Progress on retinal image analysis for age related macular degeneration

Abstract

Age-related macular degeneration (AMD) is the leading cause of vision loss in those over the age of 50 years in the developed countries. The number is expected to increase by ~1.5 fold over the next ten years due to an increase in aging population. One of the main measures of AMD severity is the analysis of drusen, pigmentary abnormalities, geographic atrophy (GA) and choroidal neovascularization (CNV) from imaging based on color fundus photograph, optical coherence tomography (OCT) and other imaging modalities. Each of these imaging modalities has strengths and weaknesses for extracting individual AMD pathology and different imaging techniques are used in combination for capturing and/or quantification of different pathologies. Current dry AMD treatments cannot cure or reverse vision loss. However, the Age-Related Eye Disease Study (AREDS) showed that specific anti-oxidant vitamin supplementation reduces the risk of progression from intermediate stages (defined as the presence of either many medium-sized drusen or one or more large drusen) to late AMD which allows for preventative strategies in properly identified patients. Thus identification of people with early stage AMD is important to design and implement preventative strategies for late AMD, and determine their cost-effectiveness. A mass screening facility with teleophthalmology or telemedicine in combination with computer-aided analysis for large rural-based communities may identify more individuals suitable for early stage AMD prevention.

In this review, we discuss different imaging modalities that are currently being considered or used for screening AMD. In addition, we look into various automated and semi-automated computer-aided grading systems and related retinal image analysis techniques for drusen, geographic atrophy and choroidal neovascularization detection and/or quantification for measurement of AMD severity using these imaging modalities. We also review the existing telemedicine studies which include diagnosis and management of AMD, and how automated disease grading could benefit telemedicine. As there is no treatment for dry AMD and only early intervention can prevent the late AMD, we emphasize mass screening through a telemedicine platform to enable early detection of AMD. We also provide a comparative study between the imaging modalities and identify potential study areas for further improvement and future research direction in automated AMD grading and screening.

1. Introduction

Age-related macular degeneration (AMD) is the leading cause of vision loss in those over 50 years old in the industrialized world. Currently, approximately 35% of adults older than 80 in the United States, for example, are estimated to have intermediate or advanced AMD (National-Eye-Institute, 2004). The number of people with this condition is expected to increase ~1.5 fold over the next ten years due to an aging population, and a higher prevalence of vascular risk factors such as hypertension contributing (Bartlett and Eperjesi, 2007). AMD can be broadly divided into early, intermediate and late stage AMD. Early and intermediate AMD, sometimes called early “dry” AMD, is associated with drusen and pigmentary changes in the retina and is associated with only minimal visual symptoms. Late AMD has two subtypes, a progressive geographic atrophy in the macula region, sometimes called late “dry” AMD; or exudative AMD due to the development of choroidal neovascularization, sometimes called “wet” AMD. While treatment of exudative AMD with anti-vascular endothelial growth factor (anti-VEGF) agents has been revolutionary and is effective in maintaining vision, such treatments are costly and may be associated with significant cardiovascular risks (Wong, 2009). Furthermore, although these treatments may be effective for the exudative or neovascular component of AMD, they have not been shown to be able to stop or reverse the atrophic processes of dry AMD (Lim et al., 2012; Meleth et al., 2011; Schmitz-Valckenberg et al, 2009). The Age-Related Eye Disease Study (AREDS) showed that specific antioxidant vitamin supplementation (AREDS formulation) reduces the risk of progression from intermediate stages (defined as the presence of either many medium-sized drusen or one or more large drusen) to late AMD, that can allow for preventative strategies (Age-Related Eye Disease Study Research Group, 2007). Thus, early identification of people with selected AMD is important, as it enables implementation of disease prevention strategies. For example, approximately 7.6% of the United States population over the age of 60 are estimated to have advanced or intermediate AMD, amenable to such treatment (National-Eye-Institute, 2004). Recent publications looking at the ten-year experience of appropriately selected patients taking AREDS formulation demonstrates that it is effective at slowing disease and improving visual acuity in approximately 25% of patients (Chew et al., 2013).

The classic method of early and intermediate stage AMD detection is retinal color photography to detect drusen (i.e., deposits of lipids and other metabolic products that form beneath the retinal pigment epithelium (RPE), (Rapantzikos et al., 2003) in the macula area. Furthermore, area quantification and classification of drusen in these color images can be accomplished through image analysis (Smith et al., 2003, 2005) (Abramoff et al., 2010). Other imaging techniques are also used in AMD detection and screening. These include non-invasive imaging techniques such as optical coherence tomography (OCT), fundus autofluorescence (FAF), and infrared (IR) imaging. More invasive techniques such as fluorescein angiography (FA) and indocyanine green (ICG) angiography, that require dye injection, are used to delineatethe AMD clinical status more accurately.

In this review, we focus our discussion on existing methods for detecting AMD lesions, e.g., drusen, geographic atrophy (GA) and choroidal neovascularization (CNV), from these retinal imaging modalities. The review also analyses current retinal imaging techniques and their relative strengths and weaknesses for detecting and classifying AMD pathologies. Topics discussed include:

• current image analysis techniques, exploring strengths and areas for further improvement

• review of existing AMD grading protocols and identification of the requirements of automation for mass screening based on AMD protocols

• review of telemedicine-based systems for AMD screening

• comparison and contrast of each of the imaging modalities and their scientific merits for any particular AMD pathology detection

• delineating future research directions for early AMD detection through mass screening of large populations.

1.1. Method of literature search

We performed a comprehensive literature review on AMD imaging modalities and telemedicine for the diagnosis and screening of AMD.A computerized search of the PubMed database and the Web of Science database was performed to identify all pertinent peer-reviewed articles published up till mid April 2013. The keywords used for the search included age related macular degeneration, imaging techniques, screening, diagnosis, imaging modalities, macula, drusen, geographic atrophy, choroidal neovascularisation, dry AMD, wet AMD, color fundus photograph, fundus autofluorescence, angiography, optical coherence tomography, spectral domain optical coherence tomography, indocyanine angiography, hyperspectral imaging, infrared imaging, telemedicine, and clinical trials.

1.2. Criteria of inclusion

All abstracts were examined to identify if articles were imaging or telemedicine, and related to the quantification of AMD pathologies, AMD screening or diagnosis.

1.3. Criteria of exclusion

Papers and abstracts were excluded from review if they were not peer-reviewed. AMD pathology segmentation papers which did not report any quantified accuracy measurement, were also excluded from consideration.

1.4. Selection of papers

Once the inclusion and exclusion criteria were examined, copies of the entire articles were obtained and citations were entered into Endnote. In the final stage, articles were reviewed and information relating to quantification of AMD pathologies or AMD screening was included in this review.

The remainder of this review article is organized as follows: A brief description of AMD stages and existing diagnosis techniques for AMD is provided in section two. Section three provides a brief description and summary of image analysis techniques for AMD pathology detection and quantification. A comparative study of different AMD imaging modalities is provided in section four. A brief overview of existing AMD telemedicine platforms for screening is provided in section five. In section six, a brief description is provided of existing clinical trials for the identification and follow-up of AMD patients. In section seven, the limitations of current imaging modalities are stated. Finally, conclusions and future research directions are drawn in section eight.

2. Classification and diagnostic and screening techniques for AMD

2.1. Classification of AMD

The early stages of AMD include soft drusen and reticular pseudodrusen, aka reticular macular disease.

2.1.1. Hard drusen and soft drusen

Drusen are the classic initial lesions of early, dry AMD. Drusen can be defined as hard and soft. Hard drusen are small lesions with sharp borders (Fig. 1 left) and soft drusen are those with indistinct borders (Fig. 1 left). Most studies agree that hard drusen alone carry no adverse prognostic significance. Soft drusen, however, have long been recognized as precursors to advanced AMD (Bressler et al., 1988). They are pale yellow deposits of cholesterol and other materials beneath the RPE, sometimes also associated with pigment clumps.

Fig. 1.

Retinal colour image showing many drusen in the macula (left) and geographic atrophy (right).

2.1.2. Reticular pseudodrusen (RPD) aka reticular macular disease (RMD)

These lesions, although described earlier, are now attracting widespread attention because of their visibility in advanced imaging modalities and their strong association with advanced AMD. Mimoun et al. (1990) described them as a peculiar yellowish pattern of “pseudodrusen visible in blue light,” because of their enhanced visibility using blue light fundus photography. In 1991, Klein et al. called them “reticular soft drusen” in the Wisconsin Age-Related Maculopathy Grading System. The first reference to “reticular pseudodrusen” was made by Arnold et al. (1995)

With the advent of autofluorescence (AF) imaging, Smith et al., 2006 made the connection between reticular pseudodrusen on fundus photography and a reticular pattern on AF imaging.

Upon further investigation with Scanning Laser Ophthalmology (SLO) imaging, including near infrared reflectance (NIR-R) and indocyanine green angiography (ICG), Smith et al. (2009), proposed “reticular macular disease” as a distinct phenotypic entity in the classification of AMD, with reticular pseudodrusen being a clinical feature of reticular macular disease. Analyzing these lesions with spectral domain optical coherence tomography (SD-OCT), Zweifel et al. (2010) found that they are localized between the RPE and the inner segment/outer segment boundary (now referred to as the “inner segment ellipsoid (ISe) band”) and referred to reticular lesions as “subretinal drusenoid deposits” (SDD). RPD are a risk factor for advanced AMD and their detection should be part of the ophthalmologic screening examination for AMD. Dry atrophic AMD, or geographic atrophy (GA) occurs when there is combined loss of photoreceptors, RPE and choriocapillaris over a well-defined area or areas (multilobular form). If this process affects the fovea, there can be severe central visual loss.

2.1.3. Early “dry” AMD

In the early and intermediate stages of AMD, the transport of nutrients and wastes by the RPE slows down and waste accumulates under the RPE forming yellowish deposits called drusen (Fig. 1 Left) (Rapantzikos et al., 2003; Bartlett and Eperjesi, 2007). As the RPE continues to slow down in its transport of nutrients and wastes, the overlying photoreceptors become damaged, up to legal blindness.

Late “dry” AMD or GA occurs with progressive atrophy of the RPE, choriocapillaris and photoreceptors (Lim et al., 2012)(Fig.1 right). It is characterized by sharply delineated areas of severe depigmentation or apparent absence of the RPE, through which larger choroidal vessels are more easily seen, with a minimum diameter of 175 μm (Baumann et al., 2010). Late “dry” AMD accounts for 80–90% of all diagnosed patients. Progression tends to be slow, and is measured over years. The fovea can be spared until late in the disease.

“Wet” AMD or neovascular AMD, refers to the proliferation of new vessels either under the RPE, breaking through the RPE, or within the neural retina, leading to death of the photoreceptors from exudative damage (fluid, lipids and blood, ultimately leading to fibrous scarring). The first two forms are types of choroidal neovascularization (CNV), whereas the third, known as retinal angiomatous proliferation (RAP), begins as intraretinal neovascularization. As a result, objects in that portion of the visual field may appear wavy or distorted, known clinically as metamorphopsia (Freund et al., 2008; Lim et al., 2012). Progression can be rapid, if untreated, resulting in permanent legal blindness.

