Automated Detection of Drusenoid Pigment Epithelial Detachments From Spectral-Domain Optical Coherence Tomography in Patients With AMD

Abstract

Purpose: This study aimed to develop an algorithm for automated detection of drusenoid pigment epithelial detachments (DPEDs) in optical coherence tomography (OCT) volumes of patients with age-related macular degeneration (AMD) and to compare its performance against traditional reading center grading on color-fundus photographs (CFPs).

Methods: Eyes with a range of AMD severities, excluding neovascular disease, were imaged using spectral-domain OCT (SD-OCT) and paired CFPs and were followed annually for up to 5 years. DPEDs were automatically identified by segmenting the retinal pigment epithelium (RPE) and Bruch's membrane (BM) layers from the SD-OCT volumes and imposing both a minimum RPE BM height (>75 µm) and a two-dimensional length requirement (>433 µm). Comparisons in detection rates and contoured areas were made between the algorithmic SD-OCT detections and manually graded and contoured CFPs.

Results: Of the 1602 visits for the 323 eyes, the automated OCT algorithm identified 139 visits (8.7%) from 50 eyes with DPED, but a reading center review of paired CFPs identified 23 visits (1.4%) from nine eyes as having DPEDs. Eyes identified with DPEDs on OCT received nine-step AMD severity scores ranging from 6 to 10, and those scores had occurrence ratios of 23/160 (14%), 89/226 (39%), 24/99 (24%), 2/63 (3%), and 1/29 (3%), respectively. On a subset of 25 visits that also underwent manual contouring of DPED lesions in CFP, the Pearson correlation coefficient for DPED areas observed by OCT and CFP was 0.85.

Conclusions: Our analysis shows the feasibility of using OCT scans to objectively detect features that historically have been detected qualitatively by expert graders on CFPs.

Translational Relevance: Automated detection and quantitation of high-risk features can facilitate screening patients for clinical-trial enrollment and could serve as an outcome metric [T1 (Translation-to-Humans) and T4 (Translation-to-Population-Health)].

Introduction

Age-related macular degeneration (AMD) is a leading cause of blindness among people over 50 years of age in high-income nations.^1–3 Progression of the disease through the initial stages to the most advanced stages is characterized by morphological and pathophysiological changes in the outer retina.^4,5 Early and intermediate stages of the disease are defined by yellow accumulations of drusen under the retinal pigment epithelium (RPE) layer and above the Bruch's membrane (BM), consisting of lipids and proteins, including complement factors and inflammatory cells.⁶ The landmark historical studies on AMD have been based on fundus features evaluated clinically on biomicroscopy and graded by human graders at reading centers using color fundus photography (CFP), including the Beaver Dam Study,⁷ Blue Mountains Study,⁸ Age-Related Eye Disease Study (AREDS),⁹ AREDS2,¹⁰ and Rotterdam Study¹¹. Very large drusen, known as drusenoid pigment epithelial detachments (DPEDs), have been evaluated on CFP and were defined in the AREDS¹² as well-defined, pale yellow or white large mounds consisting of many large drusen or confluent drusen that are at least 350 µm in the narrowest diameter and appear elevated on stereoscopic fundus photographs. In the AREDS2, these were defined as elevated mounds with one or more large soft confluent indistinct drusen with a diameter greater than 433 µm.^13,14 These lesions have been demonstrated to be important biomarkers, as patients with DPEDs are at high risk of progressing to the most advanced forms of the disease, including both geographic atrophy (GA) and choroidal neovascularization.^13,15–17 As therapies are developed to target different stages of AMD, there is a need to objectively identify high-risk features. Spectral-domain optical coherence tomography (SD-OCT) imaging has augmented our structural understanding of retinal features and, with its increased availability, SD-OCT has the potential to obtain quantitative metrics to improve disease detection and progression, especially for the localization and quantification of drusen.^18,19 

