Since October is Breast Cancer Awareness Month, we have decided to detail some information about a promising new technology coming out of Utah which was recently presented during nVidia’s GPU Technology Conference.
So why is this technology important?
Every year, over 36 million women are screened for breast cancer in the U.S.
1 in 8 women will get breast cancer and 50,000 women will die from breast cancer in 2009 (personally, I come from a high risk family and I know how it is to lose someone over breast cancer.). In addition to all of this, their imaging options for detecting cancer are limited. First off, mammography misses cancer in 30% of women who have it. In addition to that, because it is painful to some many women don’t want to do it. There are also breast MRIs but they are expensive and often require the patient to consume contrast. These are also sensitive like mammograms, but not specific as to what the problem is. This technology is most commonly used for high risk patients. Another methodology is needle biopsy, which is as painful as it sounds. There are over 2.5 million biopsy procedures each year in the U.S. and 80% or more are benign. In addition to biopsies, there are also traditional ultrasounds which are sensitive and specific, but lack image capture and archive capabilities. These often require a high degree of skill to operate which makes availability an issue.
So how is Techniscan’s technology better?
It is an automated three dimensional whole breast ultrasound which is operator independent and can be repeated time and time again. The equipment is relatively small which makes it more portable and can be moved from room to room. The reason that repetition is important is because some solutions, like mammograms, involve radiation which is harmful to people over the long term.
Techniscan achieves the 3D imagery by combining 3 different images taken by the ultrasound which gathers information at every 2 degrees as it rotates around and sends waves to the receiver on the other side of the breast. This procedure results in a volumetric image which is generated using a combination of measuring the reflection ultrasounds, attenuation of sound, and the speed of sound.
Jim Hardwick and David Robinson from Techniscan Medical Systems recently gave two talks during the nVidia GPU technology conference detailing their company’s approach towards improving the detection of breast cancer while making the entire process less painful and quicker. The way that Techniscan Medical Systems has done this is by creating a Whole Breast Ultrasound, which is currently on its way to FDA approval. They are a medical imaging company who is based in Utah that works on the research and development of medical imaging technologies.
The way that this procedure is done is by lowering the breast into a warm water solution and then the ultrasound goes around the breast and collects the necessary data. This is different from the standard reflection ultrasounds this is similar to the ultrasounds people are accustomed to. This device combines both reflection and transmission ultrasounds which means that there is one part sending the sound signals and then another receiving them from the other side of the breast. This allows for quality imagery to be created and also helps acquire enough data to create fully 3D volumetric imaging. So where does nVidia’s Tesla fit into all of this?
Initially, in order to process all of this data and to help create the 3D imagery they first used CPUs. When this device was in its first inception, they used Intel Pentium II clusters which took approximately a week to generate one image. After that, they decided to upgrade to Pentium M which brought it down to four and a half hours. But the problem was that their goal was to be able to get the scan of the patient done and processed before the patient even left the doctor’s office so that there would be no more waiting for the result. They then began to add more and more nodes [after 6 or 7 nodes, the speed gains were no longer noticeable] by 12 they were beginning to lose speed. They looked at possibly using the IBM Cell processor, but that was too expensive. They also looked at possibly implementing DSPs but they were not fast enough either.
Back in 2006, CUDA was announced they did an experiment to see whether it was worth it. They simulated the ultrasound as it passed through the tissue. The FFTs took up the majority of the processing power so they tried offloading the FFTs to the GeForce 8800GTX GPU. The FFT was 8 times faster on the GPU compared to a single core on the Intel Core 2 architecture. After that point, more optimizations were done in order to make it even faster. At one point, 2D FFTs were running at 16x the speed of the CPUs. Initially there was resistance towards this change, so they began to check all of the results that came from the GPUs with the CPUs just to prove that the data and results acquired from the GPUs was verifiable by the CPUs.
The original hardware structure was cluster based and this presented some problems. The problems were that the CPU clusters generated a lot of heat, created a lot of noise, used a lot of power, and took up a lot of space. Implementing CUDA was different in that the data acquisition structure was the same but the processing CPU node was now one CPU with two Tesla GPUs. This led to reduced noise due to the GPU fans automatically throttling up when needed. It also meant fewer processors. And when not processing and sitting idle, they draw significantly less power. For comparison, the old cluster drew 600W idle and the new Tesla-based system pulled 300W at idle, effectively pulling half the power during idle. Not only was idle power consumption improved, they also improved load power consumption as well by switching to nVidia Tesla GPUs. The power draw from the Tesla GPUs was 700W in comparison to 850-900W on the old cluster at load.
Techniscan did not stop there, once the newest Tesla C1060 cards came out, they reduced the cluster from 4 GPUs to 2 GPUs which were actually faster than the old 4 and were also cooler and quieter. They then continued with the optimizations by beginning to take out FORTRAN code out of their software and made a list of what they wanted out of the CUDA technology which runs on the nVidia GPUs. They wanted to make sure that their code performed fast, but was also maintainable to a point where it was still fast enough for a 30 minute doctor visit. It took them about 3-4 man months to port over the functions of the FORTRAN code over to CUDA code. This was done between two people, Jim the CUDA developer, and his FORTRAN coding colleague. Here is a video of Jim last year explaining this as well, covering many of the same points made here.
At the moment, the data acquisition is 12 minutes per breast and processing time is 24 minutes per breast.
This is not quite at the speed that they want it at, but with upcoming new generation of Tesla Fermi cGPUs, the speeds of computation done by the CPU could increase exponentially. In the end, they would like to be able to have accurate results given to the doctor within the standard 30 minute doctor visit so that the patient and doctor can immediately discuss the results without having to come to another visit or try to analyze the results on their own.
Using GPGPU technology is also very important for bringing the cost down, as the main intent for this device is to fit the price bracket so that even small doctors offices in outback/out-of-reach areas are able to afford it. We discussed the development of the device and the fact that it comes at a significantly lower price than radiation-based solutions means that in this case, saying that GPGPU saves lives is not a marketing gimmick, but rather the fact of the matter.
We are very excited for this information and hope that they are able to bring this product to market as soon as possible and begin to further the fight against breast cancer. Being someone who has had family members be impacted by breast cancer, I feel like we need to take every opportunity we can get to detect and treat this disease as soon as possible.