CUDA Spotlight: GPU-Accelerated Medical Image Processing
By Calisa Cole, posted Dec 02 2011 at 12:00PM
This week's Spotlight is on Anders Eklund, a SPh.D. student at Linköping University in Sweden. Anders is affiliated with the University's Center for Medical Image Science and Visualization (CMIV), a multidisciplinary research center.
NVIDIA: Anders, what is your research focused on?
Anders: I work with algorithms for medical image processing, such as image registration and image denoising. Image denoising refers to increasing the quality of an image by removing or suppressing "noise."
My research is especially focused on functional magnetic resonance imaging (fMRI), where you try to find brain activity from magnetic resonance images (MRIs) of the brain. My interests include brain-computer interfaces (BCI) with real-time fMRI, where the fMRI data is processed in real-time as the subject is in the MR scanner. A brain computer interface could help people communicate who are paralyzed or suffer from Locked-In syndrome.
NVIDIA: How did you become interested in this area?
Anders: I have been interested in image processing for a long time. As a graduate student I was involved in a project where we made a graphical user interface to connect the ITK (Insight Segmentation and Registration Toolkit) and VTK (Visualization Toolkit) libraries with MATLAB. I like the idea of helping medical doctors by giving them new tools and clearer images so they can make better and faster diagnoses.
NVIDIA: How does GPU computing play a role in your work?
Anders: GPU computing is very important for me and the research group in which I work, as many of the algorithms that we develop are very computationally demanding. To be able to develop, evaluate and improve an algorithm, it really helps if the processing time for one run can be reduced from minutes to seconds, or from hours to minutes. For example, I recently developed an algorithm for non-parametric statistical analysis of fMRI data and by using a multi-GPU implementation we saved about five years of processing time during the development and testing.
Another challenge in the medical imaging domain is the amount of data that is collected for a single patient. A 4D (3D + time) computed tomography (CT) dataset can be of the resolution 512 x 512 x 512 x 20 and require as much as 10 GB of storage. To apply image denoising to such a dataset can take several hours on the CPU, compared to 15-20 minutes on the GPU.
NVIDIA: What are some advantages of working with CUDA?
Anders: CUDA is very easy to learn. A working prototype of an algorithm can be developed in a short amount of time. I learned the basics of CUDA in about a week and then I've learned more along the way as I've implemented different algorithms.
I like the different tools that NVIDIA has developed (like the Visual Profiler) and the fact that NVIDIA continues to improve the CUDA programming model by listening to people who use CUDA.
NVIDIA: As computing becomes more powerful, what can we look forward to?
Anders: One exciting trend within medical imaging is to move the processing of the data into the surgery room, such that the medical doctors can get real-time feedback during surgery. This places high demands on the computational performance, as all the processing has to be done in real-time.
Anders Eklund holds a M.Sc. in Applied Physics and Electrical Engineering and is currently working on his Ph.D. in Medical Informatics at Linköping University. In his spare time he likes to play computer games, take photographs and travel.
Anders is currently financed by the Linnaeus center CADICS, funded by the Swedish research council, and the Neuroeconomic research group at Linköping University.