Přepis řeči - EFFICIENT IMAGE RECONSTRUCTION UNDER SPARSITY CONSTRAINTS WITH APPLICATION TO MRI AND BIOLUMINESCENCE TOMOGRAPHY

0:00:14	thanks first to me and and joe for for organising
0:00:18	very interesting
0:00:19	session which also gives us a
0:00:21	the in
0:00:22	to be prime
0:00:23	uh so it's a will be talking to you about the reconstruction algorithms under sparsity constraints and so that kind
0:00:29	of relates also to
0:00:31	that's of activity going on in signal processing it
0:00:34	a compressed sensing in give you some tuition
0:00:37	uh in "'em" are i and by means
0:00:39	nee since uh tomography
0:00:42	um
0:00:43	and the
0:00:44	also i i mean in doing so i'll give you also kind of the art of for
0:00:50	uh i turn to a reconstruction algorithms that use the
0:00:54	uh
0:00:55	uh
0:00:56	regularization constraints uh by by the way of sparsity
0:01:01	so uh first to a we i i i i i will just
0:01:04	discuss a little the variational approach to image reconstruction will need to
0:01:08	minimize so of some
0:01:10	some uh a functional and and where we introduce some sparsity constraints then
0:01:14	i will read uh read you some of the standard algorithm in particular uh it
0:01:20	soft thresholding the uh uh the algorithm that's is stuff
0:01:23	the fast version of that which is a a kind of multi step i'll grid which is
0:01:28	gives you sort of the state of the arts
0:01:30	coming from convex a analysis
0:01:32	and then uh uh in this talk i i i i want to
0:01:35	i discuss our variation of used star uh which is sort of
0:01:39	you note a like to the problem which will use step adaptation to to gain uh some speed because the
0:01:45	cool here really to get to the speed of the conventional in your taking
0:01:49	is that have been
0:01:50	use for for many years
0:01:52	and then uh application to power L or i and and by mean this nest since the tomography
0:01:58	so here is uh just the set
0:02:00	so uh okay so we we have some object X that we'd like to to image for for example
0:02:07	a proton density or some absorption map
0:02:10	and then we have these scanner
0:02:12	here that uh we will just assume we have a linear measurement system which is the the usual assumption in
0:02:18	them are i
0:02:19	X ray and so it's an integral operator but to since here we have a discrete eyes the problem so
0:02:24	we can all represent body big fat matrix H
0:02:28	that uh represents all our physics knowledge of of the problem
0:02:31	and then uh at the outputs uh oops
0:02:34	the yeah i here a see we have some
0:02:37	modeling perfection and noise and and all that i i represent by oops
0:02:41	by be uh maybe i should be problem like lawn
0:02:45	knowledge knowledge that's professes where
0:02:48	oh
0:02:49	in the room
0:02:50	uh
0:02:52	okay so so uh uh a so here was saying we we have noise
0:02:56	and so our imaging matching model here looks very easy just the this uh imaging matrix up like to the
0:03:01	object that we like to recover
0:03:03	plus some noise
0:03:04	and then in the variational approach so you just set up this
0:03:08	kind of uh uh that this criterion here where you have a
0:03:11	data consistency requirements so we you you are kind of
0:03:14	uh reconstructing X
0:03:16	and then you are you're simulating your measurements and you would like this difference to to be very small
0:03:22	spec to the true measurements
0:03:23	and then here you it should use some regularization constraint and this is the arts
0:03:28	of reconstruction here to
0:03:30	to to just choose
0:03:31	the the the right trigger is regularization
0:03:34	and and and a classically here uh for example if we put a a quadratic term your was that's say
0:03:40	sum
0:03:40	but people operator here that
0:03:42	it's that's
0:03:42	tikhonov regularization but
0:03:45	now uh
0:03:46	i with compressed sensing actually all this was started by
0:03:49	the observation that so many images natural images would represent was uh well was very few large group fusion is
0:03:57	some basis
0:03:58	in particular wavelet basis and and in fact this happens as well
0:04:02	uh very significantly with medical images since so here's some "'em" are i image
0:04:07	and and just the showing different
0:04:09	transformations and and and and and so that you can get a pretty good uh a signal restitution but just
0:04:16	keeping being a very small percentage of of large coefficients and
