HyperSLICE: HyperBand optimized spiral for low-latency interactive cardiac examination
Olivier Jaubert1 Javier Montalt-Tordera1 Daniel Knight1,2 Simon Arridge3 Jennifer Steeden1 Vivek Muthurangu1
&
Closed-loop control of k-space sampling via physiologic feedback for cine MRI
Francisco ContijochID 1,2*, Yuchi Han3, Srikant Kamesh Iyer4, Peter Kellman5, Gene Gualtieri6, Mark A. Elliott4, Sebastian Berisha4, Joseph H. Gorman, III7, Robert C. Gorman7, James J. Pilla4, Walter R. T. Witschey4
All right so while people are coming in maybe I will um kick us off so welcome again to the um 30 minutes CMR Journal Club um so um I’m Nicole cyber I’m from the University of Michigan and um Matias redrick is also with us who um does the
Journal Club on um alternate months um so today we are focusing on two papers that have a more technical Bend um and these are papers that um are look looking at how we can optimize acquisition um in order to make our Acquisitions more flexible and responsive to um either physiological
Motion or looking at um you know potentially interventions um so you know this is going away from the idea that we just collect our data you know irrespective of what exactly the patient or the heart is doing um but instead we really try to tailor our acquisition and
Then our reconstruction of course um to match what is going on with the particular um examination that we’re trying to do so I’m really excited that we have um two great speakers with us today um so we’re going to start with um a paper by Dr Francesco contio and then
We’re going to move to a paper presented by Dr Vic murungu um and let’s kick it off with um Dr Kio so um his paper is entitled close Clos Loop control of case space sampling via physiological feedback for C MRI so I’m going to hand it off to Francisco
Go ahead and and take it away uh great thanks Nicole um good morning I guess I’m in San Diego so it’s it’s the start of our day here uh hopefully people can see the slides um so I just took some screenshots because there’s a movie that I wanted to to show
Um but this is actually a really um cross- disciplinary effort and it actually took I think the biggest hurdle in terms of doing this was the actual deployment on the scanner and on the sequence which I think is hopefully getting better nowadays with things like
Uh gadgetron and fire so um I’ll try to touch on those points when I get there um the main kind of motivation of this was we had done earlier work using golden angle um Acquisitions for example in people with arhythmia and then we would Resort all the data based on what
Their ECG told us and realize that the data loses the nice sampling properties of the golden angle when you start mixing and matching uh golden angles and so basically the question here was can we somehow adapt in real time what angles we’re requiring to keep the sampling as
Uniform as possible as we go so you have to make decisions very quickly you’re in in sfps you’re talking about TRS in the two to three millisecond range so you don’t have a lot of time to kind of crunch the numbers and figure out what
To acquire next um and we wanted to do it in a way that the patient it was adaptive to the patients people have all sorts of different ECG patterns and and detecting QRS complexes and things like that can can be be a bit slow um right so basically here you know
Pulling from the introduction segmented golden trajectories it can be suboptimal in terms of their uniformity because once you start um having gaps or or taking Subs segments of them you’re not guaranteed to have any of the the uniform properties and so we wanted to do this on the fly in a closed loop
Meaning the system kind of controls itself um so um basically what we what we designed in 2D was an algorithm that would look uh would first look at the ECG in the past and find areas that were similar using the crosscorrelation and then it would look at the case Bas samples that were
Required at those points and then try to minimize the largest gap so that’s a fairly straightforward kind of a ro tool wherever the largest gap is in angles you can put a a sample right in between it and continuously try to uh Break um any of the large gaps into smaller ones
Um and so the cross correlation is great because it’s really fast and so here I’ll show kind of a schematic on the left you have the current position where your sampling is is shown in green and so you’re trying to figure out where is that case-based sample going to be in a
Radial acquisition you obviously have the most recent couple samples in yellow um that would be the typical single shot mode but then here we’ve identified one two 3 four prior beats that have a similar phase and we’ve gathered those samples into the kind of a case based
Trajectory here shown in the bottom and the green now is splitting the largest gap um and so basically the aspect of this on the right shows kind of the Clos Loop design is you have to have a log of what you scanned when you scanned it and
Kind of keep continuously querying it um as you go and so there’s kind of buffers that we had to Define we’re going to look let’s say in the last 10 seconds and we’re going to try to find something like five or six shots um so then uh here you just
Another example the cross correlation here is is what’s shown on on the top right and this is how we did this so we had it’s something quite like an autocorrelation where you have the signal itself with just zeros kind of filling most of it and then it’s matched
To its own pattern so uh it’s you know each patient has their own signature and so we don’t have to worry about different um ECG patterns and so we can basically find any of the peaks in the past and these would be periods where the the ECG matches and and that’s how
