it nice but the yeah speech can kind of thing but choose by see image and signal processing and we have here and for as the have this can who twenty in oh and that are the hot topics and two oh so as to the community at large of signal processing so um a for the yeah i and that the uh about a gene is to be a a and first that which is a the perspective of the scene or the best G to understand the can the kind for five uh_huh hmmm a i like the and the prediction some genes and proteins and a is to be able to it is is the conditions so and the prior to choose but G two the band and the size of a down to the size of a the the content was a scale and by different that from for example a pad this is medical imaging the the the some of P be to go just to be she's which i produce a information which just produced we have a number of topics which i uh i have uh come to the two and discipline which i a to to this process and and then said good friend mathematical gym which is how can put some i think so that's processing i uh uh uh i information a set uh but and then to to of us steps a first processing knows those problems can can and then to a a a a time of just first we choose to put into a just uh a just a images which type of crime so that the the fact to extract information from the background the relevant information to the classification of information and the to be able to put in in um true in this manner under images are kind of in this manner as and and uh i and in and then no side then i hi this i just to uh for size and the fact that for methods and for the processing session uh for the to look a can to the to the information to choose so uh uh uh a type is for example compressed sensing is becoming a of research but had to go because the the filter i to to spend to much time this and a a which is how does from it it's also about uh and and oh um type of research okay uh so so i could use a set have two so that was which i X uh yeah but to that to test uh i a uh a a uh and each of them so so as to have this was um a process right which the channel have to the go to a a a uh of the um a as the beaches so yeah uh i use and try and okay can give some idea is also at could me find find G Q and uh_huh i'm can some stuff so that's think that's that uh so are so T R screen D that's as far as E and a go to note some access from shown and information and uh so and my first slide okay and T because i asked to at some found notion and that function as a cat and so i don't you can still so i have um i have a sounds have stop to learn and then sit products and i've that's that's you know and and so i that so that so that asked uh_huh asked and so and then suppressed so um and that X M M M and that's consistent that's as a so that's that's so five and so and at the i-th sensor and this is yeah that sounds i a uh thanks to that construction five so so i don't function and sabrina and uh i know that have to reconstruct i have i yeah the presence of a so here and that's that's faq so that's accent i have to say that the construction masters channel i uh i don't i have a pension plan oh doesn't that so um the acquisition system and and twenty seven are expressed as a function and at times i was of uh uh optimization is convex and how a a a a a a and uh some kind of and uh or section i you a and have and it's fast but was that because we've got a a a a a a three D the mountains to kind press sound system so it's down so a five a script and oh oh can script yeah that's a um uh i i stands as a facts phones i have some so and and that's a that's it's function how that that that's facts phone so so that yeah the and that and a and the convolution "'cause" of times perhaps function as as nine that's a concepts and but a she used on the point spread function so that an absent system a so account the mouth i have just stash um a a i have a complex to that so that the structure i and uh that's a constant uh i yeah i was just a concept for example of god construct estimate costs for five example some kind of cats watson's mention that uh as a last step to for some sequences come out match well um i i hello yeah and so on have come to that to the concept of a complex when the structure that's that's a problem oh have uh statistical stuff such a complex task and five is not uh to define approach to insist substance and not rocks and uh have questions that is the kind of news and most of them to a convent but sectors in some sense yeah we propose is that we uh_huh go through an addition that the have ten minutes for discussion open discussion on all the the the topic so yeah also go on was biological imaging and perhaps web addresses a little for being very focused some biological imaging because this was also andrew was the supposed to to about medical imaging that in uh uh in through a right now i so yeah uh hmmm and it is so was so as so he of the challenge signal processing challenges for microscope okay and uh uh it has to do with the fact that this three D uh plus the and the fast so that was really thought telling you about the convolution know as it i mean as a a a uh than men's the i mean if they can be that's a can be then because um P T satisfactory this factor to to have a a a a a a lot of the challenge is and you have the set of image analysis because this is a a and the by just to have numbers and the so the a set so the to uh uh my west image shows which have to do with because because tracking so this self worse more those in images looking that's that was also a i have the show if the gene expression for a has and and you know some the some of the problem of that that to the a a a a a and to for for image processing so i i i i i will actually this is the sort of give you some example state of the art and taking a a uh algorithms that have been developed a people in a common T that she i been used by religious and i mean this is this that of the to the the challenges that i to go on the and i with sure but for example here okay so if if this you see here you have this is an example of a uh uh you know time laughs my first P where uh uh actually this is a a ms so so uh uh uh the end of the proposed um that has been the the the uh is inside the nucleus and have the are just like to to is month so for of those from more and