CN119941729B - Keratoconus analysis system based on dynamic cornea texture - Google Patents

Abstract

The invention discloses a keratoconus analysis system based on dynamic cornea texture, which comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit which are sequentially connected, and provides a multi-time-point cornea dynamic image texture analysis method which lays a foundation for high-precision identification of FFKC, wherein the screened features have optimal diagnostic capability, and feature sets with the most diagnostic value are screened by a recursive feature elimination method. The machine learning classification system comprises a segmentation module, a model training module, a model testing module and a performance evaluation module, wherein the model testing module is used for obtaining a machine learning classifier, and the model evaluates the performance of the model through an ROC curve and a confusion matrix to ensure high sensitivity and high specificity. A multi-mode identification platform integrating dynamic textures and cornea biomechanics parameters is constructed, and the platform sequentially comprises a data input module, a data processing module, a model module and an output module, so that efficient FFKC identification and auxiliary diagnosis are realized.

Description

Keratoconus analysis system based on dynamic cornea texture

Technical Field

The invention relates to the technical field of medical treatment, in particular to a keratoconus analysis system based on dynamic cornea texture.

Background

Keratoconus (KC) is a progressive keratolytic disorder characterized by progressive thinning and dilation of the cornea, resulting in irregular astigmatism and significant vision loss, severely affecting the quality of life of the patient, and even blindness if not treated in time. Researches show that the global prevalence of keratoconus is about 1/2000 to 1/1000, which is particularly common among young people and brings great burden to society and economy. With the rapid spread of refractive surgery, the amount of surgery has proliferated, with some patients developing iatrogenic keratoectasis complications after surgery, and failure to timely identify latent keratoconus is considered to be one of the main causes of its occurrence.

The contusion keratoconus (FFKC) is defined as a normal contralateral eye of a patient whose eye has been clinically diagnosed as keratoconus, which is examined normal at a slit lamp and no corneal topography abnormalities are detected. According to the global expert consensus of keratoconus 2015, "true monocular keratoconus is absent", and FFKC is currently free of clinical symptoms and signs, but there is a significant risk of developing keratoconus. Studies have shown that although the etiology of keratoconus is not completely defined, the current consensus considers its occurrence and development to be closely related to changes in regional biomechanical properties of the cornea. Even though FFKC has not found any signs of abnormalities clinically, subtle biomechanical changes may have begun to occur inside the cornea.

FFKC's early identification is of great importance 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 crosslinking, and on the other hand, accurate identification FFKC can avoid high risk patients from undergoing corneal refractive surgery that may induce corneal dilation. However, since FFKC is a subtle abnormal change, its detection and diagnosis remain a significant challenge.

The prior art relies primarily on morphological and biomechanical assessment. Wherein the cornea morphology assessment device is mainly Pentacam and Optical Coherence Tomography (OCT). Such methods identify abnormalities by measuring static parameters such as corneal curvature, thickness, etc. The method comprises the steps of corneal topography analysis, corneal percentage index (KISA%) (A/D) and wavefront aberration analysis, wherein the corneal topography analysis is used for assessing front surface morphology and providing curvature distribution and asymmetry information, the keratoconus percentage index (KISA%) (A/D) is used for integrating parameters such as corneal curvature asymmetry and central curvature value, the wavefront aberration analysis is used for measuring higher-order aberration and assessing optical quality change, and the OCT imaging is used for assessing the integrity and regularity of each layer structure, particularly Bowman layer. These techniques mainly focus on corneal morphology features and assist diagnosis by identifying morphologically abnormal conditions that are diagnostic. However, since the early stage morphological changes of FFKC are extremely small, the existing morphological parameters have limited sensitivity to their detection, and there is insufficient sensitivity and specificity to capture the early stage minute changes.

It can be seen that the current FFKC diagnostic methods rely primarily on morphological analysis and simple biomechanical parameters, and have various limitations. Morphology-based methods focus only on corneal surface features, lack sensitivity to changes in internal structures of the cornea, whereas FFKC lesions are caused by internal collagen fiber disorders, first appearing as internal microscopic lesions that change, while existing morphological parameters have not been markedly abnormal. The traditional biomechanical parameters provided by the visual cornea biomechanical analyzer (Corvis ST, CVS) enable assessment of cornea dynamic response, but are primarily affected by central cornea thickness and intraocular pressure, and are primarily global parameters, lacking in the ability to finely analyze local changes to the cornea, resulting in limited diagnostic accuracy.

