CN-122024302-A - Method for extracting image histology characteristics of cornea profile

CN122024302ACN 122024302 ACN122024302 ACN 122024302ACN-122024302-A

Abstract

A method for extracting image histology features of cornea profile includes extracting normalized region of interest (ROI) from cornea profile by image alignment, vertex positioning and adaptive clipping. And carrying out standard image histology feature extraction on the ROI image in parallel, and carrying out four types of custom feature extraction of multi-scale Gabor texture, HOG morphology, FFT frequency domain and fractal dimension optimized for cornea structure to generate a high-dimensional feature vector. Based on a machine learning strategy, the optimal feature subset is screened out sequentially through univariate pre-screening and Recursive Feature Elimination Cross Verification (RFECV), and importance sorting and diagnosis effectiveness verification are carried out. The invention solves the problems of hidden characteristics and insufficient sensitivity of traditional macroscopic parameters in the early diagnosis of keratoconus, realizes the excavation of high-dimensional microscopic characteristics from conventional images, and provides an effective auxiliary means for objective and accurate diagnosis of cornea diseases, especially early keratoconus.

Inventors

MI SHENGLI
WANG YAN
LI LONGQING
HUO YAN
Xie Ruisi
LIU YIYONG

Assignees

清华大学深圳国际研究生院
南开大学

Dates

Publication Date: 20260512
Application Date: 20260126

Claims (10)

1. A method for extracting a corneal profile image histology feature, comprising the steps of: S1, extracting a self-adaptive region of interest, namely acquiring a cornea profile image, and extracting a standardized region of interest image from the cornea profile image through pretreatment, cornea vertex positioning and self-adaptive region clipping; S2, optimized multi-dimensional feature extraction, namely, carrying out standard image histology feature extraction and custom feature extraction optimized for cornea structure on the standardized region image in parallel to generate a high-dimensional feature vector, wherein the standard image histology feature extraction comprises the steps of calculating multi-class features after carrying out multi-domain transformation on the image, and the custom feature extraction comprises the steps of calculating multi-scale Gabor texture features, gradient direction histogram morphological features, fast Fourier transformation frequency domain features and fractal dimension complexity features; and S3, feature screening and verification based on machine learning, namely sequentially performing univariate pre-screening and recursive feature elimination cross verification on the high-dimensional feature vector, screening out an optimal feature subset, and performing importance ranking and diagnosis effectiveness verification on features in the optimal feature subset.
2. The method for extracting corneal profiling image histology according to claim 1, wherein in step S1, the adaptive region of interest extraction specifically comprises: the method comprises the steps of preprocessing, namely carrying out rotary transformation on an input image according to the main structure direction of the input image to realize gesture normalization alignment, and converting the aligned image into a gray image, carrying out Gaussian convolution smoothing denoising on the gray image, obtaining an edge image by utilizing an edge detection operator, and carrying out morphological closing operation on the edge image to connect edge break points to obtain an enhanced edge image; Vertex positioning, namely extracting all contours from the enhanced edge map, screening out a preset number of candidate contours according to the contour arc length, selecting a point with the minimum longitudinal coordinate value from points on all the candidate contours in a horizontal center area of the image, and determining the point as a cornea vertex; the method comprises the steps of self-adaptive region clipping, namely taking the ordinate of a cornea vertex as an upper boundary of a region of interest, taking the difference value between the abscissa of the cornea vertex and a preset horizontal margin as a left boundary of the region of interest, taking twice the preset horizontal margin as the width of the region of interest, defining the horizontal boundary of the region of interest, calculating the average pixel intensity of each row in a vertical region corresponding to the horizontal boundary in a gray level image, dynamically determining a brightness threshold according to the overall brightness of the region, defining the row with the first average pixel intensity lower than the brightness threshold and the ordinate larger than the ordinate of the cornea vertex as a lower boundary of the cornea, determining the final height of the region of interest in a preset physiological height range based on the lower boundary of the cornea, and further extracting the standardized region of interest image.
3. The method for extracting the image histology features of the cornea section according to claim 1 or 2, wherein in the step S2, the standard image histology feature extraction specifically comprises the steps of applying a plurality of image transformations including wavelet decomposition and laplacian filtering to the standardized region of interest image to generate a set of transformed images, calculating multi-class features including shape features, first-order statistical features and texture features for the original image and each transformed image, respectively, and merging all calculation results into a standard image histology feature set.
4. The method for extracting corneal profiling image histology features according to claim 1 or 2, wherein in step S2, the custom feature extraction for corneal structure optimization specifically comprises: convolving Gabor kernels with different directions and frequencies with the standardized region of interest image, and calculating energy, mean value and variance for each convolution response graph to form a Gabor texture feature set; extracting a gradient direction histogram descriptor of the standardized region-of-interest image, and calculating the mean value, variance and shannon entropy of the descriptor to form a morphological feature set; Converting the standardized region of interest image into a frequency domain, and calculating the energy duty ratio and spectrum centroid of different radial frequency bands to form a frequency domain feature set; And calculating the fractal dimension of the standardized region of interest image as a characteristic of the complexity of the quantized structure.
5. The method of claim 1, wherein in step S2, the standard image histology feature extraction and the custom feature extraction are performed under a multi-process parallel computing framework, and the generated feature vectors are saved using an incremental strategy.
6. The method for extracting corneal profiling image histology features according to claim 1, wherein in step S3, the univariate pre-screening is specifically implemented by calculating an analysis of variance F statistic between each feature in the high-dimensional feature vector and a disease class label as a score, sorting and selecting a predetermined number of features with highest rank according to the score, and forming a pre-screened feature set.
7. The method for extracting the corneal profile image histology features according to claim 1, wherein in the step S3, the recursive feature elimination cross-validation is specifically that, starting from the feature set after the pre-screening, the following process is iteratively performed, namely training a classification model by using a current feature set and evaluating the cross-validation performance of the classification model, calculating an importance score of each feature in the current feature set, eliminating the feature with the lowest importance score, updating the feature set and entering the next iteration, and selecting a feature subset which enables the model cross-validation performance to be optimal as the optimal feature subset by recording the model performance of each iteration.
8. The method for extracting corneal profiling image histology features according to claim 1, wherein in step S3, the importance ranking and diagnosis effectiveness verification is specifically implemented by training a final classification model using the optimal feature subset, extracting importance scores of features in the model for ranking, and simultaneously quantifying the diagnosis effectiveness of the optimal feature subset with a highest cross-validation performance score obtained by the final classification model.
9. The method for extracting the image histology features of the cornea profile as set forth in claim 1, further comprising the step of S4, feature validity visual verification, wherein the optimal feature subset obtained in the step of S3 and importance ranking thereof, and the systematic visual display and verification are performed by generating at least one of a feature quantity-model performance relation curve, a feature importance bar graph, a distributed violin graph with feature values among different disease categories, and a dimension reduction scatter graph of a high-dimensional feature space.
10. A computer program product comprising a computer program which, when executed by a processor, implements a method of extracting a histology feature of a corneal profile image as claimed in any one of claims 1 to 9.

