Automated Assessment of Drusen Using Three-Dimensional Spectral-Domain Optical Coherence Tomography

Abstract

Purpose.: To compare automated assessment of macular drusen delineated by the authors' originally developed algorithm on three-dimensional (3D) spectral-domain optical coherence tomography (SD-OCT) with the assessment by certified graders on color fundus photographs in nonneovascular age-related macular degeneration (AMD).

Methods.: Automated assessment of macular drusen was performed using raster scan by 3D OCT scans in 18 eyes with nonneovascular AMD with at least one large druse (≥125 μm) and predominantly soft indistinct drusen. Drusen was defined as the regions that have the distance between the retinal pigment epithelium and calculated Bruch's membrane lines > predefined threshold distances. The agreement was assessed on maximum drusen size and drusen area within grid between 3D SD-OCT and color fundus photographs, and false-negative and false-positive drusen at each threshold distance.

Results.: There was agreement or agreement within one step in all eyes in maximum drusen size, and 15 (83.3%) of the eyes in the drusen area, except 6 pixels, regardless of threshold distances. However, the number of eyes with exact agreement in the drusen area increased when the threshold distances were smaller than 4 pixels. In the three cases with disagreement in the drusen area, false-negative drusen on 3D SD-OCT were characterized by being small in area and height.

Conclusions.: Automated assessment of drusen parameters based on the authors' algorithm on 3D SD-OCT, which was limited by the poor detection ability of small drusen, showed good agreement with the assessment by certified graders on color fundus photography in these subjects.

Drusen are extracellular deposits that accumulate between the retinal pigment epithelium (RPE) and the inner collagenous layer of Bruch's membrane, resulting in anteriorly protruding RPE over the straight Bruch's membrane. ¹ Drusen are a distinguishing feature of nonneovascular age-related macular degeneration (AMD). The parameters for estimating drusen, such as total drusen area and maximum drusen size in the macula, have been shown to have correlations with risk of progression to advanced AMD. ^{2 –8} The Age-Related Eye Disease Study (AREDS) reported that by taking a high dose vitamin and mineral supplement, AMD patients with advanced dry AMD or vision loss resulting from neovascular AMD in one eye or extensive intermediate size drusen (at least one large druse), or noncentral geographic atrophy in both eyes could delay the onset of severe AMD and accompanying severe vision loss. ⁹ This constitutes important evidence, suggesting that earlier identification of the patients at higher risk for advanced AMD allows earlier protective intervention to reduce severe vision loss. Thus, it is becoming more important to assess changes in drusen and the RPE morphology to monitor the progression of nonneovascular AMD. 

Assessment of drusen on color fundus photography may be performed according to the grading protocols established by Wisconsin Reading Center as widely used in epidemiologic studies and adapted for clinical trials, such as the AREDS. ^{10 –12} However, the assessment of drusen on color fundus photography is effort intensive, often used with a multigrader methodology for achieving objective assessment. 

Optical coherence tomography (OCT) is an interferometric technology for in vivo high-resolution, cross-sectional imaging of ocular structures, which allows automated and objective measurement of retinal structures. ¹³ Recently developed spectral-domain OCT (SD-OCT) technology allows much faster imaging (43 to 133 times faster than originally developed time-domain OCT [TD-OCT] technology), and three-dimensional (3D) analysis of macular pathologies. ^{14 –17} High-speed OCT imaging allows the elimination of eye motions in B-scans, therefore creating OCT images that better reflect the true retinal geometry than does TD-OCT, which cannot reliably reproduce the shape of the RPE. High-speed OCT imaging also enables dense 3D raster scanning, which theoretically allows detection of drusen that are small in size. Commercially available SD-OCT instruments have twice higher axial resolution (5–7 μm) than Stratus OCT. Higher axial resolution theoretically allows improved visualization of drusen that are small in height. Thus, SD-OCT instruments potentially allow depiction of small drusen. 

Recent studies showed the potential advantages of SD-OCT in assessing drusen, such as manual measurement of drusen volume ¹⁸ ; semiautomatic measurement of drusen area and drusen size ^19,20 ; automatic detection and measurement of drusen area, height, and volume ^{21 –23} ; improved visualization of drusen ultrastructures ²⁴ ; and assessment of damages of photoreceptor layer over drusen. ²⁵ Most recent studies also showed that the assessment of drusen using SD-OCT was applicable for monitoring drusen changes in volume and area over time ²⁶ and for assessing drusen in eyes with drusen and geographic atrophy. ²⁷ However, no previous studies reported the correlation of the automatically measured drusen parameters, such as drusen area and maximum drusen size, with those assessed by the gold standard drusen assessment within the Early Treatment of Diabetic Retinopathy Study (ETDRS) grid chart (6 mm) on color fundus photography. The purpose of this study was to compare automated drusen assessment using an originally developed algorithm on SD-OCT (3D OCT-1000; Topcon, Tokyo, Japan) imaging with the results by certified graders using the AREDS drusen grading system from color fundus photography. 

