2.1.

Data Collection to be Studied

The study conducted in this work was approved by the Institutional Review Board in our institutions. The research adhered to the tenets set forth in the declaration of Helsinki. All Stratus OCT study cases were obtained using the radial lines protocol (1024 $\text{samples}\times 512$ A-scans per B-scan) on a single Stratus OCT instrument (version 4.0 software). All Stratus OCT scans were taken at the macula. In addition, all subjects scanned with the Stratus OCT unit underwent visual acuity testing with refraction and a complete slit-lamp examination. Informed consent was obtained from each subject.

Macular radial line scans of the retina for each case were exported to disk with the export feature available in the Stratus OCT device. Total retinal thickness measurements were obtained as provided by the Stratus OCT built-in algorithms as well as with the OCTRIMA methodology. To demonstrate intragrader and intergarder reproducibility of OCTRIMA and the comparability of STRATUS and OCTRIMA-derived thickness results, we used the same set of ten healthy eyes (total of 60 OCT B-scans) from five normal subjects as reported previously, ranging in age from $25\text{to}34\text{years}$ (mean age $29\text{years}$ ).²² As in some Fourier-domain OCT (FD-OCT) systems, OCTRIMA facilitates the total retinal thickness calculations between the inner limiting membrane (ILM) and the inner boundary of the second hyperreflective band, which has been attributed to the outer segment/retinal pigment epithelium (OS/RPE) junction in agreement with histological studies.^{14, 23, 24} (see Fig. 1 ). Moreover, as in the Stratus OCT system, OCTRIMA’s total retinal thickness measurements can be also obtained by calculating the thickness between the ILM and the inner boundary of the innermost hyperreflective band corresponding to the inner segment-outer segment (IS/OS) junction. Accordingly, Stratus OCT retinal thickness measurements were compared with the OCTRIMA measurements in each of the nine macular regions defined by the Early Treatment Diabetic Retinopathy Study (ETDRS) using the two previously defined calculations for the total retinal thickness.²⁵ Differences in macular volume measurements were also calculated. Moreover, the agreement of the thickness measured by OCTRIMA and Stratus OCT and those recently reported in the literature as a result of using FD-OCT systems were also evaluated.

Fig. 1

OCTRIMA screenshot showing the segmentation results for an OCT B-scan obtained from a healthy normal eye. The layers have been labeled as: ILM (inner limiting membrane), RNFL (retinal nerve fiber layer), $\mathrm{GCL}+\mathrm{IPL}$ complex (ganglion cell layer and inner plexiform layer), INL (inner nuclear layer), OPL (outer plexiform layer), ONL (outer nuclear layer), OS (outer segment of photoreceptors), and retinal pigment epithelial layer (RPE). We note that the sublayer labeled as ONL is actually enclosing the external limiting membrane (ELM) and IS, but in the standard $10\text{-}\mu \mathrm{m}$ resolution OCT image this thin membrane cannot be visualized clearly, making the segmentation of the IS difficult. Thus this layer classification is our assumption and does not reflect the actual anatomic structure. Also, observe that since there is no significant luminance transition between GCL and IPL, the outer boundary of the GCL layer is difficult to visualize in the Stratus OCT image shown. Thus, a combined $\mathrm{GCL}+\mathrm{IPL}$ layer is preferable.

Complementary cases with clinically significant intraretinal features from Stratus OCT were also collected only for demonstration purposes of the image segmentation performance and error correction using OCTRIMA. Particularly, these cases included a patient with mild nonproliferative diabetic retinopathy without macular edema and a patient with neovascular age-related macular degeneration. In both cases, the macular radial lines protocol was used. Moreover, to establish the feasibility of OCTRIMA for analyzing images from advanced OCT imaging systems, two image sets from two different FD-OCT systems were also analyzed with OCTRIMA. One set of images was obtained from the Bioptigen Spectral Domain Ophthalmic Imaging System (Bioptigen Incorporated, Research Triangle Park, North Carolina), while the other set was scanned by a custom-developed ultrahigh FD-OCT adapted from our anterior segment OCT system with $\sim 3\text{-}\mu \mathrm{m}$ axial resolution. The system configurations for these two FD-OCT systems have been detailed elsewhere.^{26, 27}

2.2.

