From Utah Center for Neuroimage Analysis

Ripples

Ross and Tolga took a look at the ripples (only visible when the slices are shown in a fast sequence), and we are not sure that this noise is worth removing. It is unlikely to have any adverse effect on the tracking tool.

UINTA

(12/3/05 CBC--the following is inserted from an email exchange)

Just to give some background, UINTA is coming out of the dissertation work of Suyash Awate. It is a very powerful image filtering tool, but very computationally expensive as well. For a 500 x 900 image, Suyash guesses that it would take 8-10 hours (on a fast single processor machine) to get the same quality results I showed you. If you had only one slice no problem, but when you have a stack of 1000 slices you need a PC working on this full time for a year :) But maybe it is not completely hopeless... Suyash has been considering a potential speed-up to his algorithm (no results yet). The other option to find a large number of processors to run this on, which could also be a possibility.

--tolga Mon, 21 Nov 2005 13:29:22

Thanks for the reply. If textures are similar enough from slice to slice, might it be possible to figure out the transform on one slice, then apply to the others? Or does it not work that way?

--Chi-Bin Mon, 21 Nov 2005 14:06:53

We can do that, but unfortunately this wouldn't speed anything by itself.

--Tolga Mon, 21 Nov 2005 14:26:46

To clarify a bit more... the reason this wouldn't speed up the algorithm is because we don't learn anything explicit like a transform. What we learn is all represented in a lot of examples, so the bottleneck is applying those examples to the image. Thus, the only realistic way of speeding things up is to decrease the number of examples needed for the same quality learning. This is hard to describe in an email, if you are interested we can discuss this in a meeting.

--Tolga Mon, 21 Nov 2005 14:32:00

On Nov 21, 2005, at 2:26 PM, Tolga Tasdizen wrote: >We can do that, but unfortunately this wouldn't speed anything by itself.

Sort of. If we can use statistics from one slice and summarize them (e.g. using multipole methods), that would help.

--Ross Mon, 21 Nov 2005 14:42:51 -0700