Methods

We prospectively examined 22 patients (22 eyes) with soft drusen who were referred to the Macula Service of the Department of Ophthalmology, Kyoto University Hospital, from January 2008 through November 2008. For inclusion in this study, subjects had a clinical diagnosis of AREDS Category 3 nonneovascular AMD with at least one large druse (diameter ≥ 125 μm) in the macula of the eye under study. ¹⁰ Eyes with noncentral geographic atrophy (GA) were excluded. Eyes with ocular media opacity affecting fundus imaging were excluded from the study. Eyes with other macular abnormalities or any other condition that could cause RPE abnormalities unrelated to soft drusen were excluded. 

All patients had a comprehensive ophthalmologic examination, using the 3D OCT scans (OCT-1000) and single-field nonstereoscopic digital color fundus photographs obtained with the 3D scans (OCT-1000). All investigations of this study adhered to the tenets of the Declaration of Helsinki. This study was approved by the Institutional Review Board and Ethics Committee of the Kyoto University Graduate School of Medicine. Informed consent was obtained for all patients. 

Photographic Drusen Area Grading by Wisconsin Graders

Drusen classification from nonstereo single-field digital color fundus photographs was performed by graders using the AREDS grading protocol, with modification to account for use of a single-field digital image. ^{10 –12} Maximum drusen size and drusen area within the ETDRS grid were evaluated by certified graders at the Fundus Photograph Reading Center at the University of Wisconsin for each subfield of ETDRS grid charts (Fig. 1), according to the Wisconsin grading system. In the current AREDS grading system, the diameter of the optic disc is estimated at 1800 μm compared with the 1500 μm in the original AREDS grading system; consequently, the diameter of the current ETDRS grid is 7200 μm. However, the current SD-OCT instruments do not permit 3D imaging in areas larger than 6.0 × 6.0-mm square area, which is smaller than the current ETDRS grid. In this study, we used the conventional 6000 μm ETDRS grid to compare drusen detection on 3D imaging with the AREDS grading system within the same area. 

Figure 1.

 

ETDRS grid charts used to define subfields in the macular area. Before grading, a grid consisting of three circles concentric with the center of the macula and four radial lines is superimposed over the fundus photograph. The radius of the innermost circle corresponds to 500 μm in the fundus photograph, and the radii of the middle and outer circles correspond to 1500 and 3000 μm, respectively. The length of the green lines for the outer square is 6000 μm, which is the area examined with 128 sequential horizontal SD-OCT scans in the present study.

3D Macular Imaging with SD-OCT

A 6 × 6-mm area centered on the fovea was scanned with a standard raster-scan protocol (using the 3D OCT-1000), which included 128 sequential horizontal scans, each of which consisted of 512 axial scans with a depth of 1.68 mm. This protocol provides horizontal pixel spacings of 11.7 μm (6 mm/512 A-scans), vertical pixel spacings of 46.9 μm (6 mm/128 A-scans), and axial pixel spacing of 3.5 μm (1.68 mm/480 pixels). The image quality of the SD-OCT B-scan images that had an image quality index of >50 was used for analysis. 

Automated Delineation of Macular Drusen with Optical Coherence Tomography

Definition of the Bruch's Membrane.

