Limbal Stem Cell Deficiency (LSCD) is an eye disease that can cause corneal opacity and vascularization. In its advanced stage it can lead to a degree of visual impairment. It involves the changing in the semispherical shape of the cornea to a drooping shape to downwards direction. LSCD is hard to be diagnosed at early stages. The color and texture of the cornea surface can provide significant information about the cornea affected by LSCD. Parameters such as shape and texture are very crucial to differentiate normal from LSCD cornea. Although several medical approaches exist, most of them requires complicated procedure and medical devices. Therefore, in this paper, we pursued the development of a LSCD detection technique (LDT) utilizing image processing methods. Early diagnosis of LSCD is very crucial for physicians to arrange for effective treatment. In the proposed technique, we developed a method for LSCD detection utilizing frontal eye images. A dataset of 280 eye images of frontal and lateral LSCD and normal patients were used in this research. First, the cornea region of both frontal and lateral images is segmented, and the geometric features are extracted through the automated active contour model and the spline curve. While the texture features are extracted using the feature selection algorithm. The experimental results exhibited that the combined features of the geometric and texture will exhibit accuracy of 95.95%, sensitivity of 97.91% and specificity of 94.05% with the random forest classifier of n = 40. As a result, this research developed a Limbal stem cell deficiency detection system utilizing features’ fusion using image processing techniques for frontal and lateral digital images of the eyes.
The corneal epithelium surface is always getting repopulated by limbal stem cells. If a deficiency in these stem cells occur, visual impairment can happen. This deficiency is called Limbal Stem Cell Deficiency (LSCD) [
LSCD can be caused by various and multi factors such as environmental, genetic and as by-product of medical condition (such as chemotherapy medications) factors [
LSCD is usually diagnosed by clinical manifestations. Signs of LSCD can be speckling fluorescein staining, opacity of the epithelial layers. Signs of persistent fluorescein is also found. The most apparent one that can be detected visually is the neovascularization of the cornea [
Corneal topography is one of the classical most sensitive technique for the early diagnosis of LSCD. Le et al. [
In this paper, we are proposing a method for diagnosing LSCD using image processing and machine learning techniques. Therefore, in this research we will explore the development of automated devices that utilize image processing technology with automated analysis and processing of digital image. Automated diagnosis can be held with the aid of machine learning. We are presenting a smart mobile application to aid in the diagnoses of LSCD ocular disease. In this paper we are proposing an image processing-based method by investigating the visual features of the eye with LSCD. These features are extracted from frontal segment eye images. Medical experts can conclude all important information from the geometrical features of the frontal eye segment image of LSCD-infected eyes. In this paper, we are concerned with geometric and color feature extraction to devise an automated system for LSCD diagnosis. Such features are categorized according to their significance to differentiate between LSCD and normal eyes, especially in early stages.
The rest of the paper is divided as follows. In Section 2, we are presenting a literature review of the recent methods for eye disease diagnosis through machine learning. In Section 3, we present the proposed feature extraction algorithm. In Section 4, the feature extraction algorithm and the ranked features are tested to classify the images into LSCD or normal groups. The experiments setting and the simulation results are depicted in Section 5. In Section 6, the conclusions are presented.
LSCD diagnosis is usually performed in a medical setting through diagnostic tools such as corneal topography medical devices [
Fabijańska et al. [
Twa et al. [
Smartphones were used to produce an eye image and utilize image processing techniques to capture many corneal abnormalities. Jiao et al. [
In our research, we conducted a successful methodology to diagnose LSCD through digital images from a smartphone camera. We compare our findings to those methods using topographic map imaging. Our research utilized color and texture extraction from the digital images to identify LSCD eyes from normal eyes. Our results were validated by comparing them to images for the same cases using corneal topography and have manual diagnosis by medical experts.
In this section, we will describe the methodology used and the dataset. The proposed algorithm comprises of three major steps. The first step is the preprocessing of the acquired images of the frontal and lateral views for cases with different stages of LSCD including cornea segmentation, the second step is the processing of the digital images to obtain features relevant to the disease (learning process). The third step is the classification of new cases according to the output of the learning process, as shown in
The development of the database is attained from images in the dataset in Bengio [
The dataset contains images of resolution 2048 × 1836 for each eye. The 228 images were 102 from female populations and 126 of male populations. The age range is from 22 to 45 years old.