In recent years, with the advancement of medical imaging technology, accurate assessment of corneal morphology and biomechanical properties has become possible. CVS is used as a novel cornea biomechanical evaluation device, and a novel method for quantifying cornea biomechanical characteristics is provided by exciting cornea deformation through airflow pulse and recording dynamic response of the cornea deformation. Meanwhile, artificial Intelligence (AI) technology is increasingly used in the field of medical diagnostics, providing new possibilities for the processing and analysis of complex medical data. Image histology (Radiomics) has shown great potential in oncology and other medical fields as a technique for extracting high-dimensional quantitative features from medical images, and currently, image histology still has challenges in identifying early biomechanical abnormalities in FFKC, lacking a stable, interpretable modeling approach.

In this context, how to realize the analysis of FFKC based on the cornea dynamic deformation data is a technical problem to be solved by the present application.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a keratoconus analysis system based on dynamic cornea texture, which combines CVS, advanced image group analysis and machine learning algorithm to extract abundant texture features from cornea deformation images at a plurality of time points, thereby being capable of deeply analyzing the cornea internal structure and dynamic deformation behavior and being beneficial to the diagnosis of the contusive keratoconus.

On one hand, the invention provides a keratoconus analysis system based on dynamic cornea texture, which has the following technical scheme:

a keratoconus analysis system based on dynamic cornea texture comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit which are sequentially connected, wherein a machine learning classifier is arranged in the feature screening unit.

The image acquisition unit acquires an initial moment image, a first flattening moment image and a maximum deformation moment image of the cornea through the visual cornea biomechanics analyzer.

The image preprocessing unit extracts regions of interest (ROIs, region of interest) for each time instant image, respectively.

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 screens feature subsets by a recursive feature elimination method on texture features of images at all moments through a machine learning classifier.

In a second aspect, the present invention provides a machine learning classification system, which has the following technical scheme:

the machine learning classification system comprises a segmentation module, a model training module, a model testing module and a performance evaluation module which are connected in sequence.

A data set comprising normal eye samples and a pause type keratoconus eye sample is obtained, and the data set is divided into a training set and a testing set by a data segmentation module.

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

The model test module predicts the test set by using a trained machine learning classifier, so as to obtain a prediction result.

The performance evaluation module calculates the performance index according to the prediction result, so that the machine learning classifier with the best performance index is obtained.

In a third aspect, the invention provides a keratoconus analysis platform based on fusion of dynamic cornea textures and biomechanical parameters, which has the following technical scheme:

the system comprises a data input layer, a data processing layer, a model layer and an output layer which are sequentially connected.

The image acquisition unit is arranged on the data input layer and is used for receiving cornea dynamic images acquired by the visual cornea biomechanics analyzer and calculating biomechanics parameters, and the image acquisition unit acquires initial moment images, first flattening moment images and maximum deformation moment images of the cornea from the cornea dynamic images and transmits the initial moment images, the first flattening moment images and the maximum deformation moment images to the data input layer.

The image preprocessing unit and the feature extraction unit are arranged on the data processing layer, the data processing layer further comprises a biomechanical parameter processing module for processing biomechanical parameters, and the data processing layer is used for extracting radiological texture features from dynamic images of the cornea and carrying out standardized processing on the biomechanical parameters.

The feature screening unit is arranged on 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 used for fusing the prediction results of the first classification model and the second classification model.

And the output layer outputs a keratoconus diagnosis result, a key feature analysis result and a visual diagnosis report according to the prediction result of the fusion module.

In summary, the technical scheme has the advantages that the keratoconus analysis system based on dynamic cornea texture can remarkably improve the early detection rate of FFKC, reduce the misdiagnosis rate and provide a reliable and noninvasive diagnosis tool for clinicians. By early identifying the high risk patient with cornea expansion, serious complications after cornea refractive surgery are avoided to a great extent, and clinical decisions are optimized. Meanwhile, the simplified flow of a single device reduces the examination burden of a patient, improves the screening efficiency, provides scientific basis for early intervention of FFKC and establishment of a personalized treatment scheme, and has important significance for improving the diagnosis and treatment level of cornea diseases and the visual health of the patient.

Drawings

FIG. 1 is a schematic image of a visual corneal biomechanical analyzer capturing three key points in time of the cornea;

FIG. 2 is a schematic diagram of feature matrix output after feature extraction and normalization of all features at the initial moment;

FIG. 3 is a schematic representation of ROC curves and confusion matrices for 3 different machine learning models;

fig. 4 is a schematic representation of ROC curves of prior art CVS biomechanics parameters.

Detailed Description

The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the drawings in the embodiments, and it is apparent that the described embodiments are only some embodiments of the present invention, not all embodiments. All other embodiments, which can be made by those skilled in the art based on the embodiments of the invention without making any inventive effort, are intended to be within the scope of the invention.

