2.1.

Experimental System

The base configuration of the experimental Fd-OCT system used for acquiring 3D volumes has been described previously.⁷ For data presented in this paper, a different light source has been used: a superluminescent diode with an 855-nm central wavelength and an FWHM of $75\mathrm{nm}$ . The spectral bandwidth of this light source allows $4.5\text{-}\mathrm{\mu }\mathrm{m}$ axial resolution in the retina (refractive index, $n=1.38$ ). The high power of this light source supports use of a 90/10 fiber coupler, directing 10% of the light toward the eye, resulting in the power at the subject’s cornea equal to $700\mathrm{\mu }\mathrm{W}$ and allowing 90% throughput of the light back-reflected from the eye to the detector. With our current spectrometer design (200-mm focal length of the air-spaced doubled), the maximum axial range (seen after the Fourier transform) is about $3.6\mathrm{mm}$ in free space, corresponding to approximately $2.7\mathrm{mm}$ in the eye. Figure 1 shows a schematic of our clinical Fd-OCT system.

Fig. 1

Schematic of the clinical Fd-OCT system at UC Davis. SLD—superluminescent diode.

Each subject was imaged with several OCT scanning procedures including 3D scanning patterns centered at the fovea and ONH. We used two different scanning arrangements for 3D scanning patterns including (1) equally spaced 200 B-scans, with each based on 500 A-scans, and (2) equally spaced 100 B-scans based on 1000 A-scans. In both cases, volumes consisted of the same number of A-scans (100,000), and the time required to acquire a volume was $5.5\mathrm{s}$ for $50\text{-}\mathrm{\mu }\mathrm{s}$ CCD exposure time and $11.1\mathrm{s}$ for $100\text{-}\mathrm{\mu }\mathrm{s}$ exposure time. The longer exposure time was mainly used to increase image intensity and with 100 B-scans based on 1000 A-scans per frame acquisition mode, at the cost of more motion artifacts. Commercial Fd-OCT acquisition software from Bioptigen, Inc. permitted real-time display (at the acquisition speed) of the imaged retinal structures and saved the last acquired volume.

2.2.

Image Processing

After each imaging session, saved spectral data are processed in LabVIEW using standard Fd-OCT procedures,^{3, 7, 16} including spectral shaping, zero padding, and dispersion compensation. As a result, a set of high-quality OCT B-scans is saved in tiff format. It is possible to then analyze data and choose volumes with no eye motion or ones in which eye movement occurred only at the beginning or toward the end of the 3D acquisition, provided there was good fixation during the rest of the time (especially if the structure of clinical relevance was scanned without distortions). In these cases, the part of the image that was affected by eye motion is removed, resulting in a reduction in the number of B-scans and overall size of the reconstructed 3D volume. In addition, as a preprocessing step before volume visualization, OCT B-scans are registered using standard registration techniques. Figure 2 shows a single B-scan and also a cross section of several B-scans showing the alignment in different cross sections.

Fig. 2

Registration of the OCT frames. Top image shows one OCT B-scan. Middle image shows cross section through 98 unaligned B-scans. Bottom image shows the scans after registration with ImageJ.

To correct for axial and lateral distortions, we used ImageJ,¹⁷ with the Turboreg plug-in developed in the laboratory of P. Thévenaz at the Biomedical Imaging Group, Swiss Federal Institute of Technology Lausanne,¹⁸ which computes rigid body transformations (translation and rotation) to minimize the difference between neighboring frames, as can be seen in the lower panel of Fig. 2.

2.3.

Visualization and Analysis Software

To render the volumetric data, we used a technique similar to that described by Carbal ¹⁹ Every time the volume is rendered, it is sliced into evenly spaced view-aligned planes. Each slice is rendered through a pixel shader using the voxel’s data value, a color map, and alpha blending to create a volume rendering that is equivalent to a ray tracer that uses evenly spaced samples. This is done efficiently by implementing a highly optimized cube slicing algorithm and applying it to the bounding box of the volume. We also use adaptive slice spacing to balance performance/functionality and visual quality.