In this paper, we focus on the automatic algorithmic detection of high-risk DPEDs in OCT imaging. We build upon a previously developed retinal segmentation algorithm²⁰ and report on the OCT measures and color funduscopic grading of AMD severity and features. In the work presented here, (1) we implement our algorithm on a large dataset that consists of a cohort of over 1300 macular OCT scans, which included paired color photographs with reading center–determined AMD severity scores (AMDSCs) ranging from 0 to 10 (AREDS2 Report No. 1721) to detect DPEDs; (2) we make comparisons between the OCT-based algorithm implemented for screening of these features and color-based fundus grading of these features to enable a rigorous comparison of feature detection that has not been investigated previously; and (3) in a subset of eyes, we provide a pixel-by-pixel spatial comparison of the detections across modalities. As there is growing consensus in the field that these imaging modalities are complementary rather than replicative, understanding the differences in high-risk features is important in translating historical data based on color imaging to OCT-based detections.^22,23 The goal of this paper is to examine the differences in DPED identification using automatic algorithms with three-dimensional (3D) OCTs compared to traditional methods used by reading centers based on CFPs. 

Methods

Population

The study population was comprised of participants of an ongoing longitudinal study of dark adaptation in AMD that includes patients 50 years of age and older with and without AMD. Participants were recruited from the eye clinic at the National Eye Institute, National Institutes of Health (Bethesda, MD) between May 2011 and March 2020. The study was approved by the Institutional Review Board of the National Institutes of Health and followed the tenets of the Declaration of Helsinki. The study is registered on ClinicalTrials.gov (NCT01352975). All participants provided written informed consent after the nature and possible consequences of the study were explained. An exploratory objective of the dark adaptation study was to analyze structural risk factors of AMD progression, as conducted in this analysis. 

Image Acquisition

Multimodal images of both eyes were acquired at each study visit, including CFP images obtained using the TRC-50DX mydriatic retinal camera (Topcon Medical Systems, Oakland, NJ) and infrared reflectance (IR) and SD-OCT images obtained using the Heidelberg SPECTRALIS HRA+OCT system (Heidelberg Engineering, Heidelberg, Germany). The SD-OCT images captured were volumetric macular scans (496 × 768 × 121 pixels), spanning 30° horizontally and 25° vertically (corresponding to 1.9 × 9 × 7 mm, with a spacing of 3.87 × 11.1 × 58.7 µm between the frames of the three spatial axes). Each OCT had 121 B-scans, where each B-scan spaced 58.7 µm apart was an image of 496 × 768 pixels. 

Neovascular AMD was defined as a lesion that has demonstrated exudative activity. This was determined by the treating retina specialists (CC, TD, ATT, EYC) using multimodal imaging that may have included OCT angiography and/or fluorescein angiography. The color fundus photographs and OCTs were also graded by the reading center for evidence of exudation. We excluded scans of eyes with exudative neovascular AMD, as determined by experts from the Wisconsin Reading Center or retina specialists at the National Eye Institute (CC, TD, ATT, EYC). 