Computer-Aided Optical Coherence Tomography Grading Method

The method used for OCT image analysis (OCTRIMA) was developed using the Matlab graphical user interface design environment tool. The OCTRIMA methodology essentially provides dual functionality by combining image enhancement and denoising of Stratus OCT images along with automatic segmentation of the various cellular layers of the retina. More details of the segmentation and preprocessing process can be found in Cabrera Fernández, Salinas, and Puliafito,⁶ and Salinas and Cabrera Fernández.²⁸ Moreover, OCTRIMA has the capability to perform calculations based on measured values of corrected thickness, and reflectance of the various cellular layers of the retina and the whole macula.

OCTRIMA also facilitates semiautomatic correction of discontinuities in each detected boundary after automated segmentation, along with manual error correction using direct visual evaluation of the detected boundaries. Particularly, the methodology allows the operator to utilize simple computer mouse clicks to fix various boundaries in each of the six radial line OCT B-scans by using predefined corrective functions (see Fig. 2 ). These errors are mainly due to both the presence of high reflectivity regions in the inner retina, and loss of retinal structure information in local regions along the retinal cross section, as visualized by the commercial OCT system. Since normative data for OCT analysis are crucial to compare various treatment strategies, OCTRIMA facilitates normative data from healthy controls and also allows the user to generate a new norm using healthy or pathological subjects. The OCTRIMA’s norm is based on data from 74 healthy subjects $(35\pm 13\text{years})$ as described previously.²² In addition, OCTRIMA provides a standardized method for reporting changes in thickness as a percentage of total possible change based on normative OCT data.²⁹

Fig. 2

Flowchart showing the classification of segmentation errors and their corresponding manual corrective functions.

2.3.

Statistical Methods

The coefficients of reproducibility were calculated using the methods outlined by Bland and Altman for each of the averaged thickness measurements obtained for the total retina and intraretinal layers.^{30, 31} The coefficients of reproducibility were computed from the standard deviations (SDs) of the differences between measurements made by each grader. The statistical analysis was performed using the software package SPSS version 16 (SPSS Incorporated, Chicago, Illinois).

3.1.

Comparison of Stratus OCT Thickness Measurements and Optical Coherence Tomography Retinal Image Analysis Thickness Measurements

In contrast to Stratus OCT thickness calculations, OCTRIMA measurements of total retinal thickness were obtained using two different assumptions for the outer retinal boundary: 1. the inner boundary of the innermost hyperreflective band corresponding to the IS/OS junction (as defined in Stratus OCT), and 2. the inner border of the second hyperreflective band, assumed to be the OS/RPE junction. We note that the first assumption allows a fair comparison between the automated results of Stratus OCT and OCTRIMA algorithms for the convention used by the Stratus OCT algorithms. Table 1 shows the level of agreement between OCTRIMA and Stratus OCT measurements when the two different outer retinal border assumptions are applied on the same Stratus OCT images. Pearson correlation coefficients demonstrated ${\mathrm{R}}^2>0.98$ for all ETDRS regions (data not shown).

Table 1

Comparison between Stratus OCT retinal thickness measurements (mean values) and the OCTRIMA measurements (mean values) obtained using: 1. the inner border of the IS/OS junction, and 2. the inner border of the OS/RPE junction as the outer retinal boundary. Differences in the measurement of the total macular volume are also included and are expressed in cubic millimeters (mm3) .

When OCTRIMA calculations used the convention adopted by Stratus OCT (i.e., inner border of the innermost hyperreflective band as the outer retinal border), the mean thickness difference was $6.53\mu \mathrm{m}$ , which corresponded to 3% of the measured Stratus OCT retinal thickness. Measurements in the foveal central region (R1), inferior inner region (R4), and, nasal, inferior, and temporal outer (R7, R8, and R9) regions showed the greatest disagreement between the OCTRIMA and Stratus algorithms. Particularly, the mean difference included only 4% (except in R8) of the measured value obtained by the Stratus OCT algorithms. As expected, the foveal center point (FCP) measurements demonstrated best agreement, because no interpolation is required for these calculations (data not shown). Moreover, total macular volume, a measure derived from thickness in all data points of the macula, was 3% higher by OCTRIMA compared to Stratus OCT results, also supporting an average difference of 3% in thickness measurements. In contrast, when OCTRIMA calculations used the inner border of the second hyperreflective band (i.e., OS/RPE junction) as the outer retinal border, the mean thickness difference corresponded to 11% of the measured Stratus OCT retinal thickness. Similarly, the mean difference for the foveal region (R1) included only 17% of the measured value obtained by the Stratus OCT algorithms. Moreover, the mean difference results for the superior, nasal, inferior, and temporal inner and outer regions of the macula (R2 through R9) included 8 to 12% of the Stratus OCT measurements (see Table 1). Correspondingly, total macular volume was 10% higher by OCTRIMA compared to Stratus OCT results, also supporting an average difference of 10% in thickness measurements.

