Glaucoma diagnosis based on colour and spatial features using Kernel SVM

Anisha Rebinth, S. Mohan Kumar. Glaucoma diagnosis based on colour and spatial features using kernel SVM. Car-diometry; Issue 22; May 2022; p. 508-515; DOI: 10.18137/ cardiometry.2022.22.508515; Available from: http://www.

The damage to the optic nerve due to intraocular pressure leads to glaucoma and it is the second leading cause of vision loss globally. As glaucoma is irreversible, a computerized analysis system is needed to detect glaucoma at the earliest. A CAD system for glaucoma diagnosis is discussed in [1] based on clustering techniques. In pre-processing anisotropic diffusion filter is utilized for contrast enhancement and then the ROI is extracted. The Cup-to-Disc Ratio (CDR) is computed from the segmented region of OD and Optic Cup (OC). OD and OC are segmented by two clustering approaches; K- Means Clustering (KMC) and fuzzy C-means with hill climbing algorithm.

Convolutional Neural Networks (CNN) with Transfer Learning (TL) is described in [2] for automatic glaucoma classification. In the preprocessing stage, the fundus images with different format and sizes are processed into a same size and same standard format. Six different CNN architectures such as VGG19, ResNet50, GoogleNet etc., are used. All models are trained using Stochastic Gradient Descent (SGD) algorithm and uses binary cross entropy loss function. The data augmentation technique is used to overcome the overfitting problem.

Automatic glaucoma fundus image assessment using CNN network model with extensive validation is presented in [3]. The images used for training the networks are obtained by cropping the image around the OD region automatically. Five CNNs are employed for the classification including VGG16 and VGG19. In order to optimize the network model for glaucoma diagnosis, the SGD optimizer is used. Finally k-fold cross-validation technique is used to evaluate models. CAD based neural network using micro-statistical descriptors is discussed in [4] for glaucoma diagnosis. The first module selects the ROI around the OD region (cupping region) from the glaucoma fundus images. The second module extracts the micro-statistical estimators such as level, edge, ripple etc., using the laws kernels. Finally, the linear vector quantizer-artificial neural network classifier module is used to classify the fundus images.

Discrete Orthogonal Stockwell Transform (DOST) is discussed in [5] for automatic glaucoma classification. In preprocessing, ROI is extracted based on the maximum gray value in the green channel. Then the features are extracted from the ROI using DOST. The classification of the system employs the Machine Learning (ML) approach with the tree structured random forest classifier. ML based glaucoma diagnosis using color fundus images are described in [6]. ROI is selected manually and the green channel is utilized for the classification after normalization. VGG19 architecture with TL is employed for the classification that uses SGD optimization, dropout and data augmentation. Along with fundus image, the RNFL thickness and deviation maps, macular GCC thickness and deviation maps are also utilized independently for the classification. Final classification is obtained by a RF classifier.

Local Binary Patterns (LBP) based glaucoma detection is described in [7] using red free fundus images. The red free fundus image is obtained from the mean of green and blue channel images. Along with the classical LBP, features from the completed modelling of LBP are also utilized for the classification using K-Nearest Neighbor (KNN) [19]. The complex texture features are discussed in [8] for fundus image classification. At First, the ROI is extracted from the green channel. Features such as Gray Level Co-occurrence Matrix (GLCM) and LBP are extracted directly from the ROI and also from the optical density image of ROI. Finally, SVM with Radial Basis Function (RBF) kernel is used for the classification of glaucoma fundus images using the hybrid features.

GLCM and Markov Random Field (MRF) based glaucoma diagnosis is explained in [9]. In the preprocessing stage, the red channel is separated from the RGB fundus images and contrast enhancement is achieved by adaptive histogram equalization. OD is segmented by using k-means clustering, morphological and Hough transformation. Then, the textural features based on MRF and GLCM are extracted. Finally, SVM classifier is used Different computer vision techniques; matched filters, texture analysis, multiscale approaches, machine learning techniques and morphological methodologies are comprehensively reviewed in [10] for the identification and classification of glaucoma. Many supervised and unsupervised techniques for the localization of OD and OC are discussed along with model based, edge based and threshold based approaches.