Automated DPED Detection on OCT Volumes

A 3D segmentation algorithm was applied to the OCT macular volumes²⁰ to obtain contours for the inner boundaries of the RPE and the BM. The criteria for DPED described below were imposed on this 3D map of the inner boundary of the RPE to BM elevations.^24,25 The criteria used in this analysis required that the lesion meet a minimum drusen elevation criterion to represent a large RPE detachment in the SD-OCT B-scan (>75 µm, h in Fig. 1). The threshold height criterion for the distance of the inner boundary of the RPE to the BM was empirically set to a value of 75 µm to increase the probability of positively identifying the abruptly large RPE elevations in SD-OCTs as DPEDs, while minimizing the chance of a false detection of two contiguous drusen as one large DPED. If two DPED lesions exhibited large RPE elevations and were linked by a low connector (with height of small drusen), implementing a minimum height threshold of 75 µm (for the entire lesion) prevented the thin connector from being classified as part of the DPED. Establishing this minimum height threshold ensured that the two DPED lesions were accurately and distinctly identified, preventing inadvertent grouping of the entire structure (two drusen and one low-elevation thin connector) as a single DPED lesion. Further, this ensured that connected small drusen with low elevations were not erroneously included in the reported quantitative metrics such as mean heights and widths. When drusen with RPE elevations below the height threshold were eliminated, a minimum width criterion was imposed (W₁ or W₂ > 433 µm) as illustrated in Figure 1. Although there is no widespread agreement in the field as to the minimum RPE elevation to utilize for evidence of a drusen or DPED, a recently published article in Investigative Ophthalmology and Visual Science examined OCT-based RPE elevations in eyes with drusen of different CFP-based sizes. Eyes with small drusen had an RPE height of <72.5 µm.²⁶ In our study, we made similar observations and therefore implemented an RPE threshold height of 75 µm to identify large confluent drusenoid mounds and exclude clusters of adjacent small drusen. Other studies have also reported average OCT apical heights for small, medium, and large drusen as defined on CFPs,⁴³ and our current approach could further investigate the use of different minimal OCT apical height thresholds. A limitation to this approach would be the potential to exclude shallow DPEDs with heights < 75 µm. We investigated how drusen height threshold choices would apply to SD-OCT B-scans with small to large deposits, as shown in Figure 1. The heights of the DPEDs are reported in Table 4. Our DPED detection algorithm was designed to work on SPECTRALIS OCT B-scans but could be adapted to other devices by adjusting parameters related to the OCT pixel spacing. 

Figure 1.

 

(a) Drusen with RPE elevations of 25 µm to 150 µm are shown using different colors. The light blue and dark green arrows correspond to 25 µm and 50 µm, respectively, and denote the small drusen, which were avoided for this analysis targeting the identification of DPED. (b) Drusen from DPED patients on OCT B-scan. (c) En face OCT projection of all B-scans with drusen above 75 µm height but without enforcing 433-µm minimum size criterion. (d) With the minimum height and size criteria enforced, W₁ and W₂ are the widths in the en face image, and h is the height of the drusen in the SD-OCT B-scan.

Color Image Grading

Color fundus photographs of stereoscopic pairs were evaluated by trained graders at the Wisconsin Reading Center. Grading on each image was conducted independently of images from other visit dates. AMDSCs were assigned to each eye at each study visit, based on the CFPs, according to the AREDS nine-step severity scale. The drusen area within the Early Treatment of Diabetic Retinopathy Study grid was also graded according to AREDS Report No. 17.²⁷ The presence of DPEDs was evaluated based on the presence of an elevated mound with one or more large, soft, confluent, indistinct drusen having a minimum diameter of 433 µm.¹³ 

Comparing Drusen Lesions Detected by CFP and OCT

To compare the DPED appearance on color images to the OCT images, we manually contoured the large drusen elevated areas on the color images of a subset of 25 color photographs corresponding to OCT volumes with DPED detections. Using the color photograph stereo pair with a stereo viewer, the areas of drusen elevation were contoured manually using 3D Slicer software, version 4.^28,29 To limit the analysis to only those accumulations that would meet the definition of DPED, we imposed the same criterion: W₁ or W₂ > 433 µm. Figure 2 shows this representation of DPED drusen on CFP. The contoured color images (color image plus contour mask) were co-registered to the IR image that was simultaneously acquired with the OCT volume,³⁰ establishing a pixel-to-pixel correspondence between color and the two-dimensional (2D) en face SD-OCT map. 

Figure 2.

 

(a) CFP of a patient with DPED. (b) Drusen overlaid on color photographs after enforcing the minimum size criterion (>433 µm). W₁ and W₂ are the widths in the en face CFP image.