Description

Method for extracting image histology characteristics of cornea profile Technical Field The invention relates to a medical image processing technology, in particular to a method for extracting the image histology characteristics of a cornea section view. Background Keratoconus is a progressive corneal disorder characterized by a dilated deformation of the cornea and a conical anterior protuberance, which can lead to the appearance of highly irregular astigmatism in the patient. The progressive nature of the disease means that patient vision will gradually deteriorate over time, and some patients may also develop acute corneal edema, further exacerbating visual dysfunction, and ultimately restoring vision only by receiving a corneal transplant procedure. The four phases of keratoconus progress include "normal eye (NL)", "setback" (FFKC) "," Subclinical Keratoconus (SKC) ", and" classical cone (KC) ". Among them, early diagnosis of keratoconus presents a greater challenge because (1) the secrecy of early lesion features is that part FFKC patients have slight abnormalities in mechanical parameters, and early keratoconus morphology and biomechanical differential attenuation are easily ignored by doctors. (2) The limitations of the existing index are that the gray zone exists, and even if the instrument detects that the biomechanical parameters of the cornea are weak, the physiological characteristics of the patient (the cornea is softer) are not excluded, and the early lesions are true. The lack of an exact characterization method is not clear as to which features are subject to the variability (texture, morphology, mechanics) in the early stages of keratoconus progression. (3) The hysteresis of diagnosis relies on "follow-up observations", which are currently common practice to require periodic review by suspicious patients, with "post" confirmation by observing whether they will develop further. The classical definition of FFKC itself relies on the fact that the contralateral eye has been diagnosed as KC, which is clinically "hardly visible" for patients with both eyes in an early state. Current corneal image analysis relies primarily on a variety of imaging techniques, each of which has its specific application and limitations (1) corneal topography analysis by analysis of corneal anterior and posterior surface morphological parameters such as corneal curvature (K-value), inferior-superior (I-S) value, etc. Clinically, keratoconus diagnosis typically requires that the mean corneal curvature (K) >47 diopters or I-S value >1.4 diopters be met. However, this approach is not sufficiently sensitive to sub-clinical lesions and is difficult to detect small changes. (2) Scheimpflug imaging providing a tomographic image of the cornea, the corneal thickness profile and the posterior surface elevation parameters can be assessed. Research shows that the diagnosis of subclinical keratoconus needs to integrate multiple parameters such as the surface topography of the front and back of the cornea, cornea biomechanics and the like. However, this approach presents challenges in terms of image quality, and the low signal-to-noise ratio region affects the accuracy of the corneal profile extraction. (3) Optical Coherence Tomography (OCT) technology OCT is capable of acquiring high resolution cross-sectional images of the cornea, but image analysis presents difficulties. Manually analyzing OCT images is time consuming, laborious, subjective and poorly reproducible. In swept OCT systems, telecentric scan modes can cause artifacts in the cornea images, partial structure loss, and the like. (4) And (3) cornea biomechanics analysis, namely exciting a central area of the cornea by using an air nozzle, then extracting a motion video of the cornea after being excited, and extracting key representative parameters to represent the mechanical properties of the cornea. The main limitation of the existing methods is that they rely on a limited set of parameters defined manually, and the deep information contained in the cornea image cannot be fully mined. For early lesions, the changes in these parameters are small and not significant enough, resulting in insufficient diagnostic sensitivity. Image histology is a method for extracting and analyzing sub-visual quantitative characteristics based on medical images, and tumor heterogeneity is quantified through modeling, so that important guiding significance is shown in the aspect of accurate diagnosis and treatment of bone and soft tissue tumors. The method has the core value of extracting tiny features which cannot be identified by human eyes from conventional medical images and realizing more accurate disease diagnosis and classification. Although image histology has been successful in other medical fields (such as tumor diagnosis), no mature study has been found to apply the image histology feature to keratoconus diagnosis. In addition, existing corneal image analysis methods typically use