Open Access
Add to BibSonomy
Abstract
Qualitative and quantitative assessments of calcified drusen are clinically important for determining the risk of disease progression in age-related macular degeneration (AMD). This paper reports the development of an automated algorithm to segment and quantify calcified drusen on swept-source optical coherence tomography (SS-OCT) images. The algorithm leverages the higher scattering property of calcified drusen compared with soft drusen. Calcified drusen have a higher optical attenuation coefficient (OAC), which results in a choroidal hypotransmission defect (hypoTD) below the calcified drusen. We show that it is possible to automatically segment calcified drusen from 3D SS-OCT scans by combining the OAC within drusen and the hypoTDs under drusen. We also propose a correction method for the segmentation of the retina pigment epithelium (RPE) overlying calcified drusen by automatically correcting the RPE by an amount of the OAC peak width along each A-line, leading to more accurate segmentation and quantification of drusen in general, and the calcified drusen in particular. A total of 29 eyes with nonexudative AMD and calcified drusen imaged with SS-OCT using the 6 × 6 mm2 scanning pattern were used in this study to test the performance of the proposed automated method. We demonstrated that the method achieved good agreement with the human expert graders in identifying the area of calcified drusen (Dice similarity coefficient: 68.27 ± 11.09%, correlation coefficient of the area measurements: r = 0.9422, the mean bias of the area measurements = 0.04781 mm2).
© 2023 Optica Publishing Group under the terms of the Optica Open Access Publishing Agreement
1. Introduction
Age-related macular degeneration (AMD) is a leading cause of irreversible vision among the elderly worldwide [1]. Drusen are the hallmark feature of early and intermediate AMD and are characterized by accumulations of extracellular material that build up in the space between Bruch’s membrane (BM) and the elevated retinal pigment epithelium (RPE) [2–4]. The progression and pathogenesis of AMD have not been fully understood, but the calcification of drusen is considered an indicator of increased risk of late AMD [2,3,5–8] . Therefore, identifying and quantifying calcified drusen would be of great clinical importance in the evaluation and monitoring of AMD progression and for the understanding of the pathological basis of calcified drusen.
Optical coherence tomography (OCT) is a well-established imaging modality that has been widely utilized in ophthalmology for retinal disease diagnosis and treatment management because of its ability to provide non-invasive, depth-resolved, high-resolution information [9–12]. On OCT B-scans, drusen appear as localized elevations and disruptions of the structure of RPE [4,13], while calcified drusen can be identified from heterogeneous internal reflectivity within drusen (HIRD) [5,8]. Liu et al. [3] reported that calcified drusen appear as dark foci on en face subRPE SS-OCT images obtained from a slab defined by boundaries from 64 to 400 µm under BM. These dark foci appear due to the hypotransmission of light into the choroid and are named choroidal hypotransmission defects (hypoTDs). Compared with OCT B-scans, the en face image provides the reviewers and graders an overview of the entire scanned region at a single glance. Using a similar subRPE en face OCT imaging strategy, our group successfully identified and quantified geographic atrophy (GA) by developing an automated algorithm to detect choroidal hypertransmission defects (hyperTDs) [14]. Rather than detecting hyperTDs, we used a similar strategy to detect choroidal hypoTDs to identify and quantify calcified drusen [3]. However, hypoTDs can also result from other anatomic features in the scan other than calcified drusen, such as hyperreflective foci, a thickened RPE layer, vitelliform material, and large retinal pigment epithelial detachments [15], which complicates the grading of calcified drusen from subRPE slabs. Currently, in clinical practice, clinicians and researchers identify the various causes of hypoTDs on subRPE slabs by manually checking the corresponding B-scans. As a result, the task of identifying and quantifying calcified drusen by manual segmentation on subRPE images is labor-intensive, particularly when the corresponding B-scans need to be checked for each hypoTD, and this task is subject to grader variability. Thus, there remains a clinical need for an automated algorithm capable of identifying and quantifying calcified drusen.
Several automated algorithms have been developed and tested to automatically segment retinal anatomic features on OCT images or fundus images, including traditional computer vision [16–19] and machine/deep learning approaches [14,20–28]. However, the identification of calcified drusen presents some unique challenges, especially since the characteristic features of these drusen on 2D en face images sometimes overlap with other causes of hypoTDs. In this study, we propose a simple yet efficient strategy to automatically identify, segment, and quantify calcified drusen on OCT imaging.
2. Methods
2.1. Participants and Imaging acquisition
Patients were enrolled in a prospective SS-OCT imaging study at the Bascom Palmer Eye institute. This study was approved by the institutional review board (IRB) of the University of Miami, Miller School of Medicine and was performed in accordance with the tenets of the Declaration of Helsinki (as revised in 2013) and complied with the Health Insurance Portability and Accountability Act of 1996. All participants provided written informed consent before imaging. OCT scanning was carried out using a 6 × 6 mm2 macular swept-source OCT angiography (SS-OCTA) scanning protocol (PLEX Elite 9000, Carl Zeiss Meditec, Dublin, CA, USA). This instrument was equipped with a 100 kHz swept laser source with a central wavelength of 1050 nm and a spectral bandwidth of 100 nm, providing an axial resolution of ∼5.5 µm and a lateral resolution of ∼20 µm estimated at the retinal surface. The 6 × 6 mm2 scans consisted of 1536 pixels on each A-line (3 mm), 500 A-lines on each B-scan, and 2 repeated B-scans at each B-scan position. OCT scans were excluded from the study if signal strength was less than 7 or evident motion artifacts were observed. After being retrospectively reviewed, the eyes were included in this study only when the B-scans were present with drusen that contained hyperreflective contents with a hyperreflective cap and a hypo-reflective core (Figs. 1(A-C)) or without a hypo-reflective core (Figs. 1 (D-F)) accompanied by choroidal hypoTDs.
