Field-programmable gate arrays and other computing elements can improve imaging techniques such as optical coherence tomography.
Medical research can only advance as quickly as the technology that supports it. Medical imaging in particular plays a huge role in the entire clinical process—from diagnostics and treatment to surgery and research. In addition to the hurdles medical professionals encounter, “seeing” (in the medical sense) is one of the biggest challenges. Diseases are difficult to spot because they tend to hide deep inside the body. Techniques that enable clinicians to noninvasively see these areas are critical to ensuring progress in this field.
Optical coherence tomography (OCT) is a promising diagnostic tool that could have applications in many different medical fields. The technology takes advantage of the latest computing hardware architectures and is used to create medical instruments that can detect cancer and other conditions in a safe, simple, and effective manner.
This noninvasive imaging technique provides subsurface, cross-sectional images of materials. Interest in OCT technology has grown significantly because it provides much greater resolution than other imaging techniques such as magnetic resonance imaging (MRI) or positron emission tomography (PET). In addition, OCT requires relatively little preparation by medical staff and is safe for patients because it uses low laser outputs without the need for ionizing radiation.
To create images, OCT uses a low-power light source and the corresponding light reflections. It measures light in a way that is similar to how ultrasound machines measure sound. When the light beam is projected into a sample, much of the light is scattered. A small amount reflects as a collimated beam, which can be detected and used to create a very detailed image.
Field-programmable gate arrays (FPGAs) enable design flexibility, helping designers explore new ideas and reduce risk in the system development process. This capability is important in the medical space because it is critical to get to market quickly. Traditionally, demonstrating hardware-based processing required a custom application-specific integrated circuit (ASIC), but ASIC development is expensive and functionality is fixed.
|The world’s first 3-D OCT system, developed at Kitasato University in Japan, uses 20 FPGA-based FlexRIO modules that work in parallel to process scan data and produce 12 OCT volumes per second.|
FPGAs are reconfigurable through software. This advantage enables a designer to save development time by demonstrating hardware-based processing while preserving the option of reprogramming the FPGA to accommodate modifications that are required after initial specification. Although FPGA board designs can be complex and modular, off-the-shelf FPGA boards provide hardware to build around with infrastructure components for I/O connectivity, bus interfacing, and DRAM communication. Developing these components in-house can be time consuming and distracting.
FPGAs have rapidly grown in popularity for medical applications. With regard to medical imaging, FPGAs are primarily used in detection and image construction. The detection application involves embedded systems, with real-time performance requirements and significant hardware interface challenges. Image reconstruction, on the other hand, is similar to a high-performance computing problem.
The use of GPUs has also ramped up significantly for scientific research. One popular parallel computing architecture, called compute unified device architecture (CUDA), is used to accelerate a simulation program called Amber. The molecular dynamics simulation program is used by more than 60,000 researchers in academia and pharmaceutical companies worldwide to accelerate new drug discovery.
Neuroimaging is another one of the many medical fields directly benefitting from the computational power that GPUs provide. Using state-of-the-art medical imaging acquisition devices with advanced brain imaging techniques requires the ability to process extensive analysis and simulations with high-resolution images. By using powerful hardware devices such as GPUs, researchers can reduce the time for these large simulations, enabling faster deductions and even larger simulations to be performed.
To understand why GPUs are being adopted for these types of data-intensive applications, it’s important to consider the history of GPU hardware and the architectural design of modern GPUs. As the name implies, GPU hardware was originally designed to provide enhanced PC graphics capabilities. Early graphics hardware featured fixed-functionality logic for vertex processing and pixel operations.
However, the need for custom operations to be performed led to the inclusion of separate programmable vertex processors and pixel-fragment processors. Ultimately, the vertex and pixel-fragment processors were merged into unified processing units due to varying graphics workloads and to enable simplified, more customizable hardware designs. Today, many GPUs can be programmed to perform general-purpose computations that may not be related to graphics at all. Applications from medical imaging to complex biological simulations are benefiting from the capabilities of GPU hardware.
|A real-time rendered 3-D images of finger skin shows the fingerprint.|
Modern GPUs feature a large number of identical processor cores (several hundred is not uncommon) and are extremely efficient at solving data-parallel operations including but not limited to graphics processing. Rather than being optimized for low latency, as is the case with CPUs, GPUs are designed for high data throughput and can provide a several hundred–fold speedup when applied to the right application. GPUs enable engineers, scientists, and researchers to perform work that previously would have required an entire cluster of PCs. Among the many use cases for GPUs, medical imaging applications are appropriate because they rely on parallel data and are related to graphics processing.
Companies and researchers are currently taking advantage of these elements for novel medical imaging applications. For example, Kohji Ohbayashi of Kitasato University, supported by the Japan Science and Technology Agency, leveraged the flexibility and scalability of a PCI-based platform (called PXI) and a high-performance, off-the-shelf FPGA processing module (called FlexRIO) to develop a real-time 3-D OCT imaging system. Ohbayashi and his team used LabVIEW to program, integrate, and control the different parts of the system, combining high-channel-count acquisition with FPGA and GPU processing for real-time computation, rendering, and display.
