classifiers based on the training set.

[0018] The model testing module uses the trained machine learning classifier to predict the test set to obtain the prediction results.

[0019] The performance evaluation module calculates the performance index based on the prediction results to obtain the machine learning classifier with the best performance index.

[0020] In a third aspect, the present invention provides a keratoconus analysis platform based on the fusion of dynamic corneal texture and biomechanical parameters, and the technical solution is as follows: It includes a data input layer, a data processing layer, a model layer and an output layer which are connected in sequence.

[0021] The image acquisition unit is arranged in the data input layer, and the data input layer is used to receive the corneal dynamic image collected by the visual corneal biomechanical analyzer and calculate the biomechanical parameters. The image acquisition unit collects the corneal initial moment image, the first flattening moment image and the maximum deformation moment image from the corneal dynamic image and transmits them to the data input layer.

[0022] The image preprocessing module and the texture feature extraction module are arranged in a data processing layer, and the data processing layer also includes a biomechanical parameter processing module for processing biomechanical parameters. The data processing layer is used to extract radiomics texture features from the dynamic image of the cornea and to standardize the biomechanical parameters.

[0023] The feature screening unit is arranged at a model layer, and the model layer comprises a first classification model constructed based on texture features, a second classification model constructed based on biomechanical parameters, and a fusion module for fusing the prediction results of the first classification model and the second classification model.

[0024] The output layer outputs keratoconus diagnosis results, key feature analysis results and a visual diagnosis report according to the prediction results of the fusion module.

[0025] In summary, the above technical solution has the following beneficial effects: the keratoconus analysis system based on dynamic corneal texture of the present invention can significantly improve the early detection rate of FFKC, reduce the misdiagnosis rate, and provide clinicians with a reliable and non-invasive diagnostic tool. By early identification of high-risk patients with corneal ectasia, the occurrence of serious complications after corneal refractive surgery can be avoided to a great extent, and clinical decision-making can be optimized. At the same time, the simplified process of a single device reduces the examination burden on patients, improves screening efficiency, and provides a scientific basis for early intervention and personalized treatment plan formulation of FFKC, which is of great significance to improving the diagnosis and treatment level of corneal diseases and the visual health of patients. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Figure 1 Schematic diagram of images of the cornea at three key time points collected by the visual corneal biomechanical analyzer; Figure 2 Extract features from the image at the initial moment, normalize all features, and output a feature matrix diagram; Figure 3 Schematic diagram of ROC curve and confusion matrix of 3 different machine learning models; Figure 4 Schematic diagram of the ROC curve of the existing CVS biomechanical parameters. DETAILED DESCRIPTION

[0027] The following is a clear and complete description of the technical solutions in the embodiments of the present invention in conjunction with the drawings in the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The present invention aims to solve the key problems of insufficient sensitivity, low specificity and reliance on comprehensive evaluation of multiple devices in the early diagnosis of frustrated keratoconus (FFKC). Existing FFKC diagnostic methods mainly rely on corneal morphology and corneal biomechanics examinations, lacking in-depth analysis of the internal structure of the cornea and the dynamic deformation process, resulting in low early detection rates and high misdiagnosis rates, thereby delaying the best treatment opportunity and increasing the risk of serious complications such as corneal ectasia after refractive surgery. In addition, the current diagnostic process is cumbersome, often requiring the combination of multiple examination equipment and techniques, and relying on the clinical experience of physicians, which not only increases the burden on patients, but also reduces clinical efficiency, limiting the widespread implementation of early screening for FFKC.

[0029] Studies have shown that corneal biomechanical changes play a key role in the development of FFKC. Early lesions first manifest as changes in local biomechanical properties rather than obvious morphological abnormalities. The dynamic response process of the cornea when subjected to external forces (such as air pulses) contains rich biomechanical information, which is of great value for the identification of early FFKC. However, traditional diagnostic methods only focus on the limited biomechanical parameters extracted from corneal surface deformation and cannot fully capture the subtle changes and dynamic response characteristics of the deep corneal tissue, resulting in insufficient sensitivity to early lesions, especially in the early stages of FFKC when the morphology is normal but the biomechanics have changed.

