Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:01):
Millions of people go
blind from diabetic retinopathy
, but it is potentiallypreventable.
So what if AI could catch itearly just from a photo of your
eye?
Today we're unpacking the techthat could change the future of
diabetic vision care.
Speaker 2 (00:24):
Hello and welcome.
My name is Vasant Sarathy, I'man AI researcher and I'm with
Laura Hagopian.
Speaker 1 (00:32):
I am an emergency
medicine physician who now works
in digital health.
Speaker 2 (00:36):
We're very excited.
This podcast is going to be allabout unpacking how artificial
intelligence and emergingtechnologies in general are
going to be transforming the waywe think about our body and our
mind, and today is going to beour first episode.
Are you excited, laura?
Speaker 1 (00:53):
I'm excited.
I chose the topic, so I'm veryexcited.
Speaker 2 (00:56):
What are we talking
about?
Speaker 1 (00:57):
We're talking about
diabetic retinopathy.
That's right.
Are you excited?
Speaker 2 (01:01):
I'm excited.
I'm pretending like I didn'thear this before, but I am very
excited.
All right, what is that?
Go ahead, get started.
What is diabetic retinopathy?
Speaker 1 (01:10):
Well, this is a
complication that people with
diabetes can have, and it's aleading cause of blindness
around the world.
So it's a serious problem, andthere are lots of great
treatments available for it, andbetter to get treated before
you maybe even have symptoms.
(01:30):
But the problem is that manypeople with diabetes don't
actually get screened.
They don't actually get theregular eye exams to identify
diabetic retinopathy, and sothen, if they don't have the
exams, we don't know they haveit.
It's harder to preserve theirvision, and so that's why I
(01:52):
think this is an important topicfrom the clinical side of
things.
Speaker 2 (01:57):
But then the question
really is why is it that people
are not getting appropriatelyscreened or screened fast enough
?
Speaker 1 (02:05):
Yeah, I mean it's a
problem around the world.
Right, there aren't necessarilyenough ophthalmologists or
optometrists to do these dilatedeye exams.
But beyond that, a lot ofpeople aren't aware that they
need to happen and it can behard to go to eye appointments.
Right, you have to travel.
(02:25):
It costs money.
You have to take time off work.
It's just another thing to addto the list of all the things
you have to do with diabetes.
Some clinics may not have theright equipment, they may not
have the right support staff,they may not have the right
resources, and telehealth isused for this to some degree too
, but there may not be enoughtrained, trained physicians or
(02:50):
professionals to really gradethese images yeah, it seems like
this is a particularly toughproblem in low resource areas or
places that are hard to get toor are poorer or developing
countries and such and um, itdoes seem like it's.
Speaker 2 (03:07):
You know, like you
mentioned, there's not enough
people to do this, which isoften the starting point for
thinking about an automatedsystem to consider, uh, not
replacing people necessarily,because that's not what we're
interested in, uh, but helpingpeople and potentially providing
, you know, sort of additionalhands on the problem and
(03:27):
assisting people who areactually able to do this at a
larger scale.
So, you know, I think that'skind of where I think both of us
talked about this this pastweek and we thought about how,
you know, ai can in fact helpthis, and we're obviously not
the first people to think aboutthis.
There's quite a bit of researchthat people have already done
and even systems that peoplehave already built that have
(03:51):
started to target this issue.
So we thought this would be agreat topic, because it's an
important problem and it'ssomething that maybe AI can
actually solve effectively andhas already started making an
impact and has already startedmaking an impact.
Speaker 1 (04:06):
Yeah, and Vasant,
you're touching on a really
important point here, because itreally does need human-AI
collaboration to develop asolution like this.
Technology here isn't aboutreplacing ophthalmologists, it's
about augmenting theircapabilities.
So there's this synergy that wecan see between AI and human
clinicians, where possibly theAI could handle the mass
(04:29):
screening, identify the highrisk patients for vision loss,
and then the humanophthalmologists can focus their
expertise on those complexcases.
Speaker 2 (04:42):
Almost like a
triaging tool at some level.
Is that right?
Speaker 1 (04:45):
Yeah, exactly Almost
like a triaging tool.