0:04:20	and and so this this is really like used uh
0:04:23	to promote uh i mean to to sort of justify the you the use of wavelets and so here
0:04:29	then the idea ideas to use a regularization where you will impose a and one and that T
0:04:34	uh a on the wavelet coefficients so here are would uh represent let's say might wavelet transform
0:04:39	and one norm
0:04:40	uh uh uh promotes sparsity and why it's good its convex so so that we have like some
0:04:47	i guarantee you existence uh
0:04:49	this
0:04:50	oh
0:04:51	a solution and also it will favour sparse solution
0:04:55	okay so here is sort of uh the grandfather of of those uh a sparse a reconstruction written
0:05:01	and and and so uh what do what you want to do is to minimize this okay you have this
0:05:06	L two term
0:05:07	and this
0:05:08	sparsity constraint here i'm just as to me
0:05:11	X is the representation in the wavelet basis and i'm using
0:05:15	and augmented just the matrix that's sort of contents this
0:05:18	a a wavelet transform followed by by by by by by uh
0:05:22	i imaging device
0:05:24	so
0:05:25	then you have to very special case very easy completely classical also first if you put a day will zero
0:05:31	just the square problem so it's a a text solution is one like calculation you get that
0:05:37	uh and four she those i which may
0:05:39	way
0:05:40	a you at and
0:05:41	rob you
0:05:42	would not even want to compute that but then
0:05:44	this sort of the grand father of our britains here which is
0:05:48	but
0:05:49	and a grid
0:05:50	uh on on on to this data term
0:05:52	and and and i mean this is sort of a you know the big the backbone of many iterative out
0:06:00	and the i mean this works but the you have to two
0:06:03	to stop it after a certain number of iterations
0:06:06	but now the other interesting cases
0:06:08	H is i you or you are
0:06:11	like a wavelet transform in that case
0:06:13	fact this
0:06:14	problem you can by just going the wavelet domain
0:06:17	uh doing a a a a a a soft thresholding and
0:06:20	going
0:06:21	now it soft thresholding algorithm which was the uh discovered
0:06:25	well
0:06:26	many
0:06:27	people probably but uh
0:06:29	feed yeah they do in no of act yeah uh where where
0:06:32	among the very first here using your yeah formulation
0:06:35	and so the a kind of a start if you were just to thing those two very
0:06:39	three of
0:06:41	remarkably uh this would solve this problem
0:06:45	should look like to very difficult problem and then do a she the use them and two thousand four were
0:06:50	able to
0:06:51	sure that action
0:06:52	prove convergence
0:06:54	and so that that was uh a little revolution in the field
0:06:57	although i'm say the output myth
0:07:02	so maybe not so usable in that form
0:07:04	and now if you just look at the structure actually the fixed points uh algorithm
0:07:08	so uh what you have is a rule which is this a gradient
0:07:12	this sense followed by
0:07:14	based thresholding that will give you the new update and then there's this nice
0:07:18	mathematics that guarantees convergence
0:07:21	and fortunately it's very slow so
0:07:23	that's been shown now that uh the cost
0:07:26	will a decrease like one over and where and the number of iteration
0:07:30	but then by
0:07:32	uh building on many many years of full convex optimization then there was this proposition position in two thousand nine
0:07:39	of fist yeah
0:07:40	which is a like you control
0:07:42	over the relaxation
0:07:43	evaluation on this thing but
0:07:45	essentially what it is it's it's a an updating dating a G that not only text
0:07:49	uh
0:07:50	uh uh uh built upon the previous estimate but takes two
0:07:54	pretty
0:07:55	a so and minus one
0:07:56	and and and
0:07:58	and then will
0:08:00	a a construct a kind of
0:08:01	vacation
0:08:02	well
0:08:03	car uh the correct uh update should be
0:08:05	and then apply the is to out
0:08:08	so to to just make it
0:08:10	problem
0:08:11	trying to solve
0:08:12	and are remarkably this fist uh algorithm
0:08:14	use you uh uh and square
0:08:17	so it it's one or there
0:08:18	faster or and and and and so gives you
0:08:21	now a a a fast enough our
0:08:23	to to to apply this kind of techniques so
0:08:25	now what we have proposing here
0:08:28	it's some variation and and so we call it fast the okay
0:08:32	weighted iterative soft thresholding algorithm
0:08:35	and and the idea is pretty simple so we had the step size
