November 4, 2025 • 36 mins
In this episode of MedStar Health DocTalk, host Debra Schindler sits down with Dr. Nicole Sakla, a breast radiologist and published researcher in artificial intelligence, to explore how AI is changing the way mammograms are read and how it can help radiologists detect disease earlier, sometimes by years.
Learn how AI tools flag subtle patterns invisible to the human eye, act as a double-check system for radiologists, and may reduce unnecessary callbacks and missed cancers—all without replacing the human expertise behind every diagnosis.
Dr. Sakla also clears up misconceptions about AI in medicine, explains how it’s trained to be accurate across diverse patients, and shares what excites her most about the next generation of AI technology in breast imaging.
Listen now for a fascinating, accessible look into how human expertise and machine learning are working together to advance early detection and patient care.
For more information about breast imaging and screening at MedStar Health, medstarhealth.org/breasthealth or call the breast center at MedStar Washington Hospital Center: 202-877-2800 .
If you would like to comment on this podcast, or recommend a topic for another episode of DocTalk, send an email: DocTalk@medstar.net
For more episodes of MedStar Health DocTalk, go to medstarhealth.org/doctalk.
Listen
Watch
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
(00:01):
- Comprehensive, relevantand insightful conversations about health
and medicine happen here.
When Health DocTalk,these are real conversations
with physician experts fromaround the largest healthcare
system in the Maryland DCregion When it comes to
detecting breast cancer,we know that finding it early
makes every difference inwhat happens next in the journey.
(00:26):
For decades, radiologistshave relied on mammograms
and their own trained eyesto find even the smallest
signs of disease.
But now a new tool is changing the gain.
Artificial intelligence fromflagging subtle patterns
invisible to the human eye,to helping radiologists read images faster
and more accurately.
(00:47):
AI is beginning to play animportant role in breast cancer
detection and diagnosisearly, perhaps even by years
to better understand what people knowand how they feel about ai.
In breast imaging,Health conducted a national survey
of more than a thousand adultsand found only one
third of women surveyed,and were aware that AI was even being used
(01:08):
to help doctors read mammograms.
In today's episode ofMedStar Health DocTalk,
we are taking a closer look athow this technology actually
works, what it meansfor accuracy and detection,
and how it can ease, notadd to patient anxiety.
Welcome Health Radiologist, Dr.
Nicole SLA to help us separatefact from fear when it comes
(01:29):
to artificial intelligenceuse in mammography.
I'm your host Schindler, Dr.
Sakla. Thanks for joining me.
- It's an absolute pleasure.
Thank you for having me.
- So you have an extensive backgroundof published medical research,
including in artificial intelligence usefor early breast cancer detection in some
of the most prestigious andhigh impact medical journals.
(01:50):
So while many of us arejust learning about chat GBT
and artificial intelligence,and we haven't really heard of this,
you were already learningabout it using it in diagnosing
breast cancer, you've beenstudying it for years.
When was the writing on the wall for you?
- So it's interesting. Thankyou again for having me.
So AI has been on the docket,some people would say since
really around the eighties,but even there's been early research done
(02:13):
even prior to that.
For me personally, I had my first exposureto artificial intelligence in
medicine when I was a residentin New Jersey, a peer
of mine who's named Dr.
Duggan Deep Singh, and he was the onewho first introduced me
to artificial intelligenceand its applications in medical imaging.
What I started to do was to researchhow can AI help the doctor,
especially the radiologist,identify pathology with respect to CT MRI
(02:40):
and mammography better and faster.
As I got further on in my trainingand I started to see, okay, wait, this,
this technology may actually enable usto catch small findings that
may potentially not be ableto be seen with the naked
eye, particularly with MRI.
When I ended up going intowomen's imaging, the application
of AI for breast cancerbecame even more evident,
(03:01):
and it has really, really astounded mehow far we've come in just
even the last five years as faras the development of
artificial intelligencealgorithms and techniques.
- Let's start with the basics.
We hear the term AI all the time.
What does it actually mean inthe context of breast imaging?
- So that's a good question.
I think when people hear theword ai, they automatically
(03:21):
think of like the Jetsons,and they think that there's
gonna be like a machine poppingout of the wall and it's gonna do the job
of the doctor for you,- Right? Exactly.
- And that is what I'm actually tryingto dissuade people from
thinking of AI in that manner.
Here's why AI is moreof an assistance tool.
It's not meant to replace the radiologist.
So when it comes to a breastradiologist specifically
(03:44):
and how we use AI to findbreast cancer, we have to kind
of divide our thinkinginto two main categories.