Custom-made functions for automated delineation of macular drusen on the 3D raster scan data set (originally developed by Topcon) were based on built-in software (3D OCT-1000). We defined drusen on each SD-OCT B-scan image based on two segmented lines: the anterior boundary line of the RPE and the calculated line representing the presumed Bruch's membrane as the drusen floor. However, the reflectivity of the Bruch's membrane line is frequently weak or unseen beneath drusen, leading to segmentation errors. ^{18 –25} In histopathologic specimens, Bruch's membrane remains as a straight line beneath the drusen elevating the RPE. ^28,29 Thus, the presumed Bruch's membrane line beneath the drusen can be approximated by extrapolation from the visible line on either side of drusen. ^20,22,23 In detail, B-scan images were binarized after processing with a median filter. Canny edge detection was used to determine the inner limiting membrane (ILM) line as a boundary from low to high reflectivity (positive edge) in the inner portion of the retina. Another positive edge below the ILM line was detected using the same method, as a photoreceptor inner and outer segment junction (IS/OS) line. The line with the highest reflectivity within 30 pixels (105 μm) below the IS/OS line was determined as the line representing the RPE. The RPE line was globally fitted with straight lines as follows. The lowest point of the segmented RPE line within the B-scan image frame was determined. The nearest point of tangency on the left side of the tangential straight line through the lowest point was determined. Then, another straight line through the nearest point of tangency was used to determine the second nearest point of tangency on the left side. The same process was repeated to the end of the left side. Fitting with straight lines on the right side was carried out in the same way. A quadratic curve along the anterior boundary of the healthy RPE line was calculated using the least-squares method as a calculated Bruch's membrane by referring the fitted straight lines where the distance between the segmented RPE line and fitted straight lines were within 3 pixels. We defined drusen as regions where the distance between the two delineated lines (RPE and Bruch's membrane) was greater than the defined minimal distance (see the following text). 

Area Mapping of Drusen.

OCT fundus images were created as an en face projection image from the SD-OCT data set by integrating the magnitudes of the OCT signals at each lateral position along the axial direction, as previously described. ¹⁷ The en face SD-OCT image was registered to that of color fundus photographs based on the retinal vessel patterns and the optic nerve head as landmarks. Color fundus photographic images were converted into monochromatic images and adjusted in size to the en face SD-OCT images. Next, we performed edge detection based on signal intensity for the en face SD-OCT images. Processing in this way provides characteristic high-contrast images, in which retinal vascular patterns are highlighted. Spatial correlations of vascular patterns between the processed en face SD-OCT images and monochromatic fundus photographs were calculated to determine the optimal agreement in location for both images. If there were apparent gaps between the two images, we manually face SD-OCT images. The delineated drusen area was mapped on each color fundus photograph with the ETDRS grid chart. The drusen area was obtained by 3D mapping of basal areas for the defined drusen that were determined in the 3D volume (“cube”) scan. We did not manually correct segmentation errors on SD-OCT images for analysis of drusen size and area within the grid. 

Determination of Optimal Threshold for Detecting Drusen on B-Scan Images.

To find the optimal threshold distance for defining drusen, we calculated the maximum drusen size and drusen area within the grid by changing the threshold distances from 1 pixel to 6 pixels (1 pixel = 3.5 μm). The calculated maximum drusen size and drusen area within the grid on the 3D SD-OCT data set for each pixel category were compared with the results of the color fundus photograph grading. When the calculated values were within the range of the category determined by certified graders, we regarded this agreement as “agree.” When they were within range of the neighboring category, we regarded the agreement as “agree within 1 step.” When they were out of range of the determined category and its neighboring categories, we regarded the agreement as “disagree.” 

We also counted the number of false-positive drusen at each of the threshold distance categories (6 pixels to 1 pixel) by comparing results of the drusen mapping with the registered color fundus photographs as described in the following text. 

Evaluation of Segmentation Algorithm Failure

All the segmentation algorithm outputs were subjectively evaluated independently by two experts (SO, AS), who were masked to other clinical information. We used criteria of segmentation algorithm failure originally used by Ishikawa et al. ³⁰ In detail, algorithm failures were defined as an obvious disruption of the detected border, and/or border wandering (detected border jumping to and from different anatomic structures) for >5% consecutive (i.e., an uninterrupted error) or 20% cumulative (i.e., adding up all errors amounts to 20% of the image width) of the entire image. 

Statistics

Numbers of false-negative and false-positive drusen among the threshold pixel categories were used to define drusen on SD-OCT images, and the number of automatically detected drusen among subfields of the grid were compared by Kruskal–Wallis test, with post hoc comparisons tested with the Dunnett's rank test. Percentages of agreement and correspondence exact and within one step, were calculated and unweighted and weighted kappa scores were computed to determine intergrader agreement. ³¹ Intraclass correlation coefficients (ICCs) were computed. Statistical software (Statview 5.0; SAS Institute, Cary, NC) was used for statistical analyses. A value of P < 0.05 was considered statistically significant. 

Results

SD-OCT examinations were performed on 22 eyes of 22 Japanese patients with a clinical diagnosis of nonneovascular AMD with soft drusen. Four eyes were excluded because their quality index was <50. The remaining 18 eyes of 18 patients (13 males and 5 females) underwent analysis. These eyes, which had nonneovascular AMD with soft drusen, were found in the fellow eye of 18 patients with exudative AMD; thus, no patients met the inclusion criteria in both of their eyes. The patients were 66 to 90 years of age with a median of 78 years. The mean (±SD) spherical equivalent was −0.36 ± 1.6 (range, −3.25–1.5). The best-corrected visual acuity of the affected eyes ranged from 20/25 to 20/12.5 (median, 20/20 in Snellen equivalent). 

