Differential Vulnerability of Retinal Layers to Early Age-Related Macular Degeneration: Evidence by SD-OCT Segmentation Analysis

Abstract

Purpose.: We evaluated layer-by-layer retinal thickness in spectral-domain optical coherence tomography (SD-OCT), determined by automated segmentation analysis (ASA) software in healthy and early age-related maculopathy (ARM) eyes.

Methods.: There were 57 eyes (specifically, 19 healthy eyes under 60 years old, 19 healthy eyes over 60, and 19 ARM eyes) recruited into this cross-sectional study. The mean ages were 36.78 (SD, ±13.82), 69.89 (SD, ±6.14), and 66.10 (SD, ±8.67) years, respectively, in the three study groups. The SD-OCT scans were transferred into a dedicated software program that performed automated segmentation of different retinal layers.

Results.: Automated layer segmentation showed clear boundaries between the following layers: retinal nerve fiber layer (RNFL), ganglion cell layer plus inner plexiform layer (GCL+IPL), inner nuclear layer plus outer plexiform layer (INL+OPL), outer nuclear layer (ONL), and RPE complex. The thickness of the RNFL, ONL, and RPE layers did not show a statistically significant change across the three groups by ANOVA (P = 0.10, P = 0.09, P = 0.15, respectively). The thickness of GCL+IPL and INL+OPL was significantly different across the groups (P < 0.01), being reduced in the ARM eyes compared to healthy eyes, under and over 60 years old.

Conclusions.: The early morphologic involvement of the GCL+IPL and INL+OPL layers in ARM eyes, as revealed by the ASA, could be related to early anatomic changes described in the inner retina of ARM eyes. This finding may represent a morphologic correlation to the deficits in postreceptoral retinal function in ARM eyes.

Introduction

Age-related maculopathy (ARM) is a retinal degenerative disease involving the macula that is characterized by drusen and RPE rearrangement with secondary involvement of central visual acuity. ¹ It is a multifactorial disorder involving complex interactions among genetic factors oxidative stress and environmental risk factors. ² Numerous risk genotypes have been associated with ARM and its progression. ^3–17  

Although several imaging techniques are able to analyze the retina, such as fundus photography, ^18,19 autofluorescence, ^20–22 and scanning laser ophthalmoscopy (SLO), ²³ spectral-domain optical coherence tomography (SD-OCT) has become one of the most relevant diagnostic tools for the diagnosis of retinal diseases and ARM specifically. ²⁴ The OCT analysis is a noninvasive technique and has the capability of providing an in vivo cross-sectional view of the retinal stroma. Specifically, OCT B-scan images can show the different retinal layers. From the clinical perspective, the identification of each retinal layer opens up new perspectives for studying the correlation between the morphologic and functional aspects of retinal tissue. 

We used an experimental software analysis (automated segmentation analysis, ASA) that provides a new automated SD-OCT segmentation method to generate an accurate and objective measurement of the thickness of different retinal layers. In the present study, we used this approach to evaluate healthy and ARM eyes, and to compare the different thicknesses of the layers across groups by means of this new segmentation method of analysis. 

Methods

Subjects

The study adhered to the tenets of the Declaration of Helsinki and was approved by the Institutional Ethics Committee. Written informed consent was obtained from the participants after a detailed description of our study and the aims of the work. 

A total of 19 eyes of 19 healthy volunteers under 60 (from 22–55 years old; mean age, 36.78 years; SD, ±13.82; n = 19), 19 eyes of 19 healthy volunteers over 60 (from 60–78 years old; mean age, 69.89 years; SD, ±6.14; n = 19), and 19 eyes of 19 ARM patients (from 54–85 years old; mean age, 66.10 years; SD, ±8.67; n = 19) were included in the study. 

Six eyes of six patients with advanced age-related macular degeneration and geographic atrophy also were included in the study. Although the segmentation analysis was not applied to the SD-OCT images of these eyes, the images were used as extreme control data for the advanced ARM pathologic condition and provided a reference for comparing the expected result of a severe thinning, especially of the outer nuclear layer (ONL). ²⁵  

The eyes of the recruited healthy volunteers, both under 60 and over 60 years old, had best corrected visual acuity (BCVA) of 57.68 (±2.42) and 56.47 (±3.18), according to the Early Treatment Diabetic Retinopathy Study (ETDRS) letters. The mean BCVA of early ARM eyes was 54.15 ETDRS letters (±5.2). All patients were white Europeans. 