Fig. 1. Examples of calcified drusen on the representative OCT scans from (A-C) an eye with calcified drusen with a hyperreflective cap and a hypo-reflective core (Red arrow), and (D-F) an eye with calcified drusen with hyperreflective contents but without a hypo-reflective core (Yellow arrow). (A, D) enface subRPE OCT image obtained from a slab defined by 64 µm to 400 µm below the Bruch’s membrane (BM) shown as yellow lines in (B, E), where hypoTDs appear as the dark foci. (B, E), Representative B-scans passing through calcified drusen at the locations highlighted by dashed lines in (A, D), and (C, F) correspondingly converted OAC B-scans, respectively. Scale bar represents 500 µm.
2.2 OAC estimation from OCT volumes
The OCT signals in the scan dataset were converted to the corresponding optical attenuation coefficient (OAC) at each pixel. OAC describes the attenuation of the OCT signal with depth due to light absorption and scattering, which can reflect the structure and organization of the tissue [29–35]. Since the calcified drusen are characterized by heterogenous optical scattering properties that result in increased OAC values, the OAC can be used to localize, identify, and differentiate calcified drusen from soft drusen. To measure the OAC from OCT signal, simplified single backscattering of light is assumed. The OAC at the zth pixel along the depth, µ(z), can be expressed as [36]:
2.3 Considerations when segmenting RPE
As mentioned above, drusen are characterized by a localized elevation of the RPE on OCT B-scans, and calcified drusen are a subset of drusen in which the accumulated extracellular material possesses higher scattering property (i.e., higher OAC values) when compared with the soft drusen. Therefore, it is important to first identify and segment drusen, for which an accurate segmentation of both RPE and BM is critical. While BM segmentation has proved robust using the manufacture automated segmentation algorithm [38] (which is adopted in this study), the RPE segmentation is a more challenging task. An incorrect segmentation of RPE would lead to difficulty in the identification of the calcified drusen.
Typically, the distance between the RPE and BM at each A-line is utilized to construct a 2D drusen map and to visualize and quantify the drusen. In a previous study, the RPE was located by finding the position of the pixel along each A-line with the largest OAC value above the BM [39]. While this strategy has been successfully used to segment the RPE in many studies, it can run into problems when there are structures with higher OAC values than that of RPE located above the BM (Fig. 2), such as hyperreflective foci (e.g. migrated pigment, Figs. 2(A-C)) and calcified drusen (Fig. 2(E-G)). A possible approach to mitigate this complication is to use a median filter and a local regression with weighted linear least squares and a 2nd-degree polynomial model to physically smooth and correct the inaccurate RPE segmentation [39,40]. Such mitigation works to correct the inaccurate RPE segmentation caused by the hyperreflective foci (Fig. 2(D)), but unfortunately does not work well for the cases with calcified drusen where the corrected RPE appears lower than the actual RPE positions (Fig. 2(H)). For this reason, we further modified the RPE segmentation algorithm by introducing a RPE location correction, Zc, defined as:
where ${\Delta _{FWHM}}$ is the full-width-half-maximum (FWHM) width defined by the OAC peak alone the A-line, and ${\Delta _{RPE}}$ is the RPE width. This means that the RPE location is modified by moving the prior location up by an amount of Zc. Zhou et al found that with SS-OCT imaging, ${\Delta _{RPE}}$ is approximately between 25 and 30 µm [41]. In this study, we used a constant ${\Delta _{RPE}} = 25\mathrm{\mu }m,$ which is consistent with the measurements for older subjects [41]. This approach works because the calcified drusen is a highly scattering mass that would make the FWHM appear wider than that at the soft drusen (see Fig. 3, Fig. 3(D) vs Fig. 3(E)).Fig. 2. Effects of hyper reflective foci (HRF) and calcified drusen on retinal pigment epithelium (RPE) segmentation. (A) OCT B-scan and (B) corresponding optical attenuation coefficient (OAC) B-scan with a HRF (blue arrow). (C) A single OAC A-line through HRF with two peaks highlighted, due to HRF (blue arrow) and RPE (red arrow), respectively, which would make the direct OAC segmentation of RPE erroneous. (D) OAC B-scan with the RPE segmentation (blue line) after filtering and smoothing. (E) OCT B-scan and (F) corresponding OAC B-scan with a calcified druse (purple arrow). (G) A single OAC A-line through the calcified druse where the difficulty to segment the RPE appears apparent. (H) OAC B-scan with the erroneous RPE segmentation (blue line) at the calcified druse. Scale bar represents 500 µm.
Fig. 3. The comparison of the retinal pigment epithelium (RPE) segmentation performances on the soft drusen and calcified drusen among the manufacturer’s approach, OAC method and proposed method. (A) OCT B-scan and (B) corresponding optical attenuation coefficient (OAC) B-scan with soft drusen and calcified drusen with its location highlighted by gray dashed lines in G-I. (C) The same OAC B-scan as in (B) overlaid with the segmented RPE lines obtained by the manufacture software (blue line), OAC method (red line) and the proposed method (yellow line), respectively. (D) A single OAC A-line through the soft druse at the orange line position in (B), where the FWHM at the RPE peak is seen narrow (blue arrow). (E) A single OAC A-line through the calcified druse at the blue line location marked in (B), where the FWHM peak is seen wide (blue arrow) that was used to correct the RPE segmentation lines. (F). Another example OAC B-scan with calcified drusen with its locations marked by dashed red lines in (G-I) overlaid with the segmented RPE lines obtained by the manufacture approach (blue line), OAC method (red line) and proposed method (yellow line), respectively. (G-I) the drusen maps generated from 3D scans by displaying the distances between the Bruch’s membrane (BM) and the RPE segmented by (G) the manufacture approach, (H) the OAC method and (I) the proposed method, respectively. Note that the distance information is coded with color shown in the color bar where the dynamic range was made purposely tighter to show the differences between different methods. White and red arrows indicate the regions of calcified drusen where the RPE segmentations are more accurate by the proposed method but underestimated by either the manufacture or OAC methods. Scale bar represents 500 µm.