3.2.

Intragrader and Intergrader Reproducibility of Thickness Measurements

As a result of scanning a total of ten healthy eyes, a total of 60 OCT B-scans were collected and analyzed by two independent experienced graders (G1 and G2). Moreover, to assess the overall performance of the OCTRIMA software, the average (between the two graders) retinal thickness in each of the nine ETDRS regions obtained by OCTRIMA analysis was compared with the automated Stratus OCT results. All scans in the study had a signal strength of 9 or 10. Algorithm performance was visually evaluated by the experienced graders to detect segmentation errors. Criteria for algorithm error included evident disruption of the detected boundary (e.g., small peaks, linear and curve offsets), and/or detected boundary jumping to and from different anatomical structures (i.e., false segmentation). The average number of manual corrections needed per scan was three. The inner nuclear layer (INL) and outer plexiform layer (OPL) were the layers that required most of the manual corrections.

Table 2 shows the reproducibility attained by one grader (G2) after analyzing each of the ten eyes at two separate times (intragrader test, one week interval between analyses) using OCTRIMA software. Thickness measurements $(\text{mean}\pm \mathrm{SD})$ of the total retina and intraretinal layers are also shown in Table 2. The coefficient of reproducibility (CR) obtained for the thickness measurements was less than 0.2% for the total retina, less than 0.4% for the ONL, and less than 3% for the remaining layers. Overall, the median of the thickness differences as a percentage of the mean thickness was less than 1%. According to our results, the measurement accuracy of the OCTRIMA algorithm ranged between $0.27\text{to}1.47\mu \mathrm{m}$ (see Table 2). Excellent intragrader agreement could be observed in the Bland-Altman plots of the mean difference between both grading sessions for each of the calculated intraretinal layer thicknesses (see Fig. 3 ).

Fig. 3

Bland-Altman plots of the the mean difference between both gradings for each intraretinal layer thickness calculated by the same grader in two separate occasions using OCTRIMA. Mean layer thickness for each subject is plotted against the difference in layer thickness between the two grading procedures. Solid lines show the mean of the differences between the two gradings, while dashed lines represent the CR values.

Table 2

Intragrader reproducibility using OCTRIMA. Mean* thickness value averaged across all ten eyes. We note that the mean thickness value is a result of averaging the uninterpolated thickness measurements at every A-scan location for all six B-scans of all ten eyes. This** value is calculated by subtracting OCTRIMA thickness measurement obtained during the first grading from the OCTRIMA thickness measurement obtained during the second grading for each eye by the same grader G2, and then taking the absolute value. Median† of the differences between measurements expressed as a percentage of mean thickness across all ten eyes.

Table 3 shows the level of agreement between the two graders (intergrader reproducibility test, i.e., G1 versus G2) using the OCTRIMA software. The coefficient of reproducibility obtained for the thickness measurements was less than 0.5% for the total retina, less than 0.7% for the ONL, and less than 5% for the remaining layers. According to our results, the measurement accuracy of our algorithm ranged between $0.6\text{to}1.76\mu \mathrm{m}$ (see Table 3). Overall, the median of the thickness differences as a percentage of the mean thickness was less than 2%.

Table 3

Intergrader reproducibility using OCTRIMA. Mean* thickness value averaged across all ten eyes. We note that the mean thickness value is a result of averaging the uninterpolated thickness measurements at every A-scan location for all six B-scans of all ten eyes. This** value is calculated by subtracting OCTRIMA thickness measurement obtained during the first grading from the OCTRIMA thickness measurement obtained during the second grading for each eye by the same grader G2, and then taking the absolute value. Median† of the differences between measurements expressed as a percentage of mean thickness across all ten eyes.

3.3.