Inclusion criteria were no other eye pathologies (i.e., uveitis, glaucoma, and so forth) and no diopter opacities, to obtain a good image quality. The exclusion criterion was inadequate cooperation to obtain satisfactory images. All recruited eyes underwent ocular fundus imaging with standard photography and fundus autofluorescence, and SD-OCT evaluation by CIRRUS (Carl Zeiss Meditec, Inc., Dublin, CA). 

A validation sample of 10 additional ARM patients and 10 age-matched controls, evaluated in a blinded fashion as to the results of the first group of study patients, also was tested and included in the study. 

Imaging Methodology

Although the retinal OCT microanatomy includes several layers, ²⁶ the ASA is able to detect only specific layer complexes: retinal nerve fiber layer (RNFL), ganglion cell layer plus inner plexiform layer (GCL+IPL), inner nuclear layer plus outer plexiform layer (INL+OPL), outer nuclear layer (ONL), and RPE complex, as shown in Figure 1. 

Figure 1

 

Evaluation of all layers automatically recognized. (A) Original black and white B-scan of a healthy eye. (B) Layer segmentation with detectable layer boundaries. (C) Automated layer thickness analysis performed pixel-by-pixel.

Algorithm Design

In the following, we will briefly describe the segmentation method from a broad perspective. This method is fully automated and it can be applied to any OCT retinal B-scan from healthy eyes. The description of the algorithm will be given with reference to Figure 2A. 

Figure 2

 

(A) Original B-scan exported from the OCT device. The original image was given as input to the segmentation algorithm. (B) The image obtained after the total variation denoising. (C) The image after the shock filter. (D) The contour lines extraction. (E) An example of inclusion (see circled region). (F) A segmented image after region fusion. (G) The RPE layer identification after the complex diffusion algorithm. (H) The final segmentation superimposed onto the original image.

The algorithm consists of the following 5 steps: denoising and total variation (TV) approximation (Fig. 2B), edge enhancement (Fig. 2C), contour extraction (Fig. 2D), region fusion (Figs. 2E, 2F), and extraction of the hyperreflective layer (Fig. 2G). 

Step 1 of the algorithm provides an approximation (Fig. 2B) of the original image that is suitable for segmentation. Specifically, step 1 removes the background noise and, moreover, forms an approximation of the original image that is, tendentially, piecewise constant. In other words, this approximation shows homogeneous domains made by one or different adjacent layers. This process allows to highlight the interfaces between different layers. However, different adjacent layers having small contrast are merged into a unique layer complex; that is, they cannot be resolved, as it is evident from Figures 2A and 2B. Finally, we mention that step 1 has been carried out by a TV method. ²⁷ Step 2 (Fig. 2C) of the imaging algorithm is devoted to improve the sharpness of the boundaries of the domains found in step 1. Step 2 provides an improvement of the image especially in those parts where the boundaries between different domains are not so sharp. Step 2 has been carried out by means of the so-called shock-filter. ²⁸ Step 3 (Fig. 2D) is devoted to the extraction of the contours; that is, the boundaries between different domains. This is one of the most critical parts of the segmentation algorithm, and an ad hoc procedure has been developed (Giovinco G, Savastano MC, Ventre S, Tamburrino A. Automated detection of the retinal layers from OCT spectral domain images, submitted for publication, 2013). The idea behind this procedure is that the magnitude of the gradient of the image is maximum at the boundaries. The procedure looks for connected lines joining the points where the magnitude of the gradient achieves its maxima and has been developed ad hoc for treating images of layered structures as those of OCT. The output of step 3 may provide some artifacts. Indeed, after the contour extraction and the related segmentation, small regions may appear as inclusions in major regions (Fig. 2E). To remove those inclusions that do not have any physiological significance, we adopted an ad hoc fusion strategy based on the measure of the length of the curve between the “small” region and the surrounding regions. The fusion rule associates the inclusion with the adjacent region having the largest number of contour pixels in common with the inclusion. The effect of the region fusion algorithm is shown in Figure 2F. At this stage we have the final segmentation of the retinal layers, apart from the hyperreflective layer. This latter layer has been extracted separately. First, we remove the previously segmented regions from the image. Then, a nonlinear complex diffusion algorithm, ²⁹ is applied to mitigate the background noise affecting the hyperreflective region and the choroid. Finally, a contour extraction is applied as described in steps 2 and 3. The resulting image is shown in Figure 2G. The final segmentation result is shown in Figure 2H, where the contours are superimposed onto the original image of Figure 2A. 