[0030] In response to the above problems, the present invention proposes a keratoconus analysis system based on dynamic corneal texture, which combines a visual corneal biomechanical analyzer (Corvis ST, CVS), advanced imaging genomics analysis and machine learning algorithms to extract rich texture features from corneal deformation images at multiple time points, and deeply analyze the internal structure and dynamic deformation behavior of the cornea. The present invention screens the most diagnostically valuable features through a recursive feature elimination method, constructs a highly accurate FFKC intelligent recognition model, and achieves accurate capture of microbial biomechanical changes. In addition, the present invention integrates time domain information, and simultaneously analyzes the corneal characteristics at the initial moment, the first applanation moment, and the maximum deformation moment, comprehensively evaluates the corneal biomechanical properties, and transforms from empirical diagnosis to precise quantitative diagnosis, improves the sensitivity and specificity of early screening, and simplifies the diagnostic process.

[0031] According to a first aspect of the present invention, there is provided a keratoconus analysis system based on dynamic corneal texture, which comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit connected in sequence, wherein a machine learning classifier is arranged in the feature screening unit.

[0032] like Figure 1 As shown, the image acquisition unit acquires the cornea's initial moment image, first flattening moment image, and maximum deformation moment image through a visual corneal biomechanical analyzer. The visual corneal biomechanical analyzer uses standardized airflow pulses to record the dynamic deformation process of the cornea under the action of external forces. The specific steps include: fixing the subject's head on the device bracket; adjusting the device to the center of the subject's cornea and focusing; starting the device and applying a standard airflow pulse; recording the corneal deformation process and acquiring a high-speed camera image sequence; extracting images of three key time points from the image sequence: the initial moment, the first flattening moment, and the maximum deformation moment.

[0033] The image preprocessing unit extracts the region of interest (ROI) from the image at each moment; the extraction of the region of interest is achieved through the following process: obtaining the grayscale image at each moment according to the image at each moment; establishing a coordinate system, automatically detecting the front and back surfaces of the cornea through the visual corneal biomechanical analyzer and obtaining the coordinates; segmenting the corneal region image by edge detection and curve fitting methods: removing the corneal region image noise and enhancing the contrast, so as to obtain the region of interest of the image at each moment. The image preprocessing unit uses matlab code to implement the preprocessing steps, and ensures the consistency of the data format by parsing the original 8-bit grayscale image of the visual corneal biomechanical analyzer. Specifically, edge detection is to detect the edge of the cornea through the Canny edge detection algorithm or the Sobel operator. The corneal boundary is smoothed by polynomial fitting, and the morphological opening and closing operation is used to remove the extra noise. The corneal region is automatically cropped and the standardized region of interest is output.

[0034] like Figure 2As shown, the feature extraction unit extracts several texture features from the region of interest of the image at each moment; the texture features include first-order statistical features, shape features, grayscale co-occurrence matrix features, grayscale run length matrix features, grayscale size area matrix features, grayscale dependence matrix features, neighborhood grayscale difference matrix features and wavelet transform features; the texture features at three different moments are normalized and the feature matrix is ​​output. Specifically, the feature extraction unit uses radiomics analysis technology to extract multidimensional texture features from the image of the region of interest. Specifically, the symbolic texture is first extracted from the image of the region of interest, represented by Texture, and then the texture features extracted from the symbolic texture are represented by Feature. Furthermore, the eight types of texture features extracted from each image include 464 features, and the images of a person at three moments are 1392 feature elements.

[0035] The feature screening unit uses a recursive feature elimination method to screen out feature subsets from the texture features of the images at each moment through a machine learning classifier. The process of using the recursive feature elimination method to screen out feature subsets is as follows: input the feature matrix and the label vector of the corresponding sample into the feature screening unit; after initializing the machine learning classifier, calculate the importance score of the feature through the machine learning classifier, and sort all texture features in descending order of importance; set the range of the number of texture features, perform a five-fold cross validation to determine the optimal number of features and output the screened texture features, thereby obtaining a feature subset. Specifically, there are 51 feature elements in the feature subset, and only 51 feature elements are found in the 1,392 features of a person's three moment images after being screened by the recursive feature elimination method.

[0036] The beneficial effects of a keratoconus analysis system based on dynamic corneal texture include: Different from traditional static morphological analysis or single-time point biomechanical parameter evaluation, the analysis system of the present invention proposes a multi-time point corneal dynamic image texture analysis method. The radiomics method is used to extract comprehensive texture features from the image, including first-order statistical features, shape features, gray-level co-occurrence matrix (GLCM), gray-level dependency matrix (GLDM), gray-level run length matrix (GLRLM), gray-level size zone matrix (GLSZM), neighborhood gray-level difference matrix (NGTDM) and wavelet transform features, to comprehensively evaluate the dynamic changes of corneal microstructure. In order to ensure that the selected features have the best diagnostic ability, the recursive feature elimination method (RFE) is used to select the most diagnostically valuable feature set, laying the foundation for high-precision recognition of FFKC.