The invention aims to solve the key problems of insufficient sensitivity, low specificity, dependence on multiple equipment comprehensive evaluation and the like in the early diagnosis of the contusion keratoconus (FFKC). The existing FFKC diagnosis method mainly depends on cornea morphology and cornea biomechanics examination, and lacks deep analysis on cornea internal structure and dynamic deformation process, so that the early detection rate is low, the misdiagnosis rate is high, the optimal treatment time is delayed, and the risks of serious complications such as cornea dilatation and the like after refractive surgery are increased. In addition, the current diagnosis process is complicated, a plurality of examination devices and techniques are often combined, and the clinical experience of doctors is relied on, so that the burden of patients is increased, the clinical efficiency is reduced, and the wide development of FFKC early screening is limited.

Studies have shown that corneal biomechanical changes play a key role in FFKC progression, early lesions first appear as changes in local biomechanical properties, rather than as distinct morphological abnormalities. The dynamic response process of the cornea when subjected to external forces (e.g., air pulses) implies rich biomechanical information that is of great value for early FFKC identification. However, the conventional diagnostic method only focuses on the corneal surface deformation to extract limited biomechanical parameters, and cannot fully capture the subtle changes and dynamic response characteristics of deep tissues of the cornea, so that the sensitivity to early lesions is insufficient, particularly at the early stage FFKC of normal morphology but biomechanical changes.

Aiming at the problems, the invention provides a keratoconus analysis system based on dynamic cornea texture, which combines a visual cornea biomechanics analyzer (Corvis ST, CVS), an advanced image group analysis and a machine learning algorithm to extract abundant texture features from cornea deformation images at a plurality of time points and deeply analyze cornea internal structure and dynamic deformation behaviors. The invention screens the most diagnostic features by a recursive feature elimination method, builds a FFKC intelligent identification model with high accuracy, and realizes the accurate capture of the change of the fine biomechanics. In addition, the invention fuses time domain information, simultaneously analyzes cornea characteristics at an initial time, a first flattening time and a maximum deformation time, comprehensively evaluates cornea biomechanical characteristics, changes from empirical diagnosis to accurate quantitative diagnosis, improves sensitivity and specificity of early screening, and simplifies diagnosis flow.

According to a first aspect of the invention, a keratoconus analysis system based on dynamic cornea texture is provided, which comprises an image acquisition unit, an image preprocessing unit, a feature extraction unit and a feature screening unit which are sequentially connected, wherein a machine learning classifier is arranged in the feature screening unit.

As shown in fig. 1, the image acquisition unit acquires an initial time image, a first applanation time image, and a maximum deformation time image of the cornea through a visual cornea biomechanical analyzer. The visual cornea biomechanics analyzer adopts standardized airflow pulse to record the dynamic deformation process of cornea under the action of external force. The method comprises the specific steps of fixing the head of a subject on a device bracket, adjusting the device to the center of the cornea of the subject and focusing, starting the device, applying standard airflow pulse, recording the cornea deformation process, acquiring a high-speed shooting image sequence, and extracting images of three key time points from the image sequence, namely an initial time, a first flattening time and a maximum deformation time.

The image preprocessing unit extracts the region of interest (ROI, region of interest) from each time image respectively, the extraction of the region of interest is realized by obtaining gray images of each time according to each time image, establishing a coordinate system, automatically detecting the front and rear surfaces of the cornea through a visual cornea biomechanical analyzer and obtaining coordinates, and dividing the cornea region image by adopting an edge detection and curve fitting method, namely removing the noise of the cornea region image and enhancing the contrast, thereby obtaining the region of interest of each time image. The image preprocessing unit realizes preprocessing steps by using matlab codes, and the original 8-bit gray level image of the visual cornea biomechanical analyzer is analyzed to ensure the consistency of data formats. Specifically, edge detection is the detection of the edge of the cornea by the Canny edge detection algorithm or Sobel operator. And (3) adopting polynomial fitting to smooth cornea boundary, removing extra noise by morphological opening and closing operation, automatically cutting cornea region, and outputting standardized region of interest.

As shown in FIG. 2, the feature extraction unit extracts a plurality of texture features from the region of interest of the image at each time, wherein the texture features comprise first-order statistical features, shape features, gray level co-occurrence matrix features, gray level run length matrix features, gray level size region matrix features, gray level dependency matrix features, neighborhood gray level difference matrix features and wavelet transformation features, and normalizes the texture features at three different times and outputs a feature matrix. Specifically, the Feature extraction unit extracts multidimensional Texture features from the region of interest image by using a radiometric analysis technology, firstly extracts symbolic textures from the region of interest image, uses Texture to represent the extracted Texture features, and then uses Feature to represent the extracted Texture features from the symbolic textures. Further, eight types of texture features extracted from each graph comprise 464 features, and a graph of 3 times of one person is 1392 feature elements.