so that was a better here is the tracking i uh that will actually five true in so i'm and and i is to do that if are less fashion using oh to defined as function using the fast and i mean this is uh just works very well for a single party but the i'll the challenge is uh you may have this type of okay so you have yeah yeah of those party oh have of for and the background and so uh a real challenge design for for for for that data so it's you know you'll three a uh so next challenge uh so it has to do with self uh and uh uh so uh shape and you need it so he it is simple cells E so the violent adjust here this is the uh face got five my cross is so yeah i i is to outline of cells not just for counting them but because we want to extract the gene expression profile it's so this is also like can of it the ball of that's by by all will kind of like hmmm a fact let's images and then if you uh uh like that of oh you're so that what what you have principle here the so what's important here it's for all those cells with time so why would you want to do that because all cells actually what right thing a different genes your that'd been through recently label and and so here you you you know the outline of this the the cell but what you would like to quantify is the expression oh uh i mean the amount of fluorescence in in in both cells here and and so get that uh time course of of fluorescence and and and and and uh this this will be the date so that the biologist one two you we uh i mean to extract so what a here uh the challenges i mean the challenges is very poor contrast of those images dealing with much more complicated it's uh shapes than those used cells which are round and things cell division it should using models of time evolution uh and doing global optimization in space and time dealing with crowded images touching cells in to reducing repelling forces high throughput puts mean this needs to be done on you amounts of cells and power that imaging fast row put able and the uh above all easy to use a gram so that people uh i really applying the mean brock so now uh relating more to what to a lot was telling you uh uh you know there this a problem of extracting features from images and and so for example this was also a program developed it's the it's a program for tracing uh acts on of of neurons but at some your to me and and not i mean i mean this works well it's uh use by by people but uh uh uh i of course uh i mean life is is more complicated than that and what i see as a challenge an opportunity uh a of uh for for for a signal processing community actually coming up with uh what that would call key point for by you because key points have the huge six S for computer vision so is looking like points of interest in image so what we would need a by all we are not like we have three D data so we would not need for well we can first look at two D but optimize a for scale uh a translation rotation invariance also i mean what we're seeing we have background verses uh a structures of interest so doing detector that we might seem i uh suppress press background and probably use this kind of idea for designing wavelet type representation that can and has the feature of interest and and and and suppress the background but the real problems in three D okay because there's nothing in three oh O case to two minutes okay and and and and so i that she do that in three you so in three you have a interesting structure like sheets membranes and he the challenges developing that's a durable the vectors wavelet since really and uh above all be computationally efficient uh uh because i mean the are huge amounts of data so i just wanted to uh here are uh i i mean this is perhaps what where banning in the lab but just to show you that uh one can do a very interesting steerable wavelets so we here are i'm in in in in two D that look like has since but there a reversible uh where let's but the this seems also to be a possibility to do them in three D and this is uh pretty much uncharted so doing where it's that can what's agents and i yeah side like just like to have a size if you're working in biology so that it's a disciplinary research where there are lots of players of course the to just the medical people there's the optics the microscopy the by chemistry which is the mark and the signal processing and there were there was also like here special issue on biological imaging a few years ago but is still a good entry point for those want to get in uh in with few okay case of X yeah as so uh i will and the law speaker go a bit more into challenges that are you merging uh especially in the medical field and moving wait a little bit from the biological um so basically number one chance number one is still increasing data that now already already for treat this is going on uh the the simplest by medical signal is the electrocardiogram the second two dimensional signal is the the normal two dimensional picture then uh in the eighties came to three dimensional pictures and then the four dimensional pictures being a three dimensional close time uh more recently especially in the the last decades uh specially functional imaging on top of and apple imaging has become very cool in that case you three spatial dimensions one time dimension one functional entity for instance uh blood oxygen nation um three plus three dimensions basically meeting three a a spatial dimensions a three direction vectors uh base uh diffusion tensor imaging then there's uh nowadays very popular uh it's starting to really get popular three plus three plus one and dynamic uh imaging for where you can actually a monitor or you know that the a three dimensional flow inside the ha and especially signal processing for this is still in its infancy this sense that able try to characterise a for text flow right he's uh in he's types of images but then when you have a i thought we kind of reached a limit at uh seven dimensions here but then uh this uh and and brain database was published and a brain database is basically three dimensional data dataset plus at time aspects where it basically for a my of mice where for every voxel they have twenty thousand G which uh values for gene expression show basically but see this as it you which uh uh dimensional data set a a which is