Image Segmentation Performance and Error Correction Using Optical Coherence Tomography Retinal Image Analysis

Illustrative cases of diseases with subretinal anomalies and representative intraretinal boundary detection errors are shown in Figs. 4, 5, 6 . Compared to a free-hand correction where curves are manually drawn, OCTRIMA’s manual corrective functions facilitate the correction of segmentation errors using less reviewing time. The advantage is based on the fact that OCTRIMA manual correction is performed after the semiautomatic correction of errors, which reduces significantly the time required to redraw the curves.

Fig. 4

Segmented B-scan before and after applying manual corrections. (a) Visible segmentation errors such as small peaks, linear offsets, curve offsets, and false segmentations. Manual error removal functions are available for: (b) False segmentation: note that the RNFL is not visible on the left side of the B-scan. The outer boundary of the RNFL is not properly detected. Note that the inner and outer boundaries (outlined in magenta and red, respectively) of the RNFL overlap after manually correcting the segmentation error. (c) Small peaks: note that the small peak or overshoot in the IPL outer boundary (outlined in yellow) has been manually corrected using the “Small Peak” correction function. (d) Foveal center control: the outer boundaries of the IPL, INL, and the OPL (outlined in yellow, cyan, and green, respectively) are made to appear coincident in the central foveal region, according to the control defined in the automated segmentation algorithm. (e) Linear offsets: note that the linear offsets in the outer boundaries of the OPL (outlined in green) are corrected manually by drawing a straight line segment on the specific boundary containing the offset. (f) Curved offsets: note that the curve offsets in the inner and outer boundaries of the INL (outlined in yellow and cyan, respectively) are corrected manually using the “curve plotting” function. (Color online only.)

Fig. 5

Stratus OCT automated segmentation results for a patient with neovascular AMD showing common detection errors. The white circles indicate the areas where the Stratus OCT algorithm fails to properly detect the outer boundary. The red asterisk indicates the areas with fluid accumulation that were included in the final thickness calculations provided by Stratus OCT. (Color online only.)

Fig. 6

OCTRIMA segmentation results for a patient with neovascular AMD (nonaligned raw data for the same patient shown in Fig. 5) showing the fluid-filled regions outlined. The ONL outer border was used as the outer retinal border to compare results with Stratus OCT algorithms (see the table at the bottom). The retinal thickness map along with the ETDRS topographic map including the retinal structure and fluid-filled regions are shown in the right-bottom section.

3.3.1.

Mild nonproliferative diabetic retinopathy without macular edema

Figure 4 shows an OCTRIMA segmented B-scan before and after applying manual corrections. In this study case, the representative B-scan was taken from a set of images obtained for a diabetic patient with mild nonproliferative diabetic retinopathy without macular edema (male, $59\text{years}$ old). Characteristic intraretinal boundary detection errors such as small peaks, linear offsets, curve offsets, and false segmentations are illustrated in Fig. 4. False segmentation refers to the falsely detected inner and/or outer boundaries of an intraretinal layer. This particular error is most commonly found during the RNFL’s outer border detection. Specifically, there are certain cases in which the true anatomical thickness of the RNFL layer (or some regions of the RNFL layer) might be negligible. In other cases, one side of the RNFL layer is completely invisible in the OCT image, like for example in the temporal part of a horizontal B-scan [see Fig. 4]. In such cases, a correction is required to overlap the inner and outer boundaries of the RNFL layer in the regions of negligible thickness [see Figs. 4 and 4]. However, sometimes the boundary detection algorithm fails in such specific cases when localized bright spots of high intensity appear on some regions of the RNFL layer, and falsely displays the outer boundary of the RNFL layer as a result of the peak search algorithm, which looks for zero crossings in the structure. Hence, the RNFL outer boundary must be manually corrected to appear overlapped on the inner boundary in the invisible part of the layer. As can be seen in Fig. 4, the ILM boundary on the inner side of the RNFL is detected but no boundary is detected on the outer left side, since the RNFL is not visible on this (temporal) side for this particular scan, whereas the RNFL is bright and clearly visible on the right (nasal) side of the scan [see Figs. 4 and 4]. Figure 4 shows the manually corrected outer boundary of the IPL (outlined in yellow) using the “small peak” corrective function of the manual correction software tool, which removes the overshoots or undershoots in the individual boundaries.

