Slashdot Mirror
GPUs Helping To Lower CT Scan Radiation
Gwmaw writes with news out of the University of California, San Diego, on the use of GPUs to process CT scan data. Faster processing of noisy data allows doctors to lower the total radiation dose needed for a scan. "A new approach to processing X-ray data could lower by a factor of ten or more the amount of radiation patients receive during cone beam CT scans... With only 20 to 40 total number of X-ray projections and 0.1 mAs per projection, the team achieved images clear enough for image-guided radiation therapy. The reconstruction time ranged from 77 to 130 seconds on an NVIDIA Tesla C1060 GPU card, depending on the number of projections — an estimated 100 times faster than similar iterative reconstruction approaches... Compared to the currently widely used scanning protocol of about 360 projections with 0.4 mAs per projection, [the researcher] says the new processing method resulted in 36 to 72 times less radiation exposure for patients."
The TFA says that this tech is usually used prior to treatment, while the patient is in the treatment position.
Because processing a limited number of scans into a useful model previously took several hours, they were forced to perform many more scans to get a more accurate picture with which to build their model - because they don't want to leave the patient lying in the scanner for 6 hours prior to treatment.
With this improvement in processing power, they can produce the model from limited data in a feasable time.
So the summary does actually describe the breakthrough quite well: It's not a new image processing technique for working with limited data, it's just new hardware allowing that process to be run in a quicker way. Yes they're using a slightly new algorithm, but I doubt that is a massive breakthrough in itself.
CT scanning is associated with an increased risk of cancer in children. This development will significantly lower that risk.
Matlab is rarely ever graphically intensive...
retrorocket.o not found, launch anyway?
The technique is called iterative backprojection. The reconstruction process assumes an array of pixels which, at the beginning, are of some uniform value. It then looks at a ray of attenuation data from the CT projection (along this ray, the tissues in the target result in this degree of attenuation of the xray beam), and asks "how must the pixels along this ray be adjusted, so that their attenuation along the ray matches the data from the CT beam?". It does this for every measured ray taken during the acquisition, over many different angles. The more sparse the acquired data, the more iterations (and thus, longer) it takes to get a reliable (approximated) image.
"Our work, when extended from cancer radiotherapy to general diagnostic imaging, may provide a unique solution to solve this problem by reducing the CT dose per scan by a factor of 10 or more," says Jiang.
It's probably applicable to diagnostic cone beam scans, which are the hot item in implant dentistry. The reason it's first applied to therapy scans is because the tissue surrounding the tumor suffers radiation from scattering of the therapeutic beam, making dosage reduction highly desirable.
Apocalypse Cancelled, Sorry, No Ticket Refunds
This has been said elsewhere in this thread, the real breakthrough here is due to compressed sensing, but here are some extra information:
1- Compressed sensing basically used the idea that it is not necessary to sample an image (or a projection in this case) everywhere because natural data is fairly redundant. This is why you can capture a 10 Mpixel image in a digital camera and have it compressed to a 2 Mbyte JPEG file without losing much visible information. Compressed sensing basically does the compression *before* the sampling and not after. Researchers at Rice University for instance built a working, one-pixel camera using this brilliant principle.
2- Compressed (or compressive) sensing was proposed by Emmanuel Candes and Terence Tao respectively at Stanford and UCLA. Tao is a recent Fields medalist. I recommend reading his blog if you like mathematics.
3- This field is really less than 10 years old, it has completely turned on its head classical ideas about sampling-limited signal processing (Nyquist, Shannon, etc). It is a brilliant combination of signal, image processing and recent advances in combinatorial and convex optimization.
4- However this is only the beginning. Because compression happens before sampling, you need to make so-called sparsity assumptions about the signal ; in other words you need to know a great deal about what you are going to try to image. In interventional therapy, precise imaging of the patient is made beforehand in a classical way (CT or MRI), and this kind of technique is only used to make fine adjustments as therapy is ongoing. This is extremely useful and safe because of lower radiation output and because the physicians know what to expect.
5- Here the GPU is useful because it makes the processing fast enough to actually be used. It is an essential brick in the application, but of course not in the theory.
Best.