[0037] The keratoconus analysis system based on dynamic corneal texture of the present invention can significantly improve the early detection rate of FFKC, reduce the misdiagnosis rate, and provide clinicians with a reliable and non-invasive diagnostic tool. By identifying high-risk patients with corneal ectasia at an early stage, the occurrence of serious complications after corneal refractive surgery can be avoided to a great extent, and clinical decision-making can be optimized. At the same time, the simplified process of a single device reduces the examination burden on patients, improves screening efficiency, and provides a scientific basis for early intervention and personalized treatment plan formulation of FFKC, which is of great significance to improving the diagnosis and treatment level of corneal diseases and the visual health of patients.

[0038] Alternative imaging omics feature extraction technology: The present invention originally preferred to use imaging omics to diagnose FFKC by extracting texture features of corneal images. As an alternative, deep learning technology, especially convolutional neural networks (CNNs), can be used to automatically learn features directly from corneal images. This method does not require manual feature design, but automatically extracts deep features of corneal images through multi-layer convolution and pooling operations, which is suitable for large-scale data sets and complex image analysis scenarios. In addition, quantitative ultrasound elastography technology can also be used as an alternative to evaluate corneal elasticity by measuring the propagation characteristics of sound waves in corneal tissue, indirectly reflecting corneal biomechanical changes, and is suitable for clinical environments that require non-contact evaluation.

[0039] Alternative recursive feature elimination method: The present invention originally preferred to use recursive feature elimination method to screen the texture features with the most diagnostic value. As an alternative, principal component analysis (PCA) technology can be used to retain the maximum variance information of the data through dimensionality reduction and quickly extract the main features. Another alternative is the feature importance evaluation based on tree models, such as ExtraTrees or LightGBM. These methods evaluate the importance of features by constructing multiple decision trees and are suitable for rapid screening of high-dimensional feature spaces. In addition, LASSO regression can be used to achieve the synchronization of feature selection and model training, and important features are automatically screened through L1 regularization, which is suitable for situations where the number of features is much larger than the number of samples.

[0040] Alternative multi-time point integration strategy: The present invention preferably integrates the corneal image features at three time points. As an alternative, a time series analysis method can be used to regard the entire corneal deformation process as a continuous time series, and extract dynamic change features, such as deformation rate, rebound characteristics and other time domain information. Another alternative is to use an attention mechanism to assign different weights to features at different time points, and automatically learn the time points with the most diagnostic value, which is suitable for situations where the contributions of time points are uneven. The optical flow method can also be used to track the displacement field of each pixel during corneal deformation, calculate the local strain distribution, and more directly reflect the biomechanical properties of the cornea.

[0041] According to a second aspect of the present invention, a machine learning classification system is provided for obtaining a machine learning classifier.

[0042] The machine learning classification system includes a segmentation module, a model training module, a model testing module and a performance evaluation module which are connected in sequence; A dataset including normal eye samples and frustrated keratoconus eye samples was obtained, and the data segmentation module divided the dataset into a training set and a test set. The samples were obtained by a visual corneal biomechanical analyzer, which can provide a variety of biomechanical parameters: the first applanation time (A1T) measures the time for the cornea to reach the first applanation state; Ambrósio related thickness (ARTh) calculates the spatial distribution ratio of corneal thickness; the stress-strain index (SSI) reflects tissue hardness; the stiffness parameter at the first applanation moment (SP-A1) evaluates the ability to resist deformation; the deformation amplitude ratio (DARatio2) measures the ratio of the deformation amplitude at the vertex and the periphery at 2mm; the integrated inverse radius (IIR) evaluates the central deformation curvature; and the CVS biomechanical index (CBI) is used as a multi-parameter comprehensive score. By quantifying the external force response characteristics, these parameters evaluate the biomechanical properties of corneal tissue such as elasticity, stiffness and viscoelasticity.