Figure 3 shows such an example where at the locations of calcified drusen, the corrected RPE segmentation (yellow line) appears to match well with the actual RPE position (Fig. 3(C) and 3(F)) while the manufacture’s segmentation method (blue line) and the direct OAC with filtering method (red line) appear to be less accurate. In regions where the RPE is normal or where soft drusen are present (Fig. 3(C)), the new RPE segmentation does not deviate significantly from the prior strategies because the FWHM values are sufficiently close to the normal RPE width. Figs. 3(G)–3(I) show drusen maps from the manufacturer’s method (Fig. 3(G)), the prior OAC method (Fig. 3(H)) and the new strategy (Fig. 3(I)), respectively, illustrating this new RPE segmentation provides more accurate drusen volume quantification, particularly in the areas with calcified drusen (pointed by the red and white arrows in Fig. 3(C), (F), and (I)).
2.3 Automated segmentation of calcified drusen
The next step is to segment the calcified drusen from the drusen. Because calcified drusen contain heterogeneous high scattering material that produce higher OAC values than the soft drusen (Fig. 1), the calcified drusen from the drusen can be isolated by mapping the OAC values within the space between (but excluding) RPE and BM. In order to minimize the possibility of false positive identifications due to small deviations in the segmentations (see below), we used the hypoTDs information to constrain the mapping of calcified drusen.
The workflow schematic of the algorithm is shown in Fig. 4. Two parallel steps, each generating a binary map (Fig. 4(F) and 4(I), respectively) from a 3D SS-OCT scan, are combined to produce a final binary map (Fig. 4(J)) to represent the regions occupied by the calcified drusen.
Fig. 4. Schematic of the workflow for the proposed automated segmentation of calcified drusen from a SS-OCT volume scan. (A) Optical attenuation coefficient (OAC) B-scans with the segmentation lines of six pixels below retinal pigment epithelium (RPE, the blue line) to two pixels above Bruch’s membrane (BM, the orange line) highlighted. (B) the drusen maps resulted from displaying the distances between the Bruch’s membrane (BM) and the retinal pigment epithelium (RPE). (C) en face maximum mean projection OAC image of the slab defined by the two lines shown in (A). (D-E) Representative OAC B-scans with its locations marked by dashed red lines in (B and C) overlaid with the segmentation lines of six pixels below RPE (the blue line) to two pixels above BM (the orange line), showing small deviations in the RPE segmentations leading to false positive detection of calcified drusen (the green arrow). (F) The first binary mask derived from (C). (G) OCT B-scans with the segmentation lines (the yellow lines) located within the sclera region to define the slab thickness of 40 µm, where the first line is determined by shifting the BM segmentation line by the maximum choroidal thickness in the volume. (H) en face sub-choroid OCT image of the slab defined the two lines shown in (G). (F) the second binary mask derived from (H). (J) Final binary mask resulted from the product between (F) and (I) to indicate the regions of calcified drusen.
The first step is to obtain a binary map indicating the regions where heterogeneous scattering material may possibly be present between the RPE and BM. A drusen map (Fig. 4(B)) was first generated from the scanned OCT volume (Fig. 4(A)) using the corrected RPE segmentation discussed in the last section. Note that the drusen volume can be easily calculated from the drusen map shown in Fig. 4(B). From this drusen map, an en face OAC images (Fig. 4(C)) was then obtained from a slab taken from six pixels (∼12 µm) below RPE segmentation line to two pixels (∼4 µm) above the BM segmentation line (shown in blue and orange lines in Fig. 4(A)) using the mean of the first six maximum OAC values at each A-line. A binary image (Fig. 4(F)) is then generated from this en face OAC map by using the automated Otsu’s histogram thresholding method and removing small areas less than 125 µm in greatest linear dimension (GLD, consistent with human grader criterion). Theoretically, this binary image would be sufficient to identify the regions occupied by large drusen (≥125 µm in GLD) with high OAC values. In practice, there is no guarantee that the segmentations of RPE are perfect, especially in the area with a V-shape appearance at the border region of some drusen, for example at the positions pointed by green arrows in Fig. 4(D-E), which would make the RPE segmentation to slightly deviate from the true RPE position. Such deviation can lead to false positive identification of the calcified drusen (e.g., those pointed by the green arrows in Fig. 4(C) and Fig. 4(F)). One approach to minimize such false positive identifications is to further optimize the proposed RPE segmentation strategy by taking these special cases into the formulation of the segmentation algorithm. However, in this study, we instead introduced a constraint on the resulting binary map (Fig. 4(F)), i.e. the 2nd step below.
The second step is to obtain a binary map corresponding to hypoTDs in the sclera, which is used to constrain the possible locations of calcified drusen identified from the 1st step. Conventionally, the hypoTDs can be identified from the en face structural image generated by the subRPE slab [3]. However, the large choroidal vessels often complicate the identification of the hypoTDs using this method. To avoid this complication, we instead defined a slab located within the sclera where the OCT signal is more homogeneous, while preserving the appearance of the hypoTDs (e.g., pointed by arrow in Fig. 4(G)). The largest distance between BM and choroid-sclera boundary, denoted as X µm for simplicity, was first calculated using a validated and published automated algorithm [37]. Then the slab from X µm to (X + 40) µm below BM (shown in yellow lines in Fig. 4(G)) was used to generate an en face OCT sub-choroid image using a mean intensity projection method (Fig. 4(H)). Finally, the en face OCT sub-choroid image was binarized (Fig. 4(I)) using a locally adaptive thresholding method.
The final binary map (Fig. 4(J)) is obtained by the product between the first (Fig. 4(F)) and second (Fig. 4(I)) binary maps, showing the regions that are occupied by the calcified drusen. With this final binary map, it is then trivial to calculate the calcified drusen volumes when combined with the drusen map shown in Fig. 4(B).