The physiology of the aging macula and current understanding of the mechanisms implicated in the causes of AMD can be reviewed (Bressler et al., 1988; Ryan et al., 2013). Clinical and histopathological features of AMD, as well as related genetics and epidemiology, can be found in the review by Jager et al. (2008). However, it must be recognized that all of these topics are under active research, with new data and proposed disease paradigms appearing frequently. The natural history of AMD has been fairly well documented (Wong, 2009.

2.2. Diagnostic tests for AMD

The tests below are widely used in eye clinics for AMD diagnosis:

• Visual acuity

• Amsler grid test

• Slit lamp biomicroscopy examination

• Color fundus photography.

Details on these diagnostic approaches can be found in the review of Freund (Freund et al., 2008). These tests may be most appropriately utilized as initial tests for AMD.

The following imaging modalities can subsequently be used for further analysis:

• Fundus Autofluorescence (FAF) Imaging

• Optical Coherence Tomography (OCT)

• Fluorescein Angiography (FA)

• Indocyanine Green (ICG) Angiography

• Thermal or Infrared (IR) Imaging

• Hyperspectral Retinal Imaging (HRI).

2.3. Diagnostic imaging modalities

In this section we briefly introduce different imaging modalities that are used for AMD screening and follow up diagnostic tests.

2.3.1. Imaging modalities for AMD screening

Color fundus photography (film initially but more recently digital) has been the gold standard for AMD screening and is a rapid way to look for soft drusen, the earliest sign of AMD (Kösea et al., 2008; Sbeh et al., 2001). Fundus cameras, however, are expensive.FAFis another imaging modality which may be used to screen early AMD (Framme et al., 2005). FAF is also an excellent way to visualize reticular macular disease (Xu et al., 2013).

2.3.1.1. Fundus photography for AMD screening.

A digital fundus camera is used to capture an image of the interior of the eye, including the retina, optic disc, and macula (Fig. 1). A specialized low powermicroscope with an attached digital camera is designed for color fundus photography and uses visible spectrum light for imaging the retinal structure in vivo. In color imaging, the image intensities represent the amount of reflected red, green and blue wavelengths, as determined by the spectral sensitivity of the sensor. Since the level of ambient light normally striking the retina creates insufficient reflected illumination for digital imaging, external illumination must be projected into the eye, and the light reflected by the retina must re-traverse the pupillary plane. As a result, separate paths must be used in the pupillary plane, resulting in optical apertures of the order of a few millimeters to avoid overlap between corneal and lenticular reflections (Abramoff et al., 2010).

2.3.1.2. Fundus Autofluorescence (FAF) for AMD screening.

FAF imaging may allow for identification of retinal diseases when these are not otherwise evident through color imaging. It is an in vivo imaging method for metabolic mapping of naturally or pathologically occurring fluorophores of the ocular fundus (Fig. 2). Metabolic changes at the level of the photoreceptor/RPE complex are not visualized by digital fundus image or other routine imaging techniques such as FA in the early manifestations of macular and retinal dystrophies (Fleckenstein et al., 2010a,b). FAF may be particularly helpful to investigate patients with unknown visual loss or with a family history of hereditary retinal diseases.

Fig. 2.

An autofluorescence image of the retina with dry macular degeneration (left) and more advanced form of dry macular degeneration (right).

In FAF, the dominant sources appear to be fluorophores such as A2-E in lipofuscin granules that accumulate in the postmitotic RPE as a by-product of the incomplete degradation of photoreceptor outer segments. However, the dominance of A2-E in the macula has recently been disproved by imaging mass spectroscopy (Ablonczy et al., 2013), so additional research is needed to identify these compounds accurately. Additional intrinsic fluorophores may occur with disease in the various retinal layers or the subneurosensory space. Minor fluorophores such ascollagen and elastin in choroidal blood vessel walls may become visible in the absence or atrophy of RPE cells (Fleckenstein et al., 2010a,b). There are two noninvasive ways to obtain these images; one uses a specialized scanning laser ophthalmoscope, and the other uses special filters attached to the fundus camera.

2.3.2. Imaging modalities for diagnosis of early or late “dry” AMD

Drusen (early “dry” AMD) and GA (late “dry” AMD) are identified by imaging modalities such as color fundus photography, FAF and spectral domain optical coherence tomography (SD-OCT). Until the advent of SD-OCT, color fundus and FAF most reliably assessed drusen and GA, respectively. With SD-OCT, small and medium-size drusen may be more clearly seen as discreet areas of RPE elevation with variable reflectivity, indicating the variable composition of the underlying material (Fig. 3). With larger drusen or drusenoid pigment epithelium detachment (PED), greater elevation of RPE may be seen, often dome shaped, with a hypo or medium reflective material separating the RPE from the underlying Bruch’s membrane (Keane et al., 2012).

Fig. 3.

An OCT device (top-left) and OCT image for normal and wet AMD (top-right). Gray scale image showing the line for OCT B scan (bottom left side) and an OCT scan showing depth of the line.

2.3.2.1. Fundus photography of early or late “dry” AMD.

Color digital fundus photography has been used to grade and quantify soft drusen and GA (Smith et al., 2003; Scholl et al, 2003) as part of GA diagnosis. Most of the existing methods for evaluating color fundus-based drusen and GA grading are manual and semi-automatic. This may be because drusen subtypes such as soft distinct and soft indistinct drusen are very difficult to measure within their irregular and blurred boundaries. Fig. 1 shows soft drusen in a color fundus image.

2.3.2.2. Fundus Autofluorescence (FAF) of early or late “dry” AMD.

FAF imaging is mainly used to quantify GA for late “dry” AMD diagnosis. FAF gives information above and beyond that obtained by conventional imaging methods such as digital fundus photography. GA represents the loss of RPE, and it manifests as a markedly reduced FAF signal (Schmitz-Valckenberg et al, 2008).

2.3.2.3. Spectral Domain Optical Coherence Tomography (SD-OCT) of early or late “dry” AMD.

SD-OCT is widely used for both late “dry” and “wet” AMD diagnosis, as it is well capable of capturing the area and volume of the AMD pathologies. It is one of the most recent noninvasive technologies developed for recording the features of the retina, and it displays the information as cross-sectional views, or optical slices, i.e., mapping the layered anatomy of the retina (Fig. 3). SD-OCT functions as a type of optical biopsy, providing information on retinal pathology in real time, with a resolution vastly better than magnetic resonance imaging (MRI) or ultrasound. A major strength of SD-OCT over other imaging modalities is its ability to capture 3D-volumetric information about the pathologies. It has therefore been widely used in advanced AMD diagnosis.

The principle of OCT is the estimation of the depth at which a specific backscatter originated by measuring its time of flight. Backscatters are typically caused by differences in refractive index in transitions from one tissue to another. Backscatter from deeper tissues can be differentiated from backscatter originating at more superficial tissues because it takes longer for light to arrive at the sensor from deeper tissues. With a total retinal thickness between 300 and 500 μm, the differences in time of light relay are very small and are measured through interferometry, which OCT utilizes (Abramoff et al., 2010). The OCT imaging procedure is convenient and rapid for the patient. Huang et al. (Huang et al., 1991) describes the early OCT imaging technique, which aimed to capture 2D images. Commercially available OCT allows collection of up to 400 A-scans per second, which is not suitable for 3D imaging (Abramoff et al., 2010). In the original OCT imaging technique, very weak laser light is used to map the anatomy of the retina as the reflected image returns to a series of mirrors, with resulting reconstruction of computer processed images (Fig. 3) which are saved for analysis (Freund et al., 2008).

With SD-OCT, up to 40,000 A-scans can be acquired per second, which is enough for 3D imaging (Schimel et al, 2011). In SD-OCT the source beam is split in a similar fashion to time domain OCT (TD-OCT), but no moving mirror is required in data capture. A spectrometer is employed to analyze light frequency changes that occur from reference and subject beam interaction (Schimel et al, 2011). The Fourier transform is applied to the spectral correlogram intensities to determine the depth of each scatter signal (Abramoff et al., 2010). Spectral domain (SD) or Fourier domain allows the instrument to scan the retina much faster, providing very high-resolution 3-D images of the retina. SD-OCT is better than time domain OCT (TD-OCT) for pathology quantification (Yi et al., 2009) in accuracy, area and volume measurement, noise handling and speed, and it has become the standard of care for AMD.

SD-OCT is an excellent modality to visualize subretinal drusenoid deposits (SDD) (Zweifel et al., 2010).

A brief summary of existing commercially available SD-OCT devices is given in Kiernan et al. (2010). With SD-OCT, the ophthalmologist obtains a clearer, more accurate view of individual tissue layers and relationships among layers. SD-OCT also allows for repeated examination of the exact same areas of the retina for each patient visit, resulting in more precise measurement of the effects of treatment.

2.3.3. Imaging modalities for diagnosis of exudative AMD

Exudative or “wet” AMD represents CNV, predominantly diagnosed with SD-OCT or fluorescein angiography in current practice.

2.3.3.1. Spectral Domain Optical Coherence Tomography (SD-OCT) of exudative AMD.

SD-OCT imaging is highly suited to image the fluid associated with CNV, displaying characteristic as well as atypical images of neovascularization at the level of RPE and subretinal space. Other neovascular conditions that may be delineated with SD-OCT include polypoidal choroidal vasculopathy and peripapillary and peripheral CNV (Keane et al., 2012). Fig. 3 shows the fluid accumulation due to CNV in an SD-OCT scan of a patient with exudative AMD. OCT evaluations have greatly augmented but not entirely replaced angiography. OCT provides different information to angiography, such as whether excess fluid is present in the retina, as opposed to the site of a leak. OCT is extensively used for monitoring the progress of treatment for “wet” macular degeneration. Additional common clinical applications of SD-OCT related to CNV are to initiate diagnosis and therapy, to determine the ongoing effect of anti-angiogenic treatment and to guide the termination of anti-angiogenic therapy (Keane et al., 2012). Caution must be used in interpretation of fluid filled spaces seen on SD-OCT. The presence of fluid does not necessarily imply that CNV is active. Fluid may be present in inactive degenerative cysts, or in outer retinal tabulation, a common feature of retinal degenerations including AMD. Clinical judgment is warranted to avoid unnecessary treatment.

2.3.3.2. Fluorescein Angiography (FA) of exudative AMD.

Angiography, both fluorescein and indocyanine green (ICG), is another useful test to look for CNV. A small gage needle attached to a plastic catheter is inserted into a vein, usually in the arm, and a fluorescent dye is injected through the catheter and enters the bloodstream. As the dye enters the blood vessels of the eye, a series of pictures are taken of the retina using a fundus camera. Special filters matched for the excitation and emission wavelengths of the dye are placed in a beam splitter so that only the fluorescence of the dye in the blood vessels is visualized against the dark remainder of the background retinal tissue (Fig. 5). By looking at the pattern of the blood vessels and observing whether dye leaks from any vessels, an ophthalmologist can locate sites of CNV if they are present (Fig. 5, right).