[0043] The model training module builds and optimizes several machine learning classifiers based on the training set; building and optimizing several machine learning classifiers is achieved through the following process: inputting the feature matrix and label vector of each sample in the training set into the model training module; building classifiers such as random forest classifiers, C5.0 decision tree classifiers, and extreme gradient boosting classifiers; using five-fold cross validation to optimize the parameters of each classifier, and determining the optimal parameters based on the five-fold cross validation results; retraining and outputting the trained classifiers using the optimal parameters and training set, thereby improving the model's recognition ability for FFKC; The model testing module uses the trained machine learning classifier to predict the test set to obtain the prediction results; predicting the test set includes the following processes: inputting the feature matrix of each sample in the test set and the trained classifier into the model testing module; the trained classifier classifies each sample and outputs the prediction results; comparing the prediction results with the label vectors of the corresponding samples to obtain the prediction probability values, and outputting the confusion matrix.

[0044] The performance evaluation module calculates the performance indicators based on the prediction results, so as to obtain the machine learning classifier with the best performance indicators. The machine learning classifier with the best performance indicators is obtained through the following process: the performance evaluation module calculates the ROC curve based on the predicted probability value and the confusion matrix, which is used to evaluate the machine learning classifier; draws a visual chart based on the predicted probability value, confusion matrix and ROC curve and outputs a performance evaluation report, so as to obtain the machine learning classifier with the best performance indicators. Specifically, input the confusion matrix and predicted probability value into the performance evaluation module; calculate the accuracy, sensitivity, specificity, positive predictive value, negative predictive value and ROC curve; finally, draw a visual chart and output a performance evaluation report, so as to obtain the machine learning classifier with the best performance indicators. Figure 3 As shown in the figure, ab is the ROC curve of the random forest model and the confusion matrix made with the test set, cd is the XGBoost model, and ef is the c5.0 model. Figure 4 Shown are the ROC curves of existing CVS biomechanical parameters.

[0045] The beneficial effects of the machine learning classification system include: Based on the screening of texture features, a machine learning classifier that integrates multi-time point information is constructed, including algorithms such as random forest (RF), C5.0, and extreme gradient boosting (XGBoost). The model parameters are optimized through five-fold cross-validation to improve the model's recognition ability for FFKC. The model outputs the recognition results and probability distribution of FFKC, quantifies the diagnostic confidence, and evaluates the model performance through the ROC curve (Receiver Operating Characteristic Curve) and confusion matrix (Confusion Matrix) to ensure high sensitivity and high specificity.

[0046] According to a third aspect of the present invention, a keratoconus analysis platform based on the fusion of dynamic corneal texture and biomechanical parameters is provided, which is used to access a keratoconus analysis system based on dynamic corneal texture as an expansion module for multimodal feature integration.

[0047] It includes a data input layer, a data processing layer, a model layer and an output layer which are connected in sequence.

[0048] The image acquisition unit is arranged in the data input layer, and the data input layer is used to receive the corneal dynamic image collected by the visual corneal biomechanical analyzer and calculate the biomechanical parameters. The image acquisition unit collects the corneal initial moment image, the first flattening moment image and the maximum deformation moment image from the corneal dynamic image and transmits them to the data input layer.

[0049] The image preprocessing module and the texture feature extraction module are arranged in a data processing layer, and the data processing layer also includes a biomechanical parameter processing module for processing biomechanical parameters. The data processing layer is used to extract radiomics texture features from the dynamic image of the cornea and to standardize the biomechanical parameters.

[0050] The feature screening unit is set in the model layer, and the model layer includes a first classification model constructed based on texture features, a second classification model constructed based on biomechanical parameters, and a fusion module for fusing the prediction results of the first classification model and the second classification model. The input biomechanical parameters include stress-strain index and its improved version, stiffness parameters at the first flattening moment, Ambrósio relative thickness, deformation amplitude ratio, CVS biomechanical index, etc.; the data input layer is also used to input and manage basic patient information and other content. Specifically, the first classification model is a random forest model constructed based on texture features, and the second classification model is a linear discriminant analysis (LDA) model constructed based on corneal biomechanical parameters. Weight coefficients are assigned to the first classification model and the second classification model according to their performance indicators on the test set; the fusion module fuses the prediction results of the two classification models according to a weighted voting mechanism or a weighted probability mechanism; the performance indicators include but are not limited to the area under the ROC curve, accuracy or F1 score; the weight coefficient is calculated according to the following normalization formula: ; is the weight coefficient of the i-th classification model, is the area under the ROC curve of the i-th classification model, It is the sum of the areas under the ROC curves of the two classification models. Preferably, the two machine learning classifiers are a random forest classifier based on radiomics features and a linear discriminant analysis classifier based on radiomics features; the fusion module adopts a weighted voting mechanism to set the weight coefficient according to the relative performance of the model. The F1 score is the harmonic mean of precision and recall, which is used to measure the comprehensive performance of the model in positive class prediction.