We defined our shop and so this is kind of the movie I wanted to show so at the beginning without the Adaptive aspect we just started the algorithm with with golden angles and so here the algorithm is is at this point in ECG it’s looking in the past and identifying these prior shots
That it wants to use and then here’s a nice example of it not being very well uni uniformly distributed but the algorithm now kind of takes control and it’s putting the red line is what it’s deciding to acquire so it’s picking the largest gap and if you kind of play this
You know the movie will beat on will start beating on the right but you can see that um as those initial angles wash out you end up with a much more uniform case space um and so the image quality gets better um I if I kind of show this
This is kind of all happening um you know in real time so every time you go to acquire a new sample you match new data you get a new data set and there are times when there’s gaps uh that kind of led to some measures of uniformity
And stuff that we had to develop basically the standard deviation of these uh Ang Les is kind of what we looked at um but you can see on the top right which is an overlay of the ECG that it nicely um finds the Peaks so that was that was really
Reassuring um and then here is just another movie where on the left hand side we have the single shot you know kind of low highly under sampled movie on the right hand side it’s the same amount of data but in a single shot mode so it’s kind of blurry and and um smooth
Motion and in the middle is the arcs kind of showing Best of Both Worlds so you have more data without really a longer footprint uh these are just some still frames kind of again this is from the paper so com back to that you can see the contractile uh motion of the heart
On the right as well as the distributions and kind of the nice um boundary definition so this was still also being done under a breathhold so that’s another aspect that we didn’t handle here but what one question is whether you can also do respiratory binning in a similar fashion would be
Kind of a nice extension of this and doing this in 3D as well I’ll skip these I just pulled them in case anybody had any questions uh and then maybe the last kind of thing we we showed was what I was mentioning earlier we came at this from an arhythmic
Perspective and so if you’re trying to now gate the scan or somehow sort this data uh you might miss QRS beats or complexes you might detect extra ones and so the cross correlation does a fairly nice job here of detecting the same point in this very complicated ECG
Such that you could do some something like multi-shot Imaging uh that depends on the person’s arhythmia if it’s frequent enough and it’s a stable pattern uh this was kind of what we were showing and so you could actually think about getting not just single shot data
But multi-shot data um in in real time so we think that has applications for interventions and as well as other kind of just um diagnostic uh uses um right so the main I know one of the key challenges here is we actually did this on the scanner so we had the
Sequence kind of waiting for an angle from another computer that was crunching the numbers and that was uh a lot of asynchronous programming to figure out uh and really I don’t recommend it essentially if you can’t uh avoid it so things like um packages like gadgetron
That allow you to kind of do those connections I think are really useful um and then right here we don’t know these views in advance so if you do a typical golden angle you can write them down as a list uh or any sort of other
Trajectory here we had no idea what the angles it was going to play out it actually makes a a really interesting noise when the scanner starts playing because you lose any sort of nice periodic nature and it sounds like the machine is is kind of grinding to a halt
Um and uh but that that’s actually kind of what and so obviously maybe uh governing that with other metrics like how do you reduce Edie currents or or kind of trying to keep the smallest angle could be another um metric to optimize um but I’ll I’ll stop there and
Maybe uh we can kind of get to this in Q&A yeah great thank you than you so much for presenting this really really interesting paper um so if you have a question feel free to drop it in the chat you can also raise your hand um or
If you’re bold you can just go ahead and unmute yourself and um ask a question to Francisco I think that would be totally fine um but maybe I will kick this off so um you know I think that like it seems like magical to me that you are
Able to in real time update the system to tell it exactly what yes this is this is the question I have like how did you make that happen like what was actually the implementation of saying please go in and collect this next angle like how did how did that operate with you know
The the pul Seance program oh you’re muted yeah so I think if I if actually explain this kind of maybe it’s easier to start on the right I think that’s where we kind of all uh we write our P sequence right and and it’s going to do some RF event that
Let’s say it takes 3 milliseconds uh the first thing to kind of do is you have to be able to insert some sort of wake up so that doesn’t just pre you know populate these views and and you know it has all the instructions ready so
Basically we had to learn which was how long you know before the next event block do you need to wake up so this would be maybe hopefully as low as possible but say it’s two milliseconds so here you know you’re doing a TR of three milliseconds you know so two