Drusen Grading and Agreement on Color Fundus Photography

Table 1 demonstrates the characteristics of drusen according to conventional AREDS grading system on color photographs. There were definite, predominantly soft indistinct drusen in the macular area, except in one eye. The maximum drusen size was >300 μm except in one eye. Drusen area within the grid was categorized as grade 4 (<790 μm circle) in one eye, categorized as grade 5 (less than half of the disc area) in 2 eyes, categorized as grade 6 (from half to one disc area) in 3 eyes, and categorized as grade 7 (one disc area or more) in 12 eyes. Reproducibility was assessed by verification by a different grader. Agreement for maximum drusen size was 94% (κ = 0.64) and for drusen area within grid 83% (κ = 0.67). 

Table 1.

 

Characteristics of Drusen Graded by the Conventional AREDS Grading System

Subject Number	Maximum Drusen Size	Drusen Area within Grid	Soft Drusen within Grid	Drusenoid PED
1	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
2	≥ drusen circle C2	<1 DA	Predominantly soft indistinct	Absent
3	≥ drusen circle C2	<1 DA	Predominantly soft indistinct	Absent
4	< drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
5	≥ drusen circle C2	<1/2 DA	Predominantly soft indistinct	Absent
6	≥ drusen circle C2	<1/2 DA	soft indistinct present	Absent
7	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
8	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
9	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
10	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
11	≥ drusen circle C2	< drusen circle O2	Predominantly soft indistinct	Absent
12	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
13	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
14	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
15	≥ drusen circle C2	<1 DA	Predominantly soft indistinct	Absent
16	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
17	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent
18	≥ drusen circle C2	≥1 DA	Predominantly soft indistinct	Absent

 The drusen size measurements are based on the modern assumption. C2 = 300 μm; O2 = 790 μm; and 1 DA = 2.54 mm². DA, disc area; PED, pigment epithelial detachment.

Maximum Drusen Size and Drusen Area within Grid Calculated on Various Threshold Distances on 3D OCT

Maximum drusen size and drusen area within the grid were calculated for automatically delineated drusen on SD-OCT images based on various threshold distances (6 pixels to 1 pixel) in the definition of macular drusen (Table 2). Between the results of AREDS grading system and OCT, there was agreement or agreement within one step in all the 18 eyes in the maximum drusen size regardless of the threshold distances. With regard to drusen area within the grid, there was exact agreement or agreement within one step in 15 (83.3%) of 18 eyes regardless of the threshold distances except for 6 pixels (77.7%) as a threshold distance. In particular, exact agreement improved when the threshold algorithm used a distance smaller than 4 pixels. 

Table 2.

 

Maximum Drusen Size and Drusen Area within Grid Calculated on Various Threshold Distances on 3D Optical Coherence Tomography, and Their Agreement with Color Photograph Grading by the AREDS Grading System

Subject	Maximum Drusen Size (μm) on AGS	Maximum Drusen Size (μm) on SD-OCT: Number of Pixels Used to Define Drusen on SD-OCT
		6	5	4	3	2	1
1	300 ≤	1818	1858	1914	1983	2037	2145
2	300 ≤	775	788	804	812	830	842
3	300 ≤	716	735	750	764	780	792
4	150 ≤, < 300	190	211	228	253	332	360
5	300 ≤	517	592	611	633	723	761
6	300 ≤	632	675	704	734	754	776
7	300 ≤	1416	1457	1525	1571	1648	1690
8	300 ≤	831	846	863	877	889	1042
9	300 ≤	1182	1247	1295	1352	1461	1512
10	300 ≤	2291	2341	2413	2594	2706	2779
11	300 ≤	392	436	461	485	507	527
12	300 ≤	679	934	1077	1117	1160	1217
13	300 ≤	787	832	878	914	1203	1235
14	300 ≤	955	987	1654	1738	1992	2204
15	300 ≤	316	337	499	537	563	591
16	300 ≤	419	452	482	511	539	572
17	300 ≤	1186	1218	1248	1282	1310	1336
18	300 ≤	2501	2540	2569	2594	2632	2670
Exact agreement		18	18	18	18	17	17
Agreement within 1 step		0	0	0	0	1	1
Disagreement		0	0	0	0	0	0