So I think it's probablyimportant to talk about how have
we traditionally done this?
How have we traditionally goneabout screening people for
diabetic retinopathy?
And the traditional method iswith what's called a dilated
fundus exam, and that's when youhave someone who's well-trained
(05:08):
who dilates the eye and takes alook inside the back of it, and
that requires an expert right.
It requires someone who reallyknows what they're doing to be
able to see into the back of theeye.
It requires you to go to aphysical appointment to have an
ophthalmologist or optometristlook in the back of your eye.
(05:30):
So that's sort of thetraditional method.
But there are other methodsthat are being used.
Retinal photography is anothermethod that's being used and
that can be done, sometimes withor sometimes without dilating
the pupil, and then that's sentover via, you know, telehealth
computer to an ophthalmologistwho doesn't necessarily need to
(05:53):
physically be there anymore.
So that's the second method ishaving an ophthalmologist look
at it, but not physically be theone looking into your eye by
taking a photograph but notphysically be the one looking
into your eye by taking aphotograph.
Speaker 2 (06:06):
So can I ask a really
silly question here?
Is it the case that is there adifference between the physician
actually looking at a real eyeversus looking at an image of an
eye?
Is there some extra benefitgleaned from the physician being
there?
From this triaging standpoint?
I'm sure there is, obviouslyfrom a medical standpoint, but
(06:30):
from this sort of thinking aboutit from a purely triaging
standpoint, can effectivediagnosis happen from physicians
looking at images?
Speaker 1 (06:36):
Yeah, I mean the
simple answer is yes.
You know they've done studiescomparing these two things like
the optometrist orophthalmologists doing that
dilated look at the back of theeye, comparing that to retinal
photography and they had goodsensitivity and specificity.
I mean the studies look at.
Okay, at least 86% of cases hadagreement and when there were
(07:02):
disagreements, it wasn't likesomeone was necessarily missing,
that there was retinopathy.
It might've been a disagreementabout the grade, because you
can grade it as mild, moderate,severe, et cetera.
So for the most part, retinalphotography works really well.
Which begs the question do weneed a expert professional to be
(07:27):
looking at these photographs orcould that be piece of the
triage be automated so that wecan use an AI system?
We could use a computeralgorithm to continually learn
and update and detectretinopathy.
And for the people who haveretinopathy, now you refer them
(07:47):
into care, now you get them tosee an ophthalmologist so that
you can get the treatment onboard, whether that's laser,
vegf et cetera.
There's a lot of treatmentsavailable, but if we want to
scale something like this,potentially the AI can do that
and evaluate the retinal images.
Speaker 2 (08:07):
Yeah, that does seem
like again an opportunity for AI
systems and particularly feelsa little bit I'm feeling a
little bit more reassuredhearing that things can be
effectively done with images.
Now, this is not the case forall medical diagnosis, as we
know.
There's lots of situationswhere just piece of data like an
image or a write up orsomething is not enough.
(08:30):
But in this particular instance, for this particular purpose,
which is diagnosing and gradingdiabetic retinopathy for the
purposes of as sort of as atriage, it seems like a lot of
the boxes in terms of AI risk,you know, kind of say boxes, but
(08:51):
a lot of points about AI riskis mitigated because we worry
less about this.
We want to catch those casesearly on and anything we can do
in that front is going to bereally helpful.
And so, yeah, no, I think thatthat's exactly where I it's
tough because there's a lot ofhype about AI right now and
(09:12):
there's a lot of talk aboutgenerative AI and chat, gpt and
LLMs and all those things, butin this particular instance,
that is not the kind of AI we'relooking at.
Speaker 1 (09:23):
Oh, okay, tell me
more.
What kind of AI are we lookingat here?
Well, you, know.
Speaker 2 (09:27):
I think it's
worthwhile spending a few
seconds just thinking about AIbroadly, because AI is a term
that's heavily used right now inpublic speak and it's a
buzzword.
It's a powerful word signifyinga lot of things.
Currently it signifies chat GPTworld of AI, but before chat
GPT which chat GPT is based onwere neural networks, and neural
(09:47):
networks were used for lots ofthings like detecting photos in
your Facebook picture.
