Latest / Becker’s Healthcare Podcast / AI at the Heart of Radiology Informatics Innovation

Transcript

AI at the Heart of Radiology Informatics Innovation · Becker’s Healthcare Podcast

0:00
Philips is a health tech leader focused on
0:03
innovation that improves the health and well-being of
0:06
people. Our health care technology and informatics solutions
0:09
help care teams diagnose, treat, and manage more
0:12
patients with greater precision, speed, and confidence across
0:16
the care journey. With Philips, clinicians are empowered
0:19
with streamlined insights in the moments that matter
0:21
for every patient. Better care for more people.
0:25
Philips.
0:28
Hi, everyone. This is Lucas Voss with Becker's
0:30
Healthcare. Thanks so much for tuning in to
0:32
the Becker's Healthcare podcast series. It's great to
0:35
have you. Today, we're talking about the role
0:37
that AI plays in radiology informatics innovation. And
0:40
joining me for today's discussion, very excited to
0:42
have them, doctor Paul Yee, associate professor of
0:45
radiology and chief of the Intelligent Imaging Informatics
0:48
section at St. Jude Children's Research Hospital, and
0:51
doctor Sanjay Gandhi, chief medical officer for Philips
0:54
Enterprise Informatics. Doctor Yi and doctor Gandhi, thanks
0:57
so much for being here today. It's great
0:58
to have you.
0:59
Thank you, Lucas. It's great to be here.
1:02
Thanks for having us, Lucas. Absolutely. I do
1:04
wanna start off with introductions for those that
1:06
might not know you yet at least.
1:09
If you could just share a little bit
1:10
about yourself and your work in health care.
1:12
Doctor Yee, we'll start off with you.
1:14
Sure. Hey, everyone. My name is Paul Yee.
1:16
I'm a physician scientist and radiologist,
1:19
and I lead the intelligent imaging informatics section
1:21
at St. Jude Children's Research Hospital.
1:24
My work sits at the intersection of clinical
1:26
care, imaging informatics, and AI research. And our
1:28
mission is to build an AI driven radiology
1:31
ecosystem that improves diagnostic accuracy,
1:33
reduces clinical burden, and accelerates how quickly we
1:36
can get answers for children with cancer.
1:39
This is in line alignment with St. Jude's
1:41
mission of finding cures and saving children. And
1:44
in that vein, I focus on developing and
1:46
safely translating AI tools
1:48
that solve real workflow problems and deliver measurable
1:51
value in everyday
1:53
care in addition to helping accelerate discovery about
1:55
pediatric cancer and other catastrophic diseases.
1:59
Very excited to hear about more. Some of
2:01
this this that you've mentioned is very exciting.
2:03
Doctor Gandhi, over to you.
2:05
Thank you.
2:07
So I'm an interventional cardiologist.
2:10
I currently serve two roles at Philips. I
2:12
am the chief medical officer for our Philips
2:15
enterprise informatics business,
2:17
but also serve as the global lead for
2:19
medical affairs and clinical partnership teams at Philips.
2:23
Prior to joining Philips, I've been in academic
2:25
practice for about twenty plus years. I still
2:27
practice one day a week,
2:30
and my practice focused on medical education, clinical
2:33
research, and hospital strategy.
2:36
My passion is integrating technology and data to
2:40
improve care delivery
2:41
so we can provide better care for more
2:43
people.
2:44
Doctor Yi, radiology is facing a number of
2:47
challenges right now, staffing shortages, declining reimbursements,
2:50
and so much more certainly.
2:53
At St. Jude Children's Research Hospital,
2:55
what are some of the issues that are
2:57
most pressing for you right now,
2:59
and how is AI helping your team to
3:01
address them?
3:03
Yeah. So, Lucas, you're absolutely right that radiology
3:06
has a number of challenges, and they often
3:08
seem to be conflicting. We've got staffing shortages,
3:11
declining reimbursements, but at the same time, the
3:13
demand is just exploding nationwide.
3:16
But I will say that at St. Jude,
3:18
we have a little bit of a unique
3:19
environment.
3:21
We're a clinical trials hospital. It's built into
3:23
our name. We're not St. Jude Hospital. We're
3:25
not St. Jude Children's Hospital, but we're St.
3:28
Jude Children's Research Hospital.
3:30
So what this means is that by design,
3:32
we're really capped at a number of patients.
3:34
We don't have an emergency department. You can't
3:36
walk in off the street. By and large,
3:38
you have to be accepted for a clinical
3:40
trial to come to St. Jude to get
3:41
care. But what that means is that we're
3:44
able to provide care at an extremely
3:46
specialized level.