Subject	Drusen Area within Grid (%) on AGS	Drusen Area within Grid (%) on SD-OCT: Number of Pixels Used to Define Drusen on SD-OCT
		6	5	4	3	2	1
1	8.9 ≤	13.03	14.13	15.17	16.33	17.46	18.76
2	4.5 ≤, < 8.9	2.15	2.32	2.51	2.80	3.19	3.50
3	4.5 ≤, < 8.9	2.28	2.75	3.29	4.02	4.87	5.68
4	8.9 ≤	0.33	0.53	0.87	1.44	2.21	2.97
5	1.7 ≤, < 4.5	1.76	1.92	2.13	2.41	2.67	2.96
6	1.7 ≤, < 4.5	1.16	1.29	1.43	1.74	1.97	2.25
7	8.9 ≤	6.81	7.51	8.32	9.05	9.81	10.65
8	8.9 ≤	9.22	10.30	11.56	12.95	14.64	16.40
9	8.9 ≤	6.05	7.29	8.60	10.26	11.93	14.14
10	8.9 ≤	28.92	31.36	34.22	37.33	40.80	44.69
11	<1.7	0.62	0.78	0.87	1.00	1.16	1.52
12	8.9 ≤	4.42	5.27	6.30	7.39	8.82	10.69
13	8.9 ≤	9.99	12.17	14.46	16.58	18.62	20.87
14	8.9 ≤	12.92	14.25	15.88	17.78	19.88	22.15
15	4.5 ≤, < 8.9	0.58	0.71	0.85	1.06	1.30	1.55
16	8.9 ≤	1.95	2.19	2.50	2.86	3.19	3.62
17	8.9 ≤	6.57	7.18	7.82	8.65	9.43	10.26
18	8.9 ≤	18.39	19.07	19.88	20.83	22.46	25.05
Exact agreement		8	8	8	11	13	14
Agreement within 1 step		6	7	7	4	2	1
Disagreement		4	3	3	3	3	3

 AGS, Age-Related Eye Disease Study Grading System.

When the threshold algorithm used a distance smaller than 4 pixels, grading results for maximum drusen size obtained by Grader 1 showed exact agreement with those by OCT, whereas results by Grader 2 showed exact agreement with those by OCT in 17 (94.4%) of 18 eyes. For the drusen area, grading results obtained by Grader 1 showed exact agreement with those by OCT in 16 (88.9%) of 18 eyes, whereas results by Grader 2 showed exact agreement with those by OCT in 15 (83.3%) of 18 eyes. 

For intersession reproducibility, ICC values for maximum drusen size and drusen area calculated in selected 10 eyes were 0.999 and 0.996, respectively. 

Characteristics of Cases with Disagreement in the Drusen Area within the Grid

Three cases (cases 4, 15, and 16) showed consistent disagreement in the drusen area within the grid regardless of the threshold distances. The other eyes had varying size of drusen including confluent drusen (Fig. 2). Automated detection of drusen on SD-OCT imaging detected relatively large-sized drusen, including confluent drusen, but did not detect isolated drusen small in size on color fundus photographs. The drusen that were not detected by automated detection on SD-OCT images were characterized on SD-OCT images as being small in height (flattened) and as having a hyporeflective RPE line on the drusen. The three cases with consistent disagreement were characterized as having a large number of small drusen (Fig. 3). Even when we used smaller threshold distances, a large part of these small drusen were not detectable. 

Figure 2.

 

A representative case with exact agreement. An 81-year-old man (subject 14) who had exudative AMD in his right eye was found to have an asymptomatic, nonvascular AMD in the left eye. The visual acuity in the left eye was 20/12.5 in Snellen equivalent. (A) Color fundus photograph with ETDRS grid charts to define subfields of the macula. Soft drusen including a confluent soft drusen and small drusen were seen within the grid. (B–F) Five sectional images obtained by SD-OCT along yellow lines shown in (A). Green lines indicate automatically delineated anterior boundary of the RPE, which showed apparent protrusions of RPE that correspond to confluent soft drusen. Blue lines indicate the calculated Bruch's membrane lines. Red arrows point to the drusen automatically detected by our algorithm using 5 pixels as threshold height, and yellow arrows to drusen that were not detected. (G–L) Six small panels indicate the automatically delineated drusen area when 1, 2, 3, 4, 5, and 6 pixels were set as threshold height. The number in the lower-left corner in each panel indicates the number of pixels used for drusen definition. Drusen area was mapped in blue on the color fundus photographs. Red and yellow arrows (in B–F) point to the drusen indicated by red and yellow arrows (in G–L), respectively.

Figure 3.

