AI and Image Reconstruction, Dr. Lawrence Tanenbaum (5-29-20)
HIDEInteractive Transcript
0:02
Hello and welcome to Noon
0:03
Conferences hosted by MRI Online.
0:05
In response to the changes happening around
0:06
the world right now and the shutting down of in-
0:08
person events, we have decided to provide free daily
0:10
Noon Conferences to all radiologists worldwide.
0:13
Today, we are joined by Dr. Lawrence Tannenbaum.
0:15
He is currently the Vice President, Chief Technology
0:17
Officer, and Director of Advanced Imaging at RAD.
0:21
He previously worked at the Icahn School of
0:23
Medicine at Mount Sinai, New York, and attended
0:25
in neuroradiology while serving as an Associate
0:28
Professor of Radiology, Director of MRI, CT,
0:31
and Outpatient Advanced Imaging Development.
0:32
He's a senior member of the American
0:34
Society of Neuroradiology, and we are
0:36
excited to have him here with us today.
0:38
Reminder that there will be time at the
0:39
end of this hour for a Q&A session.
0:41
Please use that Q&A feature
0:42
to ask all of your questions.
0:44
We'll get to as many as we can before our time is up.
0:46
That being said, thank you so
0:47
much for joining us today, Dr.
0:48
Tannenbaum.
0:48
I will let you take it from here.
0:50
All right, great.
0:51
So we have the opportunity to talk about
0:53
artificial intelligence and image
0:55
reconstruction over the next 40 minutes or so.
0:59
I'm trying to make my slides advance.
1:01
That didn't work.
1:03
That didn't work.
1:06
Maybe.
1:06
There you go.
1:07
So we'll talk a little bit about,
1:09
um, what is artificial intelligence.
1:11
We'll do a little bit of the hype and
1:13
reality that you've heard around this topic.
1:16
We'll spend some time on terminology and
1:19
background, very little in this particular session. And
1:22
we'll talk a little bit about the strengths
1:24
and limitations of the tool in general.
1:27
This talk is largely about the impact that
1:29
these tools can have in terms of influencing,
1:33
uh, image reconstruction for MR and CT,
1:36
uh, reducing dose and improving workflow,
1:40
as well as the patient experience.
1:42
So we'll take it from there.
1:44
Um, I promised you a little bit of definition.
1:47
So what is artificial intelligence?
1:49
It's a branch of computer science, which essentially
1:52
allows machines to do the things that are normally
1:54
associated with human intelligence, things
1:57
like reasoning, learning, and self-improvement.
1:59
It's essentially technology that, in theory, would enable
2:02
machines to sense, comprehend, act, and learn, and you can
2:05
see we're not quite, you know, at the full fruition of
2:09
these goals, but you're starting to see some of the
2:11
things that these machines can do in healthcare. You can
2:14
see it is considered to be one of the key technologies.
2:18
You can see here between big, big data, which is
2:21
essentially an opportunity for AI, um, pattern
2:25
recognition, as well as, you know, the subtopic
2:27
of AI, you know, number one and number two on the
2:29
list of the key technologies to affect healthcare.
2:33
Back in 2018, about 50 percent of healthcare
2:36
organizations were already either involved or
2:39
planning to get involved with artificial intelligence.
2:41
So at this point, I'm not telling you anything
2:44
you don't already know.
2:46
It's a big topic, and you know that it's all around us,
2:49
you know, whether it's ads that follow me around every
2:52
website I hit or, um, crowdsourcing traffic information
2:56
on things like Waze. Um, it's really in every aspect
3:00
of our life—some form of artificial intelligence now.
3:05
I'm sure you've heard of some of the great victories
3:07
that artificial intelligence has had here.
3:10
You see an example of the poor Go champion who was
3:14
beaten by a machine, and frankly, this gave chills
3:17
to all of us and, you know, made us marvel at the
3:19
capabilities of these tools. But in reality, playing
3:22
games is one of the lowest hanging fruits for an
3:25
AI algorithm, for a machine learning algorithm, and it
3:28
frankly is really sort of an entry-level achievement.
3:31
As you'll hear over the course of the presentation, we
3:34
know we have autonomous cars out there, even autonomous
3:37
trucks with a variety of tools enhanced by computer
3:42
vision that assist the operator, uh, uh, in terms
3:47
of, you know, getting around, avoiding accidents,
3:50
and, uh, even, even, um, taking you from A to B,
3:54
you know, while you're doing the crossword puzzle.
3:57
We know, uh, from the literature that, uh, the
4:00
error rate for computers is already better than the
4:03
error rate for humans in terms of image recognition.
4:06
Uh, and it's really the same
4:07
in terms of speech recognition.
4:09
So these tools are really working reasonably well
4:12
so far in terms of these tasks. On the other hand,
4:15
there have been some fairly flamboyant failures. We
4:17
know that none of these autonomous systems is really
4:20
quite ready to be fully autonomous, and accidents have
4:23
happened. And there's certainly a role for humans in
4:26
all of these processes. You know, it's said very
4:29
often, you know, planes can fly themselves, but nobody
4:32
will get on a plane without a pilot. These days,
4:35
without a pilot with a mask on, uh, you can see this
4:38
particular circumstance just outside my office window,
4:41
uh, that, you know, Sully had to take this plane
4:44
down, override the autopilot, and land it safely in the
4:47
Hudson, um, uh, not too long ago. So really, there is
4:52
quite a long way to go in terms of what artificial
4:55
intelligence, you know, will be able to do and can do.
4:59
Uh, you've all seen robots, but with minor challenges,
5:01
they'll fall down. Uh, we talked about driverless
5:04
cars, and, you know, even though computers can do a
5:07
relatively decent job reading, at present they're
5:11
pretty much stuck at the sixth-grade reading level.
5:14
Um, and, uh, you know, not where you would expect, you
5:19
know, for something as powerful as a tool that can beat
5:22
the experts and drive a car down the highway. Really,
5:25
one of the biggest challenges is perspective in
5:28
context, and it really has a very hard time with complex
5:32
visual scenes. See, this picture strikes us all as odd.
5:35
Um, uh, you know, it just looks a little weird.
5:39
Uh, certainly we can recognize a rock, we can
5:41
recognize a tree, but you know, we, uh,
5:45
there's something odd about the picture to
5:47
us. The computer doesn't see anything odd here.
5:49
And in actuality, this is a picture of a rock floating
5:52
in a pond, and you're looking at reflections, and we
5:55
instantly interpret that in a way that computers can't.
5:58
And that really is one of the
6:00
big contextual differences.
6:02
When you see this picture on the left of a bicycle, you
6:04
know, it's easy to recognize a bicycle. For a computer,
6:07
it can recognize grass and sky, but
6:09
it doesn't recognize why that bicycle
6:12
is upside down and has its wheels
6:15
on the grass.
6:16
If you look at the right-hand side of your
6:18
screen, you can see, if you have to move my little
6:20
picture over, you can see there's a duck on a
6:23
roof here watching TV. It recognizes the TV
6:26
and the duck, but it doesn't really understand
6:28
why it's kind of odd to have a television
6:31
on a roof with a duck watching TV.
6:34
Uh, here's a cow on a lounge chair. Again,
6:37
the context is where it really, it really
6:39
starts to fail. You know, what's odd about
6:41
these pictures? Uh, that's a toothbrush.
6:43
That's a dog, but why is it odd that the
6:45
toothbrush is the size of the dog, and
6:47
what's a dog doing using a toothbrush?
6:50
So those contextual pieces
6:53
really are where things fail.
6:56
Now, computer vision works pretty well, right?
6:58
These are examples of ice cream and Dalmatians,
7:00
and, you know, as good a job as it can do
7:04
in differentiating ice cream from dogs,
7:06
sometimes it makes fairly flamboyant failures.
7:08
It does a nice job
7:10
of telling that this is a chihuahua, but every once in
7:12
a while, you know, it thinks a blueberry muffin is a
7:14
chihuahua. So there are definite limitations to these
7:17
tools, and rather than a lot of areas in society, you
7:20
know, we're pushing back now. Um, and facial recognition
7:24
software, because of all the errors that it makes and
7:26
the implications this can have for personal freedom
7:29
as well. So it's powerful, but still relatively early.
7:34
When you look at artificial intelligence out in the
7:37
enterprises, human review seems to still be essential.
7:40
I love to use this picture on the right-hand side.
7:42
This is on posters all over Los Angeles to herald
7:46
the grant that was given to one of the universities.
7:48
Um, but in this particular case, somebody wasn't
7:52
paying attention because the brain is in there
7:54
backwards. So a human would recognize that instantly,
7:57
whereas a computer might not ever recognize
8:00
the fact that there's something wrong with this
8:02
picture contextually, so to speak. Despite the fact
8:05
that AI systems are running at Facebook and Google
8:07
and Twitter, there are lots of humans involved in
8:09
this process to add the human touch and the human
8:12
perspective. Uh, and if you see the same thing
8:15
at the five largest financial institutions as well,
8:18
there are lots of humans involved there, and
8:20
uh, you know, it's not computers alone
8:23
that are doing this censorship in China.