The proposed algorithm comprises of three phases involving the frontal images of the eye. The preprocessing of the images followed by segmentation and ends with feature extraction.
Preprocessing has an important task to improve images quality. The block diagram of the preprocessing algorithm is presented in
The eyes RGB images are converted into normalized RGB encoding. We chose the RGB color space because of its ability to differentiate the color and intensity information [
Normalized RGB is calculated as depicted in
Also, we converted the normalized RGB images into HUV, where the V component is used is utilized in the preprocessing step to extract the cornea from the eye images because the V component gives better contrast between the cornea and the sclera, and can extract the cornea edges better.
We utilize the Gaussian filter of σ = 0.015 V (x, y) to yield the image symbolized by Intensity (x, y). Smoothing procedure is followed utilizing the diffusion anisotropic function (DAF) defined by Cho et al. [
The cornea images undergo segmentation process to extract color and texture as well as geometric features. The features can thus be converted into quantitative measures. The segmentation process of cornea is depicted in
We performed image normalization process to enhance the colors features in the cornea image due to degradation resultant from non-uniform illuminations. Image normalization was done utilizing Multi Scale Retinex [
The following features: Horizontal and vertical Visible Iris Diameter, curvature radius, spherical aberration and solidity are measured using our segmentation method and compared to other methods in the literature. Solidity is an important unique feature of corneal surface [
The results of the segmentation procedure of the four methods are tabulated in
Method | First doctor | Second doctor | Third doctor | Average dice similarity index |
---|---|---|---|---|
The proposed method | 0.868 | 0.867 | 0.872 | 0.87 |
Circular Hough transform | 0.622 | 0.51 | 0.523 | 0.56 |
Chanvese procedure | 0.523 | 0.55 | 0.515 | 0.53 |
Daugman procedure | 0.76 | 0.754 | 0.77 | 0.762 |
The shape, texture and color features represent the changes that can occur in the cornea due to LSCD. These features include the vertical and horizontal iris diameters. Also, Color eccentricity, cornea sphere completion shape and Cornea texture are important features. Color eccentricity is a very important feature that illustrates the course and color change of the cornea [
where,
The sphere completion shape of the front cornea exterior shows that the cornea form differs from a normal cornea versus a LSCD infected cornea.
A frontally captured image utilizing a smartphone is used in this research. The procedure consists of two stages: preprocessing stage that utilize image contrast correction, second stage of feature extraction. The LSCD cornea can be clinically detected because the cornea would be bulged slightly outward in a downward trend, which leads to a conclusion that the corneal drooping can be detected, and the differentiation can be measured in a quantitative method.
In most LSCD ailment eyes, a frail contrast is found between the cornea and sclera especially that the sclera tends to be more yellowish. Therefore, an image contrast correction method namely gamma correction will be utilized on the captured RGB converted to RGB image [
In
The shape identification of the frontal eye is done by computing texture differentiation and then the curvature is computed. The curvature is computed using an automated process which depends on third degree spline interpolation [
The proposed method probes the image in horizontally and can detect edges by the intensity differences in texture. The system changes from the Cartesian plane (x, y) to (Θ, Φ) which is polar coordinates. The origin can be detected easily which transform the edge detection problem to be straightforward. The problem is converted to an image processing one dimensional job.
The gray level intensities from 0 to 255 are converted to polar angles ranging from 0° to 360°. Therefore, curves with low to high gray levels can be obtained. The high levels are related to bright parts of the cornea texture, and low levels are related to more opaque parts of the cornea which is affected by the LCSD.
The curves should be smoothed and decrease the noise levels. In the second step we calculate the first derivative to the slopes at each point. The derivative will produce a curve with peaks characterizing the edge between dark and bright texture levels.
Color and texture analysis are used to extract iris features. Iris color palette is utilized to describe an illustration of all possible colors of the iris surface. Therefore, we defined Iris color palette to make a similar illustration of the colors on the iris surface. The iris color was used as a locus to extract color and texture features of the iris. The colors show that the iris surface can be represented by the color gamut which are not opaque in texture as opposite to iris that is affected by LSCD which is opaque in texture.
The method is based on the sensitivity of the iris texture extraction which convey important evidence that can differentiate between normal and LSCD eyes.