 

A representative case with disagreement. A 78-year-old man (subject 4) who had exudative AMD in his right eye was found to have an asymptomatic, nonvascular AMD in the left eye. The visual acuity in the left eye was 20/20 in Snellen equivalent. (A) Color fundus photograph with ETDRS grid charts to define subfields of the macula. There were many small drusen seen in the superior hemisphere of the grid. (B–F) Five sectional images obtained by SD-OCT along yellow lines shown in (A). Green lines indicate automatically delineated anterior boundary of the RPE. Blue lines indicate the calculated Bruch's membrane lines. The protrusions of RPE were mild except for those indicated by red arrows. Red arrows point to the drusen automatically detected by our algorithm using 5 pixels as threshold height, and yellow arrows to drusen that were not detected. (G) The automatically delineated drusen area using 5 pixels as threshold height was mapped in blue on the ETDRS grid charts and color fundus photographs. Red and yellow arrows in B–F point to the drusen indicated by red and yellow arrows in G–L, respectively.

Segmentation Errors

Algorithm failure was detected in 2.95%, 2.99%, and 3.13% of total B-scan images (2304 images of 18 cases) by expert 1, expert 2, and both, respectively (Table 3). Expert assessment of algorithm failure showed almost perfect agreement between experts (κ = 0.95). The number of B-scan images with algorithm failures per eye ranged from 0 (0%) to 25 (19.5%) (mean ± SD = 3.8 ± 6.7 [3.0 ± 5.2%]). 

Table 3.

 

Number of Images with Algorithm Failures Detected by Experts

Total Images	Number of Images with Algorithm Failures
	Expert 1	Expert 2	Both Experts
2304 of 18 cases	69 (2.99)	68 (2.95)	72 (3.13)

 Data are the number of images, with the percentage of total images in parentheses.

Discussion

SD-OCT provides depth information at high axial resolution, allowing detection of fine abnormal elevations of the RPE in various pathologies, such as drusen, ^{18 –23} pigment epithelial detachment, ^32,33 and polypoidal lesions. ³² High-speed imaging in SD-OCT allows dense 3D imaging of these pathologic changes. ^{18 –23} Automated drusen detection using 3D SD-OCT imaging may be a potentially useful alternative method to drusen assessment by human graders using color fundus photographs. Yi et al. ²⁰ reported a preliminary study in which SD-OCT was shown to be able to determine the drusen area and drusen volume in a semiautomated manner, although the general applicability of their method remained unclear. Jain et al. ¹⁹ semiautomatically measured maximum drusen diameter and mean drusen area within a small macular area of 2 mm in diameter in subjects with AREDS Category 3 nonneovascular AMD and found good agreement of these parameters on SD-OCT with those identified on color fundus photographs. Gregori et al. ²² and Schlanitz et al. ²³ were the first to demonstrate the completely automated detection of drusen. Gregori and colleagues ²² showed that SD-OCT allowed highly reproducible automated measurements of the drusen area and volume in subjects with nonneovascular AMD. Schlanitz and colleagues ²³ showed that their automated segmentation algorithm, based on a new SD-OCT technology, polarization-sensitive OCT, identified 96.5% of all drusen without significant error in subjects with AREDS Category 2 or 3 nonneovascular AMD. Thus, agreement between such algorithms to automatically detect drusen and conventional grading of drusen on color fundus photography remained to be determined. 

Drusen are features defined as abnormal appearances of the macular RPE on biomicroscopic examination and color fundus photography. It is currently unclear whether all the drusen determined on the basis of abnormal RPE geometry on OCT B-scans are actually drusen. ²² Abundant evidence regarding the clinical significance of drusen has accumulated based on photographic appearances of drusen; drusen parameters, such as total drusen area and maximum drusen size, categorized in accordance with the AREDS grading protocol on color fundus photographs, have been shown to have positive correlations with risks of progression to advanced AMD, and are now used as standard entry criteria and endpoints for disease progression in AMD clinical trials. ^{4 –9} Thus, photographic appearances of drusen and their grading by certificated graders are the gold standards for implicating drusen in the risk for developing neovascular AMD. Since it is difficult to repeat previous key studies on a large scale, it is important to increase our understanding of the relationship between drusen categories on SD-OCT and color fundus photographs. The present study showed somewhat limited agreement in the drusen area between the drusen parameters measured using our automated drusen detection algorithms on conventional SD-OCT images and the drusen parameters graded by certified graders on color fundus photography. 

