Detection of glaucomatous change based on vessel shape analysis

Abstract

Glaucoma, a leading cause of blindness worldwide, is a progressive optic neuropathy with characteristic structural changes in the optic nerve head and concomitant visual field defects. Ocular hypertension (i.e. elevated intraocular pressure without glaucoma) is the most important risk factor to develop glaucoma. Even though a number of variables, including various optic disc and visual field parameters, have been used in order to identify early glaucomatous damage, there is a need for computer-based methods that can detect early glaucomatous progression so that treatment to prevent further progression can be initiated. This paper is focused on the description of a system based on image processing and classification techniques for the estimation of quantitative parameters to define vessel deformation and the classification of image data into two classes: patients with ocular hypertension who develop glaucomatous damage and patients with ocular hypertension who remain stable. The proposed system consists of the retinal image preprocessing module for vessel central axis segmentation, the automatic retinal image registration module based on a novel application of self organizing maps (SOMs) to define automatic point correspondence, the retinal vessel attributes calculation module to select the vessel shape attributes and the data classification module, using an artificial neural network classifier, to perform the necessary subject classification. Implementation of the system to optic disc data from 127 subjects obtained by a fundus camera at regular intervals provided a classification rate of 87.5%, underscoring the value of the proposed system to assist in the detection of early glaucomatous change.

Introduction

Primary open-angle glaucoma is the most common type of glaucoma, which can be defined as a progressive optic neuropathy with characteristic morphological changes at the optic nerve head with associated visual field defects. Ocular hypertension (i.e. elevated intraocular pressure without glaucoma) is the most important risk factor to develop glaucoma. The diagnosis of glaucoma is based on the combination of abnormal optic disc cupping with corresponding visual field defects [1]. Since optic disc change usually precedes visual field defects [2], the focus of interest has shifted towards the quantification of optic disc parameters that can be used for early detection of glaucomatous change. Although subjective analysis of optic disc stereo photographs is still considered the gold standard to detect and monitor glaucoma [3], several objective and quantitative image processing methods could be used towards glaucoma detection based on the segmentation and analysis of vessels within the optic disc.

In the bibliography, there is a considerable number of research works focused on the detection of retinal vessels within the optic disk and the estimation of specific vessel measurements that reveal information for the state of various diseases, such as diabetes and hypertension. These methods for vessel segmentation in images of the retina can be divided into two different groups: the rule-based methods and the supervised methods. The rule-based methods comprise vessel tracking [4], [5], [6], matched filter responses [7], [8], [9], grouping of edge pixels [10], model based locally adaptive thresholding [11], topology adaptive snakes [12] and morphology-based techniques [13], [14]. The supervised methods are ridge-based [15], neural-based [16] and wavelet-based feature extraction [17]. In this paper, we present a system for the estimation of retinal vessel quantitative parameters and the classification of retinal data towards glaucoma detection after the application of a modified vessel segmentation method. The proposed segmentation method falls within the rule-based methods since there was not ground truth generated by the medical experts for the available glaucoma data which is required as a scenario within the supervised methods. Furthermore, supervised approaches to segmentation require part of the available dataset to be used for training to tune parameters result in reduction of the acquired dataset size used in this study.

For the early detection of glaucoma, techniques such as confocal laser ophthalmoscopy, optical coherence tomography, and scanning laser polarimetry are widely used for the discrimination of individuals with glaucoma and those without. In Ref. [18], the optic nerve head obtained with confocal scanning laser ophthalmoscope has been automatically analyzed for normal volunteers and patients with glaucomatous visual field defects. The analysis was based on the modeling of the optic nerve head by a smooth two-dimensional surface with a shape described by 10 free parameters, including the degree of surface curvature of the disc region surrounding the cup, the steepness of the cup walls, the goodness-of-fit of the model to the image in the cup region, and measures of cup width and cup depth. These parameters, plus others derived from the image using the proposed model as a basis, were used to discriminate between normal and glaucomatous images achieving overall classification accuracy up to 92%. In Ref. [19], image processing techniques have been applied to analyze stereochronoscopical optic nerve images acquired at two separate points in time with and without glaucomatous optic disc progression. Initially, an image alignment scheme is applied based on the manually placement of corresponding points in each image pairs and the application of a non-rigid polynomial warping algorithm to warp the later image into registration with the earlier image. Changes in vessels position and optic disc have been evaluated between glaucomatous patients and normal controls from independent observers, by viewing the registered images on a monitor, which allows rapid alternation (flickered rendering) of the two images. These evaluations were based on decisions from the observers and they were not quantitatively recorded.