[0051] Alternatives to the random forest model: In addition to the random forest model, a variety of machine learning algorithms can be used as alternatives. Support vector machines (SVMs) achieve high-precision classification by finding the optimal classification hyperplane; deep neural networks can handle more complex nonlinear relationships; integrated learning methods such as AdaBoost and Stacking can combine the advantages of multiple classifiers; Bayesian networks are suitable for probabilistic reasoning and uncertainty quantification, providing risk assessment for clinical decision-making.

[0052] The output layer outputs keratoconus diagnosis results, key feature analysis results and visual diagnosis reports based on the prediction results of the fusion module. The output layer outputs FFKC analysis results, risk assessment, key feature analysis, visual reports, clinical recommendations, etc. The platform supports local deployment or cloud service mode. The system workflow includes data acquisition, preprocessing, feature extraction, model prediction and result output. The platform adopts a modular design, which is easy to maintain and upgrade, and supports continuous learning of new data to ensure continuous optimization of analysis performance.

[0053] The beneficial effects of a keratoconus analysis platform based on the fusion of dynamic corneal texture and biomechanical parameters include: combining CVS texture analysis results with traditional biomechanical parameters to establish a comprehensive FFKC analysis platform. The platform integrates the internal texture characteristics of the cornea and the global biomechanical deformation characteristics through the model layer to form a multi-dimensional FFKC evaluation system, which effectively makes up for the limitations of the single modality method. The analysis accuracy of the system is verified by clinical data, and the algorithm model is continuously optimized to improve clinical applicability. Ultimately, the transition from surface morphology analysis to deep tissue characteristic analysis is achieved, and a scientific and precise FFKC early analysis process is established to provide an objective basis for clinical decision-making and personalized treatment plan formulation.

[0054] In summary, the beneficial effects of the present application include: 1. The present invention applies imaging genomics technology to corneal dynamic deformation image analysis for the first time, filling the gap in existing FFKC analysis technology. By extracting deep texture features from CVS images, the present invention can capture tiny changes in the internal structure of the cornea that cannot be identified by traditional parameters, greatly improving the early detection rate of FFKC and reducing the risk of missed diagnosis.

[0055] 2. The present invention uses CVS to obtain high-resolution images of the cornea at three key moments: the initial moment, the first applanation moment, and the maximum deformation moment, establishes a multi-time point integrated classifier, and extracts the texture features inside the cornea through imaging genomics technology, breaking through the limitations of traditional static morphological analysis or global biomechanical parameter evaluation, and breaking through the limitations of existing technologies that only focus on a single time point or static morphology. Experiments have shown that the multi-time point model is significantly better than the single time point model, and provides a more comprehensive evaluation of the dynamic deformation process of the cornea.

[0056] 3. The present invention uses radiomics analysis methods to extract eight types of features from corneal images: first-order statistical features, shape features, gray-level co-occurrence matrix (GLCM) features, gray-level run length matrix (GLRLM) features, gray-level size zone matrix (GLSZM) features, gray-level dependence matrix (GLDM) features, neighborhood gray-level difference matrix (NGTDM) features and wavelet transform features, which can comprehensively capture the subtle changes in corneal structure.

[0057] 4. The present invention uses the recursive feature elimination (RFE) method to screen the feature subset with the most analytical value, reducing the number of features from 1,392 to 51, significantly improving the algorithm efficiency and generalization ability while ensuring the accuracy of analysis.

[0058] 5. Based on the screened features, the present invention constructs a multi-time point integrated random forest classifier, and achieves high-precision recognition of FFKC by integrating the information of three key deformation moments of the cornea. The AUC reaches 0.989, which is significantly better than the existing evaluation parameters (SP-A1 parameter, AUC is only 0.728).