I mean, that's a form of AI,it's a form of computer vision
where an image is taken and anAI system or a neural network
let's be more broad about thatterm is able to detect certain
features in an image.
Which kind of applies here Ifyou can sort of see the
(10:09):
connection between Facebookphotos and diabetic retinopathy.
But there is a connection.
Speaker 1 (10:15):
No, I get it right,
because they're images and so
you're using image recognitiontechniques of some sort.
Speaker 2 (10:22):
Yes, and image
recognition techniques is one
sort of school of AI which allfalls into this larger school of
what's called machine learning,where you have machines
learning from data, and againthere's a lot of interesting
discussions to be had that maynot be part of this conversation
necessarily about what AI means, but here we're sort of focused
(10:44):
much more on image recognition,image classification, things
like.
There are many subtopics inthis space.
So, for instance, imageclassification might be able to
say this image is an image of adog or an image of a cat, and so
on.
Image recognition might get alittle bit more specific and say
there is a table over here inthe image and draw a little box
(11:07):
around the table, and then youhave image segmentation which
says oh, you know, this part ofthe image is the foreground and
this part is the background, andsort of drawing these blobs to
represent which parts of theimage you care about.
So there's many subtopics here,all of which apply here to
diabetic retinopathy, which thenyou know, I think the reason we
started this podcast, I thinkwas because the connection
(11:28):
between sort of the medicalaspect and the AI aspect is so
tight that in order to developgood AI systems you do need a
fairly deep knowledge in whatthe medical side of things, the
thing you're looking at, theobject you're studying.
So to that end you know, yes, Ican talk about faces and tables
and such, but in a diabeticretinopathy situation, when
(11:49):
you're looking at a fundus image, what are the types of things
that one looks at?
Like, what are you looking at?
Can you give us a sense forwhat that image is?
Speaker 1 (11:58):
Yeah, absolutely.
I mean, there are a lot ofobservable findings that are
used to kind of classify orgrade the diabetic retinopathy.
So you're looking in the backof the eye for things like
microaneurysms, but are therelittle hemorrhages inside the
(12:19):
back of the eye?
Is there beating of the veins?
Are there new blood vesselsforming?
Is there hemorrhage in otherlayers?
Is there actually anythickening of the macula or the
retina at all?
So those are the things thatyou're looking for, and it's
(12:41):
something that ophthalmologistsare certainly trained to do.
Is that something that you cantrain an AI model to do?
Speaker 2 (12:50):
Yeah, I mean that's
what people have started to do.
When I started looking at this alittle bit more, there are
these conceptual diagrams thatpeople have in their papers
showing um, the back of theretina and what is a version of
the fundus, and you can sort ofsee these like um, almost
tree-like structures that are umon this on.
You know, you sort of have anorange background imagine an
(13:11):
orange background and like redtree-like structures coming out,
blood vessels, aka bloodvessels, um, and you have other
things in the image.
So you don't not only havethose tree-like structures in a
typical fundus image, but youhave other things that are
different color and the colorsmay not mean much in terms of um
, uh, the real um, fundus, butfor our purposes we can sort of
(13:33):
distinguish them.
There's all these fuzzy areas,areas that are in the middle of
the trees.
There are these other regionsthat are different colors,
slightly different colored, butthey're also kind of fuzzy
looking.
So you have this sort of almostlike world inside of this that
has all of these features andthat's something that image
systems can recognize if, infact, we have enough data to
(13:54):
allow them to learn that abilityto recognize that.
Speaker 1 (13:56):
Okay, wait, I want to
stop you there, because what
does enough data mean?
How do you actually, how do youtrain the AI to recognize, like
I know?
As a physician, I'm trained torecognize X, y or Z.
If someone comes in withcertain symptoms, I kind of know
what to look for.
But how do you train the AI onan image?
How many images do you need totrain it?
(14:18):
How does it know to figure outwhat's mild, moderate, severe,
what's proliferative diabeticretinopathy?
That's something that, like asa physician, we might be trained
on right or ophthalmologistswould be trained on, but how
does an AI system know how torecognize that from images?