8:24
They have one human censor interpreting
8:27
context and nuance for every 100,000 users.
8:31
So humans are still a big part of this process.
8:34
Overall, if you want to get a sense of what AI
8:37
can do, what machine learning and neural networks
8:39
can do, I think this particular statement by Steven
8:42
Pinker really puts it into perspective: hard problems,
8:45
you know, beating a champion in Go is very easy for
8:48
deep learning, for AI to be able to do. Uh, easy
8:52
problems, like picking up a pencil, recognizing your
8:55
mother, uh, that's a lot harder for computers to
8:59
do, and that really is a good tenet to keep in mind
9:01
when you think about, uh, where machines are and where
9:04
they still have to go. Now, when it comes back to the
9:08
imaging enterprise, which we all care about, we've
9:11
heard some pretty alarming statements back in 2016.
9:14
Jeffrey Hinton, uh, who is the godfather of neural
9:18
networks, the godfather of machine learning,
9:20
um, he came out with a very widely publicized
9:24
statement that radiology is over. You are over,
9:29
you should stop training them right away.
9:30
I mean, they're just like the coyote who's
9:33
already run off the cliff and doesn't
9:35
understand that there's no ground beneath him.
9:37
And you just have to look down to
9:39
realize that, you know, their career is over.
9:42
And the question is, you know, is he right?
9:44
Are the doomsayers right?
9:46
Or is it really more
9:48
that, you know, as we're all starting to realize
9:51
that these tools are going to be, uh, significant
9:55
tools as augmentation, as assistance, as workflow, as
10:00
reassurance, as triage, all kinds of things that these
10:04
tools are going to be able to do for us that are
10:06
going to make us more powerful in the healthcare
10:08
enterprise, as radiologists, as brokers of information,
10:12
uh, as integrators of information, uh, than we've ever
10:16
been before. And that's the opinion, you know,
10:18
that I have right now. If you look at some of the
10:19
statistics about, um, startups in the AI space and
10:25
job loss in radiology, on the left-hand side you
10:27
see the roughly 400 global startups in artificial
10:30
intelligence, and on the right-hand side, the zero
10:32
radiologists have lost their jobs as a result.
10:35
So it may not stay that way forever.
10:37
I think in any industry, you, uh, people have
10:40
to adjust and adapt, but I think this is going
10:43
to be the bird's eye view or the 500-yard
10:45
view for quite some time, if not forever.
10:47
As a matter of fact, I think we, if you look at
10:50
the Gartner Hype Curve in terms of artificial
10:53
intelligence tools, machine learning, deep
10:56
learning, computer vision, computer diagnosis, and
10:59
the like, you can see that we're already getting
11:02
heading towards the trough of disillusionment
11:05
with deep learning of image interpretation, and
11:07
we're just starting to climb the high plateau in
11:10
terms of reading and reading comprehension and
11:13
speech comprehension, uh, and the thing that we're
11:16
we all assume can be done today, which is deep
11:19
understanding, is decades away, and may never actually
11:22
uh, come to fruition, so really interesting stuff.
11:26
So what have we learned over the last four years or so?
11:28
Maybe, you know, just about four years.
11:30
Um, uh, well, what does Geoffrey
11:33
Hinton learn? Well, in 2018,
11:35
he did a deeper dive on what it really takes to replace
11:38
a radiologist, what a radiologist really does in
11:41
the healthcare enterprise, and he realized, well, he
11:43
was a bit, if not a lot, too hasty at that particular
11:46
point. And that brings us to, you know, where are we?
11:49
Artificial intelligence can impact
11:52
the radiology enterprise.
11:54
And this slide was borrowed from
11:56
Dr. Hugh Harvey, that's his Twitter handle.
11:58
If you want to get a feel for what's actually
12:00
going on from one of the wisest, uh, and loquacious
12:04
folks in artificial intelligence and imaging,
12:08
I would follow him at his Twitter handle, Dr. Hugh Harvey.
12:11
Uh, really impressive guy.
12:13
But this slide, which I've seen in various
12:15
shapes and forms, shows the full sense of
12:18
what AI is going to be able to do for us.
12:20
Everything from improving our clinical decision
12:23
support tools, as to which exams are best used in
12:26
which circumstances, optimizing our scheduling.
12:29
Our organization is in the middle of a big
12:32
joint project, project with, with Philips
12:35
Healthcare, looking at tools to optimize our.
12:39
Late cancellation and no-show patients to find ways
12:42
to devote our efforts to reducing those in the most
12:45
efficient fashion. We're going to talk about things
12:49
like dose reduction and scan time reduction today.
12:52
But other things like post-processing segmentation is
12:55
really low-hanging fruit For um artificial intelligence
12:58
probably much of what you've heard is about triage
13:01
and detection Um, particularly with uh, COVID lung
13:04
imaging right now the more impressive apps that I've
13:07
seen are the ones that do quantification Um, I work
13:11
with a company called Icometrics, which has a really
13:13
nice quantification tool, which can give you a sense
13:16
of, uh, of prognosis and perhaps guide intervention.
13:20
Um, we've also talked about
13:22
other things like, um, reporting.
13:25
Uh, that can be optimized and actually tailored to,
13:28
uh, the, the, uh, the audience, maybe giving one report
13:32
and creating many, and even helping with communication
13:36
of urgent results, with lots of opportunities to take
13:39
the massive amounts of data, the, the daunting tasks
13:43
that we do in a repetitive, often mind-numbing way,
13:46
and optimize them and take some of the burden off
13:48
of us. But again, for the rest of our time today,
13:52
we're going to talk about how you can use artificial
13:54
intelligence for image reconstruction. We'll start off
13:57
with sort of the entry-level stuff with compressed
14:00
sensing and sparse data reconstructions. We'll
14:03
transition into iterative reconstruction for the
14:05
most part, uh, an MR-based machine learning-assisted
14:09
iterative reconstruction. Um, we'll highlight the
14:12
work of a company called Medic Vision. And then we'll
14:15
get into the true deep learning neural network-based
14:18
reconstruction that is now on the market, from
14:20
a couple of the big OEMs, as well as some of the
14:23
independent companies attacking newer systems, as well
14:26
as older systems, and essentially changing the
14:30
scan time/patient comfort equation in terms of MR
14:34
imaging. But let's start off with compressed sensing
14:37
and, you know, with typical, uh, spin echo imaging, uh, MRI.
14:41
We've got to count out each line in k-space
14:43
for each TR period. I think this can get very,
14:45
very long. I remember the old days when we tried
14:47
to do, you know, full-resolution fast
14:50
spin echo, full-resolution spin echo brain, it
14:53
would take 15, 17 minutes to get a dual-echo T2.
14:57
Over the years, we've come up with various forms of
14:59
undersampling, from half Fourier transforms to parallel
15:03
imaging with things like SENSE and, uh, um, multi-
15:07
directional SENSE, uh, like ARC and GRAPPA. Uh, the
15:11
pinnacle of where we're heading in this particular
15:13
space right now is compressed sensing, where we're
15:16
doing these pseudo-randomized but balanced sampling
15:19
of k-space, uh, really reducing the number of samples
15:22
we have to get, strikingly, to markedly accelerate.
15:26
Our actual scan times here.
15:28
You can see an application in routine practice.
15:31
We, in our network,
15:32
we have some 340 centers, somewhere in the vicinity
15:36
of 300 magnets. We probably have a hundred magnets or
15:40
so that are running the latest in terms of compressed
15:42
sensing from different manufacturers. And on all
15:45
of them, we're running something like this, where we
15:47
take a 3D mathematical acquisition and run it
15:51
faster and sharper. You, in this particular example,
15:55
we're actually using restricted field of view.
15:57
This is an example from a GE magnet.
15:59
Uh, and you see we get all this done in a
16:01
faster and sharper way by doing much more creative
16:04
sampling and faster scans. In this case, it gives us
16:07
just a little less artifact and actually even
16:09
the opportunity to raise the spatial resolution
16:12
of these studies, which is something
16:14
you should always bear in the back of your mind.
16:15
It isn't always about speed. Sometimes it's about
16:19
better quality at some of the speed trade-off
16:22
with better speed as well. As a matter of fact, some
16:24
of the manufacturers have actually gone to doing
16:26
things that I would best describe as making a
16:28
1.5T work like a 3T, and using these tools to denoise
16:33
and correct the images and get the quality where
16:34
it should be. But to give you a sense of the impact
16:37
of compressed sensing in day-to-day practice.
16:39
I show you some of our protocols from, uh, this
16:43
happens to be a Philips magnet, uh, pre and
16:45
post-implementation of compressed sensing on
16:48
top of our routine day-to-day protocols. You can
16:51
see the cumulative scan times of the current, the
16:53
cumulative scan times under compressed sensing, and
16:56
none of these are limited protocols. This is just
16:58
routine, everyday protocols with everyday quality,
17:01
and you can see just by going to a greater degree
17:04
of undersampling through this enhanced version of
17:08
compressed sensing, we're able to get savings of
17:11
20 to almost 50 percent in routine brain imaging.