Fig. 5.

Fundus camera and angiography images (left) and fluorescein angiogram shows CNV in the macula (right). ICG angiogram shows a bright area of CNV in the macula. The CNV was not visible with fluorescein dye because of abnormal fluid beneath the retina.

Sometimes, an area of CNV is not clearly defined, or it may be obscured by overlying fluid or blood. In these cases ICG dye is used instead of fluorescein (Fig. 6), as discussed in the next subsection. Fig. 4.

Fig. 6.

Red-free photograph of a patient with classic choroidal neovascularization (CNV) (left) and early-phase ICG angiogram reveals hyperfluorescence of the CNV (right).

Fig. 4.

Standard OCT scans showing choroidal vasculature: (a) OCT volume with segmented Bruch’s membrane (blue line). (b) Flattened OCT volume with segmented Bruch’s membrane. (c) Sub-volume under Bruch’s membrane containing the entire choroid. (d) Original OCT volume with superimposed segmented vessels (pink outlines) and surfaces (green line) fitting around the choroidal vasculature, used to determine choroidal vasculature thickness and choriocapillaris-equivalent thickness. Red arrow points to one of the choroidal vessels (image taken from (Abramoff et al., 2010)).

2.3.4. Other imaging modalities

2.3.4.1. Indocyanine Green (ICG) angiography.

ICG is also useful for visualizing the deeper blood vessels located in the choroid. ICG can demonstrate how the choroidal circulation interacts with other layers of the retina and whether variant forms of CNV, such as retinal angiomatous proliferation (RAP) or polypoidal choroidal vasculopathy are present (Freund et al., 2008). ICG has been useful for detecting CNV in neovascular AMD (Chang et al., 2004).

ICG is a cyanine dye and absorbs light in the near-IR range of 790 nm–805 nm and fluoresces in the IR (non-visible) range. These long wavelengths allow for visualization of the dye through overlying melanin and xanthophyll. The IR wavelengths also have the ability to penetrate the retinal layers making circulation in deeper layers (e.g., choroid) visible (Fig. 6). The detection procedure is similar to fluorescein angiography, but utilizes a dye and detection filters that allow capture in the IR (non-visible) light range.

2.3.4.2. Thermal/Infrared (IR) imaging.

IR imaging provides a non-invasive, in vivo method to image early changes in the RPE/Bruch membrane. It offers advantages over current imaging techniques by minimizing light scattering through cloudy media. It also enhances the ability to image through small pupils, retinal hyperpigmentation, blood, heavy lipid exudation, or subretinal fluid. It provides additional information regarding early CNV, and the character of drusen and sub-retinal deposits (Hartnett and Elsner, 1996).

In IR imaging with an SLO, light in the 820 nm wavelength is used to illuminate the tissue. During scanning, reflections from the tissue are acquired. IR imaging detects pathology despite the presence of hemorrhage or cataracts, that may be undetected in other in vivo studies such as angiographic dye (Schardosim et al, 2011). The Heidelberg retina angiograph2 (HRA2) has an IR scan mode. Fig. 7 shows IR retinal images (Kolar et al., 2008) for CNV. However, IR is not considered diagnostic for CNV, because these images cannot define leakage or intraretinal fluid, but only show a suggestive subretinal fibrotic ring.

Fig. 7.

Infrared frames of well-defined choroidal neovascularisation. This composite frames show nine examples of eyes affected with well-defined choroidal neovascularisations (CNVs) in exudative AMD. In infrared pictures, the well-defined CNV appear as a complex with a dark central core surrounded by a whitish reflective ring, giving a halo-like shape (black arrows). This whitish ring may have an O-shape (A, C, E, F, H) or a horseshoe aspect (U-shape) (B, D, G, I) (Semoun et al., 2009).

A limitation with using a digital fundus camera to provide useful IR images can be attributed to its failure to separate reflected and scattered light, since IR light is absorbed less than visible light and thus may scatter over longer distances. Consequently, IR fundus images are often noisy and present low detail definition. IR fundus images of static scenes taken at different time instants (i.e., infrared image sequences), tend to present pixel value fluctuations in time (Schardosim et al, 2011). This defect is largely overcome by use of a confocal SLO such as the HRA2 or Spectralis that rejects light scattered outside the plane of focus.

2.3.4.3. Hyperspectral Retinal Imaging.

Hyperspectral retinal imaging (HRI) is a new imaging technique which is capable of providing both spatial and spectral retinal information and constructing a 3D data cube for multivariate data analysis (Diaconu, 2009). It is well suited for detecting AMD pathology, such as drusen. Hyperspectral imagers commonly employ one of two methods to separate light into spectral components: dispersive optics (for example, a prism) and interferometers. In a dispersive spectrometer, light passes through the entrance slit and is collimated by a lens before it hits the wavelength selection device, typically a prism or grating which separates the light into its spectral components (Davis et al., 2007). Hyperspectral analysis has been applied to drusen as well as macular pigment and retinal oximetry. There is significant untapped potential for optical molecular biopsy of living tissue with this method (Lee et al., 2010; Fawzi et al., 2011; Gao et al., 2012).

3. Image analysis techniques for AMD

The imaging-based diagnostic techniques for AMD focus on detecting different pathologies through variations of retinal imaging. Each of the imaging modalities has strengths and limitations in capturing individual pathology. In this section, we focus on the imaging modalities and different automated and semi-automated techniques that are used to quantify the pathologies.

3.1. Automatic techniques for color fundus imaging for “dry” AMD

Existing algorithms for early and late “dry” AMD screening using color fundus image analysis mainly focus on the detection and quantification of drusen, with a few methods discussed for GA quantification. These automatic or semiautomatic image analysis methods include image normalization and enhancement, macular area mapping, and segmentation of drusen and GA. Image normalization is usually essential as the irregular shape of the retina creates shading across the field of view when illuminated by a bright flash in the fundus camera. Contrast appearance techniques have been used to improve the appearance of details in the retinal image before processing. In this discussion we categorize techniques based on the main image analysis steps such as texture, edge, mathematical morphology, intensity thresholding and clustering-based segmentation techniques.

3.1.1. Color fundus imaging for “dry” AMD

3.1.1.1. Texture-based drusen segmentation.

Texture generally describes the property of surfaces and scenes, measured over image intensities. Drusen, which are brighter in the retinal surface, have a different response than the background if the appropriate texture methods are applied. Using this hypothesis, a number of methods have been proposed which are based on Gabor filter and wavelet analysis (Kose et al., 2010), (Parvathi and Devi, 2007), (Lee et al., 2008). In Gabor filtering an input image $I(x,y)$, $(x,y)\in \Omega$, where $\Omega$ is the set of image points, is convolved with a 2D Gabor function $g(x,y)$, $(x,y)\in \omega$ to obtain a Gabor feature image $r(x,y)$ (Kruizinga and Petkov, 1999). These functions are used to extract texture information that mainly reflects a variation in response along different orientations on different surfaces. For example, the response for texture of drusen areas and background texture are different with respect to their color and intensity distributions.

The wavelet transformation is also used to achieve better results than the Fourier transformation, as wavelets have advantages of spatial localization and basis function flexibility (Brandon and Hoover, 2003). A number of basis functions for the orthonormal wavelet transform were used, which were similar in shape to a typical drusen. A combination of amplitude modulation (AM) and frequency modulation (FM) has been used to generate multi-scale features for classifying pathological structures, i.e., drusen. The AM-FM is applied on different scales to represent an input digital image as the sum of AM-FM components (Barriga et al., 2009). The AM functions characterize slowly-varying image intensity variations, whereas the FM function captures fast changing spatial variability in the image intensity, and is given by the cosine of its instantaneous phase function. This combined method is used to differentiate structures based on the gradient of the pixel-based frequency. Statistical structural information-based texture analysis techniques use texture features such as the average of standard intensity deviation, intensity distribution, and entropy of a particular area (Kose et al., 2010).

Once texture information is extracted, statistical classifiers are used to identify the region of interest, based on the texture pattern. The classifier can be defined as a device or system with n inputs, each of which is used to enter the information about one of n features, ${\mathrm{x}}_1,{\mathrm{x}}_2,\dots ,{\mathrm{x}}_{\mathrm{n}}$ that are measured from an object to be classified. An R-class classifier will generate one of R symbols C₁, C₂, …,C_R as an output, and the user interprets this output as a decision about the class of the processed object. The generated symbols C_R are the class identifiers (Sonka et al., 1999). The classifiers are of two types: supervised classifiers and unsupervised classifiers or clustering. A supervised classifier uses training feature(s) (i.e., feature vector) for a class/object to assign objects to one of a prespecified set of classes based solely on the vector of measurements taken on those objects (Hand, 2006). In contrast an unsupervised classifier measures or perceives similarities among the feature vectors (i.e., patterns) and finds the natural groupings or clusters in the multi-dimesional data (Jain et al., 2000). Both supervised classifiers such as the Artificial Neural Network (ANN) or Support Vector Machine (SVM) or unsupervised techniques such as the Expectation Maximization (EM) or Fuzzy C-Means (FCM) Clustering are used to identify the region of interest, e.g., drusen (Fig. 8) (Brandon and Hoover, 2003),(Freund et al., 2009).

Fig. 8.

Gray scale retinal image (cropped) shows drusen in the macular area (left) and segmented image (texture based) for drusen detection output (right).

A neural network (NN) is a computational technique inspired by biological neurons with the ability to adapt or learn, to generalize and to cluster or organize data. It is structured as an input layer, a hidden layer and an output layer: to accept input features, to train the network based on the features, and to classify the features into a certain number of classes. Back propagation or probabilistic algorithm-based NN are used for classifying drusen and background after analyzing the features (Ripley, 1996).The SVM algorithm is a classification algorithm which transforms the input space to a higher dimension feature space through a nonlinear mapping function, and constructs a separating hyperplane with the maximum distance from the closest value of the training set (He et al., 2006).

The EM technique obtains a partition function for a class/cluster based on the known patterns and assigns a cluster membership for a new test sample by optimizing the closest similarity match with highest probability (Lee et al., 2008). Fuzzy C-Means clustering divides data elements into classes or clusters based on similarity measures such as distance (e.g., Euclidian distance between feature vectors), connectivity and intensity. Fuzzy C-Means clustering is also referred to as soft clustering, which assigns data elements into one or more classes with membership levels (Thaibaoui et al., 2000), (Bhuiyan et al., 2007).

Table 1 summarizes the texture based methods and their accuracy.

Table 1.

Summary of texture based drusen detection methods.

