Automated anterior segment OCT image analysis for Angle Closure Glaucoma mechanisms classification

https://doi.org/10.1016/j.cmpb.2016.03.018Get rights and content

Highlights

  • A new method for automated classification of various mechanisms of angle closure disease.

  • Compound image transform features are directly extracted from the raw AS-OCT images without any segmentation.

  • Anterior chamber area measurement is not required in this method.

  • Selected morphological features are used to achieve better classification result.

Abstract

Background and objectives

Angle closure glaucoma (ACG) is an eye disease prevalent throughout the world. ACG is caused by four major mechanisms: exaggerated lens vault, pupil block, thick peripheral iris roll, and plateau iris. Identifying the specific mechanism in a given patient is important because each mechanism requires a specific medication and treatment regimen. Traditional methods of classifying these four mechanisms are based on clinically important parameters measured from anterior segment optical coherence tomography (AS-OCT) images, which rely on accurate segmentation of the AS-OCT image and identification of the scleral spur in the segmented AS-OCT images by clinicians.

Methods

In this work, a fully automated method of classifying different ACG mechanisms based on AS-OCT images is proposed. Since the manual diagnosis mainly based on the morphology of each mechanism, in this study, a complete set of morphological features is extracted directly from raw AS-OCT images using compound image transforms, from which a small set of informative features with minimum redundancy are selected and fed into a Naïve Bayes Classifier (NBC).

Results

We achieved an overall accuracy of 89.2% and 85.12% with a leave-one-out cross-validation and 10-fold cross-validation method, respectively. This study proposes a fully automated way for the classification of different ACG mechanisms, which is without intervention of doctors and less subjective when compared to the existing methods.

Conclusions

We directly extracted the compound image transformed features from the raw AS-OCT images without any segmentation and parameter measurement. Our method provides a completely automated and efficient way for the classification of different ACG mechanisms.

Introduction

Glaucoma is a kind of optic neuropathy disorder which makes a gradual loss of eye vision due to the damages of retinal cells and retinal nerve fiber layer in the eye [1]. It is mainly associated with increased intraocular pressure (IOP) of fluid inside the eye [2]. If it is left untreated, the damaged optic nerves will not be recovered, leading to eventual blindness; however, reducing IOP may slow disease progression. It is the second major cause of visual impairment and blindness worldwide. Globally, the number of glaucoma patients is estimated as 60.5 millions in 2010, which may increase to almost 80 millions by 2020 [3]. The timing of interference is a key element in glaucoma treatment and it is usually asymptomatic in its early stage; so early diagnosis is required to slow down the disease progression toward complete vision loss.

There are mainly three types of glaucoma, namely, angle closure glaucoma (ACG), open angle glaucoma and developmental glaucoma. ACG is a specific type of glaucoma in which a sudden rise in IOP is experienced, usually as a result of poor drainage due to the eye, producing more aqueous humor than it can remove, hence causing a build-up of fluids [4]. ACG is more prevalent than the other two types. An eye that is susceptible to ACG usually has noticeable features which can be identified upon visual examination [5]. Anterior chamber angle (ACA) assessment is mainly used for the detection of ACG, and it can be visualized and quantified by anterior segment optical coherence tomography (AS-OCT) imaging technique [6], [7].

It has been observed that ACG could be the result of one or more mechanisms in the anterior segment of the eye, such as exaggerated lens vault (L), pupil block (PB), thick peripheral iris roll (PIR), and plateau iris (PL) [6]. Laser peripheral iridotomy (LPI) is performed at an early and appropriate time for the eyes with anatomically narrow angles. Ophthalmologists and clinicians have found that the LPI is not always suitable for treating ACG patients with different mechanisms, and many patients continue to have appositional angle closure after treatment using LPI [7]. The optimal treatments for ACG patients with different mechanisms should be different. Therefore, it is important to classify these four mechanisms of ACG effectively, in order to not only provide different treatments for patients with different mechanisms, but also help the doctors tailor the best treatment for each mechanism accordingly.

It has been demonstrated that these four mechanisms of ACG have different patterns of angle configurations with different anterior chamber (AC) parameters [7]. AS-OCT provides excellent repeatability and reproducibility for imaging the anterior segment of eyes. AS-OCT was used to measure the AC parameters in the above mentioned four different mechanisms of ACG [6]. Wirawan et al. used some key parameters which were provided by customized software measured from the segmented AS-OCT images such as scleral spur-to-scleral spur distance, anterior chamber depth, anterior chamber angle and area [8]. Some reliable selected features were fed into Adaboost classifier and it was found to be clinically important to distinguish the four mechanisms of ACG [8]. In another study [9], the selected features are cross examined with four types of ACG mechanisms. In [10], various supervised and unsupervised features selection methods are used to find the reliable features and achieve the classification of different ACG mechanisms. However, the measurement of AC parameters relies on the accurate identification of the scleral spur in the segmented AS-OCT images by the clinicians, which is subjective and may not be reliable, introducing additional noise to the features. In [11], a new ensemble learning method based on error-correcting output code (ECOC) is proposed with application to classification of four ACG mechanisms and it is shown as an effective approach for multiclass classification. The existing methods for the classification of different ACG mechanisms are still not fully automated, requiring the help of the clinicians in feature extraction.