3:48
We also never send a bill to our
3:50
patients, to their families, so everything's covered free
3:53
of charge.
3:55
So I think that our challenges might look
3:56
a little bit different, at least the first,
3:59
look from a typical health system. But I
4:01
will say we still feel the same pressures,
4:03
staffing shortages.
4:05
I think every practice in the country right
4:06
now is struggling to find enough radiologists to
4:09
cover the shifts, to make sure that there's
4:11
adequate,
4:12
coverage for all of the different,
4:14
requirements for practice,
4:16
increasing image demand.
4:18
Imaging is the mainstay of so many disciplines,
4:20
including pediatric cancer.
4:22
And then finally, the need to maintain exceptional
4:24
quality,
4:25
and this applies whether you're dealing with kids,
4:27
adults, cancer, or noncancer.
4:30
So what I'll say is although our environment
4:31
is pretty unique, there's a lot of lessons
4:33
and commonalities that we do have with other
4:36
practices.
4:37
Where AI has helped us most, I would
4:39
say in the immediate term, is improving efficiency
4:41
and safety.
4:43
For example,
4:44
what we're using right now is AI driven
4:46
image reconstruction.
4:47
It's allowed us to cut certain scan times
4:49
dramatically.
4:50
For example, PET CT scans. These typically take
4:53
us around fifteen to twenty minutes. But using
4:56
these AI driven image reconstruction
4:59
tools, we can actually slash that down to
5:01
just about five.
5:03
And And what this means for young children,
5:04
this means less sedation,
5:07
better image quality. Believe it or not, even
5:09
though we're doing, quicker images, we actually get
5:12
better image quality,
5:13
related to the AI. And then overall, smoother
5:15
experience for families.
5:17
On the flip side, what this means for
5:19
radiologists is, hey. We get our images done
5:21
more efficiently.
5:22
They're easier to read. And, ultimately, we can
5:24
bring that value back to the patients to
5:26
make more accurate diagnoses and help their doctors,
5:30
make the best decisions for them. One final
5:32
note. On the research side, since we are
5:34
a research hospital, we're also using AI to
5:36
improve tumor measurement and response assessments. And these
5:39
have huge implications, especially as we move into
5:41
this era of precision medicine.
5:43
What radiologists do currently,
5:45
we draw a measurement. We say the tumor
5:47
is bigger because of this one measurement.
5:50
In reality, these are pretty coarse and crude,
5:52
so you can imagine that AI can help
5:54
us do more precise measurements maybe on a
5:56
volume level. And
5:58
this is, not something that's fully deployed clinically
6:00
yet, but it's shaping a future where radiologists
6:03
can spend more time integrating information and less
6:05
time on these repetitive tasks.
6:07
So, you know, Paul, maybe I can I
6:10
just have a follow-up question since you mentioned,
6:13
you know, you're a research hospital? You see
6:15
a lot of patients with cancer, especially pediatric
6:18
patients.
6:19
How do you utilize AI in forums like
6:22
TumorBoard where you have this
6:25
data that is spread out across the hospital,
6:27
and you're trying to bring it in one
6:28
place so you can make those decisions for
6:30
your patients.
6:32
Yeah. Well, I'll say this with a caveat
6:34
that I have a lot of
6:36
hopes and aspirations for AI in TumorBoard, but
6:39
it's not a clinical reality yet.
6:41
I do think that we have the, individual
6:44
pieces,
6:45
but putting it together, I think, will take
6:46
some time. But the way I see AI
6:48
shaping things like tumor boards, which for the
6:50
audience is where we have different doctors and
6:53
different specialties come together,
6:55
we discuss tough cases that might not be
6:57
so straightforward, and we try to come to
6:59
a consensus about what's the best treatment option.
7:02
What's the next best diagnostic
7:03
step?
7:04
And if you think about it, it's sort
7:06
of like the Jedi Council if you're a
7:08
Star Wars fan. You've got all these experts.
7:10
You got people, maybe you got Yoda. You've
7:12
got Obi Wan. You got people with different
7:14
experiences, different kind of, strengths,
7:17
different weaknesses as well where we can learn
7:19
to kind of prop each other up.
7:21
What I envision is,
7:23
rather than a council of Jedis, we've got
7:26
a council of so called AI agents where
7:28
each agent specializes in a particular field. So
7:31
you've got an AI agent that's an oncologist,
7:34
an AI agent that's a radiologist, a pathologist,
7:36
a surgeon.
7:37
And rather than doing these tumor boards where
7:39
we come together once a week,
7:41
where we're prone to scheduling conflicts and just
7:44
the, reality of limited time and resources,
7:47
these AI agents can talk to each other.