Method	Image processing techniques	Level of segmentation	Accuracy (%)
Brandon and Hoover (Brandon and Hoover, 2003)	Mexican hat wavelet and Feed Forward Neural Network	Drusen or non-drusen	87
		Small or large drusen	71
Kose et al. (Kose et al., 2010)	Statistical texture properties and thresholding; edge based vessel detection	Small area degeneration	84.06–92.64
		Medium area degeneration	89.89–96.06
		Large area degeneration	90.67–95.16
Parvathi and Devi (Parvathi and Devi, 2007)	Log-Gabor filter and thresholding	Drusen or non-drusen	–
Barriga et al. (Barriga et al., 2009)	Mutiscale feature through AM-FM and ROI by Partial Least Squares	Drusen and non drusen	Up to 100
		Soft and hard drusen	96
Agurto et al. (Agurto et al., 2011)	Same as (Barriga et al., 2009)	Drusen area	73
Lee et al. (Lee et al., 2008)	Multi-scale top hat filter and partition function based clustering	Drusen and non drusen	71.7–92.2
Freund et al. (Freund et al., 2009)	Multi-scale analysis and Support Vector Machine	Drusen absent or present	–
Priya and Aruna (Priya and Aruna, 2011)	Discrete wavelet transform and Probabilistic Neural Network	Dry and wet AMD	96

3.1.1.2. Thresholding-Based drusen segmentation.

Thresholding is a technique which converts an image into a binary image based on intensity or frequency to obtain the region(s) of interest. A global threshold value may be chosen automatically or on the basis of clear points in the image histogram that would allow for efficient separation. More complex intensity criteria may be used to allocate whether pixel values become white or black. For some images, adaptive or local thresholding is useful where different thresholds are applied to different sections of the image, e.g., when the image has varying levels of background illumination (Patton et al., 2006). In drusen detection methods (Smith et al., 2005), (Soliz et al., 2002), the thresholding-based techniques mainly rely on pixel intensity, derivative, histogram and local intensity probability to select the threshold for drusen regions. Threshold-based drusen segmentation methods are summarized in Table 2.

Table 2.

Table: Summary of thresholding based drusen detection methods.

Method	Image processing techniques	Level of segmentation	Accuracy (%)
Soliz et al. (Soliz et al., 2002)	Median filtering, Gaussian smoothing and probability based thresholding	Drusen and non drusen	71
		Drusen size	67
Smith et al.(Smith et al., 2003), (Smith et al., 2005)	Otsu method for background selection and intensity thresholding to segment drusen	Drusen detection sensitivity	42–86
		Drusen detection specificity	5398
		Drusen area detection	87.395.3
Rapantzikos and Zervakis (Rapantzikos and Zervakis, 2001)	Homomorphic filtering and histogram based adaptive local thresholding	Drusen and non drusen	Up to 100
Liang et al. (Liang et al., 2010)	Edge based thresholding and morphological dilation	Drusen detection sensitivity	75
		Drusen detection specificity	75
Barakat and Madjarov (Barakat and Madjarov, 2004)	Region of interest selection and intensity thresholding	Drusen detection sensitivity	86
		Drusen detection specificity	93
Morgan et al. (Morgan et al., 1994)	Median filtering, region subdivision and intensity thresholding	Drusen area detection	–

In summary, Rapantzikos and Zervakis (Rapantzikos et al., 2003), (Rapantzikos and Zervakis, 2001) showed excellent results (up to 100% accuracy) on automated hard drusen detection compared against manual segmentation by an expert grader. However, they did not specify if the method can perform with similar accuracy for soft drusen detection. Barkat and Madjarov (Barakat and Madjarov, 2004) did not categorize drusen types to measure AMD severity. Morgan et al. (Morgan et al., 1994) showed that a significant number of drusen areas were miscalculated; a major drawback of the system. Longitudinal changes were measured in Soliz et al. (Soliz et al., 2002). However, the agreement between the graders on different drusen sizes was only 67%.

3.1.1.3. Clustering-Based drusen segmentation.

Cluster analysis or clustering is the task of assigning a set of objects into groups so that the objects in the same cluster are more similar to each other than to those in other clusters. Clustering can be considered the most important unsupervised learning problem; so, as with every other problem of this kind, it deals with finding a structure in a collection of unlabeled data. Our search for existing literature on AMD detection based on clustering identified two papers, which are summarized in Table 3.

Table 3.

Summary of clustering based drusen detection methods.

Method	Image processing techniques	Level of segmentation	Accuracy (%)
Quellec et al. (Quellec et al., 2011)	Factor analysis to generate feature space and reference image distance based clustering	Drusen and non drusen	85
Hanafi et al. (Hanafi et al., 2010)	Spatial histogram and similarity based classification	AMD and non AMD image classification	77

Both methods focused on feature vector generation from intensity distributions and spatial histograms, and did not consider the validation for soft drusen detection or detection of other drusen subtypes.

3.1.1.4. Edge and template matching.

Edge detection defines the set of pixels that form the border of a region that separates it from neighboring regions. Edge detection is the approach used most frequently for segmenting images based on abrupt (local) changes in intensity. Most edge detection algorithms assess changes by finding the magnitude of the gradient of the pixel intensity values. In (Parvathi and Devi, 2007) edge-based soft drusen detection was attempted, however, it is affected by the faint boundary properties of soft drusen. The template matching method considers a template of part of or the full image to match with to find the region or image of interest, and can be performed through two approaches: feature-based and template-based matching (Gonzalez and Woods, 2008). Although edge and template matching-based methods work well for hard drusen segmentation, they fail to detect soft drusen due to the fuzziness on the edges. Table 4 summarizes the existing methods on edge and template matching-based drusen identification.

Table 4.

Summary of edge and template matching based drusen detection methods.

Method	Image processing techniques	Level of segmentation	Accuracy (%)
Mora et al. (Mora et al., 2011)	Illumination correction, normalization and Gaussian derivative	Drusen area detection	43
		Drusen pixel detection sensitivity	68
		Drusen pixel detection specificity	96
Remeseiro et al. (Remeseiro et al., 2009)	Gaussian template matching and thresholding	Drusen region identification sensitivity	83
		Drusen region identification specificity	87

Overall, the studies addressed AMD grading with drusen absence or presence, drusen pixel identification, and drusen size detection for dry AMD identification. The best methods that produced high accuracy are summarized in Table 5. However, the reported accuracies are not for the same datasets, and therefore cannot be compared to each other. Table 6

Table 5.

Drusen detection approaches and best methods on reported accuracy.

Classification of drusen	References	Accuracy
		Accuracy	Sensitivity	Specificity
Absent or present	(Agurto et al., 2011)	100%	–	–
Pixel level	(Brandon and Hoover, 2003), (Barakat and Madjarov, 2004)	87%	86%	93%
Drusen Region or size	(Barriga et al., 2009)	96%	–	–
Hard or soft drusen	–	–	–	–
Distinct or In distinct	–	–	–	–
Dry or Wet AMD	(Priya and Aruna, 2011)	96%	–	–

Table 6.

SD-OCT image analysis methods and reported accuracy.

Method	Image processing techniques	Level of segmentation	Accuracy or agreement (%)
Chiu et al., 2012	Graph and dynamic programming	Segment Retinal Boundaries	Diff. GT and Grader 1 is 4.2 ± 2.8 μm and GT and Grader 2 is 3.2 ± 2.6
		Drusen or GA	Mean diff. 1.60 ± 1.57%
Zhang et al., 2012	3D tube-like objects modeling for choroidal vessel and Markov boundary model for retinal layer thickness	Choroidal vessel	Corr. Coeff. 0.92
		Retinal layer thickness	Boundary diff. ± 10 μm
Liu et al., 2011	Multiscale texture and SVM	AMD pathologies (e.g., GA) absence or presence	Accuracy 94%
Ghorbel et al., 2011	Active contours, Markov random fields and Kalman filter	Layer segmentation	Diff. between the method and grader measurement is 4.75–12.5%.

3.1.2. Protocols for AMD grading using retinal color imaging

The international classification of grading system divides the macula area into three circles on the retina which are centered on the fovea, of diameters corresponding to 1000 μm, 3000 μm and 6000 μm (Bartlett and Eperjesi, 2007). Lesions are graded within each of the central, inner or outer circles. The predominant drusen type is classed as the most common type of drusen present within the outer circle. The drusen are classified as hard drusen with size <63 μm, intermediate soft drusen 63–125 μm, large semi-solid drusen 125–250 μm, 250–500 μm and >500 μm. Other pathologies such as hyperpigmentation, GA and neovascular AMD are graded as present or absent in the retinal image. The AREDS study defines the AMD category as 1 to 4; normal to severe AMD based on the drusen, GA, and CNV. The study defines AMD category 1 if there is no drusen existent or drusen size <63 μm and total area covered by drusen is < 125 μm. AMD category 2 or intermediate is defined as drusen size ≥63 μm and <125 μm and drusen area ≥125 μm but GA absent. AMD category 3a is defined as intermediate ≥63 μm and <125 μm, drusen area > 360 μm diameter circle if soft distinct drusen are present and 3b if soft indistinct drusen are present ≥656 μm diameter circle. AMD category 4 (4a and 4b) is advanced AMD with the presence of GA and CNV and visual acuity <20/32 due to AMD.

3.1.3. Existing AMD grading systems

Existing AMD grading systems based on color imaging (Klein et al., 1991),(Gregor et al., 1977),(Smiddy and S, 1984),(Strahlman et al., 1983) require a grader’s manual interaction, a very time consuming and tedious process. Scholl et al. (Scholl et al, 2003) proposed a semi-automated method for grading drusen and GA to measure AMD severity. AMD severity was graded using a grid to define macular subfields and standard circles to define the size of the lesions. Inter-observer variability was assessed by having three retinal specialists evaluate the color slides and re-grade the same set. The inter-observer agreement for hard drusen was 70–89% (kappa = 0.26–0.63) and intermediate soft drusen 76–94% (kappa = 0.27–0.69). Agreement ranged between 87% and 100% for larger drusen. Moderate to almost perfect agreement was reached for the presence of GA 88–98% (kappa = 0.60–0.95) and substantial to almost perfect agreement for the presence of CNV 84–100% (kappa = 0.62–1.00).

Although there has been a considerable amount of work done on drusen detection, none of it addresses drusen classification as hard and soft drusen and distinct and indistinct drusen, which are the key measures for early detection of AMD. The most challenging issues in drusen classification are the varieties of size, shape and intensity distribution for the drusen. In addition, it is often difficult to find the boundaries of drusen, and they can be difficult to reliably identify, as they have a similar appearance to other lesions, such as cotton wool spots. Some faint drusen can also appear similar to normal features of the retina, such as the background pattern caused by the normal choroidal vessels (Kose et al., 2010).

Existing techniques for AMD severity grading using color fundus imaging are mainly manual or semiautomatic. Due to the high degree of graders’ involvement, the methods may also involve inaccuracy and lack of repeatability (Niemeijer et al., 2007). Although a number of methods have been proposed for drusen detection (Rapantzikos et al., 2003),(Niemeijer et al., 2007),(Kose et al., 2010),(Kösea et al., 2008) they mainly focus on detecting the presence or absence of drusen. However, early and intermediate AMD detection require drusen classification (e.g., soft and hard drusen, distinct or indistinct drusen) along with the macular region mapping. To the best of our knowledge, no automatic method has been proposed to classify drusen as soft or hard, soft distinct or indistinct and other drusen types. An automatic method should provide higher accuracy and repeatability on grading the pathologies, which is the most important aspect for AMD screening.