A digital stereovision system for visualizing the topography of the optic nerve head from stereo optic disc images was also developed [20]. The system uses a semi-automatic method for obtaining two-dimensional (2-D) as well as new three-dimensional (3-D) measures from a number of robust signal/image analysis approaches for feature extraction, registration and 3-D visualization of the optic nerve head from digital stereo fundus image. Typical measures include the vertical and horizontal cup and disc diameter, the cup and disc areas and the cup and disc volumes. In this preliminary study involving patients with glaucoma, a strong correlation, more than 90%, is found between the computer-generated quantitative cup/disc volume metrics and the manual metrics commonly used in a clinic. Also, in Ref. [21], three measures based on global and local scales have been estimated for detecting progression related changes in polarimetric images of glaucomatous eyes. Initially, a two-step registration algorithm between the baseline image and a follow-up image has been applied based on the detected blood vessel masks of both images and their intensity values. Then, global as well as local changes of the nerve fiber layer have been measured for both healthy and glaucomatous eyes. The results obtained showed that differentiation between healthy and glaucomatous eyes could be feasible when local loss detection based on spatial coherence criterion on specific areas with minimum size and minimum decrease of nerve fiber layer has been used.

Recently, in Ref. [22], machine trained classifiers have been employed in order to improve differentiation between glaucomatous and non-glaucomatous eyes, from standard automated perimetry. The machine learning algorithms studied included multilayer perceptron, support vector machine, and linear and quadratic discriminant analysis, Parzen window, mixture of Gaussian, and mixture of generalized Gaussian. Results show an improved performance using these classifiers over the best indexes from STATPAC, a specialized statistical analyses package employed by clinicians to interpret standard automated perimetry. Also, among the machine classifiers, quadratic discriminant analysis showed advantageous characteristics compared to the others including its fast training, simple implementation global convergence, classification performance and ease interpretation. Furthermore, in a similar study in Ref. [23], eight different classification techniques such as linear discriminant functions and neural networks has been compared to glaucoma detection using 83 topographic parameters provided by the Heidelberg Retina Tomography (HRT) system. According to the obtained results, neural network techniques combined with either forward selection or backward elimination were more successful at discriminated glaucomatous patients from healthy controls. Also, three computer image analysis statistical methods, the quantile curves, the neural net, and the probability dense curves methods had been applied to describe the changes in size of the pallor area of retinal images obtained by a fundus camera [24]. Both the neural nets and the probability dense curves methods were able to discriminate various healthy and glaucomatous subjects into three specific classes.

The majority of research has focused on discrimination between normal subjects and glaucomatous patients. Reliably and objectively detecting glaucomatous progression however remains challenging. Validated techniques with which to detect changes at the optic disc using the instruments described above are not yet available. Monitoring vessel deformation is the hallmark to detect glaucomatous progression when looking at optic disc stereo photographs. This is based on the concept that elevated intraocular pressure damages the ganglion cells and the supporting tissue of the optic disc vessels causing deformation of the vessels. The purpose of this study is to extract attributes from optic disc images with diagnostic information capable for automatic identification of early glaucomatous change at the optic disc. The methodology utilizes the deformation that occurs within optic disc vessels during the progression of the disease to train a classifier to categorize vessel deformations and consequently assisting in detection of early glaucomatous change.

In this paper, a computer-based system is proposed in order to classify subjects into two classes: patients with ocular hypertension who develop early glaucoma (called hereafter converters) and patients with ocular hypertension who remain stable (called hereafter non-converters). The proposed system operates on images of the subjects acquired from a fundus camera at different periods and it consists of a number of image processing and classification modules. In particular, in Section 2, an analytical description of the system design is presented along with the algorithmic implementation details used in the specific application. Quantitative results in terms of classification is presented in Section 3 while specific issues related to the vessel shape approach used as well as the classification problem for discriminating between converters and non-converters are emphasized in Section 4.