We compared several qualitative and quantitative metrics between the OCT-detected and CFP-detected DPED lesions, including: the total area of lesions meeting the criteria across modalities, the Pearson's correlation coefficient between the DPED areas across imaging pairs, the sizes of the lesions on each B-scan compared to the equivalent linear region in the CFP, the overlap of the detected regions, and the maximum and minimum dimensions for each modality. We used t-tests to compare statistical measures or differences between the quantitative metrics. The analyses in the Results section compare DPED characteristics between SD-OCT and CFP modalities, using specific data subsets summarized in Table 1. 

Table 1.

 

Overview of Data Subsets Used for Different Analyses

 

Table 2.

 

Patient Demographics

 

Results

Patient Demographics

After excluding 374 visits with neovascular AMD out of the total 1976 imaged visits, the remaining 1602 macular OCT scans were evaluated for lesions that represented DPEDs, using the SD-OCT algorithm. The 1314 visits without neovascular AMD had corresponding AMDSCs based on CFP by the reading center. Table 2 describes the study cohort across all visits. 

Table 3.

 

DPED Parameters

 

Frequency Distributions of DPED Detections Based on Reading Center Metrics

The 1314 macular OCT scans were evaluated for lesions that represented DPEDs. The SD-OCT algorithm identified 139 OCT visits from 50 unique eyes (37 patients) as having DPEDs (mean ± SD, 2.78 ± 1.68 visits per eye). The AMDSCs of these visits spanned severity grades from 6 to 10, with the majority occurring in the group with AMDSC 7, as shown in Figure 3a. The majority (84%) of eyes detected by our OCT-based algorithm as having DPEDs present had the highest drusen area (≥2.54 mm²) graded by the reading center, as demonstrated in Figure 3b. 

Figure 3.

 

Frequency histogram of SD-OCT–based DPED detections showing the number of volume scans based on (a) reading center (RC)-based AMDSCs following the AREDS definitions, and (b) reading center provided categorical grades of drusen areas. The numbers within braces on the x-axis denote the number of OCT volumes screened in that category; the percentages above the bar plots denote the prevalence of DPEDs in that categorical grade with reference to the total number of screened volumes in that grade.

The reading center identified 23 visits from nine unique eyes (six patients) with a grade indicating the definite presence of DPEDs on CFP. Of these, 22 were also identified by the SD-OCT–based algorithm as having DPED. Inspection of the one other visit that was graded by the Wisconsin Reading Center as having DPED present but not detected by the OCT algorithm revealed the presence of only a very shallow RPE elevation. This elevation did not meet the 75-µm minimum height criterion required by the OCT-based definition of DPED and was therefore excluded from the analysis in Table 4. 

Drusen Area Comparisons Between Algorithmic SD-OCT Predictions and Annotated Contours from CFPs

A random selection of a subset of 25 visits (only five of these were deemed to be DPEDs by manual grading of CFPs at the Wisconsin Reading Center) identified by the OCT-based algorithm as having DPEDs underwent manual contouring of drusen areas in the corresponding CFPs. In Figure 4, we compare the drusen areas for DPED eyes obtained from the annotations on 25 CFPs with the corresponding SD-OCT–based algorithm-predicted drusen contours. The Pearson correlation coefficient (R value) was 0.85, demonstrating high agreement in drusen area measurements between the two modalities (Fig. 4). 

Figure 4.

 

CFP drusen areas versus SD-OCT drusen areas for DPED eyes after application of the height and size criteria to filter out the small drusens.

Point-to-Point Comparison of SD-OCT Detections and CFP Detections

To understand the spatial agreement between the SD-OCT–detected RPE elevations and the CFP-identified drusen appearances, manual contouring of the color lesions was registered to the OCT-derived 2D projection of the RPE elevation map. Figure 5 shows six cases with DPEDs meeting the threshold criterion. The color-coded overlapping labels highlight the similarities and differences between the OCT algorithm detections and the contours in CFP. In most cases, there was good agreement between the modalities as demonstrated by the shared areas labeled in blue in the color contours and the SD-OCT algorithm lesions (panels 1c to 6c in Fig. 5). There are regions, however, where the SD-OCT algorithm detected a DPED lesion, but the color contouring did not (indicated in yellow), and more areas where the color contouring identified areas of DPED but the SD-OCT algorithm did not (indicated in white). 