3.2. Fundus Autofluorescence imaging

3.2.1. Fundus Autofluorescence imaging for late “dry” AMD or GA

The literature search identified one automated segmentation method for GA that also grades manually for dry AMD severity, as proposed by Lee et al. (Lee et al., 2007). The manual method aims to segment GA to determine the AMD severity. The method first applies a non-linear adaptive smoothing operation to correct background illumination. A level set framework is then used to identify the hypo-fluorescent areas. Finally, an energy function combined with morphological scale space analysis and a geometrical model-based approach is performed to refine the hypo-fluorescent areas. The clinically apparent areas of hypo-fluorescence were drawn by an expert grader and compared on a pixel-by-pixel basis to the automated segmentation results, which achieved a sensitivity of 89% and specificity of 98%.

3.3. Spectral domain Optical Coherence Tomography (SD-OCT)

In this review, we focus on different OCT image analysis techniques for measuring AMD pathologies such as drusen, GA and CNV.

3.3.1. SD-OCT for early or late “dry” AMD

Baumann et al. (Baumann et al., 2010) proposed a method for segmentation and quantification of drusen and GA from polarization sensitive OCT. The method first locates the RPE based on depolarizing characteristics (i.e., polarization scrambling). Drusen are segmented by first locating the actual position of the RPE by its degree of polarization (DOPU). The line found by tissue-specific contrast then acts as a backbone for the subsequent calculation of the position where the RPE should be in a healthy eye. Finally, drusen are segmented and quantified by calculating the difference between the actual position and the normal ‘should be’ position (Fig. 9).

Fig. 9.

Example of drusen detection. Panel (a) presents an x–z slice running through the pathology in SD-OCT volume. Footprint of drusen probability map in x–y, are presented in b and c, respectively.

The GA zones in “dry” AMD can be detected by detecting holes in the depolarizing RPE pixels. This is achieved by simply summing the number of polarized pixels for each A-line of the 3D PS-OCT data set. The resulting thickness map of the depolarizing layer then reveals the existence and absence of polarizing scrambling tissue in an en face view. To prevent integrating choroidal pixels into computing the thickness map of the depolarizing layer, only depolarizing pixels located in a shallow band close to the photoreceptor layer are used. The method did not provide any significant analysis for accuracy measurement. The repeatability measurement shows that the variation on drusen area and volume measurements are 7.5–7.7% and 9.2% for GA.

3.3.1.1. Graph-Based drusen and GA segmentation.

Chiu et al. (Chiu et al., 2012) proposed a method for automatic segmentation of RPE and drusen from SD-OCT images for the longitudinal study of AMD. The method utilizes graph and dynamic programming to segment three retinal boundaries in SD-OCT images with drusen and geographic atrophy. The graph search (GS) and graph cut (GC) methods were synergistically combined to segment the exudates associated derangements (SEAD) i.e., drusen and GA. The segmentation problem can be formulated as an energy minimization problem such that for a set of pixels P and a set of labels L, the goal is to find a labeling $f:p\to L$ that minimizes the energy function $Enf(f)$.

The graphs are constructed considering the superior surface, inferior surface and region. The three subgraphs are then merged together to form a single graph. The segmentation accuracy was determined for layer thickness measurement and drusen or GA volume measurement. The layer thickness (Fig. 10) accuracy was measured by two certified graders which showed mean differences of 4.2 ± 2.8 μm and 3.2 ± 2.6 μm for automatic versus manual segmentation. The mean differences in the calculated total retina and drusen volumes were 0.28 ± 0.28% and 1.60 ± 1.57% respectively. The method showed very high efficiency: 1.7 s compared to 3.5 min manually.

Fig. 10.

Segmentation results of 11 retinal surfaces (ten layers).

Other methods which showed high accuracy and repeatability based on retinal layer thickness measurement and can be used to find abnormalities and/or for change detection due to drusen, GA and CNV are as follows. A fully automated 3D SD-OCT segmentation method for choroidal vessels, quantification of choroidal vasculature thickness and choriopillaris-equivalent thickness of macula is presented in (Zhang et al., 2012). A graph-based multilayer segmentation approach was utilized to segment intra-retinal surfaces. Choroidal vessels were modeled as 3D tube-like objects in a resampled sub-volume that yielded isometric (cubic) voxels. The method shows good reproducibility in choriocapillaris-equivalent (coefficient of variation 8%) for normal subjects. Koozekanani et al. (Koozekanani et al., 2001) proposed an automated algorithm for retinal thickness measurement from OCT images using the Markov boundary model. The detected boundaries from the automated method were compared with manually corrected boundaries, which showed that the method determined the thickness with 10 μm difference (from the manual method) for 74% of images (out of 1450 test images). For 98.4% of the images, the method detected the thickness within 25 μm difference.

3.3.1.2. Texture-based GA segmentation.

Liu et al. (Liu et al., 2011) proposed an automated method to identify normal macula and AMD from SD-OCT images using multi-scale texture and shape features. A fovea-centerd cross-sectional slice for each of the SD-OCT images was encoded using spatially-distributed multiscale texture and shape features. Three ophthalmologists labeled each fovea-centerd slice independently and the majority opinion for each pathology was used as the baseline decision. Two-class support vector machine classifiers were trained to identify the presence of normal macula and each of the three pathologies (macular hole, macular edema and AMD) separately. The method showed good accuracy with an area under the receiver operating characteristic (ROC) curve of 0.94.

3.3.1.3. Active contour-based drusen segmentation.

Yi et al. (Yi et al., 2009) performed a quantitative evaluation of drusen and associated structural changes in non-neovascular AMD using SD-OCT images. They used a method (Mujat et al., 2005) which determines the retinal nerve fiber layer (RNFL) thickness in OCT images based on anisotropic noise suppression and deformable splines. A qualitative assessment was performed to report the consistency of the method. However, no quantification was performed or correlation of the automated and manual grading was shown.

Ghorbel et al. (Ghorbel et al., 2011) developed an automated method for the segmentation of eight retinal layers in an SD-OCT image. Fig. 7 shows the layers. The method is based on active contours and Markov random fields. A Kalman filter is applied to model the approximate parallelism between the photo receptor segments. The method was validated by comparing against manual segmentation performed by five different graders. Results showed that the thickness measurement for the layers can vary between the automatic and manual by up to 4.75–12.5%. However, they did not use any images from AMD patients for the validation.

3.3.1.4. Gradient-Based GA segmentation.

An automated method for intra-retinal boundary segmentation in SD-OCT was proposed for fast and reliable quantification GA (Yang et al., 2010). Local and global gradient information is adopted simultaneously with a shortest path search to optimally find edges. The method was validated against manually graded images for accuracy and reproducibility, and showed high accuracy and reproducibility with faster processing time. There may be potential for AMD severity identification.

3.3.2. SD-OCT for exudative AMD

Choroidal vasculature in a standard clinical OCT scan is shown in Fig. 9. In (Giovannini et al., 1999), the authors demonstrated the use of OCT for displaying a neovascularization at the RPE. The findings suggest that the surgical removal of early age-related CNV could be performed in those cases where the OCT shows a neovascular membrane on the RPE, as in idiopathic and inflammatory CNVs. However, surgical removal of CNV is no longer state-of-the-art.

A review on recent developments in OCT for retinal imaging is provided in (Drexler and Fujimoto, 2008), (Velthoven et al., 2007) which discusses image acquisition and functionality with respect to the strengths and weaknesses of different OCT technologies for determining retinal structure and blood flow capturing capabilities. Algorithms and methodologies for retinal thickness measurement can be found in (Koozekanani et al., 2001),(Song et al., 2012),(Tatrai et al., 2011),(Ghorbel et al., 2011).

In summary, there continues to be much room for improvement in imaging retinal layer segmentation, drusen and GA volume measurement using OCT imaging. Accuracy measurements for the drusen and macular layer segmentation algorithms for AMD are provided in (Song et al., 2012),(Patel et al., 2009),(Schlanitz et al, 2011), (Loduca et al., 2010),(Ahlers et al., 2010). Manual or semi-automated algorithms for drusen volume measurement for AMD severity are described in (Khanifar et al., 2008),(Freeman et al., 2010),(Schuman et al, 2009). Among the GA quantification methods, (Fleckenstein et al., 2010a,b),(Fleckenstein et al., 2008), (Margins, 2009),(Lujan et al., 2009),(Yehoshua et al., 2010a),(Yehoshua et al., 2010b) are manual or semi-automatic and mainly focus on GA progression for AMD severity measurement. Further, an automated analysis of OCT images can be found in (Wilkins et al., 2012) which measures intra-retinal cystoids fluid. A comparison of SD-OCT devices and 3D drusen analysis software can be obtained from (Kaiser, 2011). An SD-OCT 3D volumetric analysis would also be applicable to SDDs as well as ordinary drusen, but this has not yet been implemented.

3.4. Fluorescein angiogram or angiography

3.4.1. Fluorescein angiogram for “dry” AMD

A few methods have been developed for the automated analysis of angiography images for detecting drusen and GA. Clinically, it has been shown that soft drusen and macular subretinal new vessels are associated in a statistically significant manner (Thaibaoui et al., 2000). Therefore, these methods aimed to detect drusen detection from angiographic images, utilizing one imaging modality for both “dry” and “wet” AMD pathology. Thaibaoui et al. (Thaibaoui et al., 2000) developed a method from retinal angiographic images which is based on image partition and fuzzy logic. At first, the method partitions the image into different classes with respect to a probability density function of the input gray-levels. To improve the homogeneity of the obtained classes, a local parametric gray-level transformation of the classes is applied. This mainly transforms the image into three classes, which are background, fuzzy region and drusen. Fuzzy c-Means (FCM), a data clustering technique, defines each data point as belonging to a cluster to some degree. This is specified by a membership grade. Using fuzzy clustering rules the regions are classified into background and drusen. The method was not evaluated quantitatively or gave any comparison with ground truth images.

Sbeh et al. (Sbeh et al., 2001) proposed a method for longitudinal change detection in fluorescein angiography which is based on mathematical morphology. Regional maxima components of the image which correspond to the regions inside the drusen are obtained. Shape, area and contrast of the detected regions are analyzed to decide whether these regions are drusen. Following these, non-rigid image registration is applied using the vascular structure to observe changes in the longitudinal images. A qualitative analysis was provided on change detection, however no quantitative result was provided nor any classification results of drusen given.

3.4.2. Fluorescein angiogram for “wet” AMD

A fully-automated approach for change detection of CNV has been proposed in (Narasimha-Iyer et al., 2006). In this work, the changes in CNV regions as well as the RPE hypertrophic regions were detected and analyzed to study the progression of disease and effect of treatment. Retinal features including the vasculature, vessel branching/crossover locations, optic disk and location of the fovea are first segmented automatically. The images are then registered to sub-pixel accuracy using a 12-dimensional mapping that accounts for the unknown retinal curvature and camera parameters. Spatial variations in illumination are removed using a surface fitting algorithm that exploits the segmentations of the various features. The changes are identified in the regions of interest and a Bayesian classifier is used to classify the changes into clinically significant classes. The automated change analysis algorithms were found to have a success rate of 83% in detecting the changes in CNV regions compared with manual segmentation.

