Abstract
Purpose.:
To assess the accuracy of automated classification of pigment epithelial detachments (PED) by using a software algorithm applied to spectral-domain optical coherence tomography (SD-OCT) scans.
Methods.:
HD-OCT (Cirrus; Carl Zeiss Meditec, Dublin, CA) volume scans (512 × 128) were retrospectively collected from 46 eyes of 33 patients with evidence of PED in the setting of age-related macular degeneration (AMD, n = 28) or central serous chorioretinopathy (CSCR, n = 5). In these eyes, 168 PEDs were automatically detected with a system-associated tool (Cirrus HD-OCT RPE Elevation Analysis; Carl Zeiss Meditec). Two independent, certified Doheny Image Reading Center (DIRC) OCT graders classified these PEDs into three categories—serous, drusenoid, or fibrovascular—via inspection of the B-scans. Manual classification results served as the gold standard for comparisons with automated classification. For automated classification, interindividual variation in intensities was normalized in all images. Individual A-scans within the detected PEDs were then automatically classified into one of three categories based on the mean internal intensity and the standard deviation of the internal intensity: mean intensity <30 (serous type); mean intensity ≥30 but <60 or mean intensity ≥30 and SD ≥30 (fibrovascular type); or mean intensity ≥60 and SD < 30 (drusenoid type). Individual PEDs were then automatically classified into the same three categories based on the predominant type of A-scan within the PED. For mixed PEDs (many A-scans of each type), a risk index for neovascularization was computed based on the percentage of fibrovascular A-scans. In addition, a confidence index was computed for each PED based on its mathematical distance from the PED category boundaries.
Results.:
Among the 168 PEDs, the DIRC graders classified 16 as serous, 88 as fibrovascular, and 64 as drusenoid PEDs. The automated algorithm classified 14 as serous, 96 as fibrovascular, and 58 as drusenoid PEDs. The sensitivity and specificity values for automated classification according to type of PED were 88% and 100% for serous, 76% and 64% for fibrovascular, and 58% and 81% for drusenoid, respectively.
Conclusions.:
Automated classification of PEDs using internal reflectivity characteristics appears to be sensitive for detecting serous and fibrovascular PEDs. Automated classification and quantification of PEDs may be a useful tool in future studies for stratifying PEDs according to risk and possibly predicting the risk of advanced AMD.
Retinal pigment epithelial detachment (PED) is a common feature of many chorioretinal disease processes, the most prevalent of which is age-related macular degeneration (AMD).
1 Initial studies were limited to the evaluation of these PEDs by using planar imaging technologies such as color fundus photography and fluorescein angiography. These initial studies demonstrated that PEDs may evolve over time and that identifying and classifying PEDs may be of importance. For example, some investigators observed that long-standing avascular PEDs may be associated with progression to vascularized PEDs over time, and their presence may be related to a poor prognosis.
2,3
The development of axial or cross-sectional imaging technologies—in particular, optical coherence tomography (OCT) and, more recently, high-resolution spectral domain (SD) OCT—has opened the door for more precise and comprehensive assessment of PEDs. Many investigators have noted that several types of PEDs can be observed and differentiated on SD-OCT imaging, including serous, drusenoid, and fibrovascular PEDs.
4 –7 Some groups have further identified and classified a variety of subtypes of drusenoid PEDs using characteristics such as size, curvature, and internal reflectivity.
8 The significance of these various subtypes remains to be demonstrated in future trials.
Few groups, however, have attempted to study various subtypes of fibrovascular PEDs or to determine whether the earliest signs of fibrovascular infiltration can be observed reliably on OCT. Previously, our group has correlated OCT and fluorescein angiographic (FA) findings in patients with neovascular AMD and observed that PEDs with apparent fibrovascular infiltration (evidenced by heterogenous internal reflectivity) on OCT correlated with occult choroidal neovascularization (CNV) on angiography.
9 In our previous studies, however, fibrovascular PEDs were not detected automatically, but rather they were identified and quantified by exhaustive manual segmentation by reading center experts.
10 An important attribute of OCT which has contributed to its rapid and pervasive acceptance in retinal clinical practice is that it provides automated quantitative information. Previously, automated analyses from most commercial OCT software were limited to quantification of retinal or nerve fiber layer thickness. Recently, however, investigators and OCT manufacturers have demonstrated algorithms that can reliably segment and quantify retinal pigment epithelium (RPE) elevations in patients with AMD and related diseases.
6,11,12
Despite this progress, attempts to automatically classify these areas of RPE elevation and identify PEDs with possible early subclinical fibrovascular infiltration have been limited. In this study, we describe and evaluate an algorithm that may allow automated classification and risk stratification of PEDs in eyes with AMD and central serous chorioretinopathy (CSCR) by SD-OCT.
We retrospectively reviewed OCT scans from all patients with AMD and CSCR who were referred to the Doheny Ophthalmic Imaging Unit (between March 2008 and June 2010) and underwent HD-OCT (Cirrus; Carl Zeiss Meditec, Inc., Dublin, CA) imaging with the 512 × 128 volume scan protocol (Macular Cube; Carl Zeiss Meditec). The data collection was limited to the Cirrus HD-OCT, as this was the only instrument at the time of the study with available (for research use) RPE analysis and quantification software. The 512 × 128 protocol was chosen instead of the 200 × 200 cube protocol, as the higher transverse resolution scans facilitated manual identification and classification of PEDs. All cases were scrutinized to identify eyes with good-quality scans (good signal strength, minimal or absent motion artifact, and good centration), which contained at least some evidence of RPE elevation. A total of 33 consecutive patients (46 eyes) who met these criteria were selected for further analysis. Of 33 patients, 28 (M:F, 12;16; mean age, 81.4 ± 6.6 years) had AMD and 5 had CSCR (M;F, 3;2; mean age, 57.4 ± 15.9 years). The collection and analysis of image data were approved by the Institutional Research Board of the University of Southern California. The research adhered to the tenets set forth in the Declaration of Helsinki.