We have the people who come infor their screening mammograms,
and then we have the womenwho come in for what's known
as diagnostic examinations.
So on a screening mammogramlevel, when women come in
and they're asymptomatic,they're just coming in
(04:05):
for their normal annual mammogram,what typically happens on the
day-to-day basis is they comein, they get two views of
the right breast, two viewsof the left breast, they get
the squeeze, and then they goand their imaging gets sent to a list
for the radiologist to read.
When the radiologist sits down to read allof the screening mammograms from that day,
or rather the week, there'sseveral hundred oftentimes
(04:27):
screening mammograms awaiting them.
What the AI algorithm does onthe front end is it has the
capacity to flag casesthat it finds suspicious.
So ones that it thinksneed to be prioritized.
So perhaps a patient whowas screened two days ago
has a suspicious finding,but maybe there was an
additional 50 screeners also readafter that patient, it
will flag that studyand put it in a color for us
(04:51):
and it'll say, Hey, read this study firstbecause this study may
need attention faster.
Maybe there's cancer on this exam.
So when it comes to the screening, that'show we use AI on the front end.
When the radiologistultimately opens the exam,
however, they are ahundred percent the ones
who are doing the interpretationfrom front to back.
So the AI is not doing theinterpretation for them,
(05:13):
and I think that's acommon misconception that
there's this fear that the AI is goingto replace the radiologist
in terms of interpretation,and that's, that's simply not true.
The radiologist is stillinterpreting a hundred percent.
Once they're done readingthe screening mammogram
and they've made a decisionabout the findings,
we get a little kind of cheatsheet from the AI algorithm
(05:34):
and it says, did you look at this area?
Did you look at this area?
It will circle findings thatit believes on the end side
of things that it believes are important.
So you have kind of this front endand backend use of AI to act
as on the front end as a,as a, basically a flagger to say,
Hey, this is really important.
Check this case out firstbecause maybe they have cancer.
(05:57):
And on the back end, it'sused as a double check
to make absolutely sure thatthe breast radiologist didn't,
you know, accidentally skipover calcifications or a mass
or maybe a distortion,even though they've
already reviewed the case.
So that's how we use AIwith respect to screening.
Now with diagnostic patients,these are the patients
who come into the breast reading roomand they have either a symptom,
(06:20):
they were maybe sent in from a doctorbecause they feel a new lump
or they have focal pain,or perhaps it's something
that we're following up.
We saw something ontheir screening mammogram
that we wanna investigate.
Those patients are alwaysseen by a radiologist.
Live diagnostic imaging meansthat you're going to be seen
by the radiologist.
Most practices follow that.
That's pretty much the standardof care for breast imaging.
(06:44):
So when we do our investigationwith mammography on
that day, we take specialpictures on that day.
It's not your standard screening views.
When we do those specialpictures, the AI will tell us
after we're done, and wealready kind of have an idea of
what we're gonna do withthe patient, it will again,
flag any findings that itthinks might be pertinent.
And again, it acts like a double check.
(07:06):
So overall, the thing to take away isthat it's not replacing the radiologist,
but it's helping the radiologist.
Instead of just having usgo through our examinations
and make a decision withoutany double checks, this acts
as a double check that wehave trained through a number
of studies and basically throughnumerous mammograms to act
as a double check to makesure we don't miss anything.
(07:27):
- Now, is this a standardof care now in all hospitals
who have radiology andradiologists reading mammograms?
Is there always this AI set up?
- So the current AIthat typically we refer
to nowadays is not standard yetbecause we are not at the point
where every practice in thecountry has the capacity
nor the volume to necessarily be ableto have this type of technology.
(07:50):
Here at , we're very,very lucky in that we have the,
what I consider to bethe most futuristic AI
technology and access to it.
And we really do believe thatthat's best for our patients,
and I think that we'revery lucky to have it.
However, there's always beensome version of AI that's been
around in breast imagingfor the last 20 years.
To give you an example,the earlier versions
(08:12):
of quote unquote AI that wereused on mammograms from maybe
10 years ago, what it wouldsimply do is just kind
of circle a finding, but the sensitivityand the specificity was not really there.
The AI algorithms of todayare markedly different
and they're much more improved.
- So go through your dayas a radiologist, describe
how it is that you review the imagesand where AI fits into that.
(08:35):
Is it scanned first by theAI and then you look at it,
or you look at it first andthen the AI looks at it?
And what if a hospital ora system doesn't have ai?
What's the standard of care then?
- Right, so that's a good question.