3.5. Indocyanine Green (ICG) angiogram or angiography

3.5.1. ICG for “wet” AMD

A number of automatic methods (Miki et al., 2007), (Nishiyama et al., 2001) have been reported on choroidal vascular change detection. Miki et al. (Miki et al., 2007) measured the brightness of diffuse fluorescence at the macula from ICG angiograms at four and 16 s after dye filling. The ratio of the averaged brightness of the macula to that of the disc (ch/d) was calculated for each angiogram. Groups of people: healthy young volunteers, old AMD and old non-AMD persons, were considered. There were significant differences in the ch/d ratios among the three groups four seconds after dye filling, which showed that ICG may be helpful for objective evaluation of the choroidal circulation in chorioretinal diseases. A quantitative analysis of ICG angiographic images in central serous chorioretinopathy (CSC) was presented in (Nishiyama et al., 2001). Using a Topcon IMAGEbet(R) camera, the maximum diameter of the choroidal veins and intensity of background fluorescence in the posterior fundus were measured with ICG video images. The study showed that choroidal venous dilation and the residual background fluorescence in the posterior fundus might be positive indicators reflecting the pathogenesis of CSC.

A study has been performed to identify the incidence of retinal choroidal anastomoses in patients with occult CNV and focal hot spots on ICG angiography to identify the clinical and angiographic features (Slakter JS, 2000). The study identified specific clinical and angiographic features that can aid in the diagnosis of these vascular anomalies. A case report (Chang et al., 2004) has been published to show correlation between ICG angiographic findings in patients with exudative AMD with the histological localization of ICG to the surgically excised CNV. The report showed that ICG angiography allows diagnosis of CNV with high confidence.

An examination of the in vitro interaction of human RPE cells with ICG was performed and reported in (Chang et al., 2005). A comparative study using ICG angiograms and intravenous sodium fluorescein angiography (IVFA) was reported in (Dzurinko et al., 2004) which showed that ICG is better than IVFA. Simultaneous visualization of an en face (coronal, C scan) OCT image and an ICG angiogram, displayed side by side and superimposed, permits more precise correlations between late fluorescence accumulation and structures deep in the retinal surface at the retina–choroid interface (Rosen et al., 2009). The paper also reported that SLO-based ICG is particularly effective in finding small CNV lesions masked by fluorescein staining.

3.6. Thermal/Infrared (IR) imaging

At present, IR image analysis for diagnosing AMD as neovascular or exudative is mainly performed by graders (qualitative analysis). Although a number of automatic methods have been proposed in (Schardosim et al, 2011),(Kolar et al., 2008),(Devisetti et al., 2011; Schardosim et al, 2011), they have not been tested in any clinical applications.

3.6.1. IR for “dry” AMD

Existing IR-based “dry” AMD diagnosis techniques focus on GA segmentation, change detection with image registration and motion detection. In (Schardosim et al, 2011), the motion detection method was proposed to remove background noise and illumination from IR imaging. (Kolar et al., 2008) proposed a method for IR and autofluorescence image registration which is based on noise suppression, edge detection, rigid or flexible transformation and interpolation. Devisetti et al. (Devisetti et al., 2011) developed a method for GA segmentation using intensity distribution and neural network-based GA pixel classification.

Elsner et al. (Elsner et al., 1996) performed a comparative study of IR and scanning laser ophthalmoscope for acquiring subretinal structures. Their study showed that IR provides better visibility than visible light. Kellner et al. (Kellner et al., 2010) showed that near infrared (NIR) is equally capable of detecting RPE alteration as FAF. They also showed that patterns from FAF and NIR indicate different involvement of lipofuscin and melanin in the pathophysiological process and provide further insight into the development of AMD.

A qualitative assessment for neovascular AMD was provided for near IR imaging in (Theelen et al., 2009) which was based on graders’ observation. The paper concluded that NIR imaging in neovascular AMD revealed characteristic alterations depending on the type of CNV. These changes may reflect histological differences of the lesions. Weinberger et al. (Weinberger et al., 2006) demonstrated that NIR autofluorescence is capable of detecting maculopathy as a single sensor. In (Forte et al., 2012), the authors showed that near-infrared autofluorescence might detect areas of RPE cell loss at the GA margin earlier than fundus fluorescence. Semoun et al. (Semoun et al., 2009) analyzed an IR feature for CNV, a whitish ring that is correlated to the edge of the lesion.

3.7. Hyperspectral Retinal Imaging

3.7.1. Hyperspectral Retinal Imaging for “dry” AMD

Lee et al. (Lee et al., 2010) proposed a method for capturing hyperspectral signatures of drusen and macular pigment. The method adopted non-negative matrix factorization (NMF) for finding low rank decomposition that captures the underlying physiology of drusen and macular pigment. The feasibility of hyperspectral mapping of macular pigment is also investigated in (Fawzi et al., 2011). The study showed that snapshot hyperspectral imaging in combination with advanced mathematical analysis provides a simple, cost effective approach for macular pigment mapping in vivo. Davis et al. (Davis et al., 2007) demonstrated that hyperspectral images of macular tissue are spectrally different in different forms of AMD.

Hyperspectral imaging-based techniques for monitoring oxygen saturation in the optic nerve head and overlying vessels of monkey eyes has been reported in (Khoobehi et al., 2004).

Most traditional hyperspectral retinal cameras are scanning-based systems which cause severe motion artefacts and pixel misregistration problems because of the constant micro movement of the human eyes’ (Gao et al., 2012). To overcome this limitation, a snapshot hyperspectral retinal camera which utilizes a computed tomographic imaging spectrometer (CTIS) has recently been developed for fundus imaging (Johnson et al., 2007). However, CTIS is computationally very expensive and is limited by resolution constraints. Considering this, a parallel acquisition hyperspectral imager has been developed which uses an image mapping spectrometer (IMS) (Gao et al., 2012). Integrating an IMS with a traditional retinal camera, 48 simultaneous channel imaging of a human retina was achieved, with measurement of reflectance and absorption spectra from vessels and macula. This system also allowed measurement of the spectral reflectance of an optic disc druse, which was previously unknown. In summary, this approach is still in the early stage of development, without widespread clinical application.

4. Comparative study of the AMD imaging modalities and imaging techniques

Each imaging modality has strengths and weaknesses for diagnosing individual AMD pathology. Here, we focus specifically on the strengths and weaknesses of an imaging sensor to capture AMD pathology. Currently, color fundus imaging represents the gold standard for looking at the retina for initial assessment of AMD. A number of study protocols have been established for AMD severity grading, which were discussed earlier. However, with the recent advancement of new imaging modalities, more diagnostic information can now be extracted from the retinal layers for AMD. For example, SD-OCT can capture volumetric information on drusen and other retinal fluids. IR imaging has strengths for investigating sub-retinal structures in the presence of hemorrhage or cataracts that may be undetected in other in vivo studies.

A number of studies (Lujan et al., 2009),(Jain et al., 2010),(Zweifel et al., 2010),(Rosenfeld et al., 2011),(Brar et al., 2009) have been performed on showing the correlation for AMD pathologies obtained from two different imaging modalities. Among them Jain et al. (Jain et al., 2010) and Rosenfeld et al. (Rosenfeld et al., 2011) showed a high correlation between color fundus photographs (CFP) and SD-OCT images for drusen presence and absence determination. There was also a trend towards higher rates of drusen detection with SD-OCT in eyes with larger drusen and hyper pigmentation, with CFP trending towards greater detection of smaller drusen.

4.1. Autofluorescence and IR vs. angiography/color fundus imaging

Metabolic changes at the level of the photoreceptor/RPE are complex and may not be able to be visualized using fundus cameras or other routine imaging techniques such as FA. Kellner et al. (Kellner et al., 2010) showed that NIR is equally capable of detecting RPE alteration as FAF. In (Forte et al., 2012), the authors showed that NIR autofluorescence might detect areas of RPE cell loss at the GA margin earlier than FAF.

4.2. SD-OCT vs. color fundus imaging

SD-OCT is better than other imaging modalities at capturing information non-invasively and for the quantitative assessment of all forms of AMD (Yehoshua et al., 2010b). SD-OCT enables the physicians to obtain reproducible and quantitative data to follow the disease as it progresses from drusen to both GA and CNV. SD-OCT allows the accurate measurement of drusen area, which was a drawback for AREDS (size was measured to avoid difficulty in the measurement (Yi et al., 2009)). However, SD-OCT only shows better performance for larger drusen i.e., for advanced AMD rather than for the smaller drusen: precisely identifying the borders of drusen is still a challenge (Jain et al., 2010). In contrast, CFP provides better information for smaller drusen. Therefore, CFP is useful for screening and early detection of AMD.

4.3. OCT/TD-OCT vs. SD-OCT

Non-exudative or “dry” AMD does not directly benefit from OCT in visualization. SD-OCT is better than OCT in determining details of the pathologies and faster (400 vs. 20,000–40,000 A scans) for capturing the data.

A comparative study between OCT/TD-CT and SD-OCT was performed in (Domalpally et al., 2010),(Gupta et al., 2008),(Kiernan et al., 2010). SD-OCT tends to derive increased retinal thickness and decreased nerve fiber layer thickness measurements. This is likely because of the increased depth resolution and greater volume of data acquired with each scan. SD-OCT has advantages over TD-OCT by providing better identification of normal and pathologic structure in patients with poor media clarity. SD-OCT is also faster than TD-OCT and better for noise handling.

4.4. SD-OCT vs. FAF

The size and shape of GA areas measured by SD-OCT and FAF correlate well (Lujan et al., 2009),(Brar et al., 2009),(Bearelly et al., 2009),(Sohrab et al., 2011). The strength of SD-OCT is that it provides important cross-sectional anatomical structures. In non-neovascular AMD with GA, SD-OCT provides adequate resolution for quantifying the loss of the photoreceptor (PR) layer. It may also serve as a means of tracking disease progression.

4.5. Fluorescein Angiography vs. ICG angiography

ICG angiography has several advantages over FA in imaging choroidal vasculature. Because of the poor transmission of fluorescence through ocular media opacification, fundus pigmentation and pathologic manifestations (e.g., lipid exudation), FA has limitations in imaging choroidal circulation. Recently, ICG angiography has shown that it is more useful than FA for detecting CNV in neovascular AMD (Klais and Yannuzzi, 2004).

A synopsis for the various imaging modalities for automatic analysis of AMD pathologies:

• SD-OCT is better for the measurement of size, area or volume of drusen and other pathological objects in the retina.

• CFP has strength over SD-OCT in the accurate measurement of smaller drusen. SD-OCT has better accuracy for larger drusen and hyper pigmentation measurement.