In order to validate and compare the outcomes from the proposed automated algorithm, the calcified drusen were manually outlined by two independent expert graders (J.L. and M.S.) using Photoshop (Adobe Systems, San Jose, California, USA), and agreement was reached between the two graders regarding the area of calcified drusen. In cases of disagreement, a senior grader (P.J.R.) worked as the adjudicator.
2.4 Statistical analysis and evaluation metrics
Scatter plots along with Pearson correlation analyses were used to explore the relationship between the area and volume measurements of the calcified drusen from the manual segmentations and the automated algorithm. Statistical analysis was performed using GraphPad Prism (GraphPad Software, San Diego, CA, USA).
To evaluate the performance of the algorithms, Dice similarity coefficient (DSC), pixel-wise sensitivity, and specificity were measured on the data set:
where TP indicates true positive, TN indicates true negative, FP indicates false positive, and FN indicates false negative. Here the areas outlined by the human graders were considered as ground truth or reference standard.3. Results
A total of 29 eyes with nonexudative AMD and calcified drusen from 29 patients were recruited in this study. 28 of these eyes contained hyper reflective foci lesions and 6 eyes had GA.
In the evaluation, we also evaluated the performances of the algorithm without employing the 2nd step to show the necessity of using the hypoTD information appearing in sub-choroid slab to minimize the false positive identifications, at least in the current study. The reason for this evaluation is that sub-choroid evaluation requires reliable OCT signals from deep choroid and sclera structures, which is achievable for SS-OCT imaging but often difficult for SD-OCT imaging. For simplicity in the evaluation, we named this evaluation as AccuRPE. Table 1 shows the results of our analyses. The mean DSC was 40.39% when using the manufacturer’s segmentation and 47.65% using the OAC method, respectively. Note that the manufacture and OAC methods did not involve the sub-choroidal information. This value was improved to 51.21% when using the proposed adjustment of RPE segmentation alone, i.e., the AccuRPE. However, when the sub-choroid information was used in the algorithm, the mean DSC of the proposed algorithm was significantly improved to 68.27% (AccuRPE vs Proposed approach, P value of paired t-test < 0.0001). The resulted sensitivity and specificity values are also provided in Table 1. Taken together, these analyses demonstrate that the proposed methods outperform either manufacture or traditional OAC approach, with the method being the most accurate when the hypoTD information in the choroid is considered in the algorithm.
Table 1. Performance for calcified drusen segmentation among different methodsa
Figure 5 shows the quantitative scatter plots of the areas and volume measurements of identified calcified drusen along with Pearson’s correlation analyses for the proposed algorithm vs the manual segmentations, respectively. A significant correlation was found between the measurements of the calcified drusen area (r = 0.9422, p < 0.0001, Fig. 5(A)), and the corresponding volume measurements (r = 0.9391, p < 0.0001, Fig. 5(B)). Bland-Altman analysis showed that the average bias of the area measurements was 0.04781 mm2 (95% limits of agreement [−0.1218, 0.2174], Fig. 5(C)) and the average bias of the corresponding volume measurements was 0.0023 mm3 (95% limits of agreement [−0.0105, 0.0150], Fig. 5(D)) between automated and manual approaches. Therefore, the measurements of calcified drusen computed using the proposed method tended to be a little larger than those produced by the manual segmentations.
Fig. 5. Areas and volumes of calcified drusen measured by the proposed automated methods vs the ground truth (by human graders). (A) Scatter plot of the area measurements of identified calcified drusen against manual measurements showing a strong correlation (r = 0.9422, p < 0.0001), and (B) corresponding scatter plot of the volume measurements with a correlation (r = 0.9391, p < 0.0001). (C and D) Bland-Altman agreement analysis of identified calcified drusen against manual measurements, where the solid line represents the bias, and the dashed gray lines represent the upper and lower 95% limits of agreement.
Figure 6 demonstrates these results for one representative eye with nonexudative AMD and calcified drusen. Overall, there was good agreement between the segmentations from the automated strategy (Fig. 6(H)) and those from the human graders (Fig. 6(D)), with a dice coefficient of 75.81%. The segmentations of calcified drusen can be used to visualize the calcified drusen as a subset of all drusen, which may be useful, for example, in investigating the progression from soft drusen to calcified drusen. For example, the binary map of Fig. 6(H) can be combined with the drusen map of Fig. 6(F) to generate a composite map of Fig. 6(I) to differentiate the calcified drusen (red color) from the soft drusen (green color) at a single glance. It is also trivial to show the RPE elevation map for the calcified drusen only (Fig. 6(J)), which may also be useful in the clinical investigations. Combining the binary maps (Fig. 6(D) and 6(H)) with the drusen map in Fig. 6(F), the quantitative volume measurements of the calcified drusen were 0.0211mm3 and 0.0246 mm3 for the manual and automated strategies, respectively.
Fig. 6. Representative segmentation results of calcified drusen taken from an eye in 68-year-old female patient by both the manual method (top row) and the proposed algorithm (bottom row). (A) en face OCT subRPE image. (B) An OCT B-scan with its location shown as dashed line in (A) where a calcified druse can be identified. (C) Manual outlines on the subRPE image in (A) of calcified drusen (yellow outlines), hyperreflective foci (green outlines), and choroidal hypertransmission defects (red outlines). (D) Binary image showing the regions with calcified drusen segmented by human graders. (E) en face OCT sub-choroid image obtained by a slab of 40 µm thickness with its anterior line defined by shifting the BM segmentation by the maximal thickness of the choroid, showing minimal contamination of choroidal vessels. (F) en face drusen map with the color indicating the elevation distance of RPE. (G) en face maximum mean projection OAC image between RPE and BM with the color indicating the values of OAC. (H) Binary map resulted from (G) to indicate the regions occupied by the calcified drusen using the automated method. (I) An en face composite map to visually indicate whether a druse is calcified within the drusen map (F), which was obtained with the aid of (H) when displaying (F) where the regions of calcified drusen are displayed with red color and the soft drusen are displayed with green color. Color bar represents the distance between RPE and BM.(J) en face calcified drusen map with the color indicating the elevation distance of RPE. Scale bar represents 500 µm.