17:15
And you can do this to go fast, or, I also mentioned,
17:19
you can start to do these truly isotropic acquisitions
17:23
as the ones, as you might only see in academic
17:26
settings, uh, can now be done in scan times that
17:29
are practical for everyday use.
17:32
On the other hand, you can take an exam like
17:35
this, a routine everyday MRA, which might take
17:39
4 to 6 minutes on your scanners, and get a clear
17:43
50 percent reduction across the board.
17:46
And the data is so sparse.
17:48
And every one of our machines running compressed
17:50
sensing is now doing the brain MRAs in only
17:53
about two minutes, as opposed to where we
17:55
were before, which is really kind of exciting.
17:58
When it comes to the spine, you can again see
18:01
our cumulative diagnostic scan times on the
18:03
left and where we were with compressed sensing.
18:06
And the average savings here is running
18:08
30 to 50 percent in these cases.
18:11
And that's a big boost in terms of patient comfort.
18:13
At our organization, we noticed that.
18:16
We get substantially better ratings on
18:18
our 3T than we get on our 1.5T.
18:21
And you know, all things the same, in
18:22
the same building with the same doctors,
18:24
with the same, you know, protocol goals.
18:27
Uh, and the only way we can explain that, 'cause 3T
18:29
is kind of uncomfortable, Todd, you know, it's loud.
18:33
The only way to explain why people would prefer in a
18:36
substantially, in a statistically significant way, the
18:41
3T over the 1.5T, is that those exams are
18:44
two to three minutes faster over the course of an exam
18:47
experience. So, you know, regardless of whether you're
18:49
busy or not, whether you need the volume or you don't,
18:53
uh, whether you're in, you know, reduced volume as
18:56
we're seeing in New York City due to COVID, or preparing
18:59
to come out of, uh, COVID with a surge of pent-up
19:04
demand for oncology exams that haven't been done,
19:07
the speed benefits will help you there, if not with
19:10
patient experience, at least with throughput. Now, as you
19:14
might imagine, these images are too small, I don't
19:16
think you need to see them unless you really care to,
19:18
but it shows the different approaches. You can choose to
19:20
go faster, you can choose to boost your routine quality,
19:24
or you can change the spatial resolution to compete
19:27
with the 3T across the street when you're running a 1.5T.
19:29
19:30
That's really the key piece. Safety in the lumbar
19:32
spine—take a routine four-sequence protocol, uh, from
19:35
say, seven and a half minutes down to five minutes,
19:38
or up to 11 minutes, you know, getting the better
19:41
quality or the better speed along the way. And these
19:44
are tools that we will apply creatively and perhaps
19:47
selectively depending on the patient demand. When
19:50
it comes to musculoskeletal, you see we get the
19:52
same type of boost, about 30 to 40 across the board.
19:55
So, you know, please, if you have these tools, use them.
19:58
Um, they produce image quality just as you would
20:02
expect, but do it in a way that's going to be more
20:04
resistant to motion, and it is going to be much more
20:07
patient-friendly. And in some cases, well, I'll allow
20:10
you to let your imagination roam and perhaps aim for
20:13
quality that just wasn't practical on your current
20:16
generation machine. For us, in many cases, these tools can
20:20
help with older machines that really couldn't move that
20:23
fast, particularly when I come to the next iterations of
20:26
these tools. And the next one to talk about is iterative
20:29
reconstruction. Some of you may be familiar with
20:31
that term, iterative reconstruction, from CT, and we'll
20:34
address CT, but these tools are also available for MR.
20:37
And this is a machine learning-assisted iterative reconstruction MR tool,
20:41
where they essentially took the routine exam, which
20:43
you see on the left, and created an accelerated exam,
20:50
about 33% faster, and used the tool to reduce the noise,
20:53
maintain the resolution, and put them up side by side.
20:58
We did a trial which we presented
21:00
incrementally depending on what we were doing.
21:02
We ended up doing two spines and one brain
21:05
exam on these patients, and we presented
21:09
these results in a variety of academic meetings.
21:12
But in these cases, you can
21:13
see what we're aiming at here.
21:14
You can see the blue bars on the right-hand
21:15
side. You see the red line at three and three,
21:19
which basically means the accelerated process version
21:22
is the same as the unprocessed state standard
21:25
of care version. You can see in the brain
21:27
there was no statistical significance in
21:31
ratings between the accelerated 33% faster
21:34
process version and the standard of care version.
21:36
That's really pretty cool.
21:38
Here you see an example of a lumbar spine.
21:40
Normal protocol was 249.
21:42
Here, the protocol for the trial was
21:44
1 minute 30, about a 50% reduction.
21:47
And in this particular part of the study, again,
21:50
you see there was no statistical significance.
21:52
Ratings are just about three.
21:54
Comparing the standard of care exam against
21:56
the accelerated exam with the machine
21:58
learning-assisted iterative reconstruction.
22:01
Here you see the cervical part of the exam.
22:03
We went from three minutes down to two minutes,
22:05
a 32% reduction, and again, you can see here no
22:08
statistically significant difference between the
22:12
accelerated version and the standard of care version.
22:14
In this case, the numbers above three mean that the
22:17
accelerated version was actually better, probably
22:19
because we did a better job with noise reduction
22:21
while still preserving edges, but statistically
22:24
non-inferior for all three applications across the board.
22:27
So what's the message here?
22:28
You can use iterative reconstruction commercially
22:31
available, probably adaptable to any of your
22:34
machines that don't have all the bells and whistles.
22:36
Um, maybe it doesn't have the
22:38
greatest coils for acceleration.
22:40
You can find your way to get your
22:41
acceleration, uh, in this image space.
22:46
Now, you know, the question I always ask myself is,
22:48
um, there are noise-reducing filters on our systems.
22:52
A lot of us don't use them.
22:53
We use them extensively, uh,
22:55
because we believe in them.
22:56
I think they produce, you know, entry
22:58
level versions of what you see here.
23:00
Today, uh, but the theory is that these tools
23:04
the ones you'll hear about over the next
23:05
20 minutes or so will take you beyond the ability
23:08
of a system to cope with what's built in. And this
23:11
was a little trial that we presented at RSNA.
23:14
Was it RSNA?
23:15
Uh, uh, yeah, maybe it was RSNA, um.
23:18
Somewhere over the last few months,
23:19
we presented this limited trial at a meeting, and
23:22
this is what we were after for this particular
23:24
demonstration trial. You have a typical five-minute
23:27
T1 on the left-hand side. We, uh, asked
23:30
the manufacturer to create a, uh, compressed sense
23:33
factor for the scan, which reduced the scan time in half.
23:37
If you look at the central portion of the
23:39
image here, you can see there's a little bit
23:41
of noise breakthrough, but to many folks,
23:43
radiologists actually like structured noise.
23:45
They may think this is a perfect image.
23:47
Um, if you ask the, uh, iterative reconstruction
23:52
software to process this accelerated image at two minutes and 30,
23:55
you can see we maintain really excellent edge
23:57
definition, uh, but we also do a very nice job of
24:00
selectively, uh, removing that parallel
24:03
imaging-related model in the center of the image.
24:07
So, producing an excellent image in half the time,
24:10
going beyond what the manufacturer could do in
24:13
the middle. If that wasn't obvious enough, let's
24:15
take it down to compressed sense of eight here.
24:17
You see, with the onboard noise-reducing filters,
24:20
the Philips system just kind of fails. You can't
24:23
go that hard right, and the application people.
24:28
Well, you can't push that hard. Well, it turns out if
24:30
you have this iterative reconstruction tool, you can
24:33
push that hard, and you can see we've maintained
24:35
all of our edge definition. Uh, a pleasing
24:38
relatively pleasing image appearance. You, and
24:41
we've pretty much selectively removed the central
24:43
noise in the image, which is pretty exciting.
24:45
Here's another example, uh, standard image at
24:48
2.15, uh, is that what it is?
24:51
Actually, yeah, uh, here we are at
24:54
1.09 with the Philips filter.
24:55
Here we are with 1.09.
24:57
With the, uh, iterative reconstruction processing, you
25:00
can see, again, a pretty good job reducing the noise
25:03
in the image beyond what the manufacturer can muster,
25:06
uh, taking us to the next level in this particular
25:08
limited demonstration trial. You can see that the
25:11
iterative reconstruction came closest to matching the
25:13
benchmark image quality compared to the, um, compressed
25:18
sensing with the manufacturer's noise-reducing
25:19
filter and compressed sensing with no noise-reducing
25:23
at all. All right, so that's iterative reconstruction.
25:27
Um, where is the world going now?
25:30
Well, the world is moving into, um, deep
25:33
learning reconstruction, and I'll show
25:35
you kind of why we needed to go there.
25:38
You're all familiar with iterative
25:39
reconstruction for CT, right?
25:41
It has allowed us to go to those levels that
25:44
with traditional reconstruction, we would
25:46
call that filtered back projection. We would
25:48
get an image that's non-diagnostic and noisy.
25:50
Right, but using iterative reconstruction techniques,
25:53
this is a model-based iterative reconstruction
25:55
technique, this case from General Electric, you can
25:58
see that we can actually get the image quality to
26:02
uh, restore the normal noise levels and restore the
26:04
normal image quality, yet using a markedly reduced
26:07
dose for an abdomen and pelvis if you're a dose guy.