Using design software to integrate and control the different parts of the system, they transferred data from the FPGA subsystem to a quad-core PC with a GPU to perform real-time 3-D rendering and display. The team also needed to log data for extended time periods because they wanted to do group screening tests for cancer. While the architecture didn’t limit the team’s image acquisition time, it enabled logging of up to 100 minutes on the prototype system, which required a little more than 3 terabytes of hard drive space.
The team also leveraged FPGA-based processing enabled by the FlexRIO. It computed more than 700,000 512-point fast fourier transforms per second to achieve 3-D imaging while maintaining high-channel density for the 320-channel system. The team tranferred data directly between FPGAs using PXI, high-throughput data transfers over PCI Express, accurate timing and synchronization of multiple modules, and peer-to-peer data streams without going to the host. The team also maintained a high-throughput connection between I/O and the GPU processing in the host PC, and integrated disk-array hardware for extended image data logging.
With this processing system, the team demonstrated continuous real-time display of 3-D OCT images. It could rotate the rendered 3-D image in any direction in real time. Observation of tissues such as the trachea or esophagus with good image penetration depth demonstrates the method’s applicability to optical biopsy. In addition, revealing the inside of a structure by virtually cutting the tissue surface in real time is useful for cancer diagnosis. The team observed dynamic tissue changes with the system, which surgeons could use to observe blood flow and tissue changes during surgery.
In another innovative medical imaging application, Santec Corp. demonstrated the benefits of using an FPGA on a swept source optical coherence tomography (SS-OCT) imaging system. Using FlexRIO, which combines customizable I/O with a user-programmable FPGA module, the company evaluated a new hardware architecture that gave it faster imaging rates than its existing system. It leveraged validated, off-the-shelf FPGA hardware, along with firmware for the PCI Express bus interface, which let it concentrate on algorithms and other parts of the system. For I/O, Santec used a custom adapter module design that combined a high-speed analog-to-digital convertor for image acquisition with the digital-to-analog convertor circuitry for the laser scanner control. Prototyping using modular FPGA hardware facilitated a working system and allowed it to determine required hardware changes because the I/O was decoupled from the FPGA back end.
|OCT imaging relies on the interference of light reflected from a sample and reference mirror to produce an image. Pictured is a simplified schematic of a time-domain OCT system.|
Santec acquired real data from its prototype system and compared it with images from the company’s existing system to prove the algorithms before moving code to the FPGA. Once the hardware and firmware were proven, the company moved to a more deployable PCI Express board, reusing many parts of the design and further reducing project risk. Overall, Santec achieved an image display rate of 40 frames per second with the new FPGA-based system, which was four times better than its previous system. Leveraging FPGA-based processing, the company also reduced the cost and size of the computer, enabling it to address new markets that need a small, low-cost system for portable applications.
One central challenge of using FPGAs for medical design and prototyping is that programming a system using, for example, VHSIC hardware definition language (VHDL) can be time consuming. However, recent advancements in development tools have made FPGA programming more efficient by allowing high-level graphical tools to be used for overall system design. This also helps leverage existing VHDL IP where appropriate. When used properly, these tools can enable the quick development of a prototype system so that algorithms and hardware performance can be evaluated and refined.
Similar challenges exist when incorporating GPUs into a system for general-purpose or scientific computation. In addition to requiring knowledge of specialized application programming interfaces (APIs) and GPU programming tools, developers must spend a considerable amount of time on the code required for communicating between GPUs and CPUs—a must for complex applications like medical imaging that can benefit from both computational units. While high-level tools exist for programming CPUs, linking these tools with code running on the GPU is critical to accelerating innovation. In the case of LabVIEW, a simplified, high-level API exists for passing data to and from the GPU and controlling GPU operations. It includes abstracting the process of interacting with the GPU in a way that is roughly similar to calling a multithreaded code library on the central processing unit.
GPUs and FPGAs are key elements that enable applications such as OCT. However, including these components in a system is just one step toward achieving academic or commercial success. Without a unifying software platform and off-the-shelf hardware, teams of experts in each technology must be employed to bring entire systems together.
The use of integrated hardware and software tools to accelerate discovery and development has contributed to recent scientific and engineering achievements. For example, to take advantage of FPGA and GPU technologies, systems from both Santec and Ohbayashi’s team include modular, off-the-shelf computational hardware, which reduces the amount of hardware development needed to create a functional design. Furthermore, these applications rely on system design software to abstract the infrastructural code needed to link multiple computational elements together and to simplify its programming and configuration.
Diagnostic imaging modalities continue to advance with new algorithms, high-performance processing, and better hardware. By combining modular, off-the-shelf FPGA and GPU hardware with high-level design tools, developers can create flexible and scalable systems to meet the needs of next-generation medical imaging systems.
Shelley Gretlein is director of software product marketing at National Instruments (Austin, TX). Casey Weltzin is a product manager for embedded software.