Figure 5.

 

DPED detections colocalized across SD-OCT and CFP imaging. In each panel, the leftmost column represents one B-scan from the OCT volume corresponding to the location in the fundus image indicated with a horizontal black line in column b. Column c demonstrates the spatial overlap of lesions detected with the OCT algorithm and the human annotations using CFPs. In both columns a and c, areas of overlap between the two labels (SD-OCT detection and CFP detection) are shown in blue; areas detected only on CFP contouring are shown in white; areas detected only on SD-OCT algorithm predictions are shown in yellow. The numbers on the top right in column c (in white) show the Dice similarity coefficients between the labels created by human annotators using the CFPs and predictions made by the SD-OCT algorithm.

The highlighted white lesions in panels 1c, 2c, and 4c in Figure 5 mark lesions that were detected as DPEDs by CFP but did not meet the threshold criteria for the SD-OCT algorithm. By inspecting the B-scans in panels 1a, 2a, and 4a, in Figure 5, we can observe that small RPE alterations corresponding to these areas did not meet the required >75 µm of RPE elevation in these lesions. 

Comparison of Drusen Length Between SD-OCT B-Scans and Corresponding Locations in CFP-Based Detections

To understand the pixel-by-pixel agreement between DPED lesions detected by CFP and SD-OCT, the length of the detected region was measured in each corresponding B-scan location and compared (Fig. 6). The number of B-scan locations with DPED detections on CFP was greater than that in SD-OCT (Table 3, second row). Figure 6a shows the Bland–Altman plot of the retinal pigment epithelium drusen complex (RPEDC) lengths per B-scan location between the SD-OCT drusen map and the CFP drusen map. The mean difference for the Bland–Altman plot was –380 µm, and the limits of agreements were within –1330 and 560 µm, demonstrating that, overall, longer drusen lesion lengths were detected in the CFP images. 

Figure 6.

 

(a) Bland–Altman plot comparison of the RPEDC width (in mm) per B-scan location between SD-OCT–detected lesions and CFP–detected lesions (SD-OCT – CFP annotations) color-labeled by their AMDSCs. (b) Violin plot comparison of the RPEDC width differences (in mm) per B-scan location between SD-OCT detections and CFP detections (SD-OCT – CFP) in each AMDSC category. Plots include all longitudinal visits.

Table 4.

 

Comparison of 139 DPEDs Detected by the SD-OCT Algorithm With the 22 DPEDs Detected on CFP by the Reading Center

 

Given that this analysis included lesions that were only detected by one modality, we repeated an analysis limiting the comparison to lesions where both modalities detect DPEDs, after removing non-overlapping DPED lesions. The mean difference in the Bland–Altman plot after removal of non-overlapping lesions was then –210 µm, and the limits of agreement narrowed to within –980 and 560 µm. 

In the Bland–Altman plot shown in Figure 6a, an oblique line of dots with a slope of –2 can be observed. This phenomenon arose due to the nearest neighbor interpolation applied to each OCT B-scan approximately six times to align its size with that of the corresponding CFP, enabling a direct comparison between the two imaging modalities. Nearest neighbor interpolation assigns identical values to groups of approximately six drusen widths to approximately six adjacent OCT B-scans. These groups of approximately six points collectively form the oblique lines with a slope of –2 in the plot. A comprehensive mathematical explanation is provided in the Supplementary Material. 

Table 3 demonstrates that the number of B-scan locations containing DPED detections was larger for the CFPs than the SD-OCTs, across all AMDSCs. In addition, AMDSC 7 consistently had the maximum drusen width across all severity grades, which aligns with our previous finding that eyes at this stage (AMDSCs 5–8) had the largest drusen areas among all severity grades.⁴² In Supplementary Table S1, we show that more B-scan locations were detected as having DPED lesions by CFPs compared to SD-OCTs, even for lesions that existed in both imaging modalities (after removal of non-overlapping lesions). Supplementary Table S2 presents a subset of data in Table 3, with DPED detections in OCT B-scans and CFPs intentionally matched to examine feature differences when DPED appears in both modalities simultaneously. 