So typically what happens isthe, when a patient comes in,
especially for adiagnostic exam, so we have
to remember when patients comein for a screening mammogram,
(08:57):
those patients come in areseen, they're checked in
by the front desk in most practices,and then they are seen by
a mammography specialist.
So a mam, a mammo techwho specializes in taking
mammogram pictures.
So they're not necessarily seenby a physician if they're screening,
Those are the patients wewere discussing earlier
who come in, they get twostandardized views called a CC
(09:19):
and an MLO view on each breast.
And then once those imagesare obtained, they get
to leave the centerand those images are read in
what we call not live situations.
So the radiologist can read it the nextday or the day after.
As we were discussingearlier, when they sit down
and they have a list ofscreening mammograms,
typically radiologists readthis in batches when they're
(09:39):
uninterrupted and they tend tosit there for about a couple
of hours and read severalscreening mammograms in a row.
This is the best way to honein on your interpretation
skills and to preventalso being distracted.
Now, on my typical day,what I do is I come in in the morning
and I see majority biopsypatients and diagnostic patients.
(10:00):
So as a reminder, the diagnosticpatients are the patients
who have symptoms or somethingthat needs to be evaluated.
Perhaps we saw in screeningmammogram that needs
to be interpreted and investigated,or perhaps we're following
up something that they had,we've seen it before, but we
wanna keep a close eye on it.
Or perhaps there's asymptom that they have,
like a lump or pain.
(10:20):
When I see these patients,they will get specialized
mammographic imaging mostof the time, and we use a
combination of mammographyand ultrasound.
In the many of these cases.
When we take thesespecial mammogram images,
the AI algorithm isimplemented the minute we
start taking those pictures.
So by the time theimages are sent from the,
(10:40):
what we call the machine,the mammo machine
to the reading room, and I canactually pull them up on my
screen, the AI has alreadymade a decision as to whether
or not it believes there'sanything suspicious.
So what will happen is theradiologist will typically go
through those mammo images that were sentand they're gonna make a
decision independent of the ai.
That's how most of us do it.
But before we ultimatelygo see the patient,
(11:01):
we're gonna look at that double checkand we're gonna see did
the AI flag somethingor circle something that
maybe we didn't look ator maybe we, you know, take
a double look at again.
So maybe we saw the finding,but we wanna double check.
So we always double check the AIbefore we finalize the study
and make an ultimate decision.
And that's pretty much howwe use AI in our everyday
(11:22):
practice when we're coming in-house.
Now, if we're readingscreening mammograms,
I already described to you,we use it on the front end
and on the backend totell us which studies
to prioritize firstand on the backend, again to
double check our findings.
- What's been your experience with that?
Do you find that AI often picks up thingsthat you didn't see?
- You know, to be honest with you,so I'm a a breast trained radiologist.
(11:45):
So for breast trained radiologists,I think if you ask the
majority of them today,we have extensive experience
in in reading screeningmammograms and diagnostic imaging.
You know, at manyfacilities you are doing up
to 40 diagnostics a day.
So with that level of screeningand diagnostic imaging, you kind
of become like a machine yourself.
That's actually what youultimately wanna become.
(12:06):
You wanna become the ultimatemachine that has a, a process
that you never deviate from.
Most of the time the AI is flagging thingsthat you have already seen.
However, the perfect exampleof when AI is useful is not
so much when the radiologistmisses something.
It's quite rare for the radiologistto altogether miss a finding,
if I'm being honest with you.
But when it's reallyimportant is when you're kind
(12:27):
of at a 50 50 impasse as towhat you believe a finding to be.
There are some findings that you know,you could go either way on.
Sometimes you kind of are like, you know,I wish I had a second
pair of eyes to compareand say, you know, do I
think that this is benign?
Is it probably benign, or isthis something I should biopsy?
And having an algorithmthat's ultimately trained
(12:47):
on multiple, like thousandsof additional mammograms
and is statistically backedcan help you actually make a
decision in several of those cases.
And that's what I personally findto be the most useful utility of
- Ai.
How do you make thatdetermination if, for example,
I've seen mammograms,we've all seen a mammogram,
and I'm not trained todetermine what I'm looking at,
(13:10):
or if there is a mass in all of that whitethat's showing up.
Yep. Especially when there'sheavy or dense breast tissue.
Is it most helpful then, would you say?
- So the more complicated the breast,I think the more useful the algorithm.
I agree on that. Densebreasts are notoriously harder
to read for, for radiologistsbecause they can hide small masses.
(13:31):
This is why we know that thedenser a woman's breasts are
the higher propensityand the higher risk
there is for that patientto potentially have a missed mass.