26:10
This was a CTDI of six, which is a lot lower than
26:13
our traditional techniques would do now.
26:16
The interesting thing about iterative reconstruction
26:19
is that, you know, there are a lot of complaints
26:21
about iterative reconstruction creating an odd image.
26:23
Look, there were some changes
26:25
to the quantitative values.
26:26
We'll talk about all of those.
26:27
But the good news is that
26:29
two major vendors are now
26:31
marketing FDA-cleared CE-mark software
26:35
based on neural networks that will actually
26:37
reduce the noise of CT images but maintain the
26:41
traditional look and integrity of those images.
26:43
This is an example from a GE machine. You can
26:45
see here a noisy filtered back projection, what
26:48
it looks like with traditional manufacturer
26:51
iterative reconstruction on the CT.
26:53
It's basically texture of the image a little odd.
26:56
So some of the noise is coming through at a funny
26:58
texture. When you look at the deep learning-based
27:00
image, it looks a lot more like the CT you know
27:02
your mom and pop used to love. If you look at
27:06
this particular example, um, Dr. Tim Stick gave
27:09
me these images. His Twitter handle is
27:11
on one of these slides. If you really want to
27:13
learn a little bit about CT physics every day,
27:15
please go to his Twitter page and check it out.
27:18
In fact, it's at Professor underscore Tim Stick. You'll
27:21
see that one of the slides coming up, but very nice
27:24
example of what a phantom looks like at full dose.
27:27
Okay, typical structured noise in the image on a CT
27:30
image. You know the way we can do the projections,
27:33
but otherwise, really not very high noise. So
27:35
you measure noise on CT by putting an ROI on and
27:38
looking at the variation, and the noise level here
27:40
would be reflected in the standard deviation of 15.
27:43
If we drop the dose from 80 to 8,
27:48
okay, you see the noise goes up to 32.
27:50
And you can see it in the image.
27:51
There's a lot more noise in that image.
27:53
No one likes this image.
27:54
For some reason, radiologists
27:55
love the structure of this image.
27:57
Now, if we take this image and apply what, in
28:00
this case, is ACER 50, which is a 50 percent
28:03
weighting traditional iterative reconstruction.
28:06
You can see, uh, we're still at the dose of
28:08
eight, but we're able to reduce the noise from
28:11
32 down to 21, not quite 15 of the dose that is
28:15
10 times higher, but really a respectable number
28:18
compared to where we were. Now, if I take the
28:21
same data at 8.2, excuse me, the same data at
28:25
a dose of 8.2 and apply a true deep learning
28:29
method of noise reduction.
28:31
Notice what happens.
28:33
I have an appearance very, very similar to the
28:37
routine image, not different or altered like it
28:41
is with traditional iterative reconstruction.
28:44
And by the way, the noise level is
28:45
coming right back to where it should be.
28:47
So, in this case, we're delivering with deep learning
28:50
the noise levels we expect. Okay, and the image
28:53
appearance we expect, at least in this phantom.
28:56
Now we know that regardless of which tool we use,
29:00
when we do iterative reconstruction with CT with
29:03
traditional methods, we alter the appearance of
29:05
tissue. Things look a little different in terms of
29:07
their density. This particular phantom study, done
29:10
by Dr. Tim Stick, there's his, um, there's his
29:13
handle at, you know, at Professor underscore
29:17
Tim Stick. He shows that all of these dose phantoms
29:20
have exactly the same appearance, whether they're
29:22
done with the deep learning reconstruction on the
29:25
right or the filtered back projection on the left.
29:28
So, tissues will look the same and behave the same.
29:31
The Hounsfield units will be the same here.
29:34
You see a quantitative phantom on the left-hand side.
29:36
You see the behavior of bone, acrylic, and
29:40
another material, which I forget what that is.
29:42
I think it's, um, uh, polyethylene, perhaps.
29:47
You can see that the behaviors that these
29:49
tissues will have with traditional CT reconstruction.
29:53
Look what happens when we do a, uh,
29:55
iterative reconstruction technique on this.
29:57
I'm going to go back up.
29:58
Just trying to move my picture off my slide.
30:01
Uh, whether we use a model-based iterative
30:02
reconstruction or we use the traditional iterative
30:04
reconstruction, you'll notice that there is
30:06
a significant variation in the appearance of
30:09
these tissues, the densities of these tissues,
30:12
um, uh, when it comes to, uh, or appearance
30:17
of these tissues, excuse me, when it comes to
30:19
the gold standard filtered back projection.
30:22
Um, the best way to get a feel for how the
30:25
texture of the image changes is from these slides.
30:27
Here, we're comparing any of these curves
30:29
we'll do, but focus your eyes on one.
30:32
Um, you see the red curve, the
30:33
green curve, and the yellow curve.
30:35
Uh, the blue curve is actually, uh, the curve of
30:38
what we do when we have a filtered back projection.
30:40
The overlying red curve is what we
30:42
have when we're using a deep learning
30:44
form of iterative reconstruction.
30:45
You see there's virtually no variation
30:47
in the texture of that image.
30:49
This is a, uh, a noise power spectrum image, so that's
30:52
how you get texture out of it. But notice the markedly
30:55
different noise power spectrum of the iterative
30:58
reconstruction applied in a heavy fashion with
31:01
traditional techniques. So the lesson here is one, the
31:04
densities of tissues don't change with the deep
31:06
learning tool, and two, the texture spatial for
31:10
the noise power spectrum. It doesn't change as well.
31:12
The message here is it looks the same.
31:14
Here you can see phantoms taking a close look
31:17
at the image texture from the same phantom.
31:19
Look at all that odd texture there on the left
31:22
hand side that just isn't there in the phantom
31:25
and just isn't there with the deep learning.
31:27
Again, thanks to Dr.
31:28
Stick for all these great images.
31:31
Another way that you will see demonstrated with
31:34
by some of the OEMs is something like this
31:38
where they do a subtraction experiment. Imagine
31:40
this is our phantom. Then we've just filtered
31:42
the heck out of it just to make a point, right?
31:45
And we've given up a lot of the spatial resolution
31:47
at least it seems, and then we've processed it
31:49
with deep learning here, and claiming that we
31:51
don't give up any spatial resolution when you
31:53
subtract this from this. Look at all the detail.
31:55
It's gone. It's lost.
31:58
Okay, which you don't see any structure with the
32:00
deep learning reconstruction, which is really one
32:02
of the other key points. Again, taking a, uh, the noisy
32:06
image at ultra, ultra low dose, processing it with deep
32:10
learning, subtracting the two to see if we lose any
32:12
structure, and look, not a single piece of recognizable
32:16
anatomy in there. And that's very reassuring to see
32:18
these tools are going to take us to the next level.
29:57
I'm going to go back up.
29:58
Just trying to move my picture off my slide.
30:01
Uh, whether we use a model-based iterative
30:02
reconstruction or we use the traditional iterative
30:04
reconstruction, you'll notice that there is
30:06
a significant variation in the appearance of
30:09
these tissues, the densities of these tissues,
30:12
um, uh, when it comes to, uh, or appearance
30:17
of these tissues, excuse me, when it comes to
30:19
the gold standard filtered back projection.
30:22
Um, the best way to get a feel for how the
30:25
texture of the image changes is from these slides.
30:27
Here we're comparing any of these curves
30:29
we'll do, but focus your eyes on one.
30:32
Um, you see the red curve, the
30:33
green curve, and the yellow curve.
30:35
Uh, the blue curve is actually, uh, the curve of
30:38
what we do when we have a filtered back projection.
30:40
The overlying red curve is what we
30:42
have when we're using a deep learning
30:44
form of iterative reconstruction.
30:45
You see there's virtually no variation
30:47
in the texture of that image.
30:49
This is a, uh, a noise power spectrum image, so that's
30:52
how you get texture out of it. But notice the markedly
30:55
different noise power spectrum of the iterative
30:58
reconstruction applied in a heavy fashion with
31:01
traditional techniques. So the lesson here is, one, the
31:04
densities of tissues don't change with the deep
31:06
learning tool, and two, the texture spatial for
31:10
the noise power spectrum doesn't change as well.
31:12
The message here is it looks the same.
31:14
Here you can see phantoms taking a close look
31:17
at the image texture from the same phantom.
31:19
Look at all that odd texture there on the left
31:22
hand side that just isn't there in the phantom
31:25
and just isn't there with the deep learning.
31:27
Again, thanks to Dr.
31:28
Stick for all these great images.
31:31
Another way that you will see demonstrated with
31:34
by some of the OEMs is something like this
31:38
where they do a subtraction experiment. Imagine
31:40
this is our phantom. Then we've just filtered
31:42
the heck out of it just to make a point, right?
31:45
And we've given up a lot of the spatial resolution
31:47
at least it seems, and then we've processed it
31:49
with deep learning here and claiming that we
31:51
don't give up any spatial resolution when you
31:53
subtract this from this. Look at all the detail.
31:55
It's gone. It's lost.