There has been substantial research on classification and segmentation of volumetric data. Most of this work focuses on visualization instead of explicit segmentation. The most common approach used in volume visualization is based on a one-dimensional transfer function to map scalar intensity to color and opacity. This effectively allows a region to be classified based solely on scalar intensity, hiding regions occluding areas of interest while highlighting the remaining areas with a chosen color scheme. This approach does not work for noisy data if it results in a region that cannot be classified solely by its scalar intensity (since these values occur somewhat randomly throughout the volume). This approach is also not applicable to volumes that contain volumetric objects of similar intensity values that occlude each other. A great deal of research has attempted to solve these inadequacies. Levoy ²⁰ used gradient magnitude to highlight material boundaries in volumetric data to better express structure. Kindlmann and Durkin²¹ proposed the use of a two-dimensional transfer function, represented by a scatterplot, of scalar values versus gradient magnitudes. Others have described methods that help a user better utilize this 2D transfer function.^{20, 22, 23} Iserhardt-Bauer ²⁴ combined a 2D transfer with a region growing method that expands to fill regions of interest. Region growing methods are typically ideal for large meandering regions. Multidimensional transfer functions tend to operate better than lower-dimensional counterparts since the method can use more information to segment the data. The method of Bordignon ²⁵ demonstrates a multidimensional transfer function with a user interface based on star coordinates. This is a technique that maps multidimensional data to a 2D plane so that a user can interact with it. Pfister ²⁶ compared a variety of transfer function methods including some of those listed above. Despite these advances, 2D transfer functions remain inaccessible to the common user, and their effectiveness is still predicated on noiseless data. When it comes to discrete and explicit segmentation, medical imaging research has used artificial neural networks as a means to assist in these tasks.^{27, 28} However, support vector machines^{14, 15, 29} have demonstrated better results to date than neural networks (at a larger computational cost) when applied to pattern recognition including object identification,³⁰ text categorization,³¹ and face detection.³² Tzeng ³³ compared the use of neural networks and support vector machines when trying to construct an N-dimensional transfer function that uses additional variables such as variance and color (if present) in addition to scalar values and gradient information. The ability of neural networks and SVMs to handle error (in the form of noise) makes them useful for noisy volumetric data sets. However, these methods still seek to classify a volumetric object solely based on discrete values, not on distributions. Our approach converts noise into a classifiable characteristic by using a specialized SVM that operates on distributions rather than scalar values.

3.1.

Visualization and Analysis of Volumetric Data

To illustrate the performance of our clinical Fd-OCT and our 3D data visualization and analysis software, we present four examples of visualization and analysis of the clinical data.

The first example is data acquired from a 55-year-old patient with non-exudative age-related macular degeneration (AMD). Figure 3 shows the patient’s fundus photo used for regular clinical diagnosis. The top right image shows a reconstructed en face image (virtual C-scan) from the OCT volume (100 B-scans/1000 A-scans/B-scan acquired with $100\text{-}\mathrm{\mu }\mathrm{s}$ /line exposure) superimposed on the fundus picture. This step allows precise registration of the acquired volume and estimation of the distortions caused by the subject’s eye motions during the experiment.

Fig. 3

Top left: fundus photo from a patient with non-exudative AMD; top right: same fundus photo with a superimposed C-scan reconstructed from the center depth on OCT volume. Arrows indicate the relative positions of the B-scans shown below.

To demonstrate the performance of our Fd-OCT system, four B-scans (labeled A–D) are shown in the lower part of Fig. 3. The location of the diseased structures can be easily seen in these images. As described above, this data set is used to create the interactive volumetric visualization using our software. Figure 4 shows a screen shot of our volume renderer and segmentation system. Note the points used to define the SVM segmentation.

Fig. 4

User interface of volume visualization software with SVM training data (green for structure of interest and red for background) shown on the B-scan. The result of the frame segmentation based on these inputs can be seen on the B-scan. The result of the volume segmentation is marked green on the 2D cross sections; extracted volume is shown in the upper left window panel (color online only).

Figure 5 shows an example of different 3D visualizations of this data set. We used a false color intensity based on a black-red-yellow color lookup table with black representing low-intensity voxels and yellow representing high-intensity voxels. Transparency of each voxel is based on its intensity and can be interactively set by the operator.