Milliseconds before the end of that TR you ask um you know the the sequence sends a wake up command to the pull sequence controller and that says to this other computer hey what angle should I play out um and so there there’s either an angle sitting there
Waiting for it or it’s calculated on the Fly and that was we had to kind of play around with with those parameters um and then it would send it back and so here the pull sequence controller maybe it waits half a millisecond if it doesn’t have an angle it plays the next Golden
Or else you get into these I think what you might imagine is either these Loops where you run out of instructions or you know any kind of lag in the system uh fails uh it does present a limitation for how complicated this calculation can be you know how many seconds in the past
Can you look the cross correlation is fast but it’s not you know instantaneous so if you take 50 seconds of data it might take too long um we had all sorts of issues not only on the networking side of things um uh in terms of which protocols you use but also using
Different computers and so this is kind of what we ended up landing on I think there’s a more streamlined implementation is that this adaptive controller sits on the same computer as the as the host that would be great or even on on the acquisition side um and
We also had to hack this ECG we were getting it from an Indo monitor not the system itself so uh there were lags kind of all over the place um I think we kind of stumbled upon an answer but I think if if you’re working on a particular
System where is your ECG and where is your sequence is kind of where I would start and that might be the the two things to to piece together all right perfect um yeah this is still magical I think that’s amazing that you can do this so fast and make
Decisions like that um so there’s a question from the audience um the question is are you using the prime ECG unit from NIH USA uh that’s a great question we were not we were basically using an invo monitor uh that’s MRI compatible that’s outside of the we were working on Seaman
So we couldn’t at the time get access to the ECG in real time on the seamen scanner so we had our the inv monitor has a plug in the back that we were able to kind of send to a a lab view uh and
And and use to do that way um I think things like Prime and and other units I think would be great um as as kind of alternatives to what we did or using the the ECG that the scanner itself is already kind of dealing with and and and
That’s a question of of Po sequence programming I think if is there another question from the audience so I will certainly jump in if not okay I’m GNA do this so um you said that your goal is ultimately to do this as um like to collect single shot
Data in real time for patients with arhythmia so can you kind of describe like what would your goal be is it to show like a nice clean image or to show kind of a real time you know or semi-real time acquisition over um you know the entire cardiac cycle of a non
Like a rhythmia laden beat or or what would you actually like to be able to present yeah that’s a great question I think part of it we struggle with in terms of kind of what’s the focus of the paper which was if you imagine acquiring
Let’s say um a c you would hold your breath bre for 12 seconds or something conventionally and you gather all the data and you hope it all works well or the breath hold gets extended uh the scan doesn’t converge here I think what we could do is you start scanning and then the
Images are slowly getting better and better and better because there’s more data being added to the buffer and then you could you know when you stop you now have let’s say a c beat of five or six beats that were found that kind of match the that mode and you can actually look
At the whole movie so whether you find a certain arhythmia at the beginning and it’s a single shot kind of typical real-time scan versus the data at the end would have more multi-shot characteristics and and you kind of are guaranteed to have an image like you would in a golden angle acquisition that
That’s useful but you also have kind of better control over the uniformity as opposed to just uh retrospectively sorting the data at the end so that would be kind of the the C breathhold I think for interventions this has slightly different flavor where it allows you to maybe get the best of both
Worlds and not have to stop and do a a multi-shot scan but to do it in kind of a live set set and that’s what we’re working on here now awesome thank you so much so um in the interest of time we’re going to move
To our second paper um so this is a another flavor of um you know more interactive scanning um this paper is going to be presented by Dr V mangu um and the paper is entitled hypers slice hyperband optimized spiral for low latency interactive cardiac examination
I’m going to hand over to Dr mango at this point oh you’re muted I think I’m so sorry after three years of you’re good now remember this sorry thank thanks so much Nicole and thanks for the invite um Olivia who’s the first author has got a
Great job in Industry now so he won’t be presenting today so I’m just jumping in to present this uh uh work so what’s hypers slice so basically what we’re interested in here is interactive cardiovascular Mr so that’s a combination of real time image acquisition which is pretty prevalent
But also on the Fly we construction with interactive scan plane control so you the operator needs to be able to move the scan plane and that image needs to be visualized almost immediately so I’ve just got a quick example for those of you who haven’t seen interactive before