7:49
They can scour the Internet, scour the medical
7:51
literature for all of the latest and greatest,
7:54
and simulate that
7:56
interface and that, interaction
7:58
really
7:59
in the background.
8:00
You can imagine that you could do this
8:02
not just for the patients who have the
8:04
luxury of having a tumor board accessible to
8:06
them or have the good fortune of being
8:08
accepted for the tumor board, but every patient
8:10
could potentially get this, interdisciplinary
8:13
council kind of consensus.
8:16
So that's the vision. There's,
8:18
a lot of pieces to go into it.
8:20
It's a lot easier said than done. But
8:21
that's what I hope to see,
8:23
maybe in the next ten, twenty years, maybe
8:25
even sooner.
8:27
Doctor Gandhi, very poignant follow-up question. Thank you
8:29
so much for that. I was wondering, right,
8:31
as we're seeing health system expand their use
8:34
of AI in in clinical areas like imaging,
8:36
right, informatics,
8:38
what risks should leaders be aware of and
8:40
mindful of, and and how are you working?
8:42
How is Philips working to mitigate some of
8:44
that risk?
8:46
So, clearly, when you talk of AI, there's,
8:48
you know, a lot of risk. People talk
8:50
of bias, fairness, data privacy, security.
8:54
But I think if I was to focus
8:55
on two things, I would say one thing
8:57
that the leaders need to be mindful of
9:00
is transparency and explainability of AI.
9:03
For people to trust AI, you know, they
9:06
need to understand it better. Whether it is,
9:08
you know, what goes in the algorithm, how
9:10
the algorithm was trained, what the output is
9:12
relating to, and does it apply to my
9:14
patients?
9:15
So I think having that transparency and explainability
9:18
is very important as we think of
9:21
AI, especially in health care where,
9:24
you know, patients trust the system.
9:26
They wanna trust their doctors and for the
9:28
doctors to really trust the output from any
9:31
technology.
9:33
I think the second key piece in my
9:35
mind in terms of, you know, some of
9:36
the risk is integrating
9:39
AI into the workflows.
9:41
You know, we're used to a certain workflow.
9:43
And when you disrupt a clinician's workflow, it
9:46
can really lead to,
9:48
you know, bad output at the other end
9:50
regardless of how good a tool is. So
9:52
we want solutions,
9:54
that are integrated into the workflow of clinicians
9:57
so they can support them, they can augment
9:59
them, but not replace them. And I think
10:02
those are some of the key
10:05
balance
10:06
key risk that I would kind of, keep
10:08
in mind as you look at implementing
10:10
AI at scale in health care
10:12
system. Philips has
10:15
an responsible
10:16
AI approach that has eight pillars,
10:18
but I think the first and the key
10:20
pillar is around human oversight.
10:23
We want AI that supports, not replaces
10:26
clinical judgment.
10:27
Clearly, as we design these products, as we
10:30
design these solutions, we want to put patients
10:32
and clinicians
10:33
at the center of that journey so that
10:36
when these solutions come in the market,
10:38
they're not just a tool. They're something that
10:40
is
10:41
deeply integrated with the health care ecosystem.
10:45
Doctor Yi, from the application process then and
10:48
then the application piece, what guardrails
10:51
did you put in place to ensure AI
10:54
is used safely within your teams and and
10:56
in use and and responsibly throughout an everyday
10:59
practice?
11:01
Yeah. I I think that the first thing
11:03
I'll say is that every new AI model
11:05
or use case really requires a fresh look
11:07
and, frankly, a fresh sense of humility.
11:10
We don't assume that what worked last time
11:11
is gonna teach us everything we need to
11:13
know about the next one.
11:15
Just the two examples I mentioned, when we
11:17
talk about things like image reconstruction,
11:19
that's a very different thing than talking about
11:21
tumor measurements on images. You know, one of
11:24
them has everything to do with
11:26
kind of the, behind the scenes, when we
11:28
actually get the images, when we reconstruct them,
11:30
when we send them to the radiologist.
11:32
And then the second part has everything to
11:33
do with what the radiologist actually does.
11:36
So that being said,
11:38
when I think about how do we ensure
11:40
safe and responsible use of AI in everyday
11:42
practice, I really think of three pillars. The
11:45
first one is awareness.
11:46
I think
11:47
the adage goes, you can't solve the problem
11:49
that you're not aware of. So the first
11:51
thing we do is we deeply study and
11:53
are aware of what are the known risks,
11:55
whether it's algorithmic bias and fairness,
11:57
brittleness to everyday variations and how images are
12:01
acquired,
12:02
or maybe in human computer interaction, looking at
12:04
things like automation bias and overconfidence.
