Hi, I hope that you can help me.
My problem is with the automatic recognition procedure in Brainvisa. How can I get a better automatic recognition of sulci?
Can I train BrainVisa with new brains to get a better sulci recognition?
If I can do it, somebody please can help me and tell me how can I do it?
Thanks for your time
Sincerely
Yasser
Can I get a better automatic recognition?
-
Yasser_Aleman
- Posts: 10
- Joined: Fri Oct 15, 2004 5:07 pm
- Location: Cuban Neuroscience Center
-
riviere
- Site Admin
- Posts: 1361
- Joined: Tue Jan 06, 2004 12:21 pm
- Location: CEA NeuroSpin, Saint Aubin, France
- Contact:
Hi Yasser,
> How can I get a better automatic recognition of sulci?
Well, by spending years of work into research in this field, of course
... Well, I'm joking, the recognition is not perfect, that's true.
The learning part has not been wrapped in a BrainVisa process yet because it was not something many end users would use everyday. So at the moment it's a complex procedure that I am the only one to know...
Moreover, learning a new database of brains assumes you have a database of manually labelled brains, and building it is a quite big amount of work for a specialist.
Before starting such a big work, there is something else you can do to try to improve the stability of the sulci recognitions: run it several times... Indeed the recognition is a stochastic optimisation process that tries to find a minimum in a very complex energy function. Doing it several times generally doesn't lead to exactly the same result because of the stochastic (random) part of it. Generally the results are quite close, differing on "details" that are difficult to assess. But in some cases the recognitions differ on larger parts. So it makes sense to do it several times and to keep the result that obtained the lowest final energy.
It is the opposite side of the algorithm in fact: the learning part modifies the shape of the energy function, and the optimisation part tries to use it better. What I suggest here is playing with the latter part.
We are working on a BrainVisa process that follows this idea.
Maybe I will do a process for learning a database one day, but I guess it wouldn't be useful to many people and it's a bit complicated because it needs lots of computer power, and some parallelization on a network is needed to run it in a reasonable amount of time. I've got scripts to do so on our lab network, but not in BrainVisa...
I have planned to learn the system on a larger database for a quite long time now, but never had the time to actually do it, and our database has not grown so big yet either... We also have ideas for improvements of the learning and recognition, but it's again the same problem... time...
Denis
> How can I get a better automatic recognition of sulci?
Well, by spending years of work into research in this field, of course
... Well, I'm joking, the recognition is not perfect, that's true.The learning part has not been wrapped in a BrainVisa process yet because it was not something many end users would use everyday. So at the moment it's a complex procedure that I am the only one to know...
Moreover, learning a new database of brains assumes you have a database of manually labelled brains, and building it is a quite big amount of work for a specialist.
Before starting such a big work, there is something else you can do to try to improve the stability of the sulci recognitions: run it several times... Indeed the recognition is a stochastic optimisation process that tries to find a minimum in a very complex energy function. Doing it several times generally doesn't lead to exactly the same result because of the stochastic (random) part of it. Generally the results are quite close, differing on "details" that are difficult to assess. But in some cases the recognitions differ on larger parts. So it makes sense to do it several times and to keep the result that obtained the lowest final energy.
It is the opposite side of the algorithm in fact: the learning part modifies the shape of the energy function, and the optimisation part tries to use it better. What I suggest here is playing with the latter part.
We are working on a BrainVisa process that follows this idea.
Maybe I will do a process for learning a database one day, but I guess it wouldn't be useful to many people and it's a bit complicated because it needs lots of computer power, and some parallelization on a network is needed to run it in a reasonable amount of time. I've got scripts to do so on our lab network, but not in BrainVisa...
I have planned to learn the system on a larger database for a quite long time now, but never had the time to actually do it, and our database has not grown so big yet either... We also have ideas for improvements of the learning and recognition, but it's again the same problem... time...
Denis