From these 46 eyes, 168 PEDs were automatically detected (Cirrus HD OCT RPE Elevation Analysis tool; Carl Zeiss Meditec).
10,11 The algorithm and detection thresholds have been published and are also described below.
The PED segmentation was performed using a prototype code for the Cirrus HD OCT RPE Elevation Analysis (MatLab; The Mathworks, Natick, MA), and the analysis of the image data within the PEDs was performed in the same software. The program was also used to generate the PED composition, PED classification maps, and fibrovascular PED risk and CI.
Based on mean intensity and SD for each PED, the median value of all A-scan results was calculated, giving the median mean and median SD. The mean of these values was calculated over all PEDs in each category, giving the mean median mean and mean median SD.
The sensitivity and specificity of the automated classification and the P value of the CI were calculated (Excel; Microsoft, Redmond, WA).
In this study, we evaluated an algorithm for automated classification of PEDs in eyes with AMD or CSCR. Automated classification of PEDs on OCT appeared to be both sensitive (88%) and specific (100%) for identifying serous PEDs, although only a small number of cases were included in the analysis. With the threshold points selected for this study, our automated algorithm yielded a higher sensitivity (76%) than specificity (64%) for detecting fibrovascular PEDs, but a higher specificity (81%) and lower sensitivity (58%) for detecting drusenoid PEDs. This tradeoff is not unexpected, as thresholds were intentionally chosen to increase sensitivity for detection of fibrovascular PEDs. The rationale being that early detection of fibrovascular PEDs may be of clinical value, perhaps identifying a subgroup of patients at highest risk for developing overt clinically manifest choroidal neovascularization. Indeed, several studies have suggested that OCT may be more sensitive than angiography for detecting CNV.
13,16 –18
Several points should be considered, however, when evaluating the sensitivity and specificity statistics. First, these calculations required both the human grader and the automated algorithm to choose a single best answer for each PED, despite the presence of PEDs that appeared to show characteristics of more than one type. This forced-choice approach may have ultimately compromised the sensitivity and specificity. The observation that the most of the misclassified PEDs had a low CI (indicating that the PEDs were composed of significant percentages of more than one A-scan type) is consistent with this presumption. A second limitation of these calculations is the potential inaccuracy of the gold standard. Although the reading center PED assessment protocols are based on experience from multiple CNV trials and many published studies correlating angiographic and OCT findings, histopathologic correlative data are not available to definitely prove that fibrovascular, drusenoid, and serous PEDs as determined by OCT inspection, are equivalent to the same lesions on microscopic inspection.
17,19 –22 Another limitation of this analysis is that the present automated analysis only used normalized internal reflectivity characteristics. It is possible that consideration of other features of PEDs would further improve the sensitivity and specificity results achieved in this study. However, internal reflectivity of PEDs still appears to be a major feature to classify various PED subtypes, most of the misclassified PEDs in this study were related to confounding features of PED reflectivity, such as RPE migration with shadowing artifacts and atrophy of the fibrovascular membrane within fibrovascular PEDs. Therefore, increased awareness of these features and an improved algorithm that could compensate for artifactitious reflectivity changes may further improve the automatic classification of PEDs. Finally, the classification algorithm developed in this study is ultimately limited to the PEDs that can be accurately segmented by the existing OCT instrument RPE elevation analysis. In eyes with significant segmentation errors or PEDs too shallow or small to be detected, further classification will not be possible. The clinical consequences of missing these subthreshold lesions must be redefined.
In addition, to sensitivity and specificity statistics, we attempted to explore other potentially useful parameters. The CI developed in this study appeared to be effective at identifying PEDs with a high probability of misclassification and may be helpful in identifying cases that require further scrutiny by the clinician. The fibrovascular PED risk index may also prove to be of clinical value, potentially identifying PEDs at high-risk for progression to manifest CNV. Previous studies such as that by Roquet et al.,
23 have identified that 25% of eyes with drusenoid PEDs may develop CNV over a 10-year period. One wonders whether drusenoid PEDs showing mixed features on OCT (i.e., fibrovascular type A-scans on OCT) may have an even higher percentage of CNV development. Notably, there seemed to be clear separation between the PED groups by their risk index, with serous PEDs having the lowest index (12.1), followed by drusenoid PEDs (30.6), and lastly, fibrovascular PEDs (82.6). This matched our expectation rather well, yet only prospective longitudinal data in large clinical trials will be able to evaluate and better define the potential value of this index.
Ultimately, automatic analysis of PED may be useful in detecting early development of neovascular PEDs from nonneovascular PEDs. In addition, quantitative automatic profiles of PED can be potentially advantageous in monitoring various PEDs in a clinical setting.
In summary, analysis of the internal reflectivity profiles of PEDs may allow automated classification of PEDs detected by existing OCT segmentation algorithms. Further development is needed to improve the accuracy and reliability of PED classification, and longitudinal studies are necessary to define the clinical value of this analysis. These approaches may be useful for monitoring different types of PED over time, stratifying PEDs according to risk, and predicting the risk of advanced AMD.
Supported in part by a Research to Prevent Blindness Physician Scientist Award.
Disclosure:
S.Y. Lee, None;
P.F. Stetson, Carl Zeiss Meditec (E);
H. Ruiz-Garcia, None;
F.M. Heussen, None;
S.R. Sadda, Heidelberg Engineering (S), Carl Zeiss Meditec (F), Optos (F), Optovue, Inc. (F), Topcon Medical Systems (F)