Table 4 compares the SD-OCT drusen parameters per B-scan for the 139 OCT-based DPED detections with the with the 22 CFP-based DPED detections by the Wisconsin Reading Center. This table shows that eyes that were detected by the Wisconsin Reading Center had DPEDs with maximum drusen widths and heights more than twice the minimum width and the OCT algorithm height criterion, on average. These measures were significantly greater than for the eyes detected with the SD-OCT algorithm (P < 7 × 10⁻⁸, unpaired t-test). 

Discussion

This study revealed a significant contrast between identification of DPEDs by a reading center based on CFPs and those identified using an algorithm based on SD-OCTs. We delved into the reasons behind this inconsistency by quantitatively analyzing the features responsible. Such analysis aids in delineating the advantages and drawbacks of DPED identification based on each modality. Drusen have long been recognized as a key feature in the development and progression of AMD. There has also been a history of attempts to label drusen on CFPs, both manually³¹ and automatically,²¹ using 2D algorithms. Research efforts and commercial software from OCT vendors have included segmenting RPE and depicting RPE elevation maps.^32–35 With most historical data on AMD severity and progression being based on color photographs, some analyses have correlated OCT-based drusen volumes (measured as RPE–BM distance or RPEDC) with traditional CFP-based AMDSCs.^20,22 A few studies have even compared the spatial distribution of drusen areas observed on color photography with drusen areas detected by OCT.^21,31 Studies such as the AREDS2 Ancillary SD-OCT (A2A SD-OCT) study³⁶ have demonstrated that OCT measures of RPEDC volumes and reticular pseudodrusen (RPD) graded as present or absent are associated with increased risk of progression of AMD. Certain OCT features, such as drusen volumes, hyperreflective foci, hyporeflective drusen cores, RPD, and double layer signs, have been identified as risk factors of progression from intermediate AMD to GA within a 2-year period.³⁸ Photoreceptor layer segmentation on SD-OCTs was utilized to forecast AMD progression in non-exudative eyes.³⁹ Choroidal hypertransmission in OCTs was analyzed to identify whether it can be used as a stand-alone biomarker or novel clinical trial endpoint for progression to GA.⁴⁰ Additionally, increased choroidal flow deficit and declining microperimetric sensitivity were identified as key predictors of progression to incomplete retinal pigment epithelial atrophy (intermediate AMD) in patients imaged using swept-source OCTs.⁴¹ Additional research on the association of OCT-based DPEDs as risk factors for progression and how they compare as observed in other conventional modalities such as CFPs will help inform the design and tools required for clinical trial designs based on OCT. 

Three aspects of our work enabled us to further add to the understanding of how these modalities can depict these important features. First, imaging features are better for identifying high-risk biomarkers such as DPEDs (which portend a high risk for progression or functional decline^12,37), and we propose an algorithm that does that. Second, our OCT dataset is part of a larger multimodal imaging dataset that includes CFP with reading center–based grading of AMD features (using methodologies that have been the industry standard). Finally, our work includes comparisons of features between OCT and color imaging modalities that enable a point-to-point spatial analysis of DPED appearance between modalities. 

Our multistep algorithm adds an important tool to the analytic armament of clinicians, researchers, and clinical trialists—that of being able to screen large numbers of OCT volumes for the feature of interest. Although total drusen volumes (RPEDCs) provide some quantitative data, the spatial information is lost in this metric, and identifying high-risk features such as DPEDs cannot be performed based on total RPEDC volume alone. Other studies that have utilized preselected patients have involved specific features of interest and demonstrated how quantitative algorithms perform on these eyes (either on OCT on CFPs, or both). Our approach helps fill an unmet need, as study recruitment settings or clinical research analyses would benefit from the ability to automatically and objectively identify eyes containing certain high‐risk features, including DPEDs, among others, and this would add to the efficiency and quantitative potential of such work. 