This is the interactive that you get on the scanner and actually this is during a catheterization and I think this shows all the benefits and the problems in of interactive yes you can move the scan plane this is a screen capture in real time but the images are rubbish okay low resolution
Uh low spatial resolution low tempal resolution it’s fine for the sorts of catheterizations we’re doing at the moment but probably going into the future it’s not good enough so look lots and lots of cool ways of getting high resolution real time Imaging particularly with compress sensing so we’ve got lots and lots of
Compressed sensing algorithms out there with lots of very nice interact sorry lots of nice real-time reconstructions but quite long even the ones that are fast that are run on gpus are still seconds and we need to be able to see an image in tens of milliseconds and so
There in lies the problem machine learning is interesting does speed up the Reconstruction but most of the really powerful machine learning approaches have been unrolled iterative type reconstruction so they’re still taking more than 50 to 100 milliseconds to reconstruct a frame and that’s too long so as a group we’ve always been
Interested in the postprocessing type of reconstruction so what we would call Deep artifact suppression so the idea here is that we just take the grided image with all the artifact push it through a convolutional newal network and get a a sort of artifact suppressed image out the other side so it’s very
Very quick maybe not as good doesn’t have data consistency but it does the job and it’s perfectly synergistic with low latency image Imaging so we’ve done this before we did it with radial golden angle radial much like Francisco Francesco thr with a recurrent un because one of the things with deep
Artifact suppression is it benefits from temporal redundancies you’re moving the trajectories around throughout uh time and that really adds some noise-like characteristics to the artifact uh so just once again just to show you what that looks like so this is radial and you know pretty high resolution but the big problem is that
When you do a big change in the image position you get this sort of bluring artifact over a few um over a few frames as the hidden state is updated so that didn’t quite work so what hypers slice is about is changing two things the trajectory to spiral and the
Reconstruction Network to fast DVD net which I’m going to describe the thing with spiral is that you now have more degrees of freedom because you not only have changing the the angle but you can also change the variable densi of it there’s lots and lots of parameters that
You can change with a spiral so what we had to do was find the optimum spiral trajectory and simultaneously the optimum deep artifact suppression weight of our fast DVD net and that is what the hyper band does and that’s where the hyper from hypers slice comes from so we
Took several hundred cartisian data sets uh raw data sets as our training data and what we do is they are unsampled by uh a factor of two so we just do a Grappa reconstruction we now have multicoil uh fully sampled casebase data our ground truth magnitude data so
Our Target magnitude data is just simply for your transform and then coil combination with root sum of squares so that’s the what we want and then we can then essentially this is self-supervised so we’re going to take that ground prooof data and we’re going to synthetically under
Sample it with multiple types of uh uh non-cartesian trajectory on each coil and then do root sum of square so that gives us our input data just talking a little bit about the fast DVD Network itself so this is a really simple Network and it’s been demonstrated to be very powerful in
Video D noising essentially it’s just two units there’s a first set of uh uh unet blocks and there’s a second set of uh or a second unet block and they essentially all have the same structure so very classical unet this is a 2d unet and what happens is that we we need that
Temporal redundancy so what these are is basically this is the image that we want to reconstruct and then we take the four images that come before that and we sort of put three images into the first blog three images into the second block three images into the third Block it’s like a
Sliding window they go through this block and then through another d noising block and then we get our deep artifact suppressed image so it’s still a 2d network but it does benefit from the uh images that came before it so what about the hyper band because
That’s really the main bit of this the hyper band optimization is essentially it’s not it’s not really a Brute Force it’s much clever than that but you can start with a search space of parameters and the parameters we have are what is the Inner Space radius which has got a
Specific density what is the outer space radius which has a specific density what is the transition between those and do we use Golden angle or uniform angle so we’re basically in hyper band starting off with a whole bunch of different uh options in that parameter space and then
We’re continuing training the ones that are giving us the best image quality and that essentially after several days gives us the optimum or a set of optimum trajectories and the associated Optimum reconstruction we then tested all of this both in simulation and prospective and I think rather than going through
The methods I’m just going to go through to straight to the results because I think that’s where all the bang for buck is so let’s look at some of these optimized trajectories because this was not what we were expecting and it’s quite interesting so this is just uni