Bit-Plane Slicing (BPS) and LBP techniques based automated glaucoma diagnosis is discussed in [11]. At first, the input RGB fundus images are separated into

Red, Blue and Green color images individually. Second, these individual channels are split into bit planes and LBP based statistical features are extracted. Third, the features obtained from three channels are given as inputs to three different SVM classifiers. Finally the decision level fusion is employed for the classification.

The automated digital image processing techniques for glaucoma diagnosis are reviewed in [12]. Different approaches to localize OD and OC region, blood vessel extraction along with different feature extraction techniques from the whole image are reviewed. Retinal image processing and SVM classification based glaucoma detection is discussed in [13]. After ROI extraction, histogram equalization and adaptive median filters are employed to enhance the contrast and eliminate noise. To extract the OD and OC region, colour component analysis and morphological operations are used. The quantitative features such as CDR and rim-to-disc ratio are computed and fed to SVM for classification.

2    Methods and Materials

An efficient CAD classification system aided with three stages; preprocessing, features extraction, and kernel-based classification is proposed. In preprocessing, ROI is selected automatically based on the intensity of the green channel. Followed by feature extraction using LBP and conventional colour based features from the ROI. Finally, classification step is done for the detection of the normal or glaucomatous images. Figure 1 shows the model for glaucomatous image classification.

2.1    Pre-Processing

It is well known that the OD region in the fundus images is the brightest region and thus this property is utilized for initial segmentation. Initially, the colour components of input fundus image such as Red, Green and Blue (RGB) are separated. Then, only the green channel is utilized to locate the OD region [1]. Based on the green channel’s highest intensity, ROI is extracted from the colour fundus image. Figure 2 shows the selected ROI from the input fundus images.

2.2    Features used

In the feature extraction stage, LBP [14] and the conventional colour based features are extracted from the ROI shown in Figure 2 (b). The computation of LBP is as follows:

Figure 1 Model for Glaucoma Diagnosis

(a)

(b)

Figure 2. (a) Fundus image (b) ROI image

LBP = E 7=0 2nT(imgn - imgc), where c in the sub-script denoted the center pixel and n represents the neighbourhood pixels. The definition for T (threshold function) is defined by

'1 if i n — i c >  0

0 if i n - i c <  0

For each colour component in the ROI, Eqns. (1) and (2) are applied to extract the LBP for glaucoma diagnosis. Figure 3 shows how the LBP is computed for a 3x3 input.

The conventional colour features such as mean, standard deviation, skewness and kurtosis are also extracted from each colour component. Table 1 shows the extracted colour features from each colour components.

The extracted colour based features and LBP features are analyzed independently for glaucoma diagnosis using SVM classifier [20]. The SVM classifier with RBF kernel is discussed in the next section.

ear problems can be solved by applying the kernel trick which makes the non-linearity into linearity in higher dimensional feature space. Figure 4 shows the hyperplane for a two class-problem where normal healthy eye fundus images will belong to one class (say positive class, red colored dots) and glaucoma affected eye fundus images will belong to another class (say negative class, blue colored dots).

RBF kernel is used in this study to learn mappings in supervised manner to move from data space to a high dimensional feature space. To perform a nonlinear mapping of the dataset into some high dimensional feature space in which linear operations are performed. Since the original mapping is nonlinear, any linear operation in this feature space corresponds to a nonlinear operation in data space. In machine learning, RBF kernel is commonly used in SVM classification. The RBF kernel definition is given as,

K ( x , y ) = exp

II ^{x -} y||²

2 о ²

where is the standard deviation.

2.3    Kernel-based SVM classifier

SVM classifier calculates a hyperplane that separates the feature space with maximum margin [15]. Though the basic SVM is a binary linear classifier, it can be used for multiclass problems. Also, the non-lin-

3 . Results and Discussions

The performance of colour and spatial features based glaucomatous image classification system is evaluated using ORIGA [16], RIM-ONE[17], DRISH-

654	111	1x20      1x21      1x22
512	11	1x27                1x23	223
303	101	1x26     0x25     1x24
(a) Input	(b) After thersholding ( T ) by	(c) Application of Eqn. (1)	(d) Computed LBP
	Eqn. (2)