31:58
Okay, which you don't see any structure with the
32:00
deep learning reconstruction, which is really one
32:02
of the other key points. Again, taking a, uh, the noisy
32:06
image at ultra, ultra low dose, processing it with deep
32:10
learning, subtracting the two to see if we lose any
32:12
structure, and look, not a single piece of recognizable
32:16
anatomy in there. And that's very reassuring to see
32:18
these tools are going to take us to the next level.
32:21
Again, you can actually look at the spatial resolution
32:24
of these tools, whether you apply the deep learning
32:27
reconstruction on lighter or heavier levels.
32:30
The image doesn't really look much different.
32:32
And the phantom doesn't look much different.
32:34
So you really can use this at the highest
32:36
level without fear of losing any information,
32:39
which is really kind of cool. And here's a nice
32:42
clinical case showing you what the images look
32:44
like with the deep learning reconstruction.
32:46
Nothing at all foreign about this appearance.
32:48
That really is the key thing. Again, here you see a five
32:51
millimeter traditional acer image. When you go down to
32:54
the thinner slice, you start to see the noise breaking
32:56
through, and you see the texture here, you, uh, as well.
33:01
Uh, and here you can see, um, how the deep learning
33:05
tool maintains the low noise levels, uh, and maintains
33:09
the kind of look to the image that you'd expect.
33:12
All right, so, um,
33:16
my primary goal was to talk about tools that
33:18
affect the length of an MR acquisition because
33:21
those are the ones that impact, you know, patient
33:23
comfort as well as, you know, throughput on an
33:26
exam that we do, you know, laborious scanning on.
33:31
And we're going to talk about these
33:33
tools that are now clinically available.
33:34
GE announced their AirDL recon
33:37
was FDA approved this week.
33:39
Canon had their, um, their onboard, uh,
33:42
deep learning reconstruction approved
33:43
just a couple of months before then.
33:45
And SubtleMR, which is, uh,
33:47
which is a, uh, an AI company.
33:51
They have had a iter, they have had their version
33:54
of deep learning, uh, reconstruction available for
33:59
even longer, and I'll show you examples of all three.
34:01
In this particular example, as you can see here,
34:04
you can see we've taken this 2D FASP and Echo to 0.
34:09
9 millimeters, a true 0.
34:10
9, something that you would have thought
34:12
would never possibly work, and used deep
34:15
learning to go ahead and completely denoise
34:17
this image while maintaining the structure.
34:19
Imagine that you could look at your lumbar spine in
34:22
a seamless sort of paging fashion instead of jumping
34:25
by threes or fours, jump by less than a millimeter.
34:28
How would that affect your ability to see
34:30
things like this angular fissure here at L3 4?
34:34
We have some real challenges in MR.
34:37
Sometimes it's about going faster.
34:39
Sometimes it's about raising the spatial resolution.
34:42
Sometimes it's about dealing with
34:43
something that we just struggle with.
34:45
And one of the things we struggle with pretty
34:47
consistently is getting great prostate images.
34:50
Uh, in a reasonable length of time. Uh, and here you see
34:54
this example of a patient being done with a roughly
34:57
three-minute axial scan Uh of the prostate. We
35:01
took the same data. The zeros and ones in this case
35:04
went off to a workstation, processed, sent back to the
35:07
PACS and got you the image on the right-hand side.
35:11
Okay.
35:12
It's exactly the same scan.
35:14
What do you notice first of all?
35:15
The signal-to-noise is much higher. Um, it looks
35:18
a little glossier; that might strike you initially.
35:21
But notice how the edges are actually sharper.
35:25
Look at the prostate capsule; it's sharper.
35:27
Look at these synovial cysts;
35:28
you're not losing any detail.
35:30
When you look into the tissues, you actually
35:32
can see more detail because we reduced
35:34
the background noise of those tissues.
35:36
So, the theme of deep learning reconstruction
35:39
here, now these slides say WIP, but
35:41
this is just the proof this week.
35:42
Is we can improve on routine image quality, and our
35:45
initial investigation is where do we lose anything?
35:48
And to date, we really felt that
35:49
we've lost virtually nothing. Again,
35:51
these can be applied at variable weighting. Here is the
35:53
routine image showing the routine noise levels of a 30-minute exam on the prostate, but applied at 50. You
36:02
can see that the noise levels are markedly reduced.
36:04
The edge definition is, uh, sharp
36:07
or sharper in actuality.
36:10
And look at the bladder.
36:11
Look how uniform the signal is within the
36:13
urinophilic bladder compared to this structured noise here.
36:19
Again, radiologists like structured noise.
36:21
It feels like resolution, but it's not.
36:23
It's actually structured noise.
36:26
Again, taking a closer look, here is
36:28
the zero or essentially the image which
36:29
comes right off the scanner to the PACS.
36:31
And this is the one that went through
36:33
the algorithm and then onto the PACS.
36:35
And you can see, uh, you know, a lot of
36:37
the internal architecture is better seen.
36:39
The contrast between this low signal area
36:41
and this, uh, higher signal area in the
36:45
central zone here, you can see is enhanced, actually.
36:51
Take a look at the colons.
36:52
It's enhanced.
36:53
It's cleaner.
36:54
Everything is again, you know, patient with
36:56
diverticulosis and a uterine leiomyoma. Again, the
36:59
routine image and what it looks like routinely off
37:02
the system with the deep learning reconstruction.
37:04
Like anything else I've showed you
37:06
so far, you can go either way, right?
37:08
You can go with higher quality.
37:10
You know, strain the resolution, make up the noise with
37:14
this tool that recognizes noise and eliminates it while
37:17
preserving resolution, or you can use it to go faster.
37:21
Okay, and that really is the key
37:22
piece for here across the board.
37:24
This is an example from Canon, uh, where they
37:27
actually did an exam that could not possibly
37:30
be supported at this field strength, then
37:32
used the deep learning reconstruction to take
37:34
the noise out while preserving resolution.
37:35
The structure.
37:36
This is another thing that people are
37:37
going to do with these techniques.
37:39
Um, and, uh, we, um, uh, it's another key
37:45
technique that, uh, we and others will
37:48
use to, um, to leverage these tools here.
37:52
You can see another example of the image off
37:55
the scanner on the left, the processed image
37:58
on the right, and, uh, uh, you can see really
38:03
nicely here how clearly the image is enhanced,
38:06
how the spatial resolution is not diminished.
38:09
And the edges actually may be slightly
38:11
improved on the left-hand side.
38:12
You see the example with the routine image
38:17
on the right-hand side. You see the example
38:20
with the deep learning in the bottom center.
38:22
I believe that's a deep learning example as well. Again,
38:26
look at the fifth metatarsal off the routine scanner.
38:28
Look at the fifth metatarsal and fourth metatarsal
38:30
head with the deep learning reconstruction. Again,
38:35
image on the left, image on the right, the same scan.
38:39
Okay, take a look at the articular cartilage.
38:41
Look at the contrast in this area that's denuded.
38:43
Look at the height of the articular cartilage.
38:45
It almost looks like it's bigger.
38:47
You know, our eyes, our brains react to higher spatial
38:50
resolution with the sense that structures are bigger.
38:54
And you can, you can see, you'll see this
38:55
routinely on these examples as we go along.
38:58
Here's another one: a little
38:59
full-thickness chondral defect.
39:01
Are we losing anything with these tools?
39:03
That's the question. We lose all of this, all the
39:06
salt and pepper and the soft tissues in the bone.
39:08
That's not real. Okay, uh, it disappears here.
39:12
Well, that feels odd. Well, we can get used
39:14
to this when we realize we're not losing
39:16
anything, and that's what this trial turned
39:17
up here. You see another example again.
39:20
Look at the articular cartilage here.
39:21
Look at this disease here.
39:22
Your mind might not recognize that, right?
39:25
It might not catch your eye.
39:26
Very hard to miss here.
39:28
First of all, look how much bigger the cartilage looks
39:30
and look how clearly delineated that pathology is.
39:33
And that's really what we're after.
39:34
Make the images, uh, convey the information better.
39:40
Again, you can see this chondromalacia
39:41
here in the sagittal view.
39:43
All of this structure of noise brought on by, you know,
39:45
pushing, maybe pushing the parallel imaging a little
39:47
too hard or getting close to the edges of the coil.
39:50
Notice the performance in terms of
39:52
noise top to bottom, front to back.
39:54
In this coil, it truly is revolutionary here.
39:56
You see another example on the left without
40:00
the tool, on the right with the tool.
40:02
And again, this is a zeros-and-ones processing from
40:05
the raw data. And taking a close look at this, uh,
40:08
of this side of the knee joint space. Look at how
40:10
much thicker this cartilage looks compared to that.
40:13
It looks visually like it's 30-40% thicker.
40:16
It's the same scan. All we've done is apply this tool,
40:20
reduce the noise, enhance the spatial resolution,
40:23
enhance the contrast resolution, and at least in
40:27
our limited studies, I haven't given up anything.
40:30
Here you can see this, this little
40:32
partially attached flap here.
40:34
You might not even see it here
40:36
if you weren't having a good day, but improve the
40:38
signal-to-noise ratio, preserve the spatial resolution, and it
40:41
just pops right out at you. Now, these tools can do a lot
40:46
of cool things and not simply filters. Firstly, nothing
40:49
I've shown you is filters. Um, but they're very clever.
