Several technologies—like 4D (3D over time) ultrasound imaging (Fig. 1)—have taken the medical-imaging market by storm. The medical field will continue to benefit from Moore's Law as speed and resolution continue to improve. Take for example the joint effort between engineers and scientists from IBM and the Mayo Clinic that seeks to exploit recent parallelism advances in processors such as the Cell.
The result is a dramatic acceleration in 3D medical-image processing, which significantly advances the image-fusion process. Also known as registration and overlay, this process creates 3D images by aligning two or more images captured by different devices (e.g., MRI and CT), or the same type of device on different dates. Using alignment algorithms, images are "fused" to provide more complete visual information for easier detection of tissue changes like tumor growth or shrinkage.
But University of Calgary students have taken a different approach in creating the most complete 4D model of a human yet (Fig. 2). Using a joystick, the object-oriented hologram, dubbed CAVEman, can provide a view of up to 3000 distinct body parts. This technology will help physicians plan for complex surgeries and allow patients to see a map of their body before surgery.
LET'S TALK ABOUT GOALS
Instead of using the hospital's ICU equipment, patients can be monitored at home. The patient's quality of life improves and medical costs are reduced, achieving two key goals of these new technologies. Another goal is improved accuracy—for example, imaging the heart in a single beat or the lungs in a single breath. Researchers also hope to improve diagnostic capabilities via the least invasive procedure in as close to real time as possible.
Let's not forget about reducing or eliminating false positives and false negatives. Traditional mammograms have a high percentage of false positives, resulting in the unnecessary removal of tissue in far too many patients. Added costs for false positives include evaluation costs, treatment costs (of the observed breast cancer), and the immeasurable emotional cost associated with a false-positive result. Of course, a false negative can be much worse, possibly leading to death. And for the physicians responsible for interpreting these images, the ongoing goal is to increase the potential to find anomalies in organs, tissue, and cells via the least invasive means.
These goals imply one ongoing theme for all new or redesigned medical imaging systems: the need for maximum computing power to provide the highest-resolution processed images in the least amount of time. Typically, that means maximizing the number of cores and threads for the target form factor, since many imaging algorithms are parallel-processing friendly.
But before deciding which brand of multicore processor to use (see table), carefully consider the system's scalability and upgradability. Due to jumps in performance and data rates within the semiconductor and storage industries, it's important to be able to drop in the next-generation device or add more nodes to the system (when using clustering) without redesigning and retesting the entire system. If you can get away with only a recompile, you're ahead of the game.
"Scalability of solutions is key to enabling customers' reuse of software and algorithms across products," says Bob Ghaffari, manager of the Medical Segment for Intel. "Having a silicon architecture that can address a variety of performance and power bands ranging from high-end CT equipment down to a low-power portable ultrasound product requires an architecture that can scale."
Ghaffari said Intel is focused on meeting a variety of medical application requirements by providing highly functional and flexible system-level building blocks, thereby minimizing the cost of ownership and significantly accelerating time-to-market.
When attempting to determine just how many cores and threads are needed, try to make the data path the bottleneck, because there's really no point in processing data faster than it can be stored. If a local hard drive will be used, then serial ATA (SATA) or serial attached SCSI may be the limiting factor. Otherwise, if you're writing data to a device on the network, the network connection (Ethernet or wireless) will have a known maximum bandwidth.
Intel suggests some guidelines for choosing the number and type of processors, as well as how to tweak them. First, determine board performance and form-factor criteria. Next, run the code. Continue to optimize code, and stop after a reasonable number of iterations. Then, adjust the performance. If this is adequate, you're done with architecture selection. If greater system performance is required, add external devices for acceleration offload.
Creating 4D images from a spattering of 3D images could require anywhere from 500 Mbytes to 5 Gbytes of data per patient. This is sure to grow as resolution and the number of image slices increase. Factor in the number of patients seen on a daily basis, and a thin-client network that stores all patient data on a fast central server and uses local PCs to display the images starts to look attractive.
But when portability is mandatory, a system based on the MicroTCA architecture may be the best bet. MicroTCA provides a rugged small form factor with lots of compute power, bandwidth, and built-in network connectivity. Meanwhile, the display of 4D images requires several gigabytes of temporary storage. Designers have to consider the amount, type, and speed of graphics memories like Graphics Double Data Rate (GDDR) (see "High-Speed Memory Drives Visualization").