We used an automated algorithm to detect drusen. Here, we used a simple method based on threshold height (= threshold distance between the inner boundary of the RPE and calculated Bruch‘s membrane) to define drusen. This automated segmentation method to delineate drusen is basically similar to the highly reproducible method reported by Gregori et al. ²² and the highly sensitive method reported by Schlanitz et al., ²³ in that the distance between the RPE and calculated Bruch's membrane lines (which they call interpolated RPE floor or RPE backbone) was used to delineate drusen. Successful (reproducible and accurate) performance of automated drusen detection using the threshold height depends on at least two important factors: reliable automated segmentation of the highly reflective RPE line ^{21 –23} and the threshold level for discriminating between a significant deviation of the RPE line from the calculated Bruch's membrane line and the noise that accompanies any measurement technique. ²²  

Reliable automated segmentation of the RPE line is influenced by retinal features frequently associated with drusen, such as abnormal hyperreflective lesions in the outer photoreceptor layer, ^34,35 medium internal reflectivity, ²⁵ RPE irregularities, and invisibility of Bruch's membrane. ³⁶ These features could cause segmentation algorithm failures for drusen detection. Schlanitz et al. ²¹ demonstrated detection of drusen using several types of clinical SD-OCT instruments and concluded that the commercially available automated segmentation algorithms had distinct limitations for reliable identification of drusen, especially smaller drusen; the best detection rate of drusen with negligible errors was approximately 30% in the Cirrus (Carl Zeiss Meditec Inc., Dublin, CA) 200 × 200 scan pattern. In our study, algorithm failures appeared to be minimal when inspected by graders (2.95% and 2.99% for the two graders), based on the previously reported definition of segmentation errors. ³⁰ However, it is difficult to compare our results with this previous report on segmentation errors since segmentation errors were previously defined per druse, whereas we calculated segmentation errors per B-scan. In addition, the definitions of segmentation error are different. 

Previous studies using semiautomated or manual segmentation methods to delineate drusen on SD-OCT imaging have compared the results obtained with drusen assessment on color fundus photographs. ^18,19 Jain et al. ¹⁹ compared drusen parameters between drusen segmented on SD-OCT and drusen delineated on color fundus photographs within a macular area of approximately 2 mm in diameter, centered on the fovea. Their method first identified suspected drusen areas based on irregularities in the automatically delineated RPE contour. They then made several manual adjustments to the suspected drusen areas, including adjustment of the lateral extent of marked drusen to correspond to the point at which the RPE deflection returned to baseline, and manual correction of segmentation errors. Freeman et al. ¹⁸ used manual segmentation of drusen on 96 SD-OCT images to determine drusen volumes. Although it is difficult to directly compare the results of these previous studies with those of our studies, these results showed good agreement or significant correlations with drusen assessment on color fundus photographs. 

It was difficult to automatically detect isolated and small drusen using our automated detection algorithms on SD-OCT images. This is consistent with the previous study that showed a trend for less detection of smaller drusen by semiautomated and automated drusen assessment using SD-OCT compared with that using color fundus photographs. ^19,21,22 Thus, the failure to detect some small drusen appears to be a common problem in automated drusen detection using RPE segmentation algorithms on SD-OCT images. This failure may be attributable to both the characteristics of the undetectable drusen and RPE segmentation errors. ^21,22 Flat drusen may also go undetected due to the threshold used by the algorithms. ²² In our study, three cases (cases 4, 15, and 16) showed consistent disagreement in the drusen area within the grid, regardless of the threshold distances. Both of them had similar characteristic drusen patterns on color fundus photographs; these eyes included many small drusen. The undetectable drusen were also characterized as being small in height on OCT B-scan images. Failure to detect small drusen would have some effects on results in many patients, according to the AREDS Report No. 17: 1249 of 3212 participants (38.9%) in the study had only small drusen in their right eye, and 1096 (88%) of the participants had an area less than C-1 (<125 μm). ⁷ It has been well demonstrated that larger drusen are associated with a higher risk of developing neovascular AMD. ^{2 –8} However, a larger amount of drusen increases the drusen area within the grid. It remains to be determined whether this limitation is acceptable. 

The threshold height for discriminating between a significant deviation of the RPE line from the calculated Bruch's membrane line and the noise that accompanies any measurement technique remains unknown. ²² In the present study, we tested agreement with photographic grading results (the drusen gold standard) by changing the threshold height values. We found that there were few differences in the number of eyes with disagreement in the drusen area within the grid even if the number of pixels for definition changed. This is probably because decreasing the threshold height in our algorithm was not sufficient for improving the detection of the small drusen. Thus, this disagreement appears to indicate the limitation of our algorithm. 