Data Analysis

Data from the ARM eyes, and the healthy subjects under 60 and over 60 years old were included in the primary statistical analysis. The area of the layers in pixels of RNFL, GCL+IPL, INL+OPL, ONL, and RPE were compared across the three groups by 1-way ANOVA. The post hoc Tukey honest significance difference (HSD) test was used for multiple between-group comparisons. The results from the validation sample of ARM patients and age-matched controls were evaluated in a separate analysis by ANOVA with post hoc Tukey test. In all the analyses, a conservative P < 0.01 was considered statistically significant. 

Results

The differences in the thickness of the different retinal layer complexes across ARM patients (Fig. 3), healthy subjects under 60 years, and healthy subjects over 60 years old are included in Figure 4, in which the mean layer thickness (+SD) recorded from the three groups is depicted. In the bottom of Figure 4 are shown three representative OCT B-scans with automated software segmentation from each group. 

Figure 3

 

Example of the automated segmentation of an ARM eye. (A) B-scan crossing the fovea and acknowledgment of some retinal layers. (B) Improvement of layer evaluation converted into the color false scale. (C) Final rearrangement of multiple layer segmentation with the error correction.

Figure 4

 

Effect of the pixel-by-pixel calculation of different retinal layers in ARM eyes, and healthy eyes under 60 and over 60 years old by the experimental software. The RNFL, GCL+IPL, INL+OPL, ONL, RPE complex, full pixel between RNFL and RPE (Total), showing statistically significant correlation between two groups in GCL+IPL, INL+OPL, and Total. Error bars indicate 95% confidence intervals of the means. The three OCT B-scans on bottom showed automated segmentation applied to each eye of 3 considered groups: healthy eye under 60 years old (left), healthy eyes over 60 years old (middle), and ARM (right).

One-way ANOVA indicated a significant difference across the three groups only for GCL+IPL (F = 24.66, P < 0.01), INL+OPL (F = 37.79, P < 0.01), and Total (F = 22, P < 0.01) layer thickness. No significant change in the RNFL (F = 2.37, P = 0.10), ONL (F = 2.51, P = 0.09), and RPE (F = 1.97, P = 0.14) layers was found. 

Multiple comparisons indicated that the mean values of GCL+IPL, INL+OPL, and Total thicknesses were reduced significantly in ARM patients compared to healthy subjects under 60 years old (Tukey HSD, P < 0.01) and healthy subjects over 60 years old (Tukey HSD, P < 0.01). The other between-group comparisons did not reach statistical significance. 

Figure 5 shows box plots of the mean thickness in pixel distribution (± interquartile and 99th percentile ranges) for each retinal layer, automatically detected by the ASA and recorded for the three study groups. It can be noted that, although some overlap existed between normal and affected eyes, the GCL+IPL, INL+OPL, and Total thickness of the ARM eyes were reduced substantially compared to control eyes. No such differences were found for the other measurements. 

Figure 5

 

Box plots of the distribution of the mean thickness in pixels (± interquartile and 99 percentile ranges) recorded in each considered layer in the three studied groups: ARM eyes, under 60 healthy eye subjects, and over 60 healthy eye subjects.

Figure 6 shows representative examples of six eyes with advanced age-related macular degeneration and geographic atrophy, where the absence or severe thinning of all retinal layers is shown for comparison with the ARM study eyes. 

Figure 6

 

Examples of six eyes with advanced age-related macular degeneration and geographic atrophy. The ONL was totally absent or severely reduced.

A normal range ± 95% confidence limits was determined for the thickness of each layer derived from segmentation analysis in normal old control subjects. This was done to determine the sensitivity, at 95% specificity, of thickness measurements in detecting ARM eyes. Considering the normal limits of the GCL+IPL layer thickness, 13 of 19 ARM patients (68%) showed a significantly reduced thickness. Considering the normal limits of the INL+OPL thickness, 14 of 19 patients (74%) showed a significantly reduced thickness. 