Figure 7 shows an example from a case (pointed by blue arrow in Fig. 5) with an obvious disagreement between the automated and manual segmentations (see Fig. 7(C) vs Fig. 7(F)). For this case, the dice coefficient was 50.68%, and the measured area and volume of calcified drusen were 0.1315 mm2 and 0.0081mm3 for manual segmentation, and 0.3584mm2 and 0.0252mm3 for the automated strategy, respectively. To inspect in detail the possible reasons for the disagreement, we browsed through B-scan by B-scan in the 3D volume. Arrows show the representative positions with possible true/false positive and true/false negative identifications of the calcified drusen by both the manual and automated methods. The orange arrows point to the lesions that were correctly identified as calcified drusen by both the methods. The green arrows point to a true calcified druse that was correctly identified by the automated method but was missed by the human graders, probably because of its small size and undiscernible hypoTD in choroid (Fig. 7(A)). Note that the hypoTD feature is present in the B-scan (hollow green arrow, Fig. 7(H)) and heterogenous OCT reflectivity with the druse. The yellow arrows point to a position that was falsely identified as a calcified druse by the algorithm but correctly excluded by the human graders. This false-positive identification can be explained by an incorrect segmentation of the RPE (because of its V-shape) and by the RPE pigments clumped at the apex of the V-shape that strongly attenuates the light transmission, leading to a hypoTD appearance in the choroid and below (hollow yellow arrow, Fig. 7(I)). The red arrows point to the position where there were migrated pigments attached to the RPE, which was falsely identified by the automated algorithm as a calcified drusen but was excluded by the manual graders. The highly scattering nature of the pigment clump that is attached to the RPE confuses the automated RPE segmentation algorithm, which overcorrects the RPE by the thickness of the pigment clump. To mitigate this mis-identification, one option is to use an automated algorithm of identifying the hyperreflective foci [38] to exclude the regions designated as hyperpigmentation on RPE. Another option but more laborious is to manually check individual B-scans to confirm wherever the calcified drusen are correctly identified by the algorithm.
Fig. 7. An example from an eye in 74-year-old male patient showing the areas of calcified drusen segmented by the algorithm being a relatively large disagreement with the areas segmented by human graders. (A) en face OCT subRPE image with highlighted B-scan positions for detailed inspections of segmented calcified drusen by both the manual and automated approaches. (B) en face drusen map with the color indicating the elevation distance of RPE. (C) Binary map resulted from the manual segmentation, showing the regions identified with calcified drusen. (D) en face OCT sub-choroid image showing the hypoTDs with minimal contamination of choroidal vessels. (E) en face maximum mean projection optical attenuation coefficient (OAC) image of a slab taken between retinal pigment epithelium (RPE) to Bruch’s membrane (BM), showing the OAC strength within the drusen. (F) Binary map resulted from the automated algorithm showing the regions occupied by the identified calcified drusen. (G-I) Representative OCT B-scans with their locations indicated by the dashed lines in (A); and (J-L) the corresponding OAC B-scans, respectively. In the images, the orange arrows indicate the true positive area detected by both the automated and manual methods. The green arrows point to a calcified druse that was detected by the automated method but missed by the human graders. The yellow arrows point to the position where the automated algorithm led to a false positive identification of the calcified druse due to a clear challenge of RPE segmentation at this location. The red arrows point to a pigment clump on the RPE, which was falsely identified as a calcified druse by the automated method but correctly excluded by the manual method. Scale bar represents 500µm.
4. Discussion and conclusion
Qualitative and quantitative assessments of calcified drusen and soft drusen are of great clinical importance when evaluating AMD process [4,5]. However, to the best of the authors’ knowledge, no study regarding automated segmentation of calcified drusen has been reported. In this study, we proposed an automated algorithm using OCT-derived OAC features to improve the RPE segmentation and provide a more accurate drusen area map compared to previous methods (Fig. 3, [14]). Based on this improvement, we proposed a novel automated strategy to segment and quantify the area and volume of calcified drusen. Although there is no consistent definition of calcified drusen, several groups reported that calcified drusen showed heterogeneous internal reflectivity within drusen (HIRD) on OCT B-scans [3,5,6,8,42]. Moreover, Tan et al. [5] proposed a four-stage pathway for druse progression to HIRD, stating that drusen with hyperreflective contents but without a hyporeflective core were at an early stage of HIRD. In this study, both HIRD with and without a hyporeflective cores were considered to be calcified drusen, which is the rationale of the first step in our automated strategy for segmenting calcified drusen (Fig. 4) because these HIRD are highly scattering and can be contrasted by OAC. Suzuki et al. [8] and Liu et al. [3] reported that there were shadows beneath calcified drusen, a feature which may be due to the attenuation or scattering of incident light caused by calcified nodules. These shadows give rise to the appearance of hypoTD on sub-RPE and sub-choroidal images. [3] The combination of higher OAC within drusen and hypoTDs below BM are used in our automated algorithm to segment calcified drusen from 3D OCT scans.
The proposed automated algorithm was compared with calcified drusen segmentation performed by human graders. Some differences were observed when a calcified druse was small causing less attenuation of the OCT light entering the deeper structures than larger lesions, which leads to a reduced contrast of hypoTD appearance in the en-face subRPE images by the green arrows pointing to algorithm identified calcified druse in Fig. 7. Such low contrast appearance of hypoTDs makes the human graders difficult to judge, thus often leading to missed grading of the calcified drusen. However, in this example, the automated algorithm appears more sensitive to these small lesions, but the significance of these small lesions remains to be investigated.