[0059] 5. Combining artificial intelligence and radiomics to achieve intelligent screening of FFKC and promote the development of precision medicine. Existing FFKC screening technologies mostly rely on empirical judgment of fixed thresholds, and the diagnostic results are greatly affected by the subjective influence of the operator. The present invention combines radiomics and machine learning methods to achieve data-driven intelligent FFKC diagnosis, reduce human intervention, and improve consistency. The sustainably optimized AI model continuously improves the algorithm performance as data accumulates, and promotes the development of FFKC diagnosis in the direction of precision medicine. 6. The analysis platform is easy to operate, easy to promote, and suitable for large-scale clinical screening. Traditional FFKC detection methods may require the combination of multiple devices (such as Pentacam+Corvis ST+OCT), and the inspection process is complicated and costly. The present invention only requires one CVS device to complete the detection, reducing the patient's inspection time and economic burden, improving screening efficiency, and making early screening of FFKC more feasible: It is suitable for preoperative FFKC risk screening and reducing complications of corneal dilatation after refractive surgery. Lower the threshold of clinical operation, facilitate promotion in primary medical institutions, and improve the early diagnosis rate of FFKC.

[0060] Based on previous research, the project team of this application has established a corneal dynamic image and patient database for the FFKC diagnostic model, completing the entire process from data collection to model verification. The study included a sufficient number of normal eyes and FFKC eyes, using CVS to obtain high-resolution images, and appropriately allocating training and test sets.

[0061] Through radiomics feature extraction and recursive feature elimination, we screened out the most diagnostically valuable texture features from the images of three time points of each eye. The random forest model performed well on the test set, and the diagnostic accuracy was significantly higher than that of the traditional CVS parameters. We conducted comparative experiments on the multi-time point integrated model and the single time point model. The results showed that the performance of the multi-time point model was significantly better than that of the single time point model, verifying the superiority of multi-time point information integration. In addition, the fusion of texture features and traditional biomechanical parameters further improved the diagnostic efficiency. The analysis of texture features during corneal deformation showed that there were significant differences in texture features between normal eyes and FFKC eyes at three time points, especially at the moment of maximum deformation. This shows that although FFKC has no obvious morphological abnormalities, the response pattern of its internal structure has changed after being stressed, confirming the ability of the present invention to capture subtle structural changes. Preliminary results of clinical application show that the present invention not only improves the detection rate of FFKC, but also reduces the time required for the diagnostic process, especially in pre-refractive surgery screening, showing great application value. The model's ability to identify high-risk groups early helps prevent complications of postoperative corneal ectasia and improve the safety of refractive surgery.

[0062] In general, the experimental and preliminary clinical application results fully demonstrated the feasibility and effectiveness of the present invention, provided a new technical solution for the early and accurate diagnosis of FFKC, and had broad prospects for clinical application.

[0063] The above are only preferred embodiments of the present invention. The protection scope of the present invention is not limited to the above embodiments. All technical solutions under the concept of the present invention belong to the protection scope of the present invention. It should be pointed out that for ordinary technicians in this technical field, some improvements and modifications without departing from the principle of the present invention should also be regarded as the protection scope of the present invention.

Claims

1. A keratoconus analysis system based on dynamic corneal texture, characterized in that: It includes an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit connected in sequence, wherein a machine learning classifier is arranged in the feature screening unit; The image acquisition unit acquires the cornea's initial moment image, first applanation moment image and maximum deformation moment image through a visual corneal biomechanics analyzer; The image preprocessing unit extracts the region of interest from the image at each moment; The feature extraction unit extracts a plurality of texture features from the region of interest of the image at each moment; The feature screening unit uses a machine learning classifier to screen out a feature subset using a recursive feature elimination method for the texture features of the image at each moment.

2. A keratoconus analysis system based on dynamic corneal texture according to claim 1, characterized in that: Extracting the region of interest is achieved through the following process: Obtaining grayscale images at each moment according to images at each moment; Establish a coordinate system, and automatically detect the front and back surfaces of the cornea and obtain coordinates using a visual corneal biomechanical analyzer; Use edge detection and curve fitting methods to segment the corneal area image: The image noise in the corneal area is removed and the contrast is enhanced to obtain the region of interest of the image at each moment.

3. A keratoconus analysis system based on dynamic corneal texture according to claim 1, characterized in that: The texture features include first-order statistical features, shape features, grayscale co-occurrence matrix features, grayscale run length matrix features, grayscale size area matrix features, grayscale dependency matrix features, neighborhood grayscale difference matrix features and wavelet transform features; The texture features at three different moments are normalized and the feature matrix is ​​output.

