Tuesday, January 19, 2010

CS: Dick Gordon's Op-Ed, Lianlin Li's Open Question on Compressible Priors


Dick Gordon, one of the first to develop ART for CT tomography just sent me the following (a follow up to a previous entry). I am reprinting it as is (except for some edit on the form and hyperlinks)


How to get CT out of its high dose rut: CS for CT x-ray dose reduction is a political issue



The article:

Pan, X., E.Y. Sidky & M. Vannier (2009). Why do commercial CT scanners still employ traditional, filtered back-projection for image reconstruction? Inverse Problems 25, 36p. #123009.

places the blame for what Igor Carron calls “the inability of ART to topple FBP“ on the theoreticians and the manufacturers, somehow expecting the latter’s engineers, stuck in the middle, to overcome a number of sociological problems. These are scattered through the
paper, so I pull them together here:

...it makes little sense to spend endless time and effort refining an inversion formula for an idealized imaging model when there are some compromising factors earlier in the data-flow chain....

Most of the research devoted to image reconstruction aims at developing new solutions to solving imaging model problems. The idea being that the imaging model problems evolve toward realistic situations and that this theory eventually finds its way into application. We submit that this style of research is not effective for translation based simply on the empirical evidence that not much of the work published in Inverse Problems over the past couple of decades is actually used in a CT scanner.

The bulk of the applied-mathematics research effort seems to go toward developing advanced techniques that will likely never be used in practice or address problems that have little application.

And presently users of CT are used to some of the streak artifacts in FBP-based reconstructions, and there is not really a huge motivation of exchanging FBP streaks for some other type of artifact.

Parameter explosion is one of the challenges that face optimization-based algorithms for image reconstruction in commercial products.

The real progress in moving past FBP-based reconstruction will occur when engineers have real experience with advanced image-reconstruction algorithms and can use this knowledge to design more efficient and effective CT scanners. This development will likely occur first in dedicated CT systems such as head/neck CT, dental CT and breast CT.

Among the main reasons for slow progress on introduction of algorithmic innovations into clinical CT scanners may also be the proprietary nature of CT technologies and the subjective evaluation of performance by its users.

The door to innovation in CT algorithms requires an efficient and practical route to develop and test new algorithms. It is unlikely that any investigator, mathematician or engineer could effectively correct for the raw detector measurements from the CT scanner into a form suitable for algorithm testing without the manufacturer’s blessing and assistance. However, manufacturers of CT scanners, in general, have not made the raw projection data available to anyone outside their organizations, even to customers that use their CT scanners.

There is no Digital Image COMmunication (DICOM) standard for CT raw data [86, 87]. In fact, there are no standards at all. Corrected raw projection-data sets are almost impossible to find in the public domain. Without this data, clinically realistic examples of images cannot be reconstructed. As a consequence, CT-algorithm developers almost universally use highly idealized numerical phantoms for their work and seldom show results obtained with ‘real’ experimental or clinical data. This is surely a major impediment to progress.

If the system was ‘open’ and could be refined by algorithm developers in the field, such algorithms could be tested and clinically validated. The benefits of new technology, especially in image-reconstruction algorithms, could reach patients without waiting for their possible inclusion in future generations of CT scanners (because this has not happened for the past 25 years and may not in the near future).

Perhaps the need for an open-source-community-developed CTreconstruction toolkit (CTK) with a database of corrected raw projection data from real CT scanners will be recognized.

I have been asking clinicians on and off for 40 years to put “open access” into their purchasing contracts with CT companies. But for most of them the role of the CT algorithm is invisible, and as is generally true in medicine and biology, the role of the theoretician is simply not recognized, let alone paid a salary or listened to. So how to break this impasse?

I have a simple proposal: CS (Compressive Sampling) people need to demonstrate to politicians concerned with x-ray dose that they can get the dose down by an order of magnitude or more. This, I think, is the answer to Igor Carron’s analysis:”Therefore a CS system would have to provide something else“. The politicians will then, in effect, legislate a role for CS people in CT design. The companies won’t do it themselves. Nor will their customers, the radiologists. It won’t happen in breast CT, because everyone will just use the dose ceiling already established in projection mammography.

The only radiologist that I know of who vigorously lectures other radiologists on the need for dose reduction in CT is David A. Leswick of the University of Saskatchewan. His papers are a little less fiery:

Leswick, D.A., N.S. Syed, C.S. Dumaine, H.J. Lim & D.A. Fladeland (2009). Radiation dose from diagnostic computed tomography in Saskatchewan. Can Assoc Radiol J 60(2), 71-78.

Leswick, D.A., C.S. Dumaine, N.S. Syed & D.A. Fladeland (2009). Computed tomography radiation dose: a primer for administrators. Healthcare Quarterly 12(Special Issue), 15-22.

Nevertheless, the latter starts:

“The use of computed tomography (CT) is growing, and, consequently, the associated radiation dose to patients is increasing as well. There is increasing evidence linking the radiation dose within the range of diagnostic CT with a significantly increased risk of malignancy. These two factors combine to make radiation dose from diagnostic CT a public health concern.”

Yours, -Dick Gordon gordonr@cc.umanitoba.ca

Thanks Dick, I agree with you that in the context of CT, maybe, the political whip is the only constraint that can force a dose reduction and opens the door to more efficient algorithm like ART or other nonlinear reconstruction techniques introduced with compressive sensing.

Lianlin Li mentioned this to me last week and he just wrote about this on his blog. Lianlin is particularly confused with three papers that seem to contradict each other on the subject of compressible priors for sparse bayesian estimation. The papers are:
[1]M.E. Tipping. Sparse Bayesian learning and the relevance vector machine. The Journal of Machine Learning Research, 1, 211–244, 2001.
[2] David Wipf, Jason Palmer, and Bhaskar Rao. Perspectives on sparse Bayesian learning, Advances in Neural Information Processing Systems, 16, 2003
[3] Volkan Cevher. Learning with compressible Priors, Preprint, 2009.
The slides in the blog entry are in English. The introduction is in Chinese but it is not relevant to the confusion at hand. If you want to clear up the misunderstanding between the papers as presented by Lianlin and feel you cannot comment in the blog with instructions in Chinese then you are welcome to put a comment here and I could make an entry of all y'alls responses.

Thanks Lianlin for following up on this.

Credit: NASA, Spirit sol 2147

No comments:

Printfriendly