There are several possible situations that may cause false positive segmentations of the calcified drusen. Figure 7 shows an example where a pigment clump attached to the RPE was falsely identified as calcified drusen by the automated segmentation method. This situation presents a challenge for the proposed automated RPE segmentation, which might place the RPE above the pigment clamp leading to erroneous identification of the pigment clump as calcified drusen. These foci of hyperpigmentation can also be identified using another algorithm designed to identify these lesions [38], and any overlap between the two algorithms could be investigated further. Overall, the automated algorithm appears to have a small bias compared to manual outlines, over-estimating the areas of calcified drusen compared with the human graders (Fig. 5).
The current development was based on SS-OCT imaging, and it is unknown if the proposed method can be applied to spectral domain OCT (SD-OCT) datasets since SD-OCT has a worse sensitivity roll-off along the depth and higher optical scattering property of the RPE complex compared with SS-OCT [43–45]. In this case, the use of the hypoTD information in the sub-choroid slab, even in the subRPE slab, could be problematic. For the development of an algorithm that would also be applicable to the SD-OCT imaging, the algorithm would be ideal if only the step 1 in the proposed automated approach (Fig. 4) is involved. This will be our immediate future effort to further improve and optimize the RPE segmentation algorithm so that it can cope with the special cases as we discussed above.
One important future direction is the development of deep learning algorithms to improve the performance and efficiency of calcified drusen segmentation. However, this would require annotated features from a large number of OCT datasets, which is a process that is underway. It is reasonable to believe that the proposed automated algorithm would play an important role in guiding manual graders and facilitating the annotations of calcified drusen needed to develop deep learning algorithms.
In conclusion, we have proposed a novel strategy to segment and quantify calcified drusen via the use of its high optical scattering properties and the characteristic hypotransmission defects observed in slabs below BM. Moreover, we demonstrated an improved RPE segmentation that should provide more accurate drusen maps.
Funding
Carl Zeiss Meditec Inc; Salah Foundation; Research to Prevent Blindness; National Eye Institute.
Disclosures
Dr. Gregori, Dr. Rosenfeld and Dr. Wang received research support from Carl Zeiss Meditec, Inc. Dr. Gregori and the University of Miami co-own a patent that is licensed to Carl Zeiss Meditec, Inc. Dr. Gregori and Dr. Rosenfeld received support the National Eye Institute Center Core Grant (P30EY014801) and Research to Prevent Blindness (unrestricted Grant) to the Department of Ophthalmology, University of Miami Miller School of Medicine. Dr. Rosenfeld also received research funding from Gyroscope Therapeutics and Stealth BioTherapeutics. He is also a consultant for Boehringer-Ingelheim, Carl Zeiss Meditec, Chengdu Kanghong Biotech, InflammX/Ocunexus Therapeutics, Ocudyne, Regeneron Pharmaceuticals, and Unity Biotechnology. He also has equity interest in Apellis, Valitor, and Ocudyne. Dr. Wang discloses intellectual property owned by the Oregon Health and Science University and the University of Washington. Dr. Wang also receives research support from Estee Lauder Inc, and Colgate Palmolive Company. He is a consultant to Carl Zeiss Meditec. All other authors have no disclosures.
Data availability
Data underlying the results presented in this paper are not publicly available at this time but may be obtained from the authors upon reasonable request.
Reference
1. M. Fleckenstein, T. D. L. Keenan, R. H. Guymer, U. Chakravarthy, S. Schmitz-Valckenberg, C. C. Klaver, W. T. Wong, and E. Y. Chew, “Age-related macular degeneration,” Nat Rev Dis Primers 7(1), 31 (2021). [CrossRef]
2. K. Hirabayashi, H. J. Yu, Y. Wakatsuki, K. M. Marion, C. C. Wykoff, and S. R. Sadda, “OCT risk factors for development of atrophy in eyes with intermediate age-related macular degeneration,” Ophthalmol. Retina (2022). [CrossRef]
3. J. Liu, R. Laiginhas, M. Shen, Y. Shi, J. Li, O. Trivizki, N. K. Waheed, G. Gregori, and P. J. Rosenfeld, “Multimodal imaging and en face OCT detection of calcified drusen in eyes with age-related macular degeneration,” Ophthalmol. Sci. 2(2), 100162 (2022). [CrossRef]
4. G. Gregori, F. Wang, P. J. Rosenfeld, Z. Yehoshua, N. Z. Gregori, B. J. Lujan, C. A. Puliafito, and W. J. Feuer, “Spectral domain optical coherence tomography imaging of drusen in nonexudative age-related macular degeneration,” Ophthalmology 118(7), 1373–1379 (2011). [CrossRef]
5. A. C. S. Tan, M. G. Pilgrim, S. Fearn, S. Bertazzo, E. Tsolaki, A. P. Morrell, M. Li, J. D. Messinger, R. Dolz-Marco, J. Lei, M. G. Nittala, S. R. Sadda, I. Lengyel, K. B. Freund, and C. A. Curcio, “Calcified nodules in retinal drusen are associated with disease progression in age-related macular degeneration,” Sci. Transl. Med. 10(466), eaat4544 (2018). [CrossRef]