The feature screening unit screens feature subsets by a recursive feature elimination method on texture features of images at all moments through a machine learning classifier. The process of screening the feature subsets by adopting a recursive feature elimination method comprises the steps of inputting a feature matrix and label vectors of corresponding samples in a feature screening unit, initializing a machine learning classifier, calculating importance scores of features through the machine learning classifier, arranging all texture features in descending order of importance, setting a texture feature quantity range, performing five-fold cross validation to determine optimal feature quantity, and outputting the screened texture features, so that the feature subset is obtained. Specifically, the feature subset has 51 feature elements, and the 1392 features of the 3-person time chart are screened by the recursive feature elimination method, so that only 51 feature elements are screened.

The keratoconus analysis system based on dynamic cornea texture has the beneficial effects that the analysis system provided by the invention is different from traditional static morphological analysis or single-time-point biomechanical parameter evaluation, and a multi-time-point cornea dynamic image texture analysis method is provided. And extracting comprehensive texture features from the image by adopting an image histology method, wherein the comprehensive texture features comprise first-order statistical features, shape features, gray level co-occurrence matrixes (GLCM), gray Level Dependent Matrixes (GLDM), gray Level Run Length Matrixes (GLRLM), gray level size area matrixes (GLSZM), neighborhood gray level difference matrixes (NGTDM) and wavelet transformation features, and comprehensively evaluating dynamic changes of the cornea microstructure. In order to ensure that the screened features have optimal diagnosis capability, a feature set with the most diagnosis value is screened by a recursive feature elimination method (RFE), and a foundation is laid for FFKC high-precision identification.

The keratoconus analysis system based on dynamic cornea texture can remarkably improve the early detection rate of FFKC, reduce the misdiagnosis rate and provide a reliable and noninvasive diagnosis tool for clinicians. By early identifying the high risk patient with cornea expansion, serious complications after cornea refractive surgery are avoided to a great extent, and clinical decisions are optimized. Meanwhile, the simplified flow of a single device reduces the examination burden of a patient, improves the screening efficiency, provides scientific basis for early intervention of FFKC and establishment of a personalized treatment scheme, and has important significance for improving the diagnosis and treatment level of cornea diseases and the visual health of the patient.

Instead of the image histology feature extraction technique, the original preferred method of the invention uses image histology to carry out FFKC diagnosis by extracting the texture features of the cornea image. Alternatively, deep learning techniques, particularly Convolutional Neural Networks (CNNs), may be employed to automatically learn features directly from the cornea images. The method does not need to manually design features, but automatically extracts deep features of the cornea image through multi-layer rolling and pooling operation, and is suitable for large-scale data sets and complex image analysis scenes. In addition, the quantitative ultrasonic elastography technology can also be used as an alternative scheme for evaluating the cornea elasticity by measuring the propagation characteristics of sound waves in cornea tissues, indirectly reflecting the biomechanical change of the cornea, and is suitable for clinical environments requiring non-contact evaluation.

Instead of the recursive feature elimination method, the original preferred method of the invention uses the recursive feature elimination method to screen the texture features with the most diagnostic value. Alternatively, principal Component Analysis (PCA) techniques may be employed to rapidly extract the principal features by retaining the data maximum variance information in a dimension-reduction manner. Another alternative is feature importance assessment based on tree models, such as ExtraTrees or LightGBM, which evaluate feature importance by building multiple decision trees, suitable for rapid screening of high-dimensional feature space. In addition, LASSO regression can be adopted to realize synchronous performance of feature selection and model training, important features are automatically screened through L1 regularization, and the method is suitable for the condition that the number of features is far greater than that of samples.

Instead of a multi-time point integration strategy, the invention preferably integrates the corneal image features at three time points. Alternatively, a time series analysis method may be used, and the whole cornea deformation process is regarded as a continuous time series, and dynamic change characteristics such as deformation rate, rebound characteristics and the like are extracted. Another alternative is to use a attentive mechanism to assign different weights to the features at different time points, automatically learn the time point with the most diagnostic value, and adapt to the situation of unbalanced contribution of the time points. The displacement field of each pixel in the cornea deformation process can be tracked by adopting an optical flow method, and the local strain distribution is calculated so as to reflect the biomechanical characteristics of the cornea more directly.