These all have about the same temporal resolution about 55 milliseconds uniform and you can see the grided image is pretty horrific loads and loads of aliasing artifact and then these are all the optimized ones and what you can see with all of the optimized ones is they’re heavily sampled in the middle and
They’re sparely sampled out in the outer regions and so actually our previous deep artifact suppression was really removing aliasing actually these are much less about removing aliasing and doing a super resolution reconstruction and that’s really important because that’s a slightly easier problem to do at low latency so and you can’t do this
With radial you can only do this with spiral it’s only with spiral that you can really fully sample the center and get quite nice images yes there’s a bit of artifact but much easier images for the for the unet or the fast DVD net to reconstruct and optimized one was the
Best so just to show you in simulation so one of the things that we you could see in that first image is it wasn’t very nice when you made a big change in the orientation so here’s some ground tooth simulation data we’re going from a short axis to a four
Chamber this is radial with the fast DVD net short access rubbish rubbish you’re sort of seeing a four chamber so it’s three frames before you see a four chamber that might only be 150 milliseconds but to an interventionist that’s a lifetime uh uniform spiral uh short axis short axis when it
Should be a four chamber short axis when it should be a four chamber something in between so that’s even you know even worse actually because you’re not actually seeing the right Anatomy optimized spiral short axis four chamber not a perfect four chamber but by here by the second frame in we’re getting
Really good image quality so this shows the power of optimizing your trajectory with a good Network architecture is if you compare just a simple uniform spiral with an optimized spiral much better uh image quality during those Transitions and you can see that and that’s both for fast DVD net and the
Original recurrent un but the fast DVD net is much better if you look at these ssim you get really quick up backup to high image quality um so that’s the sort of simulation word but probably what’s most interesting of course to everyone is the prospective so in a prospective we took 10
Patients we compared just regular cartisian breath old Imaging just one of the male stuff but of course high quality with uh our spiral hypers slice so reconstructed with fast DVD net but also reconstructed with storm which is a state-of-the-art compressed sensing and a spiral varnet so state-of-the-art relatively state-of-the-art non-cartesian unrolled
Ml so let me just show you some pictures of that and then I’m going to show you some videos because I think that shows you better so reference looks great this is um actually sorry this is Cartesian real time so the sort of original out of
The box out when you get your scanner that’s the real time you get it’s not bad but it’s low Tempo resolution and it’s a bit blurry this is just the simple grided which already looks okay okay that’s the point here the optimized spiral grided image looks okay storm
Actually looks quite nice has a single image varnet looks nice hypers slice looks nice but why don’t we just look at um just get the actual Right image okay so this is actually during catheterization so here we have grided image storm not bad but you know has the usual CS
Artifacts you know a little bit of Temple bluring a bit of weird spatial stuff going on vnet actually not that dissimilar to storm so very similar not not bad but not perfect and then hypers slice and it just looks great we were really impressed with how good it looks
Compared to what our state-of-the-art reconstructions um very sharp borders really nice temporal Fidelity you can see the catheter very nicely even though it wasn’t in training um so one of the other things that we were really interested in looking at here was what do it what will
It look like if you do things outside uh or distribution of training data so couple of uh things that uh we wanted to actually this is in supplemental so let me show you the video so there are two things that we’re really Keen to look at
What happens if you have loads of motion because of course our training data is breathhold it’s got no Motion in it so what happens when you have lots of motion let me just go to the right video okay here we go so what we have here is breathhold free
Breathing exercise and Peak exercise and they’re really go for it here at Peak exercise we’ve got hypers Slice versus varnet versus Storm I’ll just go back to the start so in a breath hle they all look pretty similar they’re all quite nice free breathing actually they’re all
Pretty nice maybe these are not storm is a little bit blurry but actually you know gental free breathing not bad so I think it exercise where you really see things falling apart for compress sensing and even Le maybe less so for the unrolled but but still it’s problematic and that’s probably because
It’s taking much more data in uh for reconstruction and then at Peak exercise I think hypers slice is still doing well remember this is just a single shot single frame frame by frame interactive Imaging here um and then lastly here we’ve just got different base resolution so what happens if you
Change the resolution actually it works pretty well and I just wanted to show you this is this is in action so this is a screen capture this is not a reconstruction all of this is working on the scanner the sequence is working on the scanner we have a gadget on