• GA using FAF is sufficient, however SD-OCT is better as it provides cross-sectional anatomical information.

• IR imaging is less harmful for the eye, hence increased illumination can be used that may be valuable in increasing image quality.

• ICG angiography is useful for detecting CNV in neovascular AMD.

5. Telemedicine for AMD screening and computer-aided automation

Telemedicine models in ophthalmology utilize telecommunication andinformation technologies in order to providespecialist eye care at a distance. Telemedicine-related eye care and screening could streamline the referral process, reduce waiting times and lead to overall savings in healthcare costs. It can enable early detection of sight-threatening eye diseases (Yogesan et al., 2006), (Yogesan et al., 2010) through facilitating mass screening. Several feasibility studies demonstrate the enormous usefulness of the technology (Abràmoff and Suttorp-Schulten, 2005),(Joshi and Sivaswamy, 2011),(Lorenz et al., 2009),(Liesenfeld et al., 2000),(Murakami et al., 2009).

Telemedicine has proven particularly effective for diabetic retinopathy screening. In the US and worldwide, models of diabetic care in both large and small scale health care systems, have emphasized diabetic screening clinics. Often using telemedicine and systematic organizational procedures, these models have substantially improved the ocular screening rate for retinopathy to approximately 57% in the commercial insurance sector, 66% in the Medicare population, 53% for the Medicaid insured, and 90% in integrated health care organizations such as the US Department of Veterans Affairs (Hedis, 2012), (Taylor et al., 2007), (Yogesan et al., 2012), (Li et al., 2011).

However, such screening programs for AMD are virtually nonexistent, especially in light of its high disease prevalence and social impact. Additional studies in this area of telemedicine, defining optimal conditions for AMD screening and referral, are needed. In theory, telemedicine-based screening of individuals over 55 years old could detect the presence of AMD at a level requiring referral to an ophthalmologist for further examination. However, from the viewpoint of a healthcare system or model, there is a need to minimize unnecessary referrals and increase the cost-effectiveness of screening by referring only those with a level of pathology for which intervention is currently possible. Telemedicine and ancillary tools could conceivably also be used for remote monitoring of patients with “wet” AMD post-treatment, a model that has the potential to result in close monitoring of anatomic and visual outcomes, combined with a reduction in patient travel.

Our literature search found only a few telemedicine-based AMD screening programs. Pirbhai and Sheidow (Pirbhai and Sheidow, 2004) showed that telemedicine can be used to diagnose AMD patients with high confidence using non-stereo CFP. The automatic system could achieve 89.2% sensitivity and 99.1% specificity for detecting CNV. For the presence of RPE GA the specificities were up to 86.8%. Very few treatable lesions were missed in AMD patients.

Ulrich et al. (Ulrich et al., 2009) conducted a telemedicine trial from Nepal to the USA to provide AMD and diabetic retinopathy-related consultations using fundus images. They defined early AMD as any type of non-exudative changes with a visual acuity >20/200 and advanced AMD as any exudative changes or advanced atrophic disease with a visual acuity ( ≤ 20/200).

Zimmer et al., 2005 (Zimmer-Galler and Zeimer, 2005) explored the use of a telemedicine-based semi-automated screening system, designed to be in primary care physician’s offices, to identify high and low-risk AMD patients (21 patients). They concluded that it is possible to identify low risk lesions with good agreement and high-risk lesions with excellent agreement compared to gold standard evaluation of stereoscopic film photographs.

Telemedicine based screening could conceivably be combined with automated image analysis and grading to facilitate early detection and referral by trained screeners. The addition of automated decision support could provide guidance to remote primary care providers and nurses in order to locally manage patients with eye diseases. Telemedicine technology combined with computer-aided grading may be a more effective care model for developing countries which have large affected populations and fewer ophthalmologists (Eccles, 2012). Validated tele-screening could allow prioritization of patients based on clinical urgency, with apparent “wet” AMD lesions being detected and treated sooner than patients whose conditions are rated as “dry” AMD.

OCT imaging is an ideal imaging modality for telemedicine-based screening and management (e.g., to determine whether further VEGF inhibitor injection is indicated) of AMD and other retinal diseases. It is non-invasive and relatively straightforward for a technician to use. However, the cost of the device and its relative immobility currently discourages it from being used outside eye clinics,and therefore as a component of a tele-AMD program. Low cost (probably < $10,000) portable OCT imaging devices are needed for use in telemedicine consultations. Another issue that arises with OCT imaging is that the 3D scans produce large image files that need sufficiently high bandwidth for transmission. Ongoing research opportunities exist to explore how to efficiently compress OCT scans without loss of diagnostic information. A lossless compression algorithm for maintaining detailed information in OCT scans is a key part of the telemedicine platform to enable real-time diagnosis.

Electronic medical records (EMR) are becoming an integral part of any medical or eye clinic seeking to effectively manage patient care and improve clinical outcomes. Integrating health information digitally, both data and images, reinforces the importance of ocular imaging devices that may be directly connected to the EMR. Collected digital information includes color fundus and OCT images that may be archived with other clinical data. The EMR data and images could be located in the data cloud or at a remote server. In principle, ophthalmologists could access the image-rich EMR locally or remotely to provide diagnostic advice and triage. When storing fundus images and clinical data in the cloud, appropriate encryption must be implemented to protect sensitive patient information, and this issue has somewhat limited cloud and wireless-based medical communication in large organizations.

Presently, it is not uncommon for image analysis and data processing to be performed in the data cloud, to avoid installing individual software programs on each computer. The images are transferred to the cloud for analysis. Other implications of such a model of care include the need for high bandwidth, as image transfer and processing must be rapid if one wishes to provide real-time disease grading and clinical decision support to screeners. An advantage of such a cloud-based data model is the relative ease with which developers may upgrade and maintain the software remotely.

As ophthalmology is strongly an image-based medical speciality, the store-and-forward modality is widely used in teleophthalmology systems rather than real-time video conferencing. High resolution retinal images could be stored and transmitted to a remote site for reading by an ophthalmologist. This could remove the immediate need for high bandwidth communication lines and real-time consultations. However, in some countries (e.g., Australia), only real-time video consultations are approved by Medicare for reimbursement. Therefore, unless laws are changed, a hybrid system with both store-and-forward and video conferencing would be necessary. However, it is cumbersome and relatively expensive to transmit high-resolution color images in real time and also conduct simultaneous video conferencing. One solution is to transmit high-resolution color images in advance through a store- and-forward technique. After analyzing the retinal images and other data, the ophthalmologist could speak to the patient and referring health care provider using video conferencing.

Conventional video conferencing technology, while relatively widespread in industrialized countries, carries its limitations of image resolution and bandwidth as a relative barrier between patient and ophthalmologist. The traditional video conferencing system has one camera and a screen, and does not give the impression carried by a face-to-face consultation. In the future, more advanced models of video conferencing, using multiple cameras and screens, conceivably with the implementation of augmented reality, could well provide the patient and ophthalmologist the impression of a more “real” contact, and an enhanced telemedicine experience. In principle, there is no reason that retinal images could not be viewed in a three dimensional format as well.

The development of standards for tele-AMD consultation are important to provide uniform care for patients and allow comparability of studies. Additional telemedicine studies are required for the screening and remote management of AMD in order to develop appropriate tele-AMD standards. The standards could fall into two domains:

1. Technical

   1. Data Capture (imaging devices and other diagnostic devices)
   2. Storage (image and data compression)
   3. Transmission (protocols, smart phones)
   4. Processing (image processing and automation)
   5. Quality (display – screen properties)
2. Health application

   1. Assessment (clinical guidelines, screening consultations)
   2. Diagnosis (reporting guidelines, remote testing and imaging, decision making and expert consultation)
   3. Treatment (formulation of care plans, prescribing and medication)
   4. Management (execution and modification of care plans)
   5. Monitoring (recording from medical devices, analysis of data, images etc.).

In summary, while the AREDS clinical trials validated the concept that retinal images and their grading extend direct patient examination, widespread telemedicine screening for AMD awaits advances in medical informatics, image processing, and health applications (Ferris et al., 2005), (Davis et al., 2005).

6. Clinical trials for identification and follow-up of AMD patients

A number of studies have been performed on AMD progression and the effect of vitamin supplementation as well as intravitreal drug injection (Klein et al., 2002; Bartlett and Eperjesi, 2007; Age-Related-Eye-Disease-Study-Research-Group, 1999). The Beaver Dam Eye Study showed an overall 10 year incidence rate of 12.1% for early AMD (defined as drusen and retinal pigment epithelial changes) and 2.1% for late AMD (defined as exudative AMD or central GA) (Klein et al., 1992).

The revolution in care of neovascular AMD, however, is the result of clinical trials which demonstrated statistically significant improvement in both visual acuity and anatomical results (Rosenfeld et al., 2006). This study evaluated Ranibizumab – a recombinant, monoclonal antibody, for the treatment of neovascular CNV. Results showed that intravitreal administration of ranibizumab for two years prevented vision loss and improved mean visual acuity, with a low rate of adverse effects, in patients with choroidal neovascular (CNV) AMD. In another study (Keane et al., 2008), the effects of ranibizumab on retinal morphology in patients with neovascular AMD was demonstrated using OCT. OCT images were analyzed using custom software that allows precise positioning of pre-specified boundaries on anatomical data. Changes in the thickness/volume of the retina, sub-retinal fluid (SRF), sub-retinal tissue (SRT), and PEDs were calculated at week 1 and at months 1, 3, 6, and 9 after treatment. The findings indicated that neurosensory retinal edema and SRF showed early reductions after the initiation of Ranibizumab therapy, and the effect on the retina was attenuated over time, suggesting possible tachyphylaxis. PED volume showed a slower but progressive reduction.