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

The machine learning classification system comprises a segmentation module, a model training module, a model testing module and a performance evaluation module which are connected in sequence;

a data set comprising normal eye samples and a pause type keratoconus eye sample is obtained, and the data set is divided into a training set and a testing set by a data segmentation module. The sample is obtained by a visual cornea biomechanical analyzer, which can provide various biomechanical parameters including a first applanation time (A1T) for measuring the time of the cornea to reach a first applanation state, ambr% by weight of the cornea thickness (ARTh) for calculating the spatial distribution ratio of the cornea thickness, a stress-strain index (SSI) for reflecting the hardness of the tissue, a stiffness parameter (SP-A1) at the first applanation time for evaluating the deformation resistance, a deformation amplitude ratio (DARatio 2) for measuring the deformation amplitude ratio at the vertex and the periphery of 2mm, a comprehensive reciprocal radius (IIR) for evaluating the central deformation curvature, and a CVS Biomechanical Index (CBI) as a multiparameter comprehensive score. By quantifying the external force response characteristics, these parameters evaluate the biomechanical properties of the corneal tissue, such as elasticity, stiffness, and viscoelasticity.

The model training module builds and optimizes a plurality of machine learning classifiers based on the training set, wherein the building and optimizing of the machine learning classifiers is realized by inputting a feature matrix and a label vector of each sample of the training set in the model training module, building random forest classifiers, C5.0 decision tree classifiers, extreme gradient lifting classifiers and the like, optimizing parameters of each classifier by adopting five-fold cross validation, determining optimal parameters according to the five-fold cross validation result, retraining and outputting the trained classifiers by using the optimal parameters and the training set, thereby improving the recognition capability of the model to FFKC;

The model test module predicts the test set by using a trained machine learning classifier to obtain a prediction result, wherein the prediction of the test set comprises the steps of inputting a feature matrix of each sample of the test set and the trained classifier in the model test module, classifying each sample by the trained classifier to output the prediction result, comparing the prediction result with a label vector of a corresponding sample to obtain a prediction probability value, and outputting a confusion matrix.

The performance evaluation module calculates the performance index according to the prediction result, so that the machine learning classifier with the best performance index is obtained. The machine learning classifier with the best performance index is obtained through the following processes that a performance evaluation module calculates an ROC curve according to the prediction probability value and the confusion matrix and is used for evaluating the machine learning classifier, and a visual chart is drawn according to the prediction probability value, the confusion matrix and the ROC curve and a performance evaluation report is output, so that the machine learning classifier with the best performance index is obtained. The machine learning classifier with the best performance index is obtained by inputting a confusion matrix and a prediction probability value into a performance evaluation module, calculating accuracy, sensitivity, specificity, a positive prediction value, a negative prediction value and an ROC curve, drawing a visual chart and outputting a performance evaluation report. As shown in fig. 3, ab is the ROC curve of the random forest model and the confusion matrix with test set, cd is XGBoost model, and ef is c5.0 model, respectively. The ROC curve of the prior art CVS biomechanics parameters is shown in fig. 4.

The machine learning classification system has the beneficial effects that on the basis of texture feature screening, a machine learning classifier integrating multi-time-point information is constructed, and the machine learning classifier comprises algorithms such as Random Forest (RF), C5.0, extreme gradient lifting (XGBoost) and the like. And optimizing model parameters through five-fold cross validation, so that the recognition capability of the model to FFKC is improved. The model outputs FFKC recognition results and probability distributions, quantifies diagnostic confidence, and evaluates model performance via ROC curve (Receiver Operating Characteristic Curve) and confusion matrix (Confusion Matrix), ensuring high sensitivity and high specificity.

According to a third aspect of the invention, a keratoconus analysis platform based on fusion of dynamic cornea texture and biomechanical parameters is provided, and is used for accessing a keratoconus analysis system based on dynamic cornea texture as an expansion module of multi-modal feature integration.

The system comprises a data input layer, a data processing layer, a model layer and an output layer which are sequentially connected.

The image acquisition unit is arranged on the data input layer, the data input layer is used for receiving cornea dynamic images and calculated biomechanical parameters acquired by the visual cornea biomechanical analyzer, and the image acquisition unit acquires initial moment images, first flattening moment images and maximum deformation moment images of the cornea from the cornea dynamic images and transmits the images to the data input layer.

The image preprocessing unit and the feature extraction unit are arranged on the data processing layer, the data processing layer further comprises a biomechanical parameter processing module for processing biomechanical parameters, and the data processing layer is used for extracting radiological texture features from dynamic images of the cornea and carrying out standardized processing on the biomechanical parameters.