In this paper, our study is motivated by providing a fully automated expert system for the classification of different ACG mechanisms. We attempt to directly extract the features from the raw AS-OCT images without segmentation and parameter measurement. It is less subjective without the intervention of doctors. Specifically, in the proposed method, compound transform [12] is applied to the raw AS-OCT images to extract around three thousand different morphological features, from which a small set of discriminative features is selected for classification.

In the rest of this paper, Section 2 reviews related literature on general glaucoma diagnosis methods. The proposed method is presented in Section 3, followed by experimental results in Section 4. The conclusion is made with future works highlighted in the last section.

Section snippets

Related methods

During the past one decade, several studies have investigated the usefulness of the computer-based expert support systems for the early detection of glaucoma using different imaging modalities as listed in the literature [13], [14], [15], [16], [17], [18]. From the fundus images, using image processing techniques, the optic disk and blood vessels were extracted which could provide useful information to diagnose glaucoma. Higher order spectra features and textural features of fundus images

Proposed method

The proposed method for the classification of different ACG mechanisms is shown in Fig. 1, including the acquisition of raw AS-OCT images followed by the extraction of a large number of morphological features using compound image transforms and image statistics [19]. A small set of discriminative and inter-dependent features selected by using maximum relevance-minimum redundancy (mRMR) algorithm [20] is fed into a Naïve Bayes Classifier for classification.

Experimental results and discussion

Our experiment was conducted using Matlab Toolbox from the original authors for feature selections with mRMR method [19]. Each feature in the original feature set was normalized to have zero mean and a standard deviation of unity and discretized in order to compute the mutual information for feature selection. Prior to classification using NBC, ranked features are selected amongst extracted 2919 features using mRMR algorithm with mutual information difference. Leave-one-out cross-validation

Conclusions

Angle-closure glaucoma, which is a specific type of glaucoma, has been observed to be the result of one or more mechanisms in the anterior segment of the eye. Traditional methods for classification of the different mechanisms are based on calculation of the key parameters, such as the angle opening distance (AOD), trabecular-iris space area (TISA) and angle recess area (ARA) from the segmented AS-OCT images with the help of doctors, which are not fully automated. Since the manual diagnosis

Conflict of interest statement

None declared.

Acknowledgement

This work was supported by Ministry of Education (MoE) AcRF Tire 1 Funding, Singapore, under Grant M4010981.020 RG36/11.

References (46)

  • A. Wirawan et al.

    Feature selection for computer-aided angle closure glaucoma mechanism detection

    J. Med. Imaging Health Inform.

    (2012)
  • S. Issac Niwas et al.

    Cross-examination for angle-closure glaucoma feature detection

    IEEE J. Biomed. Health Inform.

    (2016)
  • S. Issac Niwas et al.

    Reliable feature selection technique for automated angle closure glaucoma mechanism detection

    J. Med. Syst.

    (2015)
  • X.S. Bai et al.

    Learning ECOC code matrix for multiclass classification with application to glaucoma diagnosis

    J. Med. Syst.

    (2016)
  • U.R. Acharya et al.

    Automated diagnosis of glaucoma using texture and higher order spectra features

    IEEE Trans. Inf. Technol. Biomed.

    (2011)
  • J. Nayak et al.

    Automated diagnosis of glaucoma using digital fundus images

    J. Med. Syst.

    (2009)
  • M.J. Greany et al.

    Comparisons of optic nerve imaging methods to distinguish normal eyes from those with glaucoma

    Invest. Ophthalmol. Vis. Sci.

    (2002)
  • M. Balasubramanian et al.

    A framework for detecting glaucomatous progression in the optic nerve head of an eye using proper orthogonal decomposition

    IEEE Trans. Inf. Technol. Biomed.

    (2009)
  • J. Xu et al.

    3D optical coherence tomography super pixel with machine classifier analysis for glaucoma detection

  • J. Xu et al.

    Automated assessment of the optic nerve head on stereo disc photographs

    Invest. Ophthalmol. Vis. Sci.

    (2008)
  • L. Shamir et al.

    Wndchrm – an open source utility for biological image analysis

    Source Code Biol. Med.

    (2008)
  • H. Peng et al.

    Feature selection based on mutual information criteria of max-dependency, max-relevance, and min-redundancy

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2005)
  • L. Shamir et al.

    Impressionism, expressionism, surrealism: automated recognition of painters and schools of art

    ACM Trans. Appl. Percept.

    (2010)
  • Cited by (0)

    View full text