40:52
They can be trained on an artifact
40:55
and learn to remove it. Here,
40:57
you see an example. We've got truncation artifact
40:59
and these multiple amylations in the cord. We can train
41:02
the algorithm to remove this artifact and get this.
41:05
So why would I want to remove that artifact?
41:07
Well, first of all, traditionally, we
41:09
remove the artifact by turning on a
41:12
Fermi filter or an apodization filter.
41:15
We do this on anything that has a rectangular
41:17
view, uh, to avoid the ringing, uh, that goes
41:20
into the image and the overall, uh, deterioration
41:23
of the image. To use these optimization filters,
41:26
we actually throw out the edges of k-space,
41:28
essentially throw spatial resolution in the garbage. If
41:31
I could turn off that filter, I would get the artifacts
41:34
back. But if I use the deep learning algorithm to remove
41:37
the artifact, I could actually win the battle, and that's
41:39
what I've done here. I've turned the Fermi filter off,
41:42
bringing my resolution back up, but then I get all of
41:44
these artifactual ringings. This artifactual ringing,
41:48
when I use the deep learning algorithm,
41:50
even though the filter is off, I can get rid of
41:52
the artifact and maintain the spatial resolution.
41:54
That's what these examples are, by the way.
41:56
Fermi filter off here.
41:59
What the deep learning algorithm does for this scan.
42:01
See all these little echoes: one, two,
42:04
three, they just don't show up when you
42:06
use the tool, when you train them out.
42:07
Here again, a typical image without the optimization
42:10
filter on, and what it looks like with the
42:13
tool, which recognizes that artifact as well,
42:15
as well as denoising and enhancing the spatial
42:17
just a little bit, as I showed you before.
42:20
This is a paper that we were
42:21
going to present to the ASNR.
42:23
If you go online to the ASNR,
42:24
you can see this right now.
42:25
We also presented in a different format
42:27
the artifacts made before we actually looked
42:30
at 93 image pairs of 28 patients at 1.5
42:33
and 3T. We looked at the routine images and
42:37
looked at the deep learning processed images
42:40
with the error recon DL tuned to 75% noise
42:43
reduction, and here are just some examples.
42:45
That's the starter came off the machine
42:46
on the left. There's the stir on the right.
42:51
I'm sorry.
42:53
There's the axial T2 on the left, very respectable.
42:56
Notice, much lower noise, slightly sharper edges.
43:00
And really respectable axial on the left, slightly
43:03
better, slightly sharper image on the right.
43:06
Typical T1 on the left, enhanced T1 on the right.
43:09
Notice all this noise out here, really
43:11
starting to see that structured noise.
43:13
You see how the fat pulses, you
43:14
can see noise shining through.
43:16
You can see how you just don't see that here.
43:18
Air is air, it's black.
43:20
It really is quite nice, and
43:21
again, what are we giving up?
43:22
I don't think we're giving up
43:22
anything of any note here again.
43:25
Look at all of this structured noise. Look at
43:27
how well we manage that noise yet still maintain
43:29
the tiniest little bits of structure in the
43:32
abdominal tissues, and look at this gorgeous, you
43:35
know, sharp image with really a very impressive T1
43:38
contrast. Not bad here. You know, this was a trial
43:40
to see, did we lose anything with this tool again?
43:44
Same type of example without the tool. With the
43:47
tool, you know, if you can do this in the same scan
43:51
time, what could you do in terms of acceleration?
43:54
I mean, that's really the key piece.
43:55
Again, here's just a couple of views of the same thing.
43:59
Off, 25, 50.
44:01
It's going to be 50, 75, and 100.
44:03
I like 100.
44:04
The study was done at 75.
44:06
But what are we losing between here and
44:08
here, other than the noise in the image?
44:10
As a matter of fact, if you look at the
44:11
end plate, see this little jaggedness here?
44:14
We don't see it here.
44:15
Seems like the algorithm recognizes
44:17
these eddy current type artifacts
44:19
and just removes them. Again, taking a closer look.
44:23
What have we lost?
44:24
The end plates are sharper.
44:25
The disc-sack interface is sharper.
44:27
We haven't lost any contrast resolution.
44:30
It doesn't look like we've
44:30
introduced any interesting artifacts.
44:33
Look at this annular fissure here.
44:35
That we see when you take a close look at that annular
44:37
fissure. You can argue clearly bigger here, clearly
44:41
less likely to be missed because of the enhancement,
44:44
spatial resolution, and look at all this noise.
44:48
Okay, it's just gone. So without belaboring the point,
44:52
the abstract would show that there was a statistically
44:55
significant improvement of the routine images when
44:59
the deep learning recon was applied for all of
45:02
these measures that I won't read through here. You
45:05
know, but they're all the things you'd ordinarily look at.
45:07
Now, to be fair to something I don't
45:09
have personal experience with, deep
45:11
learning is also available from Canon.
45:13
They call it ACE.
45:14
Um, you can see that they did the same thing
45:16
where they took a 3D flare, which took, um, in
45:21
100 seconds, the original image had an SNR of 7.
45:25
Um, and you can see it doubled the SNR
45:29
by processing it with their algorithm.
45:32
Uh, and again, it can show you how you
45:34
can start to aim at something on a 3T
45:36
that's closer to what people do on a 7T.
45:40
Well, you can't read that, right?
45:41
Well, what if you could denoise that stir and get that
45:44
kind of internal hippocampal architecture?
45:47
That's the key point.
45:48
So aim at something that's not practical.
45:50
Use the tool to make it so.
45:53
And that really is another thing
45:54
that folks will do with these tools.
45:55
Probably more likely to do those at academic
45:57
sites and sites like the ones where I
45:59
practice, where I'm trying to create, uh,
46:02
superb quality in a patient-friendly way.
46:05
Well, these slides, which are DTI, are just really there
46:08
to tell you that we don't lose any quantitative value
46:12
with these tools. The fractional anisotropy values are
46:15
exactly the same whether they process it with deep
46:17
learning or not. Very much like you had in the CT, in
46:20
the CT data earlier when the household unit stayed
46:23
the same across the board. Now, one last piece that
46:27
really is applicable to any system out there. You, uh,
46:29
that we have a fair amount of experience with in our
46:31
institution is, uh, subtle MRIs, deep learning-based
46:35
image processing, and here you can see it took an
46:37
image that looked pretty crude that was accelerated.
46:40
This is an axial reconstruction of a sagittal 3D
46:42
and turning it into a higher resolution version. So
46:45
for the most part, we really haven't talked about
46:47
resolution enhancement. What SubtleMR can do, which
46:50
is a true convolutional neural network processing
46:53
in the image space, is they can actually train in
46:55
higher spatial resolution with their algorithm.
46:58
You'll see some of that.
46:59
We actually have a pilot going
47:01
on right now, uh, at our centers.
47:03
We're in the middle of designing a financial
47:05
ROI model to determine how we can integrate this
47:08
into our practices, and if it comes through,
47:10
we'll find a way to make this stuff work.
47:12
But in this particular circumstance, um, we
47:15
actually did this particular abstract, which we are
47:18
presenting I think for the second time at the um,
47:23
at the American Society of Neuroradiology meeting
47:25
over the next couple of days. If you log on to
47:27
their servers, you can actually see the data in a more
47:31
meticulous fashion, but these are the highlights.
47:34
We wanted to see how well their aftermarket deep
47:38
learning tool did in terms of matching our routine
47:41
standard of care quality at highly reduced scan times.
47:45
In this particular trial, I
47:46
believe the number is an average of 33%.
47:48
And here you'll see the typical
47:50
triptych of images to evaluate. On the left-hand side,
47:53
you see standard of care; on the right-hand side, you
47:56
see what the images look like with a 30 acceleration.
47:59
And in the middle, what happens when you take the fast
48:01
image on the right and process it, uh, for, you know, and
48:05
process it with deep learning, incorrect, and create
48:08
the middle image. And the real goal of this study was
48:10
to see how well the DL fast emulated the standard of
48:15
care image. Again, fast on the right; standard of care
48:19
on the left. The image on the right, processed with deep
48:21
learning in the middle, uh, to create the same resolution
48:24
and noise levels across the board. Um, we've just done a
48:28
larger set of reads in this particular, um, experiment.
48:32
This is our original reads. Uh, but on the larger
48:35
set of reads with three readers of a larger data
48:37
set, there was, interestingly enough, superiority
48:41
of, uh, the deep learning FAST over
48:45
the standard of care for certain select criteria,
48:49
uh, and no statistical significance for the rest.
48:52
So the key message here, we were aiming
48:54
for non-inferiority; we clearly had it. The
48:56
standard of care was in no
48:58
way superior, uh, to the processed FAST images,
49:04
but in actuality, the processed
49:06
FAST images had a few areas of improvement.
49:08
Minor superiority over what we were
49:10
doing with routine standard of care.
49:12
You see another example: fast image on the right,
49:14
standard of care on the left. You can see the middle
49:17
image is a little lower and a little higher in
49:19
SNR, while not giving up any spatial resolution.