The feature screening unit is arranged on 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 used for fusing the prediction results of the first classification model and the second classification model. The input biomechanical parameters comprise stress-strain index, improved version of stress-strain index, rigidity parameter at the first flattening moment, ambr degrees si o relative thickness, deformation amplitude ratio, CVS biomechanical index and the like, and the data input layer is also used for inputting and managing basic information of patients and the like. Specifically, the first classification model is a random forest model constructed based on texture features, the second classification model is a Linear Discriminant Analysis (LDA) model constructed based on cornea biomechanical parameters, weight coefficients are distributed to the first classification model and the second classification model according to performance indexes of the first classification model and the second classification model on a test set, the fusion module fuses prediction results of the two classification models according to a weighted voting mechanism or a weighted probability mechanism, the performance indexes comprise but are not limited to area under ROC curves, accuracy or F1 fraction, and the weight coefficients are calculated according to the following normalization formula;For the weight coefficient of the ith classification model,The area under the ROC curve for the ith classification model,Is the sum of the areas under the ROC curves of the two classification models. Preferably, the two machine learning classifiers are random forest classifiers based on radiological features and linear discriminant analysis classifiers based on radiological features, and the fusion module adopts a weighted voting mechanism to set weight coefficients according to the relative performance of the model. The F1 score is a harmonic mean of Precision and Recall, and is used for measuring the comprehensive performance of the model on positive class prediction.

Instead of the random forest model, a variety of machine learning algorithms may be employed as alternatives to the random forest model. The Support Vector Machine (SVM) realizes high-precision classification by searching an optimal classification hyperplane, the deep neural network can process more complex nonlinear relations, the integrated learning method such as AdaBoost, stacking and the like can be used for fusing the advantages of a plurality of classifiers, and the Bayesian network is suitable for probability reasoning and uncertainty quantification and provides risk assessment for clinical decisions.

And the output layer outputs a keratoconus diagnosis result, a key feature analysis result and a visual diagnosis report according to the prediction result of the fusion module. The output layer output content comprises FFKC analysis results, risk assessment, key feature analysis, visual reports, clinical advice and the like. The platform supports either a local deployment or cloud service mode. The system workflow comprises data acquisition, preprocessing, feature extraction, model prediction and result output. The platform adopts a modularized design, is convenient for maintenance and upgrading, supports continuous learning of new data, and ensures continuous optimization of analysis performance.

The keratoconus analysis platform based on the fusion of dynamic cornea textures and biomechanical parameters has the beneficial effects that a FFKC comprehensive analysis platform is established by combining CVS texture analysis results with traditional biomechanical parameters. 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, so that the limitation of a single-mode method is effectively overcome. The analysis accuracy of the clinical data verification system is utilized, the algorithm model is continuously optimized, and the clinical applicability is improved. Finally, the transformation from surface morphology analysis to deep tissue characteristic analysis is realized, a scientific and accurate FFKC early analysis flow is established, and objective basis is provided for clinical decision making and personalized treatment scheme formulation.

In summary, the beneficial effects of the application include 1, the application applies the image histology technology to the cornea dynamic deformation image analysis for the first time, and fills the blank of the existing FFKC analysis technology. By extracting deep texture features from the CVS image, the application can capture the micro cornea internal structure change which cannot be identified by the traditional parameters, greatly improves the early detection rate of FFKC and reduces the missed diagnosis risk.

2. According to the invention, CVS is adopted to acquire high-resolution images of the cornea at three key moments, namely an initial moment, a first flattening moment and a maximum deformation moment, a classifier integrated at multiple time points is established, texture features in the cornea are extracted through an image histology technology, the limitation of traditional static morphological analysis or global biomechanical parameter evaluation is broken through, the limitation that only a single time point or a static form is concerned in the prior art is broken through, experiments prove that a multiple time point model is obviously superior to a single time point model, and more comprehensive evaluation of the dynamic deformation process of the cornea is provided.

3. The invention applies an image group analysis method to extract eight types of characteristics from cornea images, namely, a first-order statistical characteristic, a shape characteristic, a gray level co-occurrence matrix (GLCM) characteristic, a Gray Level Run Length Matrix (GLRLM) characteristic, a gray level size area matrix (GLSZM) characteristic, a Gray Level Dependence Matrix (GLDM) characteristic, a neighborhood gray level difference matrix (NGTDM) characteristic and a wavelet transformation characteristic, and can comprehensively capture the tiny change of the cornea structure.