These are also parts of a boundary that form a straight line segment, but are incorrectly detected as a peak or an elevated or depressed line segment by the automated segmentation algorithm. This detection error is classified as a linear offset. To resolve this class of error, the user has to manually select two points to draw a straight line segment on the specific boundary containing the offset. For example, a straight line segment was manually drawn to correct the linear offset in the outer boundary of the OPL [see the boundary outlined in green in Fig. 4]. Figure 4 shows the manual corrections for the inner and outer boundaries of the INL (outlined in yellow and cyan, respectively) that had segmentation errors as a result of curve offsets [see Fig. 4]. Curve offset is a term given to the curved portion of a boundary that has not been recognized as a curve, but instead has been incorrectly segmented as an elevated or depressed curve. Curve offsets have been rectified using a function based on a customized contour model originally introduced to identify nonconvex shapes in OCT images.³ This function allows the user to select multiple closely spaced points that will be joined to trace a curve and remove the offset. Figure 4 shows the result of the “curve plotting” function applied to correct the inner and outer boundaries of the INL (outlined in yellow and cyan, respectively).

Additionally, an OCTRIMA predefined control is also in place for the inner retinal layers [RNFL, ganglion cell, and inner plexiform layer complex $(\mathrm{GCL}+\mathrm{IPL})$ and INL] and OPL in a $1.5\text{-}\mathrm{mm}$ -diam zone in the fovea, where retinal reflections are minimally visible.⁶ The control forces the ILM, the inner and outer side of the $\mathrm{GCL}+\mathrm{IPL}$ complex, and the outer side of the INL and OPL to be coincident in this region [see Fig. 4]. Sometimes, small peaks appear at the periphery of this controlled foveal region. In such a case, it appears that the coincident layers deviate from the true foveal visible boundary and need to be corrected so that they overlap in the periphery of the controlled foveal region. Thus, the “overlap” function of the manual correction software tool is useful to rectify the segmentation at the fovea [see Fig. 4].

Figure 7 shows OCTRIMA thickness maps for the overall macula and each intraretinal layer obtained from the patient image data shown in Fig. 4. The ETDRS-like regions are based on nine sectional thickness values in three concentric circles obtained from interpolation of the six linear scans, with diameters of 1, 3, and $6\mathrm{mm}$ . These maps are obtained according to the standards set by the ETDRS, similarly to the Stratus OCT analysis software, and can be easily exported to a PDF document along with the numerical results in tabulated format. The output data includes three main quantitative measures: thickness, volume, and reflectance. The sectional measurements in the retinal thickness map are calculated from the averaged data from the six individual scans. The norm used in OCTRIMA was obtained from 74 healthy eyes $(35\pm 13\text{years})$ , as described previously.²²

Fig. 7

OCTRIMA thickness maps for a diabetic patient with mild nonproliferative diabetic retinopathy without macular edema (male, $59\text{years}$ old). Note that thickness maps for the total retina and the intraretinal layers are shown. We also note that an OCTRIMA macular map is divided into nine zones that correspond to the ETDRS regions: fovea within a diameter of $1\mathrm{mm}$ centered on the foveola; pericentral ring, the circular band from the central $1\text{to}3\mathrm{mm}$ , divided into four quadrants, i.e., superior, inferior, temporal, and nasal; and peripheral ring from $3\mathrm{mm}$ up to $6\mathrm{mm}$ , divided into the same quadrants.

3.3.3.

Advanced optical coherence tomography imaging systems

Our methodology is also suitable for FD-OCT images that have better resolution, and as such, ensure robust segmentation. To prove this, we applied OCTRIMA segmentation to raw images taken from FD-OCT devices using the same parameters optimized for Stratus OCT images. Particularly, Fig. 8 shows the OCTRIMA segmentation results for an image obtained with the Bioptigen Spectral Domain Ophthalmic Imaging System (Bioptigen Incorporated, Research Triangle Park, North Carolina). Figure 8 shows OCTRIMA segmentation results for an OCT image from a healthy eye obtained with a custom-developed FD-OCT adapted from our anterior segment OCT system with $\sim 3\text{-}\mu \mathrm{m}$ axial resolution.²⁶ We note that significant improvement has been obtained for the delineation of the INL, OPL, and OS in the advanced OCT images analyzed.

Fig. 8

Feasibility of OCTRIMA for analyzing images from advanced OCT imaging systems. (a) OCTRIMA segmentation results for an OCT image obtained with the Bioptigen Spectral Domain ophthalmic imaging system. (b) OCTRIMA segmentation results for an OCT image obtained with the custom developed FD-OCT adapted from our anterior segment OCT system.