Data mining technique for automated diagnosis of glaucoma using higher order spectra and wavelet energy features

Abstract

Eye images provide an insight into important parts of the visual system, and also indicate the health of the entire human body. Glaucoma is one of the most common causes of blindness. It is a disease in which fluid pressure in the eye increases gradually, damaging the optic nerve and causing vision loss. Robust mass screening may help to extend the symptom-free life for the affected patients. The retinal optic nerve fiber layer can be assessed using optical coherence tomography, scanning laser polarimetry (SLP), and Heidelberg Retina Tomography (HRT) scanning methods. These methods are expensive and hence a novel low cost automated glaucoma diagnosis system using digital fundus images is proposed. The paper discusses the system for the automated identification of normal and glaucoma classes using Higher Order Spectra (HOS) and Discrete Wavelet Transform (DWT) features. The extracted features are fed to the Support Vector Machine (SVM) classifier with linear, polynomial order 1, 2, 3 and Radial Basis Function (RBF) to select the best kernel function for automated decision making. In this work, SVM classifier with kernel function of polynomial order 2 was able to identify the glaucoma and normal images automatically with an accuracy of 95%, sensitivity and specificity of 93.33% and 96.67% respectively. Finally, we have proposed a novel integrated index called Glaucoma Risk Index (GRI) which is made up of HOS and DWT features, to diagnose the unknown class using a single feature. We hope that this GRI will aid clinicians to make a faster glaucoma diagnosis during the mass screening of normal/glaucoma images.

Introduction

Glaucoma is one of the most common causes of blindness with a mean prevalence of 2.4% for all ages and 4.7% for ages above 75 years [11]. It is estimated that more than 4 million Americans suffer from glaucoma, and half of them are unaware that they have it. Approximately 120,000 of Americans are blind as a result of glaucoma, thus accounting for 9–12% of all cases of blindness in the US. About 2% of the population between 40 and 50 years old and 8% over 70 years old have elevated intraocular pressure (IOP) [63], which increases their risk of significant vision loss and even blindness [50]. Glaucoma leads to (i) structural changes of the optic nerve head (ONH) and the nerve fiber layer and (ii) simultaneous functional failure of the visual field. Glaucoma is generally diagnosed based on the patient’s medical history, IOP, visual field loss tests and the manual assessment of the ONH via ophthalmoscopy or stereo fundus imaging [37]. The structural changes are manifested by a slowly diminishing neuroretinal rim indicating a degeneration of axons and astrocytes of the optic nerve.

Preventive medicine and mass screening of patients can reduce the cost involved in the surgical treatment (mean trabeculectomy, laser surgery, drainage implants) of end stage glaucoma and prevent from disease progression [13]. Significant loss of the optic nerve fibers lead to irreversible vision impairment, making glaucoma one of the most common causes of blindness [43]. The increase in intraocular pressure due to a blockage of the outflow of aqueous humor damages the optic nerve which is transmitting information from the retina to the brain [1]. The disease is characterized by progressive degeneration of optic nerve fibers and astrocytes presenting a distinct pathogenetic image of the ONH.

Glaucoma is commonly diagnosed by examination of the ONH [43] using ophthalmoscopy and assessment of the visual field [3]. However, visual field defects may only appear after damage to the optic disk. Various image processing techniques were applied on stereoscopic and non-stereoscopic optic nerve images from glaucoma patients. This method extracted the diameter, area and volume of the optic cup and disk, to evaluate any deformation [43].

During eye examinations, ophthalmologists examine specific regions and identify possible markers of diseases [27]. Various algorithms have been used to identify typical features such as abnormality of blood vessels [48] and ONH [17], [28], location and quantification of microaneurisms or drusen [58], [46]. Digital fundus images are widely used for the automated detection of diabetic retinopathy stages [61], [40], however, visual inspection needs to be conducted in glaucoma evaluation [43]. There are a number of techniques available for the identification of glaucoma based on image segmentation [51], [6], [41]. The drawback of these segmentation-based techniques is that small errors in localization and/or delineation may lead to significant errors in the measurements and may affect the diagnosis outcome.