0:08:39	or
0:08:40	but not your
0:08:41	yeah
0:08:42	getting factor for for the great so we replace that by a diagonal matrix
0:08:47	and actually we can also uh some
0:08:49	flexibility on the
0:08:51	a side of the regularization so
0:08:53	with the slightly more channel
0:08:55	regularisation matrix
0:08:57	and you she uh the this is uh or or style with my are derived using
0:09:01	uh max a uh uh uh and M
0:09:04	a a a a a a a as a sort of constructing a a a quadratic bound on on the
0:09:09	cost functional
0:09:10	and and then all using a you know always trying to go and the bottom of this quadratic bound no
0:09:15	if you have more flexibility here with
0:09:17	introducing this matrix
0:09:19	fact but you can do you can construct
0:09:21	better a quadratic bound that
0:09:23	re tailored to it to your problem
0:09:26	and so hopefully this should allow us to go faster
0:09:29	and then actually we we can do some have here
0:09:32	and and so we obtain a now this uh this kind of uh in a quite the on the cost
0:09:38	so no if you look at it superficially for mathematician doesn't look so much better because we still like what
0:09:44	are and where one or and square
0:09:46	but the important thing here we instead of having
0:09:48	a in a norm we have
0:09:51	weighted
0:09:52	and actually this i for the call
0:09:54	oh
0:09:55	and now it means that clever engineering can really like from down those calls
0:10:00	quite significant and and then
0:10:02	can uh and that a be much fast
0:10:04	so the application here i
0:10:06	just want to
0:10:07	a show you two examples by by the way i forgot to knowledge my
0:10:11	cool all is okay and who the want to
0:10:14	did all the work and and this is the first one matt i'm much should
0:10:17	just can K on is working on product and more i actually in collaboration
0:10:21	with the guy who invented the pollen and the rice so it's cuss was month from it T H three
0:10:28	so it here the idea is just to to have the a bunch of colours here
0:10:32	and for the
0:10:33	soon or a little basics about "'em" or rice so M are i it's she we compute samples of of
0:10:38	a a a a measure samples of the for yeah transform
0:10:41	but if you have colours in multiple colour they also this sensitivity function so it's
0:10:46	it's you are object X multiplied by a C
0:10:50	a function that's a sense it to you but you of your cord and doing the for transform all all
0:10:54	that
0:10:54	and in power and the right you just put several core
0:10:58	so you you can do the the measurement in parallel and hopefully this
0:11:01	allows you
0:11:02	to reconstruct images using
0:11:04	fewer samples
0:11:06	and so now here how can we exploit the structure of the problem so we need really to
0:11:10	they those degrees of freedom that that left which is this diagonal weighting matrix
0:11:15	so if you look at the pollen and more i what we have
0:11:18	so for for each call here or you would have this sense
0:11:21	T matrix which is just a a i uh a diagonal matrix followed by full you
0:11:26	and now the problem here is this sensitivity can vary a not
0:11:30	and
0:11:30	on on you on your on your course
0:11:33	and it's a pretty in a gene use and that makes the the problem not so well conditioned because for
0:11:38	many years
0:11:39	actually i are i would mean you
0:11:40	and inverse for your transforms so with the unitary for
0:11:43	vol
0:11:44	oh rate in the
0:11:46	problem for inverse
0:11:47	a the imaging uh if inverse problem
0:11:50	people but uh i mean with was that it becomes much uh uh less less well uh condition
0:11:56	so now the idea of sensitivity based a precondition conditioning
0:12:01	here uh
0:12:02	okay so it's uh try to uh inverse he respect respect to to to the sensitivity
0:12:08	just tried to get the better condition
0:12:11	uh a system matrix
0:12:12	uh in in this uh precondition
0:12:16	them
0:12:16	variables
0:12:17	and so now if you think of it in the wavelet domain you could almost to the same but okay
0:12:23	uh you just have to transpose your send
0:12:26	i
0:12:26	in the wavelet domain and an intuitive is just the way that is more less local so you would just
0:12:31	i agree
0:12:32	a sense C P T V T oh over the support or of the where let's
0:12:37	and and and then get this kind of precondition E
0:12:40	matrix
0:12:41	here here and and and so that makes a a much better version