49:22
And the standard of care, maybe this is one of the
49:23
ones that got the higher ratings. Now,
49:29
I mentioned, um, the last piece I want to talk about,
49:32
which is super resolution. Um, these deep learning
49:35
tools, properly trained, I told you can remove
49:38
artifacts, but these deep learning tools, properly
49:41
trained, can actually train in spatial resolution.
49:44
And this is really fascinating.
49:46
Look at this image, which is a routine day-to-day case
49:48
done 30 slices a minute, 39 at one of our facilities.
49:51
If we try to accelerate that scan, we can
49:54
only do it by dropping the spatial resolution.
49:56
We drop the spatial resolution; we
49:57
actually increase the signal to noise,
50:00
but we clearly have a fuzzier image.
50:02
Imagine we could borrow the signal to noise,
50:04
borrow the contrast to noise that makes this
50:06
little lesion stand out more here than here, but
50:09
then train the spatial resolution of this on this.
50:12
That's what this represents.
50:14
Okay.
50:14
This is the subtle MR commercial available software.
50:17
We essentially denoise, accelerate, and maintain spatial
50:22
resolution by training what we've learned from the
50:24
high-res images onto the low-res images of the site,
50:27
allowing us to do arguably a better image than we
50:30
had before in less time than we had used in the past.
50:34
This is an example from an outside institution,
50:36
but the same thing: routine clinical quality
50:38
scan, accelerated low-res, high-contrast res, low
50:41
spatial res. Take the spatial res from the high
50:44
quality scan, train it onto the high-contrast
50:47
resolution scan, and you get the acceleration and
50:50
the quality you're used to. This is from my shop.
50:52
These are 3D FLAREs. The standard of care here is going
50:57
to be a little noisy because, first of all, we're
50:59
looking at thin partitions. Second of all, we're
51:01
trying to create an isotropic scan that makes good
51:03
reformats. So we will get a little noisy when we
51:07
take that scan and make it faster by dropping the
51:09
spatial resolution. In-plane, it looks pretty good.
51:11
But when we go through-plane, it's going to
51:13
suffer. Imagine that we could train the spatial
51:16
resolution of this onto the signal to noise
51:18
and contrast to noise of this, and then come up
51:20
with this, which is pretty, really pretty cool.
51:22
I can't see the scan time here, 346 to 246 in this
51:26
particular example. Here's an example from our
51:30
trial, uh, which we'll talk about in a moment.
51:32
You can see here.
51:33
I'm sorry.
51:33
There it is.
51:34
I'm going to go back up.
51:36
This was the standard of care scan.
51:37
A little noisy.
51:39
Okay.
51:40
This was the fast scan, a little low spatial
51:42
resolution, and this is what happens when you train
51:44
the spatial resolution of this onto this; you get this.
51:47
So a high spatial, high-contrast
51:49
resolution, high SNR example.
51:51
Pretty, pretty impressive stuff, I think.
51:53
So this is the trial we're presenting at the ASNR.
51:55
You can find this on the ASNR website, where we
52:00
essentially went for a 30% acceleration and super
52:03
resolution and aimed for a better result than our
52:05
standard of care. And here you see the examples: the
52:08
clinical standard of care scan, the clearly lower
52:11
resolution fast scan, and the net result with deep
52:14
learning with denoising and super resolution.
52:18
Look at what we get here: a much better result than
52:21
either one of these two alternatives. Again, clinical
52:25
standard of care scan, lower resolution fast scan,
52:30
deep learning fast scan, clearly higher spatial, clearly
52:33
higher SNR. And if you look closely, you know, these
52:36
two are exactly the same scan. This is a different one.
52:38
So there will always be a
52:39
slight difference in position.
52:41
So the fact that it looks bigger, you know, could
52:43
be a reflection of a slight difference in position,
52:46
but clearly, you can see higher SNR, higher spatial
52:49
resolution, and faster than the standard of care scan.
52:53
This is one of my favorites because I've
52:54
now seen this example a couple of times.
52:56
Clinical standard care looks pretty good.
52:58
Clinical, excuse me, fast scan starts to look a little
53:01
rough, and who knows what the reformats could look like.
53:05
Look at what the deep learning scan looks like.
53:07
It's really sharp.
53:08
It's really high SNR.
53:09
Now I want you to take a look.
53:11
There are two lesions here that are clearly seen
53:14
on the fast scan that, even though we raised the
53:16
spatial resolution of the deep learning scan, are still
53:19
beautifully seen, but are really hard to see.
53:22
I mean, we're pretty well registered
53:23
here on this gyrus, and we really
53:25
have a hard time seeing these. Yet,
53:27
we do see them here.
53:29
And I think that's largely due to the fact that
53:31
we have the inherently high contrast resolution of
53:36
the fast scan trained into the higher resolution
53:41
resolution-enhanced process fast scan.
53:44
So in our trial, which you can see the full results
53:46
on, uh, on the web, we found that these
53:51
tools produced a boost in quality, SNR, and resolution.
53:57
Yet allowed us to go roughly 30% faster.
54:01
Now the same tools are also applied to things like
54:04
ASL. You can take an ASL scan from eight minutes
54:07
clinically; we would do four minutes, take it down to
54:09
one minute. And you can actually make the one-minute
54:12
scan look very much like the eight-minute scan
54:13
here. Kind of cool! If you're doing synthetic MRI in
54:17
your practices, you know, there are some trade-offs.
54:19
With synthetic MRI, one complex acquisition, you then
54:22
retrospectively calculate any contrast you like. You see
54:26
just moving the mouse can actually change the contrast.
51:55
You can find this on the ASNR website where we
52:00
essentially went for a 30 acceleration and super
52:03
resolution and aimed for a better result than our
52:05
standard of care. And here you see the examples: the
52:08
clinical standard of care scan, the clearly lower
52:11
resolution fast scan, and the net result with deep
52:14
learning with denoising and super resolution.
52:18
Look at what we get here: a much better result than
52:21
either one of these two alternatives. Again, clinical
52:25
standard of care scan, lower resolution fast scan,
52:30
deep learning fast scan, clearly higher spatial, clearly
52:33
higher SNR. And if you look closely, you know, these
52:36
two are exactly the same scan. This is a different one.
52:38
So there will always be a
52:39
slight difference in position.
52:41
So the fact that it looks bigger, you know, could
52:43
be a reflection of a slight difference in position,
52:46
but clearly you can see higher SNR, higher spatial
52:49
resolution, and faster than the standard of care scan.
52:53
This is one of my favorites because I've
52:54
now seen this example a couple of times.
52:56
Clinical standard care looks pretty good.
52:58
Clinical, excuse me, fast scan starts to look a little
53:01
rough, and who knows what the reformats could look like.
53:05
Look at what the deep learning scan looks like.
53:07
It's really sharp.
53:08
It's really high SNR.
53:09
Now I want you to take a look.
53:11
There are two lesions here that are clearly seen
53:14
on the FAST scan that, even though we raised the
53:16
spatial resolution of the deep learning scan, still
53:19
beautifully seen, that are really hard to see.
53:22
I mean, we're pretty well registered
53:23
here on this gyrus, and we really
53:25
have a hard time seeing these, yet
53:27
we do see them here.
53:29
And I think that's largely due to the fact that
53:31
we have the inherently high contrast resolution of
53:36
the FAST scan trained into the higher resolution,
53:41
resolution enhanced process FAST scan.
53:44
So in our trial, which you can see the full results
53:46
on the web, we found that these
53:51
tools produced a boost in quality, SNR, and resolution,
53:57
yet allowed us to go roughly 30 percent faster.
54:01
Now, the same tools are also applied to things like
54:04
ASL. You can take an ASL scan from eight minutes.
54:07
Clinically, we would do four minutes, take it down to
54:09
one minute, and you can actually make the one minute
54:12
scan look very much like the eight minute scan.
54:13
Here, kind of cool if you're doing synthetic MRI in
54:17
your practices. You know, there are some trade-offs
54:19
with synthetic MRI. One complex acquisition, you then
54:22
retrospectively calculate any contrast you like. You see,
54:26
just moving the mouse can actually change the contrast.
54:28
It's a wonderful tool. We showed in an ASNR paper
54:32
that it essentially can fill in for routine day-to-day
54:34
clinical individual scans. The only real trade
54:38
off is a slightly odd look to the FLAIR. By the way,
54:41
the top row is synthetic and the bottom row
54:43
is the traditional individual example. You
54:47
can see this is a pretty much a FLAIR horror show.
54:50
It gets about this bad. You get some speculation
54:53
in the CSF; you get some hyperintensity of the
54:56
sulfide G. You know, if this is predictable,
54:59
could I use deep learning to get rid of it?
55:01
Well, with subtle medical folks in one of their
55:03
publications show that they actually can take,
55:06
you know, pretty much an awful synthetic MRI scan and
55:10
with a variety of different methods of processing
55:12
it restore the traditional quality we expect. So this
55:16
may be one of the limiting factors. We're removing
55:19
one of the limiting factors to routine implementation.