Fig. 5

3D visualization of the volumetric data shown in Fig. 4. (a) whole volume; (b) left part of the volume removed by $x\text{-}z$ clipping plane; (c) left part of the volume removed by $y\text{-}z$ clipping plane; (d) visualization of segmented part of the volume RPE and photoreceptor-outer segment complex.

To test the performance of our SVM segmentation algorithm, we used it to segment the retinal pigmented epithelium (RPE) and photoreceptor layers, structures that show the main changes associated with disease progression in AMD. The bumpy surface is due to large drusen, a defining feature of early stage AMD. A thickness map of these segmented structures is created from the SVM segmentation. Figure 6 shows the corresponding thickness maps from the SVM segmentation and from manual segmentation. As can be seen from the images in Fig. 6, the SVM method leads to clear separation of these layers, allowing its visualization as well as estimation of the thickness of the area of interest. The differences between the SVM and manual segmentations suggest that some refinement of the SVM may be necessary to better segment high density structures.

Fig. 6

Left: false color representation of the thickness map extracted from SVM-based segmentation of the RPE and photoreceptor-outer segment complex. Right: false color representation of the thickness map created from manual-based segmentation of the same structure (color online only).

Figure 7 shows another SVM-based segmentation of the RPE-photoreceptor-outer segment complex from the retinal foveal region of a 56-year-old patient with early-stage dry AMD. In contrast to the first patient presented, the RPE-photoreceptor-complex disruptions are subtle. Figure 7 shows results generated with our custom visualization software (including $z$ -axis intensity projection).

Fig. 7

Evaluation of the volumetric data set acquired over foveal region of a volunteer with early stage dry AMD. Upper left: 3D rendering of the data; upper right: screen shot from the user interface of the SVM-based segmentation software with segmented structure (RPE $+$ photoreceptors outer segments) highlighted in green; middle left projected on $z$ -axis intensity of the original volume; bottom left projected on $z$ -axis intensity of the SVM segmented structure (RPE $+$ photoreceptors outer segments); bottom right: false color thickness map of the SVM segmented layer (color online only).

Our algorithm was able to segment these structures. However, since the size of these disrupted regions was small and there was intensity variation across the volume, mainly at the corners of the image, not all drusen were identified in the thickness map. When the size of disruption is bigger, as in the previous example, the SVM segmentation algorithm has been able to pick out the features of interest very accurately.

In the next two examples, we tested the SVM segmentation algorithm in segmenting the RNFL around the ONH. Thinning of RNFL is known to be a good indicator of possible onset of glaucoma; therefore, being able to accurately segment and measure its thickness and follow it over time would be a good diagnosis and monitoring tool in glaucoma management. Figure 8 shows the segmentation results of the ONH from a healthy 30-year-old volunteer with our SVM segmentation software.

Fig. 8

Evaluation of the volumetric data set acquired over ONH region of the healthy volunteer. Upper left: 3D rendering of the data; upper right: screen shot from the user interface of the SVM-based segmentation software with segmented structure (RNFL) highlighted in green; middle left projected on $z$ -axis intensity of the original volume; bottom left projected on $z$ -axis intensity of the SVM-segmented structure (RNFL); bottom right: false color thickness map of the RNFL extracted from the SVM segmentation (color online only).

We also segmented the same structure in a 48-year-old glaucoma patient with a moderate amount of visual field loss. The results are shown in Fig. 9 .

Fig. 9

Evaluation of a volumetric data set acquired over the ONH region of a volunteer with moderate glaucoma. Upper left: a 3D rendering of the data; upper right shows a screen shot from the user interface of the SVM-based segmentation software with segmented structure (RNFL) highlighted in green; middle left projected on $z$ -axis intensity of the original volume; bottom left projected on $z$ -axis intensity of the SVM-segmented structure (RNFL); bottom right; false color thickness map of the RNFL extracted from the SVM segmentation (color online only).

As can be seen from these last two examples, SVM-based segmentation was able to successfully differentiate RNFL from other retinal structures, allowing visualization of the RNFL thickness map that shows thinning of this layer for the subject with glaucoma. It confirms good performance of the SVM-based segmentation for thick and well-defined structures.