An earlier study has also been performed to determine whether intravitreal bevacizumab (Avastin, a monoclonal antibody from which ranibizumab is derived) could improve OCT results and the visual acuity outcome in a patient with neovascular AMD. (Rosenfeld et al., 2005). Within one week, OCT results demonstrated the resolution of SRF, resulting in a normal appearing macular contour, and the macular appearance and visual acuity was maintained for at least four weeks. Subsequently, a number of studies were performed on safety, biological effect, and a possible mechanism of action of intravitreal bevacizumab in patients with neovascular AMD (Avery et al., 2006),(Bashshur et al., 2006). These studies reported that the intravitreal injection has no significant ocular or systemic side effects. One week after injection, most patients had a reduction of baseline macular thickness, and after four weeks 30 of 81 eyes demonstrated complete resolution of retinal edema, SRF, and PED. Of the 51 eyes with an eight-week follow up, 25 had complete resolution of retinal thickness, SRF and PED. A number of comparative studies have been performed to evaluate the relative efficacy and safety of treatment of neovascular AMD with ranibizumab (Lucentis) and bevacizumab (Avastin) on a fixed and variable schedule (Group et al., 2011), (Catt-Research-Group et al., 2011), (Rosenfeld, 2006). These studies utilized ranibizumab as the trial agent and demonstrated that it is an effective treatment for neovascular AMD.. These and other studies demonstrate that while the risk of injection is low, ocular complications such as endophthalmitis, vitreous hemorrhage, and retinal detachment are possible, as are slightly increased rates of subsequent myocardial infarction and cerebrovascular accidents, compared to those not undergoing injection. Following the encouraging clinical trial results with ranibizamab, several investigators began evaluating intravitreal bevacizumab for the treatment of CNV. Given its molecular similarity to ranibizamab, its low cost ($50 compared to $1,950), and its availability, the interest in Avastin has been considerable (Catt-Research-Group et al., 2011). However, cautionary reports of widespread endophthalmitis resulting from inappropriate pharmacy preparation have made adherence to compounding standards mandatory (Goldberg et al., 2012)

There is currently no approved treatment for the “dry” type of AMD. However, the numerous AREDS studies (AREDS, 2001), (AREDS, 2007) have early or late “dry” AMD. However, the AREDS has shown that in patients with intermediate and advanced AMD there is a statistically significant reduction in the risk of progression to advanced AMD with antioxidant and zinc supplementation, and this effect persists for 10 years in approximately 25% of appropriately treated individuals (Chew et al., 2013).Clinical innovation in the area of “dry” AMD is ongoing. A selected investigational drug, ACU-4429 (Acucela, Bothell, WA) has been used for investigating the downregulation of photoreceptor activity. The ongoing phase 1 study has thus far shown that the drug is safe and well tolerated in healthy volunteers. A clinical trial looking at the preservation of photoreceptors and the RPE was followed by a therapeutic strategy used for the treatment of Alzheimer’s disease (Rosenfeld, 2009). An antibody against amyloid β was used in a phase 1 study as an intravenous treatment for GA in AMD patients. This antibody, known as RN6G (Pfizer, New York, NY), was shown to decrease the amount of amyloid β in the eye from a mouse model of AMD when given as a systemic therapy (Ding et al., 2008). Other studies are evaluating newer treatments to prevent the progression from nonexudative to exudative AMD.

7. Limitation of current imaging modalities and image analysis techniques

The following limitations were highlighted within the current systems by the symposium on advances in optical imaging and biomedical science (2009):

• need for faster camera

• need for eye motion correction

• need for improved estimates of light safety for the eye

• integration of imaging technologies in the next generation of super cameras

• improved, less noisy SD-OCT imaging.

The development of SD-OCT alone has increased speed by roughly two-fold and frequency-swept OCT provides additional gains. These increases notwithstanding, image acquisition remains a fundamental limitation to the scientific and clinical value of the new imaging technologies. One reason that speed is so important in ocular imaging is because the eye is always in motion due to naturally occurring eye movements. The problem can be exacer-bated in patient populations that have inferior fixation ability. The shorter the image acquisition time, the less image blur and distortion due to eye motion. A complementary approach to increasing speed is to correct for eye motion. Eye motion correction can be achieved either through active tracking and image stabilization during image acquisition, or through clever post-processing to remove eye movement artifacts.

Given the need for speed outlined above, parallel development of automated analysis software is also required. This includes software for eye motion correction, visualization, and feature analysis such as the development of improved segmentation algorithms for identifying specific retinal layers (particularly from OCT images) that are disrupted in retinal disease. Software that allows repeatability and reproducibility from the same patient would facilitate longitudinal studies of disease progression and/or efficacy of therapy.

Another need for improvement is the estimation of light safety. The more light one can put onto the retina, the faster an image can be acquired and the better the quality of the image. However, there are very important safety limits for light intensity. Recent evidence suggests that the light safety standards do not adequately protect the eye under some conditions.

8. Conclusions and future directions

In this review, we described current ocular imaging modalities and their strengths and weaknesses for extracting different pathological information relevant to the clinical care of AMD. We also discussed existing methods for AMD screening through the use of automated or semi-automated measurement of pathologies. Our study found that the majority of automated or semi-automated methods focus predominantly on grading drusen using different imaging techniques, as drusen are among the first clinical signs of AMD. Multi-modal imaging studies and their strengths and weaknesses were also discussed. Use of these tools have the potential to increase the sensitivity and specificity of screening in determining the severity of AMD. We also outlined the potential for future retinal imaging systems and the next generation of AMD grading tools for monitoring disease progression.

The potential for widespread AMD screening and diagnosis would likely benefit with a focus on research in automated grading and mass screening using a telemedicine platform. Such a validated tool would be highly effective for the early detection of AMD by nurses or trained screeners, as it is for diabetic retinopathy. Currently, the sole established treatment option for “dry” AMD is to detect the disease at a stage where AREDS formulation may provide a measure of treatment to prevent further degeneration and vision loss. Therefore an important interventional focus should be on the logistics of large-scale screening for AMD detection, targeting both urban and rural populations. Although there are few studies on tele-AMD, the AREDS studies did establish the usefulness of standardized images in grading disease and the effectiveness of AREDS formulation in treatment of selected individuals. Additional studies are needed to demonstrate the utility of telemedicine for the widespread screening and management of AMD. OCT imaging could be an excellent ancillary tool for tele-AMD applications, but their role in remote and rural clinics remains to be defined.

To date we have not identified publications related to automated analysis of drusen classification (hard and soft, distinct and indistinct), GA and CNV identification that can be key measurements for AMD severity identification. Automated clinical decision support for AMD with a telemedicine platform may be a potential option to achieve this goal. A highly accurate AMD pathology classification system, similar to what was used in the AREDS-related publications, is needed in order to develop a telemedicine based AMD screening system integrated with automated clinical decision support. This could guide screeners to refer patients in a timely manner.

The current clinical standard of care, SD-OCT, has been shown to be the best analytical tool to non-invasively extract pathological size and volumetric information from retinal imaging, as well as to provide quantitative measurement of macular changes over time. However, it has known limitationsin detecting certain pathologic changes, while at the same time CFP has shown strength in the extraction of small drusen information but lacks large drusen boundary accuracy. One could consider using a multi-modality image fusion approach, which could use CFP images and SD-OCT images, with the application of novel feature extraction and feature fusion techniques. Since OCT is currently not practical for screening due to its expense, an OCT/CFP fusion approach may not be the best current option for early detection, but such a tool may be viable in the future for a screening role. Thus, a complement of such sensors may provide a better method for AMD severity identification, monitoring of AMD progression, and clinical intervention.

The Optical Imaging and Biomedical Science Conference 2009 recommended a number of areas for improvement and future works that are summarized as follows:

• The value of integrating imaging technologies in the next generation of super cameras by combining OCT and adaptive optics, OCT and multiphoton microscopy, or adaptive optics and fluorescence.

• The promise of functional imaging: There is increasing excitement about combining new functional imaging with structural imaging. For example, we can currently monitor blood flow with Doppler OCT and analyze the response of single photoreceptors to light with interferometric methods. These methods can detect subtle changes in the refractive index inside single cells.

• For clinical applications, we need to develop non-invasive ways to assess function. These methods include measuring the intrinsic scattering changes of retinal cells in response to stimulation, localized sensitivity thresholds (microperimetry), and the combination of electroretinography with high resolution imaging.

In addition, the following areas can be expanded for accurate and efficient AMD screening.

8.1. Expert system for AMD screening

A clinical decision support system for AMD based on telemedicine, automated image analysis, artificial intelligence and machine learning techniques (e.g., ANN or SVM) using multiple imaging modalities could be a huge advance in disease management. First, the system could be trained for the automated identification of AMD pathologies and other biomarkers using established protocols to grade disease severity. In the analysis phase, such a tool utilizing imaging and biomarker sensitivity might be used in conjunction with ophthalmic expertise to further refine the staging and treatment of disease, further exploiting imaging and biomarkers.

8.2. 3D imaging

Besides OCT, 3D imaging of color and other 2D imaging modalities can provide better information on the retinal surface, and consequently improved pathology detection. 3D visualization of the retina could assist a remote ophthalmologist in viewing the retina with a similar clinical confidence as an in-person examination.

8.3. Portable camera with multimodal imaging options and computer-aided decision support

Portable, low-cost non-mydriatic fundus cameras capable of producing diagnostic quality retinal images would be a great boost in achieving the goal of community screening by trained nurses or primary care providers. Fundus cameras with fully automated focusing and image capturing would be invaluable to imagers. Finally, while presently only a theoretical possibility, a portable fundus camera with in-built image analysis and automated clinical decision support for diseases such as diabetic retinopathy, AMD and glaucoma would be a goal worth striving for in detecting major causes of treatable blindness.

8.4. Imaging kiosk of the future

As mentioned above, it is possible to envision a single system for the screening of diabetic retinopathy, glaucoma and AMD. It could be designed to be fully automated, from taking a photograph to automated image analysis. Many of the engineering and software components of such a “holy-grail” ocular health project either already exist today or are in their infancy. Such a goal is worth pursuing for its immense public health value, and these kiosks could be located at optometry practices, pharmacies and primary care clinics.

In the clinical setting of an eye clinic, an automated system that facilitates grading of age-related macular change would probably not add significantly to the patient outcome or evaluation. However, as new medications, genomic medicine innovations, and other interventions become available to identify and treat those at greatest risk of AMD, the ability to detect early stage AMD will allow improved clinical outcomes for this disease. A final thought for a person at risk for AMD, therefore, is that their visual prognosis may rest on making an investment in an automated system that useslow cost OCT and/or fundus imaging, allowing population-based screening to identify such individuals at an early stage.

List of abbreviations

AM

Amplitude modulation

AMD

Age-related macular degeneration

ANN

Artificial Neural Network

AREDS

Age-Related Eye Disease Study

CFP

Color fundus photograph

CNV

Choroidal neovascularization

CSC

Central serous chorioretinopathy

CTIS

Computed tomographic imaging spectrometer

EM

Expectation Maximization

EMR

Electronic medical records

FA

Fluorescein Angiography

FAF

Fundus Autofluorescence Imaging

FCM

Fuzzy C-Means Clustering

FM

Frequency modulation

GA

Geographic atrophy

GC

Graph cut

GS

Graph search

HRA2

Heidelberg retina angiograph2

HRI

Hyperspectral Retinal Imaging

ICG

Indocyanine Green (ICG) Angiography

IMS

Image mapping spectrometer

IR

Thermal or Infrared (IR) Imaging

IVFA

Intravenous sodium fluorescein angiography

NIR

Near infrared

NMF

Non-negative matrix factorization

OCT

Optical coherence tomography

PED

Pigment epithelium detachment

PR

Photoreceptor

RAP

Retinal angiomatous proliferation

RNFL

Retinal nerve fiber layer

ROC

Receiver operating characteristic curve

RPE

Retinal pigment epithelium

SD-OCT

Spectral domain optical coherence tomography

SEAD

Segment the exudates associated derangements

SLO

Scanning laser ophthalmoscopy

SRF

Sub-retinal fluid

SRT

Sub-retinal tissue

SVM

Support Vector Machine

TD-OCT

Time domain OCT

VEGF

Vascular endothelial growth factor