Speaker 2 (14:37):
Yeah.
So when humans look at theseimages, especially trained
humans, you're looking ateverything at the same time.
You're looking at what thethings are that you're seeing.
You're also looking atanomalies and oddities in the
image at the same time as you'relooking at what grades you have
to give it, and so on.
With an AI system it's a littlebit more specific.
So you're specific and computerscientists think about this this
(14:59):
way, in terms of specific tasks.
So, for instance, maybe thecomputer is trained only to
recognize all of the bloodvessels.
Maybe the computer is trainedonly to recognize all of the
blood vessels.
Maybe the computer is trainedonly to recognize a hemorrhaging
blood vessel.
And in order to train that, forinstance, you would give it
images of blood vessels that arenormal versus blood vessels
that are hemorrhaged, and thoseimages are going to look
(15:21):
different and those differencesare represented in the pixels of
the image.
Pixels then convert to numberswhich the AI system is able to
process, and those differencesare represented in the pixels of
the image.
Pixels then convert to numberswhich the AI system is able to
process and a neural network,which is often what's used for
these purposes.
You can think of it as a machinewith a lot of knobs, and what
you do is you give it theseimages, pixels, which, like I
(15:44):
said before, is converted tonumbers, and those numbers you
also give it the right answer.
So you also say this is ahemorrhage or this is a regular,
normal blood vessel, and so on,and those images and the right
answers allow the machine to gosort of back and forth and
adjust its own knobs and onceit's done, it now has, you know,
the machine has all the knobsettings to be exactly what it
(16:05):
needs to be, so that when a newimage comes in, it's able to
tell you one or the otherwhether it's hemorrhaging blood
vessel or it's just a regularblood vessel.
And this that's at the core ofthe neural networks itself.
Now there is a whole bunch ofdifferent methods and
architectures and ways ofarranging these neural networks
that make one better than theother for different purposes,
(16:25):
and researchers have exploredthat, you know, in various
fields, but also specificallyfor diabetic retinopathy as well
.
But that's just image detection.
And the grading is anotherissue as well, which is, instead
of just saying whether it'shemorrhaging or not, you could
give it images and you couldtell it this is severe, this is
moderate, this is light, okay.
Speaker 1 (16:42):
And we know that
right, like if there's, if
someone has no signs ofproliferative retinopathy and
they have more than 20intraretinal hemorrhages in each
of the four quadrants,indefinite venous beating in two
or more quadrants, those arethe things that make would make
us say, okay, that's severe.
So if we're able to sort ofhave an algorithm from it on the
(17:02):
clinical side, that's somethingthat could be translated into
sort of the AI interpretation ortriaging of what's going on in
this image?
I have a question, though so youtrain it on a set of images and
then does it keep learning fromnew images, or it's just that
initial set that it was trainedon?
Speaker 2 (17:21):
It's often the case
that you have an initial set
that it's trained on and it'sdeployed and then it's no more
learning, it's not not learninganymore.
But that doesn't mean that thepeople who are developing it are
not making it better.
Uh, they could be taking in thedata and working and updating
their models, updating the, thetraining itself, and then
releasing new versions that are,uh, trained to allow for new
(17:42):
data to come in.
And actually, the new data isan interesting point you
mentioned because I didn't quiteget into this, but it's you
know and we can.
Next, which is the idea of thedata drifting and changing over
time.
So, if you look at the results,there's a lot of techniques out
there and big companies likeGoogle have tools out there, ai
(18:03):
tools out there that can do thiseffectively.
And if you look at the toolsout there AI tools out there
that can do this effectively andif you look at the results that
have been reported, they'revery high accuracy, very high
performance.
So you're looking at modelsthat can detect and grade these
systems at over 98, 99% accuracy, which obviously is massively
(18:23):
helpful from a triagingstandpoint.
Speaker 1 (18:25):
Right For the more
severe forms of diabetic brain
apathy.
Speaker 2 (18:27):
That's right, that is
exactly right.
So one of the challenges iswhat do you do when you have
weaker forms?
And also, relating to the pointwe mentioned earlier, what do
you do when it's a littledifferent, when things start to
change a little bit?