Qualitative assessment of the ONH using stereo photographs (ONHSPs), confocal scanning laser ophthalmoscopy (CSLO), scanning laser polarimetry (SLP), and optical coherence tomography (OCT) were used to distinguish normal eyes from those with early-to-moderate glaucomatous visual field defects [25]. Receiver operating characteristic (ROC) curves were generated from discriminant analysis of CSLO, SLP, and OCT measurements and from ONHSP scores to test the performance. It was found that the quantitative methods CSLO, SLP, and OCT did not perform better than the qualitative assessment of disk ONHSPs conducted by observers experienced at distinguishing normal eyes from those with early-to-moderate glaucoma. They proposed that the combination of the imaging methods might significantly improve this capability. The geometric parameters that change due to glaucoma disease are optic disk diameter, optic disk area, cup diameter, rim area, and mean cup depth [25].

Nowadays glaucoma is generally diagnosed using HRT images. Swindale et al. [51] and Adler et al. [6] have modelled a smooth two-dimensional surface that fitted to the ONH of topography images. Damages in the glaucomatous eye were detected using optic disk measures (cup and disk area, height variation using HRT images) [55]. This global shape approach was compared with a sector-based analysis by Iester et al. [32]. Zangwill et al. [62] have automatically diagnosed glaucoma using optic disk parameters, additional parapapillary parameters and SVM classifier. It was shown that the detection of glaucoma by separately applied, shape-based methods on different modalities (CSLO, SLP, OCT) did not perform better than qualitative assessment of the optic disk by ophthalmologists [25]. Most of these shape approaches assumed a valid segmentation of the optic disk. However, a small error in these segmentation based techniques may result in significant change in the measurements and error in the diagnosis.

Texture is an important widely used feature for analyzing medical images [8], [52], [44], [53], [54]. In ultrasound medical images [3], [49] peripapillary chorioretinal atrophy (PPA) is considered as one of the glaucoma risk factors. It can be identified as bright regions in retinal fundus images, and therefore, incorrectly included as the part of the optic disk regions during the automated disk detection scheme.

Recent studies show that higher order spectra concept was used to diagnose the epilepsy using electroencephalogram (EEG) signals [19], sleep stages [2] and cardiac abnormalities using heart rate signals. Higher order spectra invariants have been used for shape recognition [14] and to identify different kinds of eye diseases [5], [3]. HOS is a non-linear method helps to capture the subtle changes in pixels of the image, which can be used as features for the automated classification. It can be used as a powerful tool for the non-linear dynamical analysis of the biomedical signals. It was observed that HOS techniques would be a better approach than traditional time domain and frequency domain methods in analyzing the bio-signals. It performs better even for weak and noisy signals. HOS is useful in detecting non-linear coupling, deviation from Gaussianity and features derived from it can be made invariant to shift, rotation and amplification [20]. These features can be explored for numerous healthcare applications.

Fig. 1 shows the block diagram of the proposed system. During the pre-processing stage, colored image is converted into the grayscale image and contrast of the image is increased by histogram equalization. Radon transformation is used to convert the 2D image into 1D signal. Then the important features namely, phase entropy, bispectrum entropy using HOS and average energy of wavelet coefficients are extracted from the image.

Further, statistically significant features are evaluated from the extracted features using independent sample t-test. Subsequently, these features are fed to SVM classifiers for automated diagnosis. We have devised an integrated index (Glaucoma Risk Index, GRI), from the extracted features which can be used as an adjunct tool for the diagnosis of normal/glaucoma classes. The layout of the paper is as follows. Section 2 explains the image acquisition process. Section 3 discusses the pre-processing and extraction of the features using HOS and wavelet. A brief description of the SVM classifier is discussed in Section 4. In Section 5, we proposed a new integrated index called glaucoma risk index. Section 6 of the paper presents the results of the proposed method. The results are discussed in Section 7 and finally, the paper concludes in Section 8.