12:07
I think second, after we're aware of what
12:09
we're looking out for, we can then figure
12:11
out how do we best measure them.
12:13
Some of the things that we do is
12:14
build validation pipelines
12:16
that look not just that
12:18
how do we measure these things from a
12:19
statistical standpoint, but also ask, will these things
12:22
bear out clinically?
12:24
One example is that I like to use
12:27
is if you have a nationwide sample and
12:29
you're looking at something like a quality metric,
12:32
you might be able to say that, hey.
12:33
A point 001%
12:34
difference is statistically significant,
12:37
but I think it's questionable if that actually
12:39
means anything clinically. So I think it's important
12:41
that whatever we measure, it's placed in the
12:43
clinical context. When we think about how does
12:45
this impact radiologists
12:47
interpreting the study, how does this impact the
12:49
patient's outcome, Maybe how does it impact the
12:52
patient's experience?
12:53
And then I think finally, once you have
12:55
an idea of what you're measuring, how to
12:57
measure it, then it's figuring out a governance
12:59
system.
13:00
Meaning, how do we have a process and
13:02
a framework where we can actually operationalize
13:04
these,
13:05
metrics?
13:06
And so every tool that we use at
13:08
St. Jude goes through interdisciplinary review, and this
13:10
includes radiologists, physicists, informaticists,
13:14
IT and clinical operations.
13:16
And the idea is, again, this council kind
13:18
of analogy is
13:19
none of us know everything. And in fact,
13:21
it really does take a village because I
13:23
might be a radiologist interpreting studies, but there's
13:25
a technologist who's interacting with the patient and
13:27
actually bring them onto the scanner.
13:29
They're actually acquiring the images.
13:32
There's gonna be a physician at the bedside
13:35
communicating these results to the patient themselves.
13:37
So I think it's important that we keep
13:39
in mind everyone's unique vantage point, their unique,
13:42
perspectives.
13:43
And that combination, I think, is enough to
13:45
keep us humble enough to recognize new risks,
13:47
but also confident enough to deploy AI safely
13:50
where it adds real value and feel like
13:52
we've got enough,
13:54
checks and balances to make sure we're,
13:57
pushing the boundaries, but doing so in a
13:58
way that's gonna be safe and trustworthy.
14:01
Yeah. We come back to the council approach.
14:03
Right? There is a lot of folks involved,
14:05
a a governance council, so to speak. A
14:08
lot of organizations are are
14:10
employing them right now to be able to
14:12
do exactly what you just described, which is
14:14
really, really key.
14:16
I wanna close this off with a little
14:18
bit of a forward looking perspective here on
14:20
things, and I'm really excited about this question
14:22
because it's always yields great responses, I feel
14:25
like. How do you both envision
14:28
AI
14:29
reshaping
14:30
radiology over the next three years? And most
14:32
importantly, what excites you
14:35
the most about that future? Doctor Gandhi, we'll
14:37
kick it off with you. So I don't
14:39
know, Lucas, if you're a glass full or
14:41
a glass healthy kind of a person.
14:44
Glass full.
14:45
So it's it's it's always full. No. I'm
14:48
actually there's a lot of hype, around AI,
14:51
but, truly, there's a lot of hope, as
14:53
well.
14:54
I believe that the next few years would
14:56
be truly transformative.
14:58
AI will move from isolated tools that we
15:01
have now where people are looking at isolated
15:04
algorithms for implementation
15:06
to deeply integrated systems that can automate workflows,
15:10
enhance decision making, and personalized care.
15:13
So if you imagine end to end automation
15:16
from image acquisition to reporting,
15:18
plus the kind of multi agentic work that
15:21
doctor Yi was describing
15:23
with multiple agents working together to provide advanced
15:26
insights at point of care. And what this
15:29
would mean for the health care system would
15:31
be less administrative burden, faster diagnosis,
15:34
and more time for clinicians to truly focus
15:37
on their patients.
15:38
To answer your question about what excites me
15:40
most, I I think I'm most excited about
15:43
AI's ability to enable proactive personalized care
15:47
while keeping clinicians and patients firmly.
15:50
Doctor Yi, same question for you. How are
15:52
you seeing this evolve? And then, most importantly,
15:54
what are you excited about?
15:57
Yeah. Well, first, I think Sanjay,
15:59
really put it super well. I I think
16:01
that it will be transformative. I think AI
16:04
is gonna reshape every step step, not just
16:06
the radiology
16:07
life cycle, if you will, but really the
16:09
entire patient care ecosystem over the next few
16:11
years.