now uh just about to be explored to correlate it's uh with all kinds of for imaging so this is number one the dimensionality of the data is still increasing rapidly and signal processing challenge are the uh plentiful there seconds is that um imaging uh until about ten years ago was mostly about imaging structure and it to me with ct anymore yes and function with for instance uh a nuclear technique it would highlight some aspect function and past decade all kinds of technique became available to also look at biochemistry and so now we can look at diseases on a biochemical level and then see what happens first in biochemistry uh what goes wrong in a particular disease by chemically and then study effects all structure and function now the data acquisition is now getting so far that we can acquire these triangle structure function and by chemistry but the signal processing and image processing is still largely to we think this on as separate entities and i i think there still a big challenge in trying to come up with in to great if signal processing it actually looks at all data at the same time there at very important trends uh i i think shows also that the few of by a middle image and that was is is you ring at at this point in time used to be the case that you could get away with a validation of a segmentation algorithm for instance just by citing some uh uh uh accuracies that others have achieved and then uh a benchmark your algorithm to that but the problem with that is that all those those uh results published in these papers they are usually done different patient different test state that different test protocols different error metrics different gold standard everything is if what you see and this make it very difficult to really make an object is mark and especially in the pattern recognition community there's already a lot of standardised databases we known classification result we see now in the medical image processing field more and more trends words we objective quantification they make fable a centralized data evaluation scripts are run sense really and everybody can benchmark their data their algorithm against that particular that's it i i think this trendy so input and that nowadays if you try to publish uh and algorithm on a topic which has already been that smart in one of these challenge a you cannot get away without uh benchmarking marking it against it and there's kind of a default rejection if you do not include validation experiments on these standardised uh um data bases so very important field of of of yeah it shows that the the the the field is but your in a that's for if very important new topic uh is a key to more data analysis right the main question to answer is something is changing but what now and you can imagine that if you look at an in fit individual patient uh yeah it still manageable uh you look at follow up uh is easy is getting worse yes or no uh but nowadays one tends to look at groups of patients anyway between a two and one thousand that this study D V disease each development in two weeks writes treatment affect but even more data is acquired normal subjects there and now you huge population studies right there every year scanning people healthy people and they're is getting every here and then basically wait until people get sick at some point in time and then they can backtrack through the date that whether they can find early signs of that you can imagine that that finding these are signs is really a needle in a a a a stick and uh well there's a huge uh image processing challenges trying to really combine all that data and to backtrack to bit data um well another trend to actually here is uh a fact that as that already showed that that these biochemistry a chemistry became available uh i think there is a there's a lot to gain in the combination of chemistry tree and specified as signal sensor specific signal sensors that for can look at multi spectral imaging as we can see here and and let's say one of the applications that we are working on a second probe a specific for a probe which is injected for four it to more a and that uh basically that broke can be used to identify this the lymphatic uh system and then it's very easy for and some parts of particular search you really need to remove lymph nodes and this is only possible because of a very delicate balance between and so point acquisition device and the probes that are being used in here you can actually see that's removing the lymph node becomes very easy using these combines a in in finally uh especially uh uh yeah recently there there is immersed a lot of attention for personalised integrative modelling modelling and simulation um perhaps some of you have heard of the virtual physiological human project well especially you this has been a whole public for the past few years and more and more this to these uh very integrated models go from the cell like organism they start to percolate we into clinical practise uh yeah we're actually they use this this uh uh uh patient specific data tuned to mobile to the observations and from there um they can actually do predictions and simulations so that bout wraps up my uh my uh overview of uh what i perceive as very important to trends in i think we need them both basic uh i think of some i L some algorithms uh are so generic that they can basically for for for a lot but i think we also need dedicated times algorithms of four particular application perhaps you i actually my my advice there is uh for you know for those are not been working too much uh in in the body oh area my advice what would be a good start working on the specific algorithm because you look at the very big challenge you know the ultimate segmentation algorithm you'll have to compete against the all those benchmarks that that are out there so if you if you're working on the specific yeah a problem so you'll have people very happy at the other end and and and and so it will really help you also to learn what what the issues and of course i mean if along the way you happen to stumble on something really general well then you go also for for for the more general albert but i think specific is good there has been it's went let's say in the nineties to look for this uh big general super segmentation