The analysis of the validation sample showed that the mean GCL+IPL layer thickness was significantly (P < 0.01) reduced in ARM eyes compared to age-matched control eyes (controls 140.9 ± 9.7 × 10⁹, ARM 105.9 ± 33 × 10⁹). Similarly, the mean INL+OPL thickness was significantly (P < 0.01) reduced in ARM compared to age-matched control eyes (controls 109.5 ± 8.9 × 10⁹, ARM 54.5 ± 31.7 × 10⁹). No significant between group differences were observed for the ONL thickness (controls 147.9 ± 11.4 × 10⁹, ARM 152.4 ± 13.7 × 10⁹). 

Discussion

The advanced stages of ARM involve a thinning of the retinal layer thickness in all retinal layers. ^30,31 In the early phases of ARM, the initial retinal morphologic changes remain unknown. The early detection of retinal layer abnormalities associated with ARM could be helpful to understand disease pathophysiology better, and to determine potential biomarkers for disease progression and targets of pharmacological treatment. 

To investigate the early ARM morphologic retinal thickness reduction, the ASA software provided precise and unambiguous recognition of the boundaries of the retinal layers in all SD-OCT B-scans. 

Although several SD-OCT devices incorporate the automated partial retinal layer thickness analysis, they have a dedicated software program integrated into the instrument and are not available for the analysis of other exported images. ^32–37 Moreover, our calculated algorithm is different from those reviewed in previous studies. Specifically, among the previous methods we mention those based on several algorithms, such as the nonlinear complex diffusion, edge detection via proper kernels, active contours, Markov Random Field, Kalman filtrering, and level sets. ^32–37 In this contribution, we have proposed a segmentation method based on a combination of the TV approach, with an edge enhancement through the shock filter, and an ad hoc contour extraction and region fusion. ^27,28  

In our study, we showed that inner retinal layers were affected early in the disease process. These changes could not be accounted for by an aging effect because no differences were found between healthy young and old eyes. The analysis of a validation sample of early ARM patients and age-matched controls showed results similar to those of the primary analysis, with a selective reduction of inner layers, including the INL and GCL. The estimated sensitivity of our segmentation analysis for the detecting ARM eyes, at 95% specificity, was 68% for GCL+IPL and 74% for INL+OPL thickness measurements. 

The involvement of inner retinal layers in early ARM (specifically of the GCL+IPL and INL+OPL complex) could be related to a postreceptoral functional loss in ARM attributed to synaptic malfunction and/or the ischemia postreceptoral hypothesis. ³⁸ According to Arden et al., ³⁹ ischemic factors could be the basis of postreceptoral damage in early ARM. Moreover, a hypoxic disorder may create a “vicious circle” with the secondary main oxygen request by the photoreceptors. Vascular endothelium growth factor (VEGF) can be produced secondary to this hypoxic condition, and proinflammatory factors with choroid suffering may be created. ⁴⁰ The inadequate choroid circulation can compromise the RPE function with insufficient photoreceptor disc degradation and the accumulation of extracellular amorphous material. ⁴¹ As reported by Yu and Cringle, ⁴² and Cringle et al., ⁴³ the photoreceptors are more resistant to ischemic distress than the postreceptoral cells, especially those located in the IPL; that is, the amacrine and, to a lesser extent, the horizontal cells. The vulnerability of the postreceptoral region to ischemia similarly was apparent in a response reduction of cone-mediated flicker sensitivity with the focal electroretinogram (FERG) temporal response frequency protocol. ^44,45 The morphologic defect found in the ARM eyes of the present study, corresponding typically to GCL+IPL and INL+OPL layers, can be considered a morphologic correlate of these previously described functional defects. 

In conclusion, our data showed an early abnormality in the postreceptoral retinal layers in the early stages of age-related macular degeneration, and point at OPL, INL, IPL, and GCL as major vulnerable loci for damage as a result of chronic ischemia, inflammation, and predisposing genetic factors. 

Acknowledgments

The authors alone are responsible for the content and writing of the paper. 

Disclosure: M.C. Savastano, None; A.M. Minnella, None; A. Tamburrino, None; G. Giovinco, None; S. Ventre, None; B. Falsini, None