Using both minimum height and length criteria for DPEDs, we identified a subset of eyes in a large study of AMD with a spectrum of disease severity that harbored the high-risk feature of interest. In comparing the metric analysis of the lesions identified with the automated SD-OCT algorithm to the much fewer eyes identified as having DPEDs with CFP reading center grading, we observed that the reading center–detected eyes contained much larger DPEDs (diameter > 1000 µm) than the minimum criterion (diameter > 433 µm), thus providing insight into the factors responsible for the difference in numbers identified between the modalities, as the graders seemed to impose a criterion significantly greater than the minimum required. Although the exact workings of the reading center's methodology are not explicit, it seems that the factors influencing greater DPED lesion identification may include overestimation of sizes and cues from the color, shape, or pigmentation which could suggest the presence of multiple smaller drusen rather than a DPED. The advantages of automated detection include the ability to easily adjust the specifics of lesion characteristics, depending on study design and desired inclusion thresholds. Adjusting the minimum size criterion for the lesion to have a 433-µm horizontal width to a criterion requiring both a horizontal and vertical minimum length of 433 µm changed the identified number of visits with DPED detections from 139 to 117. Alternatively, altering the minimum RPE–BM height threshold from 75 µm to 23 µm changed the number of identified visits from 139 to 772. Such findings demonstrate the ability to easily tailor feature definitions for a particular study and the ramifications of these threshold criteria. 

This number is notably low compared to the 139 detections made by our algorithm from OCTs. The analysis and findings are presented in Tables 3 and 4. Table 3 reveals that, in a subset of 25 eyes manually graded in CFPs, the DPED width appears smaller in CFPs compared to OCTs. Thus, for a detection of a lesion measured on CFPs versus OCTs, it will appear smaller on CFPs and reach the criterion width less often. However, the larger underlying discrepancy results in our observation suggest that the lesions may not be detected on CFP by reading centers unless they reach a threshold far beyond that of the 433 µm, as demonstrated by the comparison of OCT-measured lesions in Table 4. We obtained DPED detections in CFPs from the reading center (23 eyes) and in OCTs from our algorithm (139 eyes). We then extracted SD-OCT–measured dimensions in these detection groups, as presented in Table 4, which compares the dimensions of the lesions detected across the detection groups between reading center CFPs and algorithm-predicted SD-OCTs, not on the same lesions in the two modalities, which is shown in Table 3. The analysis demonstrates that the lesions detected by the reading center based on CFP grading were much larger than the minimum threshold of 433 µm. Despite the reading center using a 433-µm minimum width threshold, the strict criteria regarding the appearance of DPED lesions on CFPs seem to require much larger lesions and lead to a reduced number of DPED detections by the reading center. 

Our results identified DPED lesions in eyes with corresponding CFP grades of AREDS severity grade 7, which is the AMDSC that also has the CFP reading center–based highest category of drusen area.²³ Two distinct comparisons were analyzed. First, the agreement in dimensions of the DPED lesions detected independently on both CFP and OCT was assessed (Table 3). Although Table 3 indicates that DPED lesions appear smaller on CFPs, the quantity of B-scan locations identified as having DPEDs was significantly greater in CFPs compared to OCTs (second row of Table 3). Despite the drusen width being smaller in CFPs than in OCTs according to Table 3, the much larger number of B-scan locations containing DPEDs in CFPs resulted in a higher weighted-average drusen width (weighted by the number of positive DPED B-scan locations) compared to OCTs. Consequently, the comparison of drusen width between OCTs and CFPs yielded negative mean limits of agreement, as shown in the Bland–Altman plot in Figure 6a. The limits of agreement metric incorporates the number of B-scan locations in addition to the drusen width, and the higher number of B-scan locations with DPEDs in CFPs is the underlying reason for this observation. Second, an examination of DPED lesions in eyes identified by the reading center as having DPED via CFP was compared to those identified using an OCT algorithm on the OCT scans. There was a notable discrepancy in the number of eyes identified by the reading center as having DPEDs, with the lesions within these fewer eyes being significantly larger (refer to Table 4). The instances of disagreement between the modalities were insightful and demonstrated that, in general, more lesions and larger lesion areas were detected on CFP than observed on OCTs. Examining these areas with a pixel-wise comparison demonstrated that, in most cases, these areas represented regions on the color image that did not meet the height criteria on OCT. There were also some instances, though less frequent, of areas of detection on SD-OCT that were not contoured on CFP (yellow highlighted region to the right in the B-scan in Fig. 5, panel 3a). Analyzing these areas demonstrated in many cases that the pigment contrast on CFP imaging obscured the appearance of elevation (corresponding to the yellow highlighted region to the right in CFP in Fig. 5, panels 3b and 3c) that was evident on OCT imaging. 