Reconstruction but it’s all being piped back into the scanner in real time and you can see here during the transitions I might just go back to that transition so we’re actually monitoring as we pull a catheter back through the patient you can see thing and then when you get
These big moves it almost immediately gets you good image quality so that’s what was quite exciting uh about this work it’s ready to use on the scanner the Reconstruction time per frame so the the frame time the acquisition time is 55 milliseconds the Reconstruction time is 33 Mill seconds
There’s about 19 milliseconds of other stuff going on during communication so we are reconstructing quicker than acquisition which is of course completely necessary when you’re doing low latency real time uh visualization so I think that’s my 10 minutes great thank you so much for showing us these gorgeous images um so
I’m first goingon to ask the audience if you guys have any questions again you can raise your hand drop something in the chat or go ahead and just unmute and shout it out okay in the meantime I’m going to ask a couple of questions so um as far
As I think I understood you used a cartisian data set so you had a whole bunch of cartisian case-based data and then you um like sampled that along different spirals to kind of understand you know what spirals were working well with your network and um and you
Know the spiral length so how long were your readouts for each of the spirals yes so let me just go to that yeah don’t expect you to necessarily have top of my head we we so I think the maximum we allowed was 3.8 milliseconds pretty short yeah these are
Short spirals because 1.5 ssfp so but they I think the final acquisition was less than 3.8 millisecond but we allowed it to go up to 3.8 milliseconds got it so the question I had about that I’m gonna Matias has raising his hand I’ll get to you in just a sec so um you
Didn’t like have to worry about um like off-resonance artifacts or any like fat blurring or anything because the readout was so short yeah I mean we as you know we do a lot of spiral Imaging and the way we get around all these problems is to keep the even when we’re doing
Gradient Echo we try to be less than 10 milliseconds I think that keeps a lot of these problems away from us got it okay go ahead Matas yeah thank you very much fantastic work as usual V so um I have a question which is slightly off topic but
I’m because I have been interested in the in the trajectory of Interventional Mr in in a while and there’s so much great work being done but it’s just sometimes it appears so difficult to actually get people to clinically do that where do you think are in in in a
Nutshell the most important hurdles one should uh take uh to to make it let’s say used more often which would make total sense yeah so we we so if people are interested in how to do this in a cheap and cheerful way we’re doing a
Live case at SC so on Friday so pleas do come to that but our whole philosophy is don’t buy lots of expensive stuff don’t buy a you know $60,000 fiber optic microphone so we actually do it in a very routine scanner we don’t have anything new so the first from almost
All of our first two years we purposely used only things that you could only sequences that you could buy on the scanner and only very only kit that you would have in the scanner except the screen inhouse in room screen was the only additional bit of Kit we bought and
We want it as a clinical service we do four to five Mr cats in a morning session and it without any without any clever stuff when we have the clever stuff then we can do other things so I think I think and I’ve been in the intervention game for quite a long time
I think we made it look too hard because we had all these XMR Suites and we spent loads of money and we said well you can do actually what we needed to do is make it look easy and and actually it doesn’t have to be too complicated as long as
You’re not trying to do something very complicated like interventions we’re just doing right heart casts and we truly believe for right heart cat if you have an MR scanner and an interventionist who will push the uh catheter you’re good to go um so I I think partly it’s our fault I think we
Made it look difficult thank you very much uh it’s very good to hear from your side thank you all right if there are no additional questions from the audience then we are going to wrap up our CMR Journal club for today um two plugs so one um we have
A journal Club um our next one is on February 7th I believe um that Matias will be running um and oh go ahead Mattias if you have something to to say about that one I just wanted to add the details have to be but it will be about my
Cardial in fortun it will be about clinical uh role of certain prognosticators that are derived from CMR uh so there will be two papers to moved on that details to follow awesome um and the second thing I want to say is that we are not affiliated with scmr in
Any way but I think a lot of us will be at the scmr and so um I think you know I speak for myself and and Matias I hope that um you know if you’re part of this CMR Journal club and you’re at the semr please come say hello because we love to
Hear from people who are involved in the journal Club um so yeah definitely we hope to see you all there um and at you know anytime you can drop us a line but this is a special place where we’re the in person can say hi yeah and if you enjoy the journal
Club spread the word um uh more input and also we’re also always welcoming um ideas and comments on uh how we can make this even better thanks and bye everybody bye guys bye bye thank you