55:22
Mind you, if you can improve signal to noise,
55:25
in MR, you can improve it in MR functional imaging,
55:28
you can improve it for PET imaging. And this is
55:31
another FDA-approved tool called subtle PET. In this
55:34
particular case, where they actually took a, we did
55:36
a one-minute scan. Noisy compared to the four-minute
55:40
scan standard. Well, they actually didn't do it.
55:42
They actually simulated it through list mode, but
55:45
nevertheless, using list mode simulation, they were
55:47
able to take their tools and make the one-minute scan
55:50
look like the four-minute scan. When you're talking
55:52
about multiple beds for a patient, you can take an
55:54
exam time, you know, down to a very respectable short
55:58
amount of scan time compared to what it normally
56:00
takes. Here's a 20-minute for a beta-PET scan. Here's
56:04
what happens when we simulate five minutes with list
56:06
mode. And this is what happens when we process a list
56:08
mode scan with the subtle PET software. Now, being
56:11
an MR guy more than a PET guy, I'm really kind of
56:14
excited about things like this. And this is not FDA-
56:16
approved as yet. This is some work in progress.
56:19
They published this, so it's coming from the published
56:21
literature. But I'm sharing this under my own volition.
56:25
You can see they've got a conscious enhancing left
56:27
thalamic lesion here. When you use only 10% of the
56:30
contrast, you can't perceive the enhancement.
56:33
But we train the algorithm on full-dose enhancement.
56:36
It learns to bring the enhancement out of this 10%
56:39
and then can synthesize an image that looks
56:41
very much akin to what the routine image looks like,
56:44
essentially allowing you to go to 10% of the dose.
56:47
Here's a full-dose scan. This is what it looks
56:49
like at a tenth of the dose, and here's what
56:51
it looks like when you simulate or synthesize
56:54
using deep learning this image from that image.
56:58
It's pretty impressive, I think.
56:59
Particularly if you're concerned about serial
57:02
administration of gadolinium, the patients who are
57:04
going to get multiple exams over the course of their
57:07
lifetime, pediatric patients, patients who may
57:12
actually be seen outside of the neural area for the
57:15
inflammatory bowel disease, high-risk cancer screens.
57:17
It's, you know, the potential to use less contrast.
57:21
Pretty exciting here.
57:21
One more time.
57:22
I don't see anything here around the right frontal
57:24
horn at 10%, at full dose I see forward
57:27
enhancement, and just using deep learning to take
57:30
this image and make it look into that image. The question
57:33
is what do we miss, and clinical trials need to
57:35
be done to see whether you can fully emulate
57:37
this approach. But it is still pretty exciting.
57:41
So I have no idea where I am on time.
57:42
I expect I'm running long, so I hope I gave you
57:45
some sense of background in AI. But I hope
57:49
you walked away with excitement over what these
57:52
tools can do to get us over the limitations
57:55
of the tools we use today and the limitations of
57:58
scan times and radiation doses we have today.
58:01
Thank you very much for your attention.
58:04
Perfect.
58:04
Thank you so much for that.
58:05
I really appreciate you joining
58:06
us for this noon conference today.
58:07
I do see it's one o'clock.
58:08
I'm not sure what your schedule looks like.
58:09
But if you have time, we do have a
58:10
few questions in this Q&A feature.
58:12
I can stay.
58:14
Okay, perfect.
58:14
Um, if you open up that Q&A feature,
58:16
it might be best if you read it.
58:17
If you move your mouse to the top, you'll,
58:19
oh, now you'll see it at the bottom.
58:21
Okay, uh, under Q&A.
58:23
Got it.
58:24
I gotta hide this little button here.
58:28
Uh, we will change our carriers soon.
58:30
I don't know what that means.
58:31
Okay, I'm looking for a question I can translate.
58:33
Um, okay, a real quick one while I'm
58:36
looking for something else to talk about.
58:38
Are medical school students dropping out
58:40
of radiology, losing interest in radiology?
58:42
Well, I don't have last year's data, but
58:44
over the previous two years, we had essentially
58:47
100% filling of radiology spots, up from
58:51
only about 85% just a few years earlier.
58:54
So while undoubtedly some people may not see
58:56
the field as attractive as they did in the past,
59:00
clearly,
59:01
they are not running away from the field right now.
59:04
And it still is a very, very attractive field. As
59:07
a matter of fact, I think personally it is one of the
59:09
best times to be a radiologist, where you'll be more
59:12
central to everything that goes on. We've always been
59:14
the center of diagnostic information, creating the
59:17
next step in the pathway. But since we're naturally
59:20
suited sitting where we are in the center of all this
59:23
data, uh, and we now have tools and assistance that
59:27
will help us filter through the data, scraping the
59:29
medical record for key points that we need to find,
59:31
I'm bringing it all to our fruition, reminding us that
59:34
hey, you know, this patient has got a high risk of
59:38
cancer. Do you want to take a second look at this?
59:41
You know, non-descript findings—those
59:43
things are going to make it a lot better.
59:45
Um, uh, here's a quick one.
59:48
Can you apply deep learning to cone beam CT?
59:50
I suspect there's no reason you can't. I don't
59:52
have much experience with it. Um, another question
59:56
that says could AI be cleaning up too much?
59:58
That's a great question. Nobody really knows. I would
60:01
say that quoting one of my mentors, good is good.
60:04
Uh, you can recognize it when you see it, um, and, uh
60:08
thus far. In our research work and our clinical work,
60:11
we have not seen loss. As a matter of fact, all of my
60:14
experience in my career suggests that if you reduce the
60:16
noise, you reduce the noise. You actually get a big boost
60:21
in your contrast resolution, and that greatly overwhelms
60:26
any possibility of blurring out any pathology. But
60:30
obviously, you know, ideally, you would have controlled
60:33
trials in large-scale trials and multi-centered trials,
60:37
bearing in mind that we have that to virtually nothing
60:38
in MRI. But I, you know, I love the idea of having it.
60:41
Quick question, is the subtle MRI FDA approved?
60:44
It is.
60:45
Um, I mentioned several times during the talk,
60:47
which ones were approved, which ones weren't.
60:49
I really didn't show you anything except for maybe
60:51
the subtle GAD, uh, today that is not FDA approved.
60:57
Can the software now potentially
60:59
introduce more artifact?
61:00
There have been some phantom demonstrations, you know,
61:03
on ridiculous use cases that have shown the fear, but
61:06
nothing in these commercialized products
61:09
has led any credence to danger using them so far.
61:15
Will there be a learning curve
61:16
in learning to use these images?
61:17
Yeah, you know, it's hard to, you know, when you
61:19
go to, if you've ever experienced going from 1.5 to 3T,
61:22
the first reaction everybody has is suspicion.
61:25
Yeah.
61:26
What have you done to these images?
61:28
They look filtered.
61:29
Okay, I said that myself.
61:31
And why, what is that reaction?
61:33
You are reacting to the absence of noise.
61:36
So, yeah, it took a while to get used to 3D.
61:39
Not just the contrast differences, but you
61:41
actually look and the image is different.
61:43
It's cleaner.
61:43
There's less noise.
61:45
It does take some time to get used to less noise.
61:47
Um, but I imagine it's something I'd be happy
61:51
to get over if the scans were half as long,
61:53
the contrast resolution was
61:54
better, and the noise was lower.
61:56
The curve appeal of these images is extremely high.
62:00
Uh, let's see.
62:02
Uh, I think I've answered most of the other questions
62:04
in one way or the other, looking for another one.
62:08
Um, yeah, all right.
62:11
I think that's basically it.
62:12
I don't see anything else coming
62:13
up that I haven't really answered.
62:15
Um, somebody asked,
62:20
can I describe spatial resolution?
62:22
You know, I'm a Fred Flintstone physicist,
62:23
but I imagine you could describe that as
62:26
the smallest structure you can resolve.
62:28
To me, it's sharpness.
62:30
Um, you have an HDTV at home, you have a 4K TV if
62:33
you're lucky. Um, you know immediately it's better than
62:36
the old 13-inch screen you had on the cathode ray tube.
62:40
You can appreciate the sharpness.
62:42
You see how you feel details better, the image,
62:44
the screen is more intimate, the pictures are more human.
62:48
I think the same thing applies when you look at a
62:49
higher spatial resolution. You can see MR images.
62:52
You don't need to prove to me, at least, that higher
62:54
resolution is better.
62:55
Perfect.
62:56
All right.
62:56
As we bring this to a close, I want to thank you,
62:58
Dr. Tannenbaum, for your time today.
62:59
We really appreciate it.
62:59
And thanks to all of you for
63:00
participating in this Noon Conference.
63:02
A reminder that it will be made available
63:04
on demand, complimentary, at MRIonline.
63:06
com in addition to all previous Noon Conferences.
63:09
Uh, Monday, please join us for a Noon
63:11
Conference on Imaging of the Tinnitus with
63:14
Dr. David Pastel.
63:15
And follow us on social media for updates and
63:16
reminders on all upcoming Noon Conferences.
63:18
Thanks and have a wonderful day.
Lawrence Tanenbam, MD, FACR
Chief Technology Officer, Director of Advanced Imaging, Vice President
RadNet
© 2026 Medality. All Rights Reserved.