I don't know from a medicalstandpoint what that actually
means, but what ends uphappening is these models.
If you remember the knobturning example I gave you
(18:50):
earlier, all it's doing islearning patterns in those
pixels, it's figuring out andassociating patterns it's seeing
in those pixels with namedgrades or named conditions, and
so that's all it's doing.
It's identifying patterns.
So that's all it's doing, right, it's identifying patterns, and
so that means it assumes thatthe new pattern it's going to
see is going to be similar tothe pattern, the collection of
(19:11):
patterns it's seen before.
But now if, instead, if you havea new pattern, that changes and
scientists and computerscientists call that out of
distribution, when you havesomething new that was not in
your original pattern collectionor distribution of models,
distribution of images and sothat new thing is now going to
be difficult to classify.
(19:33):
And this is a really funnyexample, actually, of this that
came to mind, which was that thesomebody trained a COVID lung
images to distinguish betweenimage detector, to distinguish
between COVID and non COVIDlungs, and it did great and it
performed very well, and yougave it a COVID lung.
It would tell you that and soon.
And then to that same system.
Somebody gave it a picture of acat and it said with 100%
(19:56):
confidence that it was a COVIDlung.
Now to the machine.
It doesn't know that somethingis a cat.
It's just looking at pixels,right.
But that's an example of acompletely out of distribution
image that it has to give you ananswer.
It doesn't say I don't knowwhat is this, it just says it
has to give you one of the otheranswer cat or, you know, covid
(20:16):
lung or not, covid lung, cat isnot even in its vocabulary okay,
so like, basically, you'retelling me we should not send
pictures of cats to the ai.
Speaker 1 (20:27):
That interprets
diabetic retinopathy.
Speaker 2 (20:29):
That is correct.
Speaker 1 (20:30):
But can I play this
back to you in all seriousness
If, for example, these AI modelswere trained on people, say in
the United States, and then youtook something like this to a
different country, Could thatchange in geography or change in
that screen population orchange in ethnicity?
(20:54):
Could all those things affectthe algorithm so that you would
need to sort of reprogram it ortitrate it or change it a little
bit, based on using it in adifferent location with a
different set of people?
Speaker 2 (21:12):
It could.
But, that said, it doesn'tdepend.
It depends much more on thedifferences in the images
themselves.
You know, I think one way tothink about it is if a doctor
looking at an image in theUnited States you know is able
to do it correctly, if youtransported that doctor to
another country, what changesfor them when they look at those
(21:33):
images?
Do those images lookqualitatively different,
quantitatively different in some?
Speaker 1 (21:38):
way.
Speaker 2 (21:39):
You know, do those
cultural differences or genetic
differences play out in theactual pixels, in the actual
image?
Speaker 1 (21:46):
Yeah, I got it okay,
and you're actually bringing up
another really interesting point.
Um, I don't know if you meantto or not, but I'm gonna, I'm
gonna hook on it which is likeoh, the camera that you use is
important, right?
Um, like the manufacturer, themodel, if there's wear and tear,
if there's a smudge on it orit's also and this is true no
matter where you do it but it'sthe technician who's taking the
(22:09):
images.
I know there were some in someof these papers we looked at
where the AI was like oh, wecan't grade this image, we can't
tell you.
You know, is this mild,moderate, severe?
Like you have to go see anophthalmologist or an
optometrist because we can'ttell you?
And potentially, if you don'tdilate the eye or if you have
someone with less training orless experience or the room
(22:32):
lighting is off, then all ofthose things can affect the
image that is produced.
Speaker 2 (22:37):
That is a great point
.
So we've been sort of talkingabstractly about culture and
genetics and so on asinfluencing how it manifests.
But in reality what the machineis looking at is a bunch of
pixels which is sourced from acamera which is placed in a
location that is influenced bylighting and all these other
things and the quality of thepixels being produced.
(22:58):
And all of that plays into thecorrect image detection and
performance.
And so, yes, these systems canperform really well on datasets
that are nice and clean and wecan show 98% performance.