It is difficult to completely compare grading on color fundus photography and quantitative drusen assessment using SD-OCT because the former reduces the continuous drusen property into simple categorical data. Our comparison means that we reduced the quantitative drusen property measured with SD-OCT imaging into categorical data. Such a coarse scale, with only three groups for drusen size and four groups for drusen area, can cause more agreement between drusen parameters measured using SD-OCT images and color fundus photographs. We also included the neighboring category (agreement within one step), as a category other than disagreement. 

In clinical practice, patients who have drusen are usually aged and often have media opacity due to cataracts. Greater media opacity often causes poorer OCT B-scan signal strength, which leads to unreliable measurements; lower signal strength is associated with decreased thickness of the macula and retinal nerve fiber layer (RNFL), as suggested by previous studies. ^37,38 Cataract surgery increases both signal strength and RNFL thickness. ³⁸ This could also be the case for drusen assessment, since our method to detect drusen is based on the segmentation of the anterior boundary of the RPE line, similar to measurement of thickness between two boundaries. Therefore, we used eyes with good B-scan images that had an image quality (signal strength) index of >50 for analysis. 

SD-OCT may be complementary to the grading of color photographs for drusen; definition of drusen on color fundus photographs is based on macular pigment abnormalities, whereas on SD-OCT images it is based on abnormal RPE geometry. ²² The clinical significance of drusen detected only on SD-OCT images remains unknown. In addition, SD-OCT imaging can provide new drusen parameters, such as their height and volume of drusen, ^{18 –20,22,23} drusen ultrastructure, ²⁵ and abnormalities of outer retinal layers over drusen. ²⁸ Longitudinal studies are required to determine the relationship between drusen detectable only on SD-OCT and novel drusen parameters visualized by SD-OCT with disease progression in AMD. 

Theoretically, the algorithm that we used for determination of the presumed Bruch's membrane beneath drusen may have an inaccurate approximation of retinal geometry by a quadratic curve when B-scans include unusual retinal configurations, which may occur in eyes with high myopia. However, we did not encounter such issues for the determination of the presumed Bruch's membrane, probably because our subjects did not have high myopia. 

The limitations of this study are small sample size and the bias present in study subject selection. Subjects were limited to those with nonneovascular AMD and at least one large druse (≥125 μm) (AREDS Category 3 nonneovascular AMD); these subjects had predominantly soft indistinct drusen. We focused only on AREDS Category 3 nonneovascular AMD because eyes with AREDS Category 3 nonneovascular AMD have a much increased risk of progression to neovascular AMD. ³⁹ The bias in study subject selection may have caused higher agreement than was actually present, because in clinical practice, some patients have only small drusen that can go undetected by our algorithms. Another limitation of this study is the use of a specific algorithm on a single type of SD-OCT instrument with one specific imaging protocol. The reproducibility and accuracy of drusen detection will differ with algorithm used, imaging protocols, and the type of instruments used. A third limitation of this study was that we used the conventional ETDRS grid with a diameter of 6000 μm. This is because the 3D imaging in the current SD-OCT instrument we used did not include a circle area wider than 6000 μm. However, we found similar agreement with the AREDS grading results on the new ETDRS grid with a diameter of 7200 μm regardless of the differences in the grid area (Iwama D, Hangai M, Yoshimura N, unpublished data, 2011). The development of SD-OCT instruments that allow wider 3D imaging would resolve this limitation. 

Although limited to our particular algorithm for the detection of drusen, and to study subjects with AREDS Category 3 nonneovascular AMD and predominantly soft indistinct drusen, our study successfully showed that SD-OCT allowed automated assessment of drusen area and size based on the threshold distances of the delineated RPE and calculated Bruch's membrane, with minimal segmentation algorithm failures, in good agreement with the categorized drusen parameters assessed by certified graders according to the established AREDS grading protocols on color fundus photography. The advantages of this method as a tool for evaluating drusen remain to be determined in a longitudinal study. 

Footnotes

 Supported in part by Grant-in-Aid for Scientific Research 21249084, Japan Society for the Promotion of Science, Tokyo, Japan; and the Japanese National Society for the Prevention of Blindness.

Footnotes

 Disclosure: D. Iwama, None; M. Hangai, Topcon Corp. (C), Nidek Co. Ltd. (C); S. Ooto, None; A. Sakamoto, None; H. Nakanishi, None; T. Fujimura, Topcon Corp. (E); A. Domalpally, None; R.P. Danis, GSK (C), CoMentis (C), Sangamo (C); N. Yoshimura, Topcon Corp. (C), Nidek Co. Ltd. (C)