4. The invention screens the feature subset with the most analysis value by a Recursive Feature Elimination (RFE) method, reduces the number of features from 1,392 to 51, obviously improves algorithm efficiency and generalization capability, and ensures analysis accuracy.

5. The invention constructs a random forest classifier integrated at multiple time points based on the screened characteristics, realizes high-precision identification of FFKC by integrating information of three key deformation moments of cornea, and has AUC reaching 0.989 which is obviously superior to the existing evaluation parameters (SP-A1 parameters, AUC is only 0.728).

5. By combining artificial intelligence and radiohistology, FFKC intelligent screening is realized, the conventional FFKC screening technology for promoting accurate medical development is based on the empirical judgment of fixed threshold values, and the diagnosis result is greatly influenced by the subjective effect of an operator. The invention combines the image histology and the machine learning method to realize intelligent FFKC diagnosis based on data driving, reduce human intervention and improve consistency. The sustainable and optimized AI model, along with the continuous improvement of algorithm performance of data accumulation, promotes FFKC diagnosis to develop to the accurate medical direction. 6. The analysis platform is simple and convenient to operate and easy to popularize, and is suitable for large-scale clinical screening, the traditional FFKC detection method possibly needs to be combined with various devices (such as Pentacam + Corvis ST +OCT), the detection process is complex, and the cost is high. The invention can complete detection by only one CVS device, reduces the examination time and economic burden of patients, improves the screening efficiency, makes the early screening of FFKC more feasible, is suitable for pre-operation FFKC risk screening, and reduces the complications of cornea dilation after refractive surgery. The clinical operation threshold is reduced, the promotion of basic medical institutions is facilitated, and the FFKC early diagnosis rate is improved.

Based on the previous study, the project group of the application establishes FFKC a cornea dynamic image of a diagnosis model and a patient database, and completes a complete process from data acquisition to model verification. The study included a sufficient number of normal eyes and FFKC eyes, acquired high resolution images using CVS, and appropriately assigned training and test sets.

By means of image histology feature extraction and recursive feature elimination, we screened the most diagnostically valuable texture features from images at three time points per eye. The random forest model has excellent performance on a test set, and the diagnosis accuracy is obviously higher than that of the traditional CVS parameter. The result shows that the performance of the multi-time-point model is obviously superior to that of the single-time-point model, and the superiority of multi-time-point information integration is verified. In addition, fusion of texture features with conventional biomechanical parameters further improves diagnostic efficacy. Analysis of the texture features during the deformation of the cornea shows that there are significant differences in the texture features of the normal eye and FFKC eye at three time points, particularly at the moment of maximum deformation. This shows that FFKC has no obvious abnormality in morphology, but the response mode of the internal structure after stress has been changed, confirming the ability of the invention to capture micro structural changes. The clinical application preliminary result shows that the invention not only improves the detection rate of FFKC, but also reduces the time required by the diagnosis process, and particularly has great application value in the screening before refractive surgery. The early recognition capability of the model to the high-risk group is beneficial to preventing postoperative cornea dilation complications and improving the safety of refractive surgery.

In general, experimental and clinical preliminary application results fully prove the feasibility and effectiveness of the invention, provide a new technical scheme for FFKC early and accurate diagnosis, and have wide clinical application prospects.

The above is only a preferred embodiment of the present invention, and the protection scope of the present invention is not limited to the above examples, and all technical solutions belonging to the concept of the present invention belong to the protection scope of the present invention. It should be noted that modifications and adaptations to the present invention may occur to one skilled in the art without departing from the principles of the present invention and are intended to be within the scope of the present invention.

Claims (9)

1. The keratoconus analysis system based on dynamic cornea texture is characterized by comprising 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;

Applying airflow pulse to cornea, and collecting an initial moment image, a first flattening moment image and a maximum deformation moment image of cornea by the image collecting unit through a visual cornea biomechanics analyzer;

the image preprocessing unit extracts the interested areas from the images at each moment respectively;

The feature extraction unit is used for respectively extracting a plurality of texture features from the interested region of the image at each moment;

the feature screening unit screens feature subsets by adopting a recursive feature elimination method on texture features of images at all moments through a machine learning classifier;

normalizing the texture features at three different moments and outputting a feature matrix;

inputting a feature matrix and a label vector of a corresponding sample in the feature screening unit;

After initializing a machine learning classifier, calculating importance scores of the features through the machine learning classifier, and arranging all texture features in descending order of importance;

Setting a texture feature quantity range, performing five-fold cross validation to determine the optimal feature quantity, and outputting the screened texture features, so as to obtain a feature subset;

The method comprises the steps of receiving biomechanical parameters calculated by a visual cornea biomechanical analyzer, constructing a first classification model based on texture features, constructing a second classification model based on biomechanical parameters, fusing the prediction results of the first classification model and the second classification model, and outputting keratoconus diagnosis results.

