SURGICAL GUIDANCE/ULTRA HIGH-SPEED PROCESSING: The next step in real-time OCT

KANG ZHANG and JIN U. KANG

Compared to other 3-D imaging modalities such as magnetic resonance imaging (MRI), computed tomography (CT) and ultrasound, Fourier-domain optical coherence tomography (FD-OCT) is more compact and produces imagery with significantly higher resolution. These attributes make FD-OCT suitable for applications such as medical diagnostics and provide promise for surgical guidance (see Fig. 1).

In recent years, the image acquisition rate of FD-OCT systems has reached multi-hundreds-of-thousands to multi-millions of A-scan images per second. Such ultra-high acquisition speed has enabled time-resolved volumetric (4D) recording and reconstruction of dynamic processes such as eye blinking, papillary reaction to light stimulus and embryonic heart beating.

Unfortunately, though, OCT data processing and image rendering/display technology have not kept up with the OCT data acquisition technology. This means that despite their ability to acquire imagery in real time, OCT systems must store data for later processing. Thus, while the approach can provide valuable pre-, intra-, and post-operative images for evaluation at a later time, it is incapable of providing real-time, intraoperative imaging for surgical guidance and visualization, and on-site diagnostic evaluations. This has kept FD-OCT systems from reaching their full potential.

Besides the speed issue, standard FD-OCT systems have another limitation: They suffer from spatially reversed complex-conjugate ghost images that could severely misguide users. As a solution, full-range complex FD-OCT (C-FD-OCT) removes the complex-conjugate image by applying a phase modulation on interferogram frames. However, C-FD-OCT processing is even more time-consuming and presents an extra burden in providing real-time images during surgical procedures.

Solving for speed

FD-OCT signal processing usually includes two stages: spectral signal processing and volumetric visualization. Because of the huge data volumes involved, both are demanding computing tasks and constitute two bottlenecks. To solve this issue, the group has been working on modeling and implementing computationally intensive FD-OCT data processing and visualization using general-purpose graphics processing units (GPGPU).^{1, 2} Compared to CPU-based signal processing in current ultra-high speed FD-OCT systems, the method is low-cost and requires no optical modification to the system, and processes and displays 3-D images (both standard FD-OCT and C-FD-OCT) in real time.

The system consists of a high-speed, line-scan camera-based OCT spectrometer, a broadband light source, and a reference arm and an imaging arm with scanner and lenses (see Fig. 2). For the C-FD-OCT mode, phase modulation is applied to each B-scan's 2D interferogram frame by slightly displacing the probe beam off the first galvanometer's pivoting point. A workstation with a frame grabber (PCIE-x4 interface), and GPU (PCIE-x16 interface) on the same mother board processes the OCT signal. The GPU is programmed through NVIDIA's Compute Unified Device Architecture (CUDA) technology.

FIGURE 2. The GPU accelerated FD-OCT system includes a high-speed, line-scan camera-based OCT spectrometer, a broadband light source, a reference arm and an imaging arm with scanner and lenses. In C-FD-OCT mode, phase modulation is applied to each B-scan's 2D interferogram frame by slightly displacing the probe beam off the first galvanometer's pivoting point. A workstation with a frame grabber (PCIE-x4 interface), and GPU (PCIE-x16 interface) on the same motherboard processes the OCT signal.

A flow chart of the system's signal processing shows three major threads: one for raw data acquisition (Thread 1), another for GPU accelerated FD-OCT data processing (Thread 2), and the last for display (Thread 3). The threads synchronize in the pipeline mode; solid arrows specify the main data stream while hollow arrows indicate the internal data flow of the GPU (see Fig. 3).

FIGURE 3. The system processes signals using three major threads, which synchronize in the pipeline mode: Thread 1 is for one for raw data acquisition, Thread 2 is for GPU accelerated FD-OCT data processing and Thread 3 is for display. Solid arrows specify the main data stream, while hollow arrows indicate the internal data flow of the GPU.

In comparison

In most FD-OCT systems, the raw spectral data is sampled linearly in wavelength space and nonlinearly in k-space. When fast Fourier transform (FFT) is applied directly to such data to speed processing, image quality is seriously degraded. Both hardware and software solutions have been applied to address this issue: Hardware solutions such as linear-k FD-OCT have been successfully implemented, but these methods generally increase system complexity and cost. Software solutions include various interpolation methods such as simple linear interpolation, oversampled linear interpolation, zero-filling linear interpolation, and cubic spline interpolation. Alternatively, the non-uniform fast Fourier transform (NUFFT) has been applied to FD-OCT signal. Compared with the interpolation-FFT method, NUFFT is immune to the interpolation-caused errors such as increased background noise and side-lobes in the system point spread function. Moreover, NUFFT has improved sensitivity roll-off.

In a benchmark line rate test of different FD-OCT processing methods, all algorithms were tested on the GTX 480 GPU with 4,096 lines of both 1024-pixel spectrum and 2048-pixel spectrum (see Fig. 4). For each case, both the peak internal processing line rate and the reduced line rate, considering the data transfer bandwidth of PCIE x16 interface, are listed.

FIGURE 4. Results of a benchmark line rate test comparing processing methods: (a) 1024-pixel FD-OCT, and (b) 2048-pixel FD-OCT. LIFFT is standard FD-OCT with linear spline interpolation, LIFFT-C is C-FD-OCT with linear spline interpolation, CIFFT is standard FD-OCT with cubic spline interpolation, CIFFT-C is C-FD-OCT with cubic spline interpolation, NUFFT is standard FD-OCT with NUFFT, and NUFFT-C is C-FD-OCT with NUFFT.

FIGURE 5. Real-time C-FD-OCT images, generated using GPU-NUFFT and displayed at 29.8 fps with original frame size of 4096 pixel (lateral) × 1024 pixel (axial), show (a) fingertip (coronal); (b) finger palm (coronal); (c-d) fingernail fold (coronal); and (e-f) fingernail (sagittal).

Compared to standard FD-OCT, GPU-NUFFT C-FD-OCT in-vivo imagery of a human finger is free of conjugate artifact, DC noise and autocorrelation noise (see Fig. 5). These noises are difficult and time-consuming to remove in standard FD-OCT. Moreover, due to the implementation of the complex OCT processing, the image depth is effectively doubled, with the highest SNR region in the zero delay point.

Real-time C-FD-OCT could play an important role in surgical visualization, guidance, and intervention applications, and the achievement of this goal need not be cost-prohibitive.

REFERENCES
1. K. Zhang and J. U. Kang, Opt. Express 18, 11772–11784 (2010)
2. K. Zhang and J. U. Kang, Opt. Express 18, 23472–23487 (2010)

Kang Zhang is a Ph.D. candidate, and Jin U. Kang is a professor in the Photonics and Optoelectronics Laboratory at the Department of Electrical and Computer Engineering in the Johns Hopkins University, Barton Hall, Baltimore, MD, www.ece.jhu.edu/photonics/index.htm. Contact Kang Zhang at [email protected].