But the real test is beingdeployed in a real situation,
with the cameras available tothose people and seeing if these
systems can perform well inthat Now, as a developer or a
(23:21):
computer scientist or an AItechnician, you can sort of
think about this in advance andsay, okay, these are the cameras
, this is the kind of settingthose people are going to be at.
Let's create a data set in ourlab that, you know, mimics that
situation.
You want to capture as much asyou can of the real problem that
you're solving so that yourpixels look very similar to the
real situation.
But that's a difficult problemand it's not straightforward.
(23:43):
But you know, I think thisparticular application of
diabetic retinopathy at leastfrom the literature it seems
like there's people havedeployed it in these places
successfully.
I think that this is an exampleof a successful use of AI in a
medical setting.
Speaker 1 (24:16):
And some of the stuff
we were looking at was, yeah,
in sort of more resource limitedsettings, and it's like you
don't want to overburden theophthalmology services there
because there just aren't enoughof them.
But even in the United Statesthis is underscreened, and being
able to have an image taking atlike sort of a routine
appointment would be amazing forso many people to figure out.
Okay, do I actually need to bescreened or not?
Do I actually need to go in andsee an ophthalmologist or not?
Do I need laser or othertreatment or not?
Speaker 2 (24:39):
Absolutely.
I mean, and we, you know, welooked at a little some of the
literature on this and some ofthe new stories on this, so
we'll link all that in the shownotes as well.
So people can, you know, kindof dig deeper, to the extent
that they want to dig deeper, um, but I think we are, um, you
know.
So what are your sort oftakeaways here, laura?
(24:59):
What do you think are some ofthe things that, uh, folks can
take away from this example ofthe use of AI in a medical
setting?
Speaker 1 (25:08):
Yeah, I mean, I think
this is something that could be
really big right.
This is something that, fromthese early studies, looks like
okay, we're actually ready forprime time in some locations.
Obviously, in the US there aresome, are some more regulations
etc by the fda, and ongoingresearch is something that's
(25:29):
needed right to prevent drift.
But this is a great example ofhow the technology can be used,
how the ai can be used tobasically collaborate with
humans, and we know that notenough people are screened and
we know that this is a hugecause of blindness.
So if AI could handle that massscreening, if AI could identify
(25:53):
people who are at high risk andget those people into care, it
could reduce the workload onproviders.
And then you would have thissecond step where you need a
human, obviously in the loop forthings that are complex or
things that are ungradable, tooptimize efficiency and accuracy
(26:14):
and then decide on next steps.
If someone needs treatment,they really do need to see a
provider.
And so the AI is one step inthe process here.
It's not the end step, but it'ssomething that can help scale
screening for this significanthealth condition.
Speaker 2 (26:34):
That's great, I mean,
yeah, and so I want to sort of
close on this note that we haveonly touched the surface of this
really broad and very excitingto be honest, exciting topic
within the medical AI space, andobviously there's a whole bunch
of questions, both from atechnical standpoint how what
are the different AItechnologies that are used?
(26:55):
What is the next step coming up, what are the biggest AI
challenges but also medicallyspeaking.
You know, how do we get thisout there to people?
How do we get this deployedeffectively?
There's so many questionsaround this and I would love to
hear from you guys if you knowwhat questions you want answered
or is there anything that isparticularly that you want us to
dig deeper into?
(27:16):
But I think that you know we'vesort of scratched the surface
and it looks really interestingunderneath and I'm very excited
that it's actually being used asan AI person myself.
Really interesting underneath,and I'm very excited that it's
actually being used as an AIperson myself.
Yeah, and so I, you know I wantto do.
You have anything else to addon to this, laura?
Speaker 1 (27:34):
No, I'm, I'm just
excited.
I convinced you to do thistopic because you weren't so
sure at first, and now I thinkyou're fully converted.
That it was.
It was a pretty cool topic,yeah.
Speaker 2 (27:46):
I really am, because
we hear a lot of stories about
AI being used in differentsituations and people are not
sure.
People are nervous and peopledon't trust it and they should
be cautious and they should nottrust everything that you see
out there.
But here's an example where youcan, and it's exciting and
there's so much more to read.
We'll put most stuff on theshow notes, but that's it for
(28:07):
now.
Thank you so much for joiningus and see you next episode.