Texture portrayal operator utilizes Local binary pattern (LB) for classification [
a. The setting of the threshold
b. Encoding.
Threshold setting
c. In the first iteration, the pixels group of 3 × 3-pixel zone is usually set as the initial threshold value.
d. The other pixels are compared to the initial threshold value
e. Each pixel value will take a 0 or 1 binary value. If the value of the pixel is superior to the initial threshold, it will take a value of 1, otherwise, it will take the value of 0.
f. The eight pixels with values 1 or 0 will be arranged in a clockwise order not including the central pixel. The calculated value is depicted as the Local binary pattern code of the pixel at the center.
Local binary pattern is depicted in the example given in
The Local binary pattern code is a texture feature. The Local binary pattern code is characterized by its statistical histogram which is considered as the final texture feature vector that is computed from the original Local binary pattern code image.
As a summary of the preprocessing step, the image is transformed into binary image of 0 and 1 according to the threshold. The Local binary pattern code of every pixel is computed and the output would be texture feature image and the Local binary pattern histogram is depicted as the feature vector of the cornea image.
Support vector machine is an influential classification solution to many problems. The key advantages of support vector machines are the generalization capability and the strong learning method. Support vector machine usually leads to identifying global minimum.
The Support vector machine is a linear high-dimensional feature space machine [
The hyperplane is depicted in
where,
The Support vector machine’s learning process is the classification to the maximum of the margin between any two classes for separation.
It can be depicted as a procedure to minimize the cost function as defined in
Subject to:
ξi ≥ 0 is a representation of the slack variable,
p is the count of the learning pairs xi, di.
The optimization problem is solved by the Lagrangian function [
Subject to
In this research, the training of the Support vector machine’s classifier utilizes the extracted features from training set.
In the following subsections, we present two algorithms to compute the curvature and characterize the ratio features of the horizontal and vertical diameter of the iris as the change in Cornea curvature will increase these measures. We utilize ratio computation and the numerical integration computation method [
For ratio computation, the initialization process is first normalized by extracting the farthest point of the cornea curve followed by calculating the reference ratio, R. the reference ratio is defined as the ratio of the width, w, to the width of the image,
The required ratio is computed as follows:
The ratio “Ratio” is a characteristic of the droopiness of the cornea downwards and is based on the fact that the smaller the ratio is, the more is the dropping of the eye cornea downwards as shown in the following
We computed the ratio of the length to the width of the curve for a normal eye on the left and for an LSCD eye where the cornea is more dropped on the right.
The features from the curve extraction, are fused to select the most preeminent features. The selection methodology in the suggested system aids the evaluation of the impact of the implemented corneal feature selection based on the ranking of the features method. The adopted corneal features are sphericity, eccentricity, horizontal and vertical diameters. The adopted features are measured from the cornea utilizing the ratio calculation, and the numerical integration technique.
The features are selected to attain stability, scalability and minimize feature variability for classification purposes. The data size is the main reason for the instability or stability of the feature selection. Data size which is small with a high dimensionality can induce feature instability. Therefore, we select and rank the features accordingly. The most challenging problem is choosing the optimal features to form the feature model. In this research, Rclassify, latent feature selection algorithm methods will be utilized to define the impertinence of the features.