Although the automated SD-OCT–based detection offers several advantages over qualitative manual CFP-based grading, there are limitations to our analyses. Challenging cases, such as B-scans with artifacts, poor signal to noise ratio, poor tracking, and misalignment, as well as a few cases with outer retinal atrophy, confounded our published segmentation algorithm.²⁰ We had to visually inspect and eliminate 33 visits that had such segmentation errors from the entire cohort. Also, there could be other confounding and unexpected cases that could lead to segmentation errors in unseen test datasets. Although this is a limitation of our segmentation model,²⁰ this limitation of a lack of generalization to largely differing test domains holds true for most machine learning models. Utilizing manual grading from SD-OCTs could prospectively avoid some of these shortcomings, but doing so requires manual expertise and significant time, making the process very expensive and not feasible when dealing with large studies. Additionally, although the manual contouring of the color images was done by retina specialists, they were not contoured by Wisconsin Reading Center–trained graders. 

The threshold for drusen height was determined empirically and was similar to that reported for small drusen based on apical heights in OCTs,²⁶ whereas the threshold for drusen width from CFP diameter was obtained from previous definitions.^13,14 We understand that there is potential room for fine-tuning these parameters with respect to OCTs to determine the best height and width representing DPEDs.⁴³ 

The lifecycle of drusen—from their development, growth, progression to DPEDs, regression, and possible atrophy—represents a dynamic process that poses significant challenges in longitudinally correlating DPEDs due to the continuously adapting biomarker. OCT and CFP pairs collected at the same visit were analyzed without adjusting for correlations between visits, which is a limitation of our present study. 

Because we identified DPED eyes by thresholding RPE elevations with respect to BM, our algorithm cannot distinguish between deposits that occur due to drusenoid PEDs versus fibrovascular PEDs. In future work, we aim to differentiate between these deposits (drusenoid vs. fibrovascular) by identifying other discriminative features between those deposits. This is a limitation of our algorithm, and we did not evaluate eyes with neovascular AMD. 

Three-dimensional OCT volumes are becoming the mainstay of AMD monitoring in current retinal clinics. As there is a huge amount of information contained in OCTs, the ability to detect high-risk features, such as DPEDs, would help the clinician in assessing the risk of a patient progressing to advanced disease. Additionally, as clinical trials and treatments develop, the ability to identify high-risk features from the 3D OCT volumetric data will be even more valuable for determining eligibility criteria for clinical trials and, ultimately, to monitor patient progression and response to intervention. Further algorithmic validation on additional datasets will be necessary for these algorithms to be integrated into clinical trial planning. 

Acknowledgments

Supported by a grant from the National Eye Institute Intramural Research Program, National Institutes of Health (EY000509-10). 

Preliminary results from this work were presented at the ARVO 2022 Annual Meeting. 

Disclosure: S. Mukherjee, None; C. Duic, None; T. De Silva, None; T.D.L. Keenan, None; A.T. Thavikulwat, None; E.Y. Chew, None; C. Cukras, None