0:12:45	after precondition version of the great
0:12:47	so here are some examples i could also have shown you was real data are but here were just using
0:12:53	the ship and look and don't because
0:12:55	just the to you know compare the performance of the R so
0:12:58	we had taking a a pretty realistic you a set up here with the
0:13:03	uh for receiving corps some
0:13:05	some noise here and and uh a certain number of of a radial a trajectory
0:13:11	and where using uh a a reconstruction in the in the hard to domain
0:13:15	using a two level haar transform and so this would be the traditional is to out
0:13:20	so this chart soft thresholding
0:13:24	and so this is the cost it but it would should
0:13:26	oh of the cost function and this is the uh this is the the error with respect
0:13:32	this solution
0:13:33	actually
0:13:34	exact
0:13:34	sure
0:13:35	optimization problem
0:13:36	and so you see this kind of
0:13:38	uh a typical uh it pollution of the classical uh i i mean the typical out with is not very
0:13:44	fast
0:13:45	and then
0:13:45	if you do fist uh it's much five
0:13:48	this is
0:13:49	and square
0:13:50	type of performance
0:13:52	but now if you do this trick conditioning as as as we propose
0:13:56	in this work
0:13:57	we get more less the same slope but okay
0:14:00	we get a
0:14:00	or earlier
0:14:01	and actually in practice this makes a huge difference because we may be happy here
0:14:06	K with this kind of uh performance
0:14:10	and so it can
0:14:10	means we can
0:14:11	get there uh at that much fun
0:14:14	"'cause"
0:14:14	as the number
0:14:15	iteration
0:14:17	and and same a here so maybe just to show you some uh uh uh examples of reconstruction so here
0:14:23	have just showing you the evolution of of the solution
0:14:26	was a number of iterations so that's the T star
0:14:29	so it or too soft thresholding goes there very slowly
0:14:32	fist fist close them very a much faster and if we have those appropriate weight we will there even faster
0:14:38	so
0:14:39	she here we chose a such that a
0:14:41	uh we get yeah this quality here
0:14:44	with this number of iterations so so we can
0:14:47	a essentially the uh or or or here here's it what's okay to actually this is better quality of that
0:14:53	so
0:14:53	we can go go there for five
0:14:56	this particular
0:14:58	so why not
0:15:00	yeah
0:15:01	times
0:15:01	that's sir and
0:15:02	a a faster and and and
0:15:05	anyway all those methods that uh
0:15:08	optimising the same cost function
0:15:10	okay okay so then these second application was
0:15:13	but you wouldn't since tomography
0:15:16	and so this is a a a a much harder inverse problem of this
0:15:20	very very badly conditioned
0:15:22	and hear what you have your you your you have some uh uh by
0:15:26	hmmm hmmm this sense uh uh
0:15:28	uh
0:15:30	a Q markers within the specimen
0:15:33	and and then a you are you i have like detector
0:15:36	just the around the object of interest
0:15:39	and here i mean you you have okay optical uh optical setup but
0:15:44	i i i mean here the like this
0:15:46	just
0:15:46	if you
0:15:48	and and in fact you have the diffusion equation so you can shall solve this diffusion equation using the green
0:15:53	function
0:15:53	and this will give you a a a a a system matrix and
0:15:57	voice once to a a
0:15:59	and near
0:16:00	uh a model but
0:16:01	three need badly condition because was green function
0:16:05	so here the green functions actually look like that so it is it K like one over
0:16:09	uh a a a normal hard to the power to
0:16:12	and it happens
0:16:13	be very well condition also
0:16:15	so the idea of now is a by second call third row shot
0:16:19	was try to
0:16:21	you know use this flexibility of the precondition conditioning tried to try to
0:16:25	you mode get a much better conditioning by just multiplying this by this
0:16:29	diagonal matrix and so
0:16:30	yeah the inside that
0:16:32	if you few where a more this normal
0:16:34	uh uh this a system matrix
0:16:36	to make it like constant and energy
0:16:38	it would produce a much better condition
0:16:41	and and and so this
0:16:42	this results in
0:16:44	type of of
0:16:45	set up for
0:16:47	uh reconstruction of group
0:16:49	and in this case of so here we have also some experiments use syntactic force just to demonstrate that it
0:16:55	really works so we just have to for uh
0:16:57	a by lee
0:16:58	mean S and sources point sources