2. The keratoconus analysis system based on dynamic corneal texturing according to claim 1, wherein the extraction of the region of interest is achieved by:

obtaining gray images at each moment according to the images at each moment;

Establishing a coordinate system, automatically detecting the front and rear surfaces of the cornea through a visual cornea biomechanics analyzer and acquiring coordinates;

Dividing the cornea region image by adopting an edge detection and curve fitting method:

the image noise of the cornea region is removed and the contrast is enhanced, so that the interested region of the image at each moment is obtained.

3. The keratoconus analysis system of claim 1, wherein the texture features include first-order statistics features, shape features, gray level co-occurrence matrix features, gray level run length matrix features, gray level size area matrix features, gray level dependency matrix features, neighborhood gray level difference matrix features, and wavelet transform features.

4. A machine learning classification system for obtaining the machine learning classifier of any one of claims 1-3, the machine learning classification system comprising a segmentation module, a model training module, a model testing module, and a performance evaluation module connected in sequence;

acquiring a data set comprising a normal eye sample and a pause type keratoconus eye sample, wherein the segmentation module divides the data set into a training set and a testing set;

The model training module builds and optimizes a plurality of machine learning classifiers based on a training set;

the model test module predicts the test set by using a trained machine learning classifier, so as to obtain a prediction result;

And the performance evaluation module calculates the performance index according to the prediction result, so as to obtain the machine learning classifier with the best performance index.

5. The machine learning classification system of claim 4 wherein constructing and optimizing a plurality of machine learning classifiers is accomplished by:

inputting a feature matrix and a label vector of each sample of the training set in the model training module;

optimizing parameters of each classifier by adopting five-fold cross validation, determining optimal parameters according to the five-fold cross validation result, retraining by using the optimal parameters and the training set, and outputting the trained classifier.

6. The machine learning classification system of claim 5 wherein predicting a test set comprises:

inputting a feature matrix and a trained classifier of each sample of the test set into the model test module;

classifying each sample by the trained classifier, and outputting a prediction result;

and comparing the prediction result with the label vector of the corresponding sample to obtain a prediction probability value, and outputting a confusion matrix.

7. The machine learning categorization system of claim 6, wherein the machine learning categorizer that achieves the best performance index is implemented by:

the performance evaluation module calculates an ROC curve according to the prediction probability value and the confusion matrix and is used for evaluating the machine learning classifier;

And drawing a visual chart according to the predicted probability value, the confusion matrix and the ROC curve, and outputting a performance evaluation report, thereby obtaining the machine learning classifier with the best performance index.

8. A keratoconus analysis platform based on fusion of dynamic cornea texture and biomechanical parameters for accessing the keratoconus analysis system based on dynamic cornea texture as claimed in any one of claims 1-3,

The system comprises a data input layer, a data processing layer, a model layer and an output layer which are sequentially connected;

The image acquisition unit is arranged on the data input layer and is used for receiving cornea dynamic images acquired by the visual cornea biomechanics analyzer and calculating biomechanics parameters, and the image acquisition unit acquires initial moment images, first flattening moment images and maximum deformation moment images of the cornea from the cornea dynamic images and then transmits the initial moment images, the first flattening moment images and the maximum deformation moment images to the data input layer;

the image preprocessing unit and the feature extraction unit are arranged on the data processing layer, the data processing layer also comprises a biomechanical parameter processing module for processing biomechanical parameters, and the data processing layer is used for extracting radiological texture features from dynamic images of the cornea and carrying out standardized processing on the biomechanical parameters;

The feature screening unit is arranged on 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 used for fusing the prediction results of the first classification model and the second classification model;

And the output layer outputs a keratoconus diagnosis result, a key feature analysis result and a visual diagnosis report according to the prediction result of the fusion module.

9. The keratoconus analysis platform based on dynamic cornea texture and biomechanical parameter fusion according to claim 8, wherein the first classification model is a random forest model constructed based on texture features, the second classification model is a linear discriminant analysis model constructed based on cornea biomechanical parameters, and weight coefficients are assigned to the first classification model and the second classification model according to their performance indexes on a 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, area under ROC curve, accuracy, or F1 fraction;

The weight coefficient is calculated according to the following normalization formula: ;

Wherein, the For the weight coefficient of the ith classification model,The area under the ROC curve for the ith classification model,Is the sum of the areas under the ROC curves of the two classification models.