The proposed classification algorithm: Rclassify, which is an extension to the Relief algorithm. The major drawback of Relief algorithm is its limitation to binary classification and also its jeopardizing of losing data [
Classification methodology represents the class-predicting procedure for the data in question. In this research, the classification of LSCD eyes versus normal eyes are based on extracted and ranked features. The classification utilizes the feature selection methods. How well the classification performs is the actual representation of the performance of the whole system. It also characterizes the appropriateness of the data type and size. In this research, the classification study takes account of SVM classifier including linear, quadratic, and radial basis function as shown in
Average results for SVM kernels | ||||
---|---|---|---|---|
Kernel type | Accuracy % | Sensitivity % | Specificity % | Error rate |
Linear | 91.2 | 87.9 | 96.1 | 0.096 |
Quadratic | 85.1 | 80.22 | 90.1 | 0.1433 |
Cubic | 83.6 | 79.1 | 87.9 | 0.1678 |
Radial Basis Function | 95.43 | 91.9 | 96.99 | 0.0483 |
Average results for number of trees for the random forest classifier | ||||
---|---|---|---|---|
Number of the trees in the forest | Accuracy % | Sensitivity % | Specificity % | Error rate |
10 | 95.2 | 88.9 | 99.7 | 0.056 |
40 | 91.4 | 90.22 | 95.1 | 0.0733 |
200 | 91.6 | 89.1 | 95.9 | 0.088 |
400 | 91.4 | 90.22 | 95.1 | 0.0733 |
Average results for distance function | ||||
---|---|---|---|---|
Function | K = 5 | K = 7 | K = 9 | |
Euclidian | Accuracy % | 89.9 | 89.7 | 97.7 |
Sensitivity % | 70.22 | 75.1 | 92.1 | |
Specificity % | 99.9 | 99.9 | 99.9 | |
Error rate | 11.22 | 10.1 | 3.51 | |
Chebychev | Accuracy % | 89.9 | 93.7 | 91.7 |
Sensitivity % | 78.12 | 71.3 | 85.5 | |
Specificity % | 100 | 100 | 100 | |
Error Rate | 10.22 | 14.1 | 7.321 |
The ranked features are: <L5, L7, L6, L1, L4, L9, L8, A6, A7, A4, A1,A2, L3, A5, L2, A3> and are utilized to optimize the parameters for each classifier. The performance comparison among the algorithms are the accuracy, sensitivity, specificity, and error.
The experimental results in
The simulation results, of the Computing Ratio that indicate the diagnosis of infected eyes with LSCD, are presented in this subsection. The experiments are designed such that the correlation between the RC of the LSCD infected eyes and the Actual Ratio from an actual physician diagnosis, are extracted.
Simulation results presented in
The Bland-Altman Plot, of the actual Ratio of LSCD in patients from an actual medical expert diagnosis versus the Computation Ratio of LSCD computed using our proposed algorithm, is depicted in
The correlation between the Actual Ratio and the computation Ratio, as plotted in
The simulation results depict resilient strong linear relationship between the physician Actual Ratio and the predicted Computation Ratio diagnosis.
Positive correlation is indicated using Pearson’s r formula as presented in
“r” was calculated to be equal 0.925, which denotes high correlation between the Actual Ratio of LSCD in patients from an actual medical expert diagnosis versus the Computation Ratio of LSCD computed by our algorithm (utilizing the data presented in
Bland-Altman metric usually denotes the positive correlation of two quantities. Bland-Altman plot is used constantly in correlating the predicted medical diagnosis (P) with actual diagnosis (D). The presented Bland-Altman between A and D in our case found that the values of the two tests are of highly similar.
Results: We compared the diagnosis done by a medical expert for 228 cases. Their medical evaluation was found fully in the dataset:
Comparison of the experimental results of actual medical diagnosis with our proposed method as shown in
Predicted cases by our proposed classifier | Total | True positive | False negative | |||||
---|---|---|---|---|---|---|---|---|
Mild | Moderate | Severe | Normal | |||||
Actual Cases diagnosed | Mild | 96 | 2 | 0 | 22 | 120 | 80% | 20% |
Moderate | 0 | 46 | 2 | 2 | 50 | 92% | 0% | |
Severe | 0 | 0 | 20 | 0 | 20 | 100% | 0% | |
Normal | 8 | 2 | 0 | 28 | 38 | 73% | 21% |
We presented the development of an automated LSCD detection technique (LDT) approach using frontal and lateral eye image views captured by smartphones. Very few studies of LSCD automated diagnosis are found in the literature. Our study is one of the initial an automated LSCD detection techniques that utilize image processing techniques. A 3-step preprocessing and the cornea region segmentation algorithms were proposed. We developed a feature extraction algorithm from the training data set.
We developed a method for LSCD detection technique utilizing frontal eye images. A dataset of 280 eye images of frontal and lateral LSCD and normal patients were used in this research. First, the cornea region of both frontal and lateral images is segmented and the texture features are extracted using the feature selection algorithm. The experimental results exhibited that the combined features of the geometric and texture will exhibit accuracy of 95.95%, sensitivity of 97.91% and specificity of 94.05% with the random forest classifier of n = 40. As a result, this research developed a Limbal Stem Cell Deficiency detection system utilizing features’ fusion using image processing techniques for frontal and lateral digital images of the eyes.