4. A keratoconus analysis system based on dynamic corneal texture according to claim 3, characterized in that: The process of selecting feature subsets using recursive feature elimination method is as follows: Inputting a feature matrix and a label vector of a corresponding sample into the feature screening unit; After initializing the machine learning classifier, the importance score of the feature is calculated by the machine learning classifier, and all texture features are arranged in descending order of importance; The range of texture feature quantity is set, and a five-fold cross validation is performed to determine the optimal number of features and output the filtered texture features to obtain a feature subset.

5. A machine learning classification system, characterized in that Used to obtain the machine learning classifier according to any one of claims 1 to 4, wherein the machine learning classification system includes a segmentation module, a model training module, a model testing module and a performance evaluation module connected in sequence; Acquire a data set including normal eye samples and frustrated keratoconus eye samples, and the data segmentation module divides the data set into a training set and a test set; The model training module constructs and optimizes several machine learning classifiers based on the training set; The model testing module uses the trained machine learning classifier to predict the test set to obtain the prediction result; The performance evaluation module calculates the performance index according to the prediction results, thereby obtaining the machine learning classifier with the best performance index.

6. A machine learning classification system according to claim 5, characterized in that: Building and optimizing several machine learning classifiers is accomplished through the following process: Inputting the feature matrix and label vector of each sample in the training set into the model training module; Five-fold cross validation is used to optimize the parameters of each classifier, and the optimal parameters are determined based on the five-fold cross validation results; the trained classifier is retrained using the optimal parameters and training set and output.

7. A machine learning classification system according to claim 6, characterized in that: Making predictions on the test set involves the following process: Inputting the feature matrix of each sample in the test set and the trained classifier into the model testing module; The trained classifier classifies each sample and outputs the prediction result; Compare the prediction result with the label vector of the corresponding sample to obtain the predicted probability value and output the confusion matrix.

8. A machine learning classification system according to claim 7, characterized in that: Obtaining the best machine learning classifier performance metric is achieved through the following process: The performance evaluation module calculates the ROC curve based on the predicted probability value and the confusion matrix to evaluate the machine learning classifier; Draw visual charts based on the predicted probability values, confusion matrix, and ROC curve and output a performance evaluation report to obtain the machine learning classifier with the best performance indicators.

9. A keratoconus analysis platform based on the fusion of dynamic corneal texture and biomechanical parameters, used for accessing the keratoconus analysis system based on dynamic corneal texture as described in any one of claims 1 to 4, characterized in that: It includes a data input layer, a data processing layer, a model layer and an output layer which are connected in sequence; The image acquisition unit is arranged at the data input layer, and the data input layer is used to receive the corneal dynamic image acquired by the visual corneal biomechanics analyzer and calculate the biomechanical parameters. The image acquisition unit acquires the corneal initial moment image, the first flattening moment image and the maximum deformation moment image from the corneal dynamic image and transmits them to the data input layer; The image preprocessing module and the texture feature extraction module are arranged in a data processing layer, and the data processing layer further comprises a biomechanical parameter processing module for processing biomechanical parameters, and the data processing layer is used to extract radiomics texture features from the dynamic image of the cornea and perform standardization processing on the biomechanical parameters; The feature screening unit is arranged in a model layer, and the model layer comprises a first classification model constructed based on texture features, a second classification model constructed based on biomechanical parameters, and a fusion module for fusing the prediction results of the first classification model and the second classification model; The output layer outputs keratoconus diagnosis results, key feature analysis results and a visual diagnosis report according to the prediction results of the fusion module.

10. The keratoconus analysis platform based on the fusion of dynamic corneal texture and biomechanical parameters according to claim 9, characterized in that: The first classification model is a random forest model constructed based on texture features, and the second classification model is a linear discriminant analysis model constructed based on corneal biomechanical parameters, and weight coefficients are assigned to the first classification model and the second classification model according to their performance indicators on the test set; The fusion module fuses the prediction results of the two classification models according to a weighted voting mechanism or a weighted probability mechanism; The performance indicators include but are not limited to the area under the ROC curve, accuracy or F1 score; The weight coefficient is calculated according to the following normalized formula: ; in, is the weight coefficient of the i-th classification model, is the area under the ROC curve of the i-th classification model, It is the sum of the areas under the ROC curves of the two classification models.

Images