6. Y. Ouyang, F. M. Heussen, A. Hariri, P. A. Keane, and S. R. Sadda, “Optical coherence tomography-based observation of the natural history of drusenoid lesion in eyes with dry age-related macular degeneration,” Ophthalmology 120(12), 2656–2665 (2013). [CrossRef]
7. C. Bonnet, G. Querques, J. Zerbib, H. Oubraham, R. B. Garavito, N. Puche, and E. H. Souied, “Hyperreflective pyramidal structures on optical coherence tomography in geographic atrophy areas,” Retina 34(8), 1524–1530 (2014). [CrossRef]
8. M. Suzuki, C. A. Curcio, R. F. Mullins, and R. F. Spaide, “Refractile drusen: clinical imaging and candidate histology,” Retina 35(5), 859–865 (2015). [CrossRef]
9. M. Adhi and J. S. Duker, “Optical coherence tomography – current and future applications,” Curr Opin Ophthalmol 24(3), 213–221 (2013). [CrossRef]
10. E. A. Swanson and J. G. Fujimoto, “The ecosystem that powered the translation of OCT from fundamental research to clinical and commercial impact [Invited],” Biomed. Opt. Express 8(3), 1638–1664 (2017). [CrossRef]
11. C. L. Chen and R. K. Wang, “Optical coherence tomography based angiography [Invited],” Biomed. Opt. Express 8(2), 1056–1082 (2017). [CrossRef]
12. A. H. Kashani, C. L. Chen, J. K. Gahm, F. Zheng, G. M. Richter, P. J. Rosenfeld, Y. Shi, and R. K. Wang, “Optical coherence tomography angiography: a comprehensive review of current methods and clinical applications,” Prog. Retinal Eye Res. 60, 66–100 (2017). [CrossRef]
13. R. F. Spaide and C. A. Curcio, “Drusen characterization with multimodal imaging,” Retina 30(9), 1441–1454 (2010). [CrossRef]
14. Z. Chu, L. Wang, X. Zhou, Y. Shi, Y. Cheng, R. Laiginhas, H. Zhou, M. Shen, Q. Zhang, L. de Sisternes, A. Y. Lee, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Automatic geographic atrophy segmentation using optical attenuation in OCT scans with deep learning,” Biomed. Opt. Express 13(3), 1328–1343 (2022). [CrossRef]
15. R. Laiginhas, J. Liu, M. Shen, Y. Shi, O. Trivizki, N. K. Waheed, G. Gregori, and P. J. Rosenfeld, “Multimodal imaging, OCT B-scan localization, and en face OCT detection of macular hyperpigmentation in eyes with intermediate AMD,” Investigative Ophthalmology & Visual Science 2(2), 100116 (2022). [CrossRef]
16. S. Niu, L. de Sisternes, Q. Chen, T. Leng, and D. L. Rubin, “Automated geographic atrophy segmentation for SD-OCT images using region-based C-V model via local similarity factor,” Biomed. Opt. Express 7(2), 581–600 (2016). [CrossRef]
17. A. K. Feeny, M. Tadarati, D. E. Freund, N. M. Bressler, and P. Burlina, “Automated segmentation of geographic atrophy of the retinal epithelium via random forests in AREDS color fundus images,” Comput. Biol. Med. 65, 124–136 (2015). [CrossRef]
18. D. J. Ramsey, J. S. Sunness, P. Malviya, C. Applegate, G. D. Hager, and J. T. Handa, “Automated image alignment and segmentation to follow progression of geographic atrophy in age-related macular degeneration,” Retina 34(7), 1296–1307 (2014). [CrossRef]
19. I. P. Okuwobi, Z. Ji, W. Fan, S. Yuan, L. Bekalo, and Q. Chen, “Automated quantification of hyperreflective foci in SD-OCT with diabetic retinopathy,” IEEE J. Biomed. Health Inform. 24(4), 1125–1136 (2020). [CrossRef]
20. Z. Ji, Q. Chen, S. Niu, T. Leng, and D. L. Rubin, “Beyond retinal layers: a deep voting model for automated geographic atrophy segmentation in SD-OCT images,” Trans. Vis. Sci. Tech. 7(1), 1 (2018). [CrossRef]
21. R. Xu, S. Niu, Q. Chen, Z. Ji, D. Rubin, and Y. Chen, “Automated geographic atrophy segmentation for SD-OCT images based on two-stage learning model,” Comput. Biol. Med. 105, 102–111 (2019). [CrossRef]
22. T. D. Keenan, S. Dharssi, Y. Peng, Q. Chen, E. Agrón, W. T. Wong, Z. Lu, and E. Y. Chew, “A deep learning approach for automated detection of geographic atrophy from color fundus photographs,” Ophthalmology 126(11), 1533–1540 (2019). [CrossRef]
23. Z. Hu, Z. Wang, and S. R. Sadda, “Automated segmentation of geographic atrophy using deep convolutional neural networks,” Medical Imaging 2018: Computer-Aided Diagnosis (SPIE, 2018), 10575, pp. 229–237. [CrossRef]
24. M. Wu, X. Cai, Q. Chen, Z. Ji, S. Niu, T. Leng, D. L. Rubin, and H. Park, “Geographic atrophy segmentation in SD-OCT images using synthesized fundus autofluorescence imaging,” Comput. Methods Programs Biomed. 182, 105101 (2019). [CrossRef]
25. L. Varga, A. Kovács, T. Grósz, G. Thury, F. Hadarits, R. Dégi, and J. Dombi, “Automatic segmentation of hyperreflective foci in OCT images,” Comput. Methods Programs Biomed. 178, 91–103 (2019). [CrossRef]
26. O. Mukherjee, T. De Silva, P. Grisso, H. Wiley, D. L. K. Tiarnan, T. A. Thavikulwat, E. Chew, and C. Cukras, “Retinal layer segmentation in optical coherence tomography (OCT) using a 3D deep-convolutional regression network for patients with age-related macular degeneration,” Biomed. Opt. Express 13(6), 3195–3210 (2022). [CrossRef]
27. D. A. Y. Cahyo, A. P. Yow, S. M. Saw, M. Ang, M. Girard, L. Schmetterer, and D. Wong, “Multi-task learning approach for volumetric segmentation and reconstruction in 3D OCT images,” Biomed. Opt. Express 12(12), 7348–7360 (2021). [CrossRef]