16:12
When I think about this as a radiologist,
16:14
I don't think it's just gonna
16:16
impact image interpretation,
16:18
but everything from planning the studies, figuring out
16:22
when to schedule a patient, how to protocol
16:24
or get figure out which type of study
16:27
to obtain,
16:29
down to image acquisition reconstruction,
16:31
reporting, follow-up, and communication ultimately with patients and
16:34
clinicians.
16:35
What excites me most, is not just automation,
16:38
but I would say twofold. One is augmentation.
16:40
The idea that AI can help clinicians do
16:43
more than what we could do before.
16:45
For example, there's a lot of things that
16:47
I do in my daily practice
16:48
that are pretty rote. I have to drag
16:51
and drop images into a particular
16:53
arrangement on my screen. I can imagine that's
16:56
not exactly the most fun part of my
16:57
day. If AI can do this, it just
17:00
gives me more brain space and bandwidth,
17:02
the focus on the things I enjoy and
17:04
the things that honestly are gonna make a
17:06
difference in a patient.
17:07
It's gonna be things like actually interpreting the
17:10
images, making sure that we're getting the right
17:12
diagnosis once we found the findings.
17:14
I think also in addition to augmentation, I'm
17:16
exciting excited about this
17:19
possibility of orchestration.
17:21
We've talked about agentic workflows. I think it's
17:23
incredibly exciting, as Sanjay put really well,
17:26
that these AI tools have been very, shall
17:28
we say, isolated.
17:30
They do things extremely well, but they're kind
17:32
of like one trick ponies. They might be
17:33
able to find a pneumothorax on a chest
17:35
x-ray, or they might be able to tell
17:37
you if a report contains an incidental finding
17:40
to get follow-up on. But with these agentic
17:42
workflows, we're gonna have something closer to a
17:44
true ecosystem
17:45
where it's not just, hey. Here's a tool.
17:48
It does this extremely well. But here's a
17:50
tool that does it extremely well, and now
17:52
this other AI agent can make sense of
17:54
it and decide on the next step.
17:56
And that might sound like a little bit
17:58
scary, but I think that, again, if we
18:00
use these safe and trustworthy
18:02
kind of, guidelines,
18:04
we really have an exciting future ahead.
18:06
And so,
18:07
if I had to summarize it, I think
18:08
we're moving from point solutions to real AI
18:10
ecosystems
18:11
that are gonna really make a big difference.
18:15
Yeah. Beautifully said. I I wanna close this
18:17
off again with with giving the floor over
18:19
to you both.
18:20
Doctor Ghani, we'll start with you. Anything else
18:23
to add that we didn't touch on? Any
18:24
final thoughts you'd like to share?
18:26
I think one thing that we didn't touch
18:28
on is that none of us can do
18:30
it alone. I I know we came tangentially
18:32
where Paul said, you know, we all need
18:34
to work together, whether it is academia,
18:36
whether it's industry, whether it's regulators,
18:39
whether it is private sector,
18:41
and technology companies. I think the future of
18:44
AI to be successful
18:46
needs to be collaborative.
18:48
I think it needs to be data driven,
18:50
and it needs to be patient centered.
18:52
So if we do it well, we can
18:54
truly democratize
18:56
access. We can reduce disparities
18:58
and truly empower physicians. And this is, you
19:01
know, what we are truly committed to do
19:03
as Philips.
19:04
Doctor Yi, anything else to add from your
19:06
end here? Yeah. I I'm incredibly optimistic about
19:08
the future of AI and imaging,
19:10
but I I really think it takes a
19:12
village. You know, I believe that the organizations
19:14
that will succeed are ultimately the ones that
19:16
bring together
19:17
clinicians, engineers,
19:19
IT professionals, physicists, administrators, industry partners, all of
19:23
the above.
19:24
But really, collaboration is how we turn AI
19:26
from exciting prototypes into tools that genuinely improve
19:29
care. And so I think the technology is
19:31
incredibly exciting.
19:32
I think that
19:34
at face value, one could say that it's
19:36
ready. But now it's about people, process, and
19:38
collaboration
19:39
and making it work in this amazing health
19:41
care ecosystem that we're blessed to be a
19:43
part of.
19:44
Doctor Yee and doctor Ghani, thanks so much
19:45
for being here today and taking some time
19:47
for us. What a great conversation. Thanks for
19:49
being here. Thank you, Chris. This was a
19:51
wonderful conversation, and thank you, Paul. Yeah. Thank
19:54
you both. Absolutely. And we also want to
19:56
thank our podcast sponsor, Phillips. You could tune
19:58
into more podcasts from Becker's Healthcare by visiting
20:01
our podcast page at beckershospitalreview.com.