framework for instance that could be applied just to everything and now i see kind of more dedication again to single applications so that and and only some of those very generic uh segmentation framework yeah they're they're still uh actively researched yeah i mean yeah in most of areas like segmentation i mean the generated stuff is already in that i well maybe not all music you're very lucky you'll find your generic i'll uh maybe uh yeah since we don't have many many uh questions uh so we i guess in our presentation we did it's just too much on the inverse problems because we thought every someone else would talk about "'cause" the you will compress and you or uh a those L one type of to as a a pretty much in the focus of of the signal processing community and and the a search this stuff with say actually those a a a a a a very hot topic in in in imaging in general so essentially essentially people are read is it in all the classical uh uh uh image reconstruction algorithms the you for "'em" C T uh in in the medical imaging area a the same uh for for for for different a could be source so that's a and and and not just the the traditional uh but that each is that also like lots of people and optics so uh designing all always new new a new modalities and then you mode that it is uh ten to work very closely with signal processing because traditionally for example be plenoptic optics they would just want to to to to do a microscope so that you see an image but now uh uh once you have signal processing you can do lot more like a tomography start a acquire measurements that you may not necessarily see but once the goes through a signal processing group then that the can can we construct images hopefully using that's measurements and so forth and so this all obviously they hot so if you go to medical imaging call you have lots of compressed sensing people the same on on the microscope beside so so that's so uh a there are yeah where the problems up pretty you well be fine but but then uh the guys also a very interested in a quantitative measure so suisse showed in our example just to just to elaborate on a michael was saying is that but but for than them by your logic logical um can the goal has always been to do a nice image and nice image as you may know is probably don't the base one in terms of uh optimising devising the the rate of acquisition the rate of uh information it's very but the opposite i i actually if you a through signal processing are able to extract the meeting for information even if there is a lot of bad run or clutter to means that you can either improve the the timing of your acquisition you can expose more time you're your sample or you can uh also spend more time on doing um different ranges of uh wavelength so you can have a more dense a field of acquisition if we were able to reduce the time that you exposing example for each of the the top in this is something that only signal processing related method can bring to the community so there is a whole range of application enough possible uh topics of priesthood for these comedians icassp if only you are interested to i mean uh the other with uh some challenging but also interesting problem not to mention the con the also the the fact that we are more immoral maybe use of the plantation there is some kind of a coming together of uh medical imaging in biological imaging and they are really merging in one of the key uh but all that is is molecular imaging which is in the the use of medical imaging techniques to have resolution which is similar to uh microscopy uh imaging in there there's also a lot of challenges still opened for the crib well there are no more question i think needs O uh i i just saw on the medical imaging side uh maybe not everyone will agree with me but uh it's pretty much your because uh i mean the big revolution happened about thirty years ago uh invention of summarise so i mean M R I still going very very strong and always getting better stronger magnets et cetera uh a new type of modality is drink different kind of measurements but it's pretty sophisticated now what's happening on the other hand on the biological side we are more or less experiencing this state of thirty years ago in medical imaging because they're all kinds of new modalities actually coming out almost every year and for example in microscopy they've been able to beat the the relay team it's by a factor of one hundred going you know like below what physics still but but by using some tricks of course and and they're like a novel microscopy is being developed the current the and very often hand in hand with with signal processing so so that's very important actually the the other thing we could even say it's all the medical imaging couldn't exist without signal processing because a mirror i the first thing it is it's the fourier transform and and and the so uh so i see that uh there there i mean there's improvement in in modality especially uh going out to find the resolution uh and and very much happening now in in the area of biology but i suppose also it equal so i don't know if few and this all of this small animal imaging yeah or uh uh so so where we getting always uh need for higher resolution source so it's very very act oh i i i mean this is very very hard because of for example with my cross could be a good you use of your essence and there's lots of uh a uh you know like naturally fluorescent so the stuff inside you used specimen so usually i i mean i mean this uh this will be a uh uh you will interact with some biologists to do the the lot labelling and and and i mean it's never a very simple but of course i mean the goal a goal of biochemistry is is finding mark is that they extremely selective and and also with this molecule image in principle i mean they're very good marketers soak is just sometimes the the the resolution is it's so great yeah so uh so we like with but um they very good mark but very poor was should with a by you them in this since very good mark "'cause" but that terrible resolution and and K i think we have to close the session the "'cause" the next uh station is a starting and uh i think you all for your or of coming here and uh hopefully will see you in some more dedicated uh conference is of stations to uh imaging thank you