Figure 3 LBP computation

Table 1

Colour based Features

Mean	Standard deviation
nn h = - EE ROI ( i , j ) n i = 1 j = 1	о = 1	nn - EE ( ROI ( i , j ) -h )2 n , = 1 j = 1
Skewness	Kurtosis
nn - EE ( pi . -h )3 <         " i = 1 j = 1 __________________ 3/ n_n_          /2 1 EE ( pi . -h )2 < n i = i j = 1             ;	s =- V	nn 1 EE ( pi . -h )4 n . = 1 j = 1 3 X 2    3 nn 1 EE ( pi . -h )2 n . = 1 j = 1                J

Figure 4 SVM -hyperplane

TI-GS1 [18] and HRF [21] database images. Also, k-fold cross validation with k =10 is used to validate the model. Figure 5 shows the samples images in the databases. Figure 6 shows the k -fold approach.

To evaluate the performance metrics such as accuracy, specificity and sensitivity, update T_P and F_N if glaucoma image is correctly classified and incorrectly classified respectively while testing each fold. Similarly, update T_N and F_P if normal image is correctly classified and incorrectly classified. Finally, the performance of the system for glaucoma diagnosis is evaluated based on these parameters. Table 2 shows the performance metrics used.

(a)                                                                   (b)

(c)                                                                   (d)

Figure 5 Sample images in the database (a) HRF (b) RIM-ONE (c) (b) DRISHTI-GS1 (d) ORIGA

Figure 6 k-fold technique in CV

Table 2

Performance metrics

Sensitivity	Specificity	Accuracy
T P	T N	T P + T N
T P + F N	T N + F P	T P + F P + T N + F N

Table 3 shows the performances of the glaucomatous image classification system using SVM-RBF classifier by colour and LBP features.

It can be seen from Table 3 that the LBP features outperforms the colour features to classify the abnormal fundus images form glaucoma diagnosis. For HRF database images, LBP based system provides 70% accuracy whereas colour features based system gives only 56.67%. Similarly, approximately 9% improvement can be seen for DRISHTI-GS1 database images by LBP (79.21%) over the colour features (70.30%). Also, it is observed that the performance of the colour based system for RIM-ONE and ORIGA database im- ages are 67.18% and 77.85% whereas it is 75.57% and 80.77% for LBP based system. Figure 6 plots the performance of the fundus image classification for glaucoma diagnosis using colour and LBP features. It can be seen from Figure 7 that the overall accuracy and sensitivity of the LBP based system is higher than the colour features based system.

However, the colour based system gives highest specificity than LBP which means that more normal images are correctly classified by the colour features than LBP based system.

4 . Conclusions

In this paper, a CAD system is developed for glaucoma diagnosis using colour and LBP based features. The performances of the colour based and LBP based systems are validated using 10-fold cross-validation with three important performance metrics; accuracy, specificity and sensitivity. Initially, the color components such as Red, Green, and Blue are separated and

Table 3

Performances of the glaucoma image classification system

Performance metrics	Features	Database
		HRF	DRISHTI-GS1	RIM-ONE	ORIGA
Accuracy	Colour	56.67	70.30	67.18	77.85
	LBP	70.00	79.21	75.57	80.77
Sensitivity	Colour	53.33	74.29	58.97	70.24
	LBP	66.67	82.86	71.79	77.38
Specificity	Colour	73.33	70.97	77.17	81.95
	LBP	60.00	61.29	70.65	80.50

• ■    Accuracy-Colour

• ■    Sensitivity-LBP

• ■    Accuracy-LBP

• ■    Specificity-Colour

• ■    Sensitivity-Colour

• ■    Specificity-LBP

Figure 7 Performance of the system for glaucoma diagnosis using colour and LBP features

then the ROI is obtained. Features are extracted from only the ROI area from each component. Finally, the glaucoma classification is done by using the RBF kernel based SVM classification system. Results show that the efficacy of LBP based system is better than colour based features for all four databases; ORIGA, RIM-ONE, DRISHTI-GS, and HRF used in this study.

Statement on ethical issues

Research involving people and/or animals is in full compliance with current national and international ethical standards.

Conflict of interest

None declared.

Author contributions

The authors read the ICMJE criteria for authorship and approved the final manuscript.