28. T. T. Yu, D. Ma, J. Lo, M. J. Ju, M. F. Beg, and M. V. Sarunic, “Effect of optical coherence tomography and angiography sampling rate towards diabetic retinopathy severity classification,” Biomed. Opt. Express 12(10), 6660–6673 (2021). [CrossRef]
29. S. Chang and A. K. Bowden, “Review of methods and applications of attenuation coefficient measurements with optical coherence tomography,” J. Biomed. Opt. 24(09), 1 (2019). [CrossRef]
30. W. J. Choi, Y. Li, and R. K. Wang, “Monitoring acute stroke progression: multi-parametric OCT imaging of cortical perfusion, flow, and tissue scattering in a mouse model of permanent focal ischemia,” IEEE Trans. Med. Imaging 38(6), 1427–1437 (2019). [CrossRef]
31. T. M. Cannon, B. E. Bouma, and N. Uribe-Patarroyo, “Layer-based, depth-resolved computation of attenuation coefficients and backscattering fractions in tissue using optical coherence tomography,” Biomed. Opt. Express 12(8), 5037–5056 (2021). [CrossRef]
32. B. Ghafaryasl, K. A. Vermeer, J. Kalkman, T. Callewaert, J. F. de Boer, and L. J. van Vliet, “Attenuation coefficient estimation in Fourier-domain OCT of multi-layered phantoms,” Biomed. Opt. Express 12(5), 2744–2758 (2021). [CrossRef]
33. J. Kübler, V. S. Zoutenbier, A. Amelink, J. Fischer, and J. F. de Boer, “Investigation of methods to extract confocal function parameters for the depth resolved determination of attenuation coefficients using OCT in intralipid samples, titanium oxide phantoms, and in vivo human retinas,” Biomed. Opt. Express 12(11), 6814–6830 (2021). [CrossRef]
34. S. Chang, C. Murff, T. Leng, S. L. Groth, and A. K. Bowden, “Depth-resolved extraction of optical attenuation for glaucoma assessment in clinical settings: a pilot study,” Biomed. Opt. Express 13(8), 4326–4337 (2022). [CrossRef]
35. A. A. Moiseev, K. A. Achkasova, E. B. Kiseleva, K. S. Yashin, A. L. Potapov, E. L. Bederina, S. S. Kuznetsov, E. P. Sherstnev, D. V. Shabanov, G. V. Gelikonov, Y. V. Ostrovskaya, and N. D. Gladkova, “Brain white matter morphological structure correlation with its optical properties estimated from optical coherence tomography (OCT) data,” Biomed. Opt. Express 13(4), 2393–2413 (2022). [CrossRef]
36. K. A. Vermeer, J. Mo, J. J. A. Weda, H. G. Lemij, and J. F. de Boer, “Depth-resolved model-based reconstruction of attenuation coefficients in optical coherence tomography,” Biomed. Opt. Express 5(1), 322–337 (2014). [CrossRef]
37. H. Zhou, Z. Chu, Q. Zhang, Y. Dai, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Attenuation correction assisted automatic segmentation for assessing choroidal thickness and vasculature with swept-source OCT,” Biomed. Opt. Express 9(12), 6067–6080 (2018). [CrossRef]
38. H. Zhou, J. Liu, R. Laiginhas, Q. Zhang, Y. Cheng, Y. Zhang, Y. Shi, M. Shen, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Depth-resolved visualization and automated quantification of hyperreflective foci on OCT scans using optical attenuation coefficients,” Biomed. Opt. Express 13(8), 4175 (2022). [CrossRef]
39. Z. Chu, Y. Shi, X. Zhou, L. Wang, H. Zhou, R. Laiginhas, Q. Zhang, Y. Cheng, M. Shen, L. de Sisternes, M. K. Durbin, W. Feuer, G. Gregori, P. J. Rosenfeld, and R. K. Wang, “Optical coherence tomography measurements of the retinal pigment epithelium to Bruch membrane thickness around geographic atrophy correlate with growth,” Am. J. Ophthalmol. 236, 249–260 (2022). [CrossRef]
40. I. Gijbels and I. Prosdocimi, “Loess,” WIREs Computational Statistics 2(5), 590–599 (2010). [CrossRef]
41. H. Zhou, Y. Dai, G. Gregori, P. R. Rosenfeld, J. L. Duncan, D. M. Schwartz, and R. K. Wang, “Automated morphometric measurement of the retinal pigment epithelium complex and choriocapillaris using swept source OCT,” Biomed. Opt. Express 11(4), 1834–1850 (2020). [CrossRef]
42. I. Byon, Y. Ji, A. R. Alagorie, L. Tiosano, and S. R. Sadda, “Topographic assessment of choriocapillaris flow deficits in the intermediate age-related macular degeneration eyes with hyporeflective cores inside drusen,” Retina 41(2), 393–401 (2021). [CrossRef]
43. A. R. Miller, L. Roisman, Q. Zhang, F. Zheng, J. R. d. Oliveira Dias, Z. Yehoshua, K. B. Schaal, W. Feuer, G. Gregori, Z. Chu, C.-L. Chen, S. Kubach, L. An, P. F. Stetson, M. K. Durbin, R. K. Wang, and P. J. Rosenfeld, “Comparison between spectral-domain and swept-source optical coherence tomography angiographic imaging of choroidal neovascularization,” Invest. Ophthalmol. Vis. Sci. 58(3), 1499–1505 (2017). [CrossRef]
44. Q. Zhang, C.-L. Chen, Z. Chu, F. Zheng, A. Miller, L. Roisman, J. R. d. Oliveira Dias, Z. Yehoshua, K. B. Schaal, W. Feuer, G. Gregori, S. Kubach, L. An, P. F. Stetson, M. K. Durbin, P. J. Rosenfeld, and R. K. Wang, “Automated quantitation of choroidal neovascularization: a comparison study between spectral-domain and swept-source OCT angiograms,” Invest. Ophthalmol. Vis. Sci. 58(3), 1506–1513 (2017). [CrossRef]
45. J. Xu, S. Song, S. Men, and R. K. Wang, “Long ranging swept-source optical coherence tomography-based angiography outperforms its spectral-domain counterpart in imaging human skin microcirculations,” J. Biomed. Opt 22(11), 1 (2017). [CrossRef]