If you segment out your architecture properly, with an overall goal of designing only what isn't readily available, chances are you're in good shape. The major building blocks include the analog front end, the digital back end, the graphics display renderer, and a system controller with optional networking (Fig. 3).
Data acquisition and image pre-processing make up the analog front end. They rely heavily on the imaging modality, which may require one or more DSPs, FPGAs, or ASIC ICs. The digital back end includes the image reconstruction and post-processing blocks. Depending on the modality's complexity, this block could be a simple processor (GPU) or one or more advanced processors (CPU and/or GPU) containing multithreading capabilities with multiple cores. For demanding tasks like image processing and reconstruction, when top performance is needed, processors like the Cell Broadband Engine may be more appropriate (Fig. 4).
If your future involves multiple cores, seriously consider software-based decisions, such as the operating system, message passing interface, parallel programming language, and so on. Even otherwise trivial decisions like which type of file system to use become substantially more important and should be made carefully (see "Parallel Programming And Multicore Environments" and "Multicore My Way").
TechniScan's UltraSound CT Imaging System produces fully digital breast images based on transmission ultrasound. This type of ultrasound can be used to produce two images of the breast based on both the speed and attenuation of sound (Fig. 5).
"When a vendor says that they can replace a major component in my system that doubles the performance of the original component and requires the same power and cooling as the original component, I get really interested," says Frank Setinsek, system architect for TechniScan (see "Advances Trigger An Ultrasonic Boom,").
Except for X-rays, which are recorded directly on film, all medical imaging modalities use similar basic principles and rely on a similar data flow (Fig. 6). The process starts with the imaging machine building an analog "image." It does so by applying one stimulus or more to the patient (subject) and then recording the response to the stimulus. Then the raw data is usually pre-processed and "scrubbed" to both suppress noise and enhance signal quality.
Next, the pre-processed image is typically reconstructed by converting (e.g., using a Fourier transform) thousands of transmission measurements into a pixel map that makes up a physically meaningful image or volume. The image or volume then is postprocessed to improve its appearance and usefulness. The image display may be standalone or a composite built using overlaying images captured with different technologies, like MRI and PET. If slicing techniques were used, the slices may be viewed one at a time or combined for a 3D view.
Finally, computer-aided diagnosis (CAD) may be employed to aid in analysis and interpretation of images. CAD works by using the post-processed data and applying segmentation, followed by feature selection for the regions of interest and feature classification using pattern-recognition algorithms. The physician or radiologist then enters the equation as the final interpreter. After analyzing the images and optionally using historical data as a base for comparison, the physician delivers the diagnosis or update to the patient (see "Video Processing Brings New Meaning To Motion,").
ADDITIONAL WEB RESOURCES
One Web site, www.rtstudents.com, was designed with radiology students in mind. This portal to other useful sites also contains a plethora of great links for research, discussion, and resources to aid learning.
GE's Medcyclopaedia includes a medical-imaging encyclopedia, a glossary, and an outstanding interactive e-learning section with a complete anatomy breakdown (www.medcyclopaedia.com). With this site, you'll never get the cerebellum confused with the temporal lobe again; the elearning module also includes a virtual index-card-by-picture or -by-name learning system for medical-imaging terms.
If you're looking for information on high-performance computing (HPC) using clusters, some helpful Web sites include IEEE's Computer Society Task Force on Cluster Computing (www.ieeetfcc.org), the Linux HPC site (www.linuxhpc.org), the Windows HPC site (www.winhpc.org), the Sun HPC site (sun.com/hpc), and IBM's deep computing site (www.ibm.com/servers/deepcomputing/).
Also, be sure to read GPU Cluster for High Performance Computing by Zhe Fan, et al. Other written resources include white papers such as Intel's Optimizing Software for Multi-core Processors and How Much Performance Do You Need for 3D Medical Imaging?, Toshiba's The Next Revolution: 256-Slice CT by Richard Mather, PhD, and Altera's Medical Imaging Implementation Using FPGAs.
DIAGNOSTIC FOOD FOR THOUGHT
Before designing your next medical imaging system, there's one last thing to consider. With the correct image analysis and diagnostic programming, is it possible for a computer to "out-diagnose" a physician or radiologist? It certainly seems feasible for some ailments even now, and this possibility grows stronger with each generation of processor power and knowledge.