0:17:00	so here we are imposing sparse
0:17:03	state of because it's uh
0:17:05	a very small sources here
0:17:07	and here's a comparison of this three algorithms by the way also
0:17:10	uh physically realistic modeling of different optical palm which is uh
0:17:15	number of sources detectors et cetera also some noise
0:17:19	and so here it's to a bit of course
0:17:21	this slow
0:17:22	the fist act was much faster and our i'll get
0:17:25	was at the same the rate
0:17:27	i the fist are but gets there much faster and
0:17:30	in this case we have a ten fold
0:17:33	you
0:17:34	but essentially using the same out with just with this
0:17:37	clever
0:17:38	uh precondition E
0:17:39	here are uh some is also here
0:17:41	is the uh the source
0:17:44	and and so this is the evolution of the
0:17:46	standard i'll go so it has a
0:17:48	no it takes a C is very badly conditioned so it's very very low
0:17:52	it will take a long time to get there this is the fist a
0:17:56	goes faster and this is the weighted fist or
0:17:59	and and essentially here we also set it up so that you have have a correspondence your so it's a
0:18:05	bit and fine
0:18:06	or ten times is a lot
0:18:08	oh
0:18:08	for
0:18:09	image
0:18:10	so so then this makes it much more practical so that really gets me to the and here
0:18:14	so the conclusion here
0:18:16	uh in a principal we hope we can hope we get better quality
0:18:20	taking advantage of of all those methods based on on sparsity so it has been demonstrated in the field that
0:18:26	you get better
0:18:27	instruction now we can also
0:18:29	it might think speed and this is this
0:18:31	a she's
0:18:32	T Q was of
0:18:33	different next generation strategy so the multi step update the adaptive weights
0:18:38	and and so now present we we
0:18:41	with a the factor
0:18:42	so
0:18:43	or even a little less of the best the
0:18:46	you here taken
0:18:47	we which are conjugate way
0:18:49	and and so now it becomes very applicable
0:18:52	now that when used for signal processing this you like it's not like you can our
0:18:56	to requires you expert is to just find the right
0:18:59	i i that this step and conditioning so that you algorithm
0:19:03	we will work for us and finally generality
0:19:06	because it's transpose able to many
0:19:08	modalities as uh meal and was saying that X rate
0:19:11	tomography more breathy action the lab we also use a uh working face
0:19:15	thus the more graffiti
0:19:16	also or S
0:19:17	my process
0:19:18	other modality
0:19:20	so that's just you and
0:19:22	thank you for your
0:19:23	uh attention
0:19:28	i have time for maybe one quick question
0:19:36	just one if you can time i H i seven wavelet sparse in seven like total variation a their air
0:19:42	you technique can be used X a total variation a construct
0:19:45	yeah actually is uh a a a very good question which just the actually yeah i have a paper on
0:19:49	that so that hopefully should come out in journal
0:19:52	i
0:19:53	the that so also
0:19:55	uh uh there's a trick with wavelet uh a because where that's and not completely shift invariance so so
0:20:01	uh two
0:20:02	in fact if you just a use a a conventional wavelet and poke the valuation of the valuation works a
0:20:07	little but
0:20:09	but now
0:20:09	if you allow for sure
0:20:11	wavelet wavelet they something
0:20:12	that's called
0:20:13	cycle spinning
0:20:15	we show that we can get
0:20:17	as as good or even a little better than total the variation
0:20:21	and it a in its much faster uh uh to a rate
0:20:25	the way that you bed
0:20:26	the total valuation where is all this
0:20:29	a you do you are but of course people are also working on total variation and there also fast version
0:20:34	of total variation
0:20:36	but uh i i think it
0:20:38	easier to get five our written
0:20:40	uh with with
0:20:41	the the sort of uh since this this formulation rather than and that is
0:20:45	formulation but
0:20:47	uh and and
0:20:48	so the good news is this with way that we can it
0:20:50	yeah that's with that
0:20:52	oh
0:20:52	so it's
0:20:53	we have documented it with the M or i
0:20:57	it's like my so once again things

EFFICIENT IMAGE RECONSTRUCTION UNDER SPARSITY CONSTRAINTS WITH APPLICATION TO MRI AND BIOLUMINESCENCE TOMOGRAPHY

Medical Imaging

Přednášející: Michael Unser, Autoři: Matthieu Guerquin-Kern, Jean-Charles Baritaux, Michael Unser, Ecole Polytechnique Fédérale de Lausanne, Switzerland