So again, having that second pairof eyes is incredibly important
because when you're goingthrough such mammograms,
especially with increased complexity,maybe the fiber glandular
tissue pattern is a little bitmore complex than someone
who has a homogeneous patternthat's kind of the same
(13:52):
tissue pattern everywhere.
AI becomes more useful asopposed to the easy cases
where you know, the tissue'svery homogeneous appearing
and maybe has more fat content.
Those cases are a little biteasier, but AI really comes
and makes a difference for caseswhere there is increased complexity.
- You mentioned, and you werevery emphatic about this as
(14:14):
as we want to be in this podcast,that AI isn't replacing a radiologist
and it's not making a diagnosis.
Right. It's just a helping tool.
- Correct. - Right. And thatdistinction is very important
because in that survey that Ireferenced earlier, we found
that while younger generationsare a little bit more aware
of AI's role, most people,especially women over 45,
(14:36):
still aren't sure how itworks or what it does, and,
and the uncertainty can fuel fears.
- Right.
- Do you ever havethat conversation with patients?
Do they even know thatyou're using this tool?
I've, how would that come out- Really?
Yeah, so I've had a mixedexperience with this.
Some patients have, are now asking,I believe this year
actually more so than any,that they're starting
(14:57):
to realize, okay, wait,I have this AI thing.
They're starting to hear aboutthis thing called artificial
intelligence that's beingused in breast imaging,
but not just in breast imaging,but in radiology in general.
And they are asking, which I appreciatebecause it gives kind of the opportunity
to disavow any fears associated with it.
But basically the thingthat I think is the most
important takeaway isthat AI is used as a helpful double check,
(15:20):
but it's never goingto, it's never intended
to replace the radiologistas far as interpreting,
and I'll give an example.
So when we talk about patientswho come in, patients go
through several differentdiagnostic histories.
Some patients have a historyof having maybe breast
reduction or surgery,or maybe they had breast
cancer in the pastand they've had a lumpectomy.
Those simple examples canresult in people having,
(15:43):
for example, scars or maybecalcifications in their breasts
because of the changes from surgery.
- And a calcification canlook like a, maybe a cancer.
- Well, so that's the thing.
Sometimes these post-surgical changes,and I'm just using this as an example,
there are several benign findings,which when you don't have
the correct history in mind,can look like cancer.
So it may be flagged, forexample, by the AI saying, whoa,
(16:05):
whoa, whoa, look at this areaof maybe weird fibro glandular
tissue or distortion.
However, we as the radiologists,we know our patients all throughout.
So holistic medicineand making sure you know
your patient, the story,the surgical history,
their pertinent medical findingsis incredibly important.
- And it, - It doesn't reallyget replaced by an algorithm
because at the end ofthe day, we may know, oh,
(16:27):
it marked an area thatwas their surgical scar.
They've had it, you know, for the last 30years, no big deal.
- Right? - So while I thinkAI is incredibly useful,
its limitation is thatit's still not a human.
It's meant to be ahelp, not a replacement.
- 36% of women in our survey aged 40and older, told us they
wouldn't feel comfortablewith AI assisting their doctor.
(16:50):
Why do you think that is?
- I think the inherentfear is that there's a,
a depersonalization.
I think we see it not justin medicine, unfortunately.
You know, we all know whatit's like when you're trying
to like return somethingto an organization
and like you get like a voicemailor you get like a voice machine.
I think most patients want to knowthat there's still a doctor there,
that we still have a support system,especially when we are
dealing with cancer.
(17:11):
And I think my most important takeaway,and what I hope most people
take away from this podcast isthat we are never deserting our patients.
The interpreting physician isa hundred percent interpreting
their mammogram from start to finish.
That will never change.
This algorithm simply helps usto make less mistakes in the event
that we are maybe on our50th study of that day.
(17:31):
And we're looking at a scanthat's maybe particularly
complex, it acts as double check.
And I think the fear comes from a worrythat the doctor's gonna
be removed one day,and that now I'm gonna
have a machine lookingand maybe the machine's
gonna miss something,or maybe the machine won't know
that I had a family historyof breast cancer, or that
I felt a lump last week.
(17:53):
So my hope is that I can reassure patientsthat we are not trying to
replace the doctor in any way.
I love my patients, I love my job.
I would never want to giveaway the opportunity to act,
to participate in apatient's care to a machine,
because we're just notthere in terms of the,
at the sensitivity specificity.
(18:13):
But not only that, history is importantand physical exam is still important,
and talking to patientsis still important.
So it's more of a help,not a replacement. Right.
- I think when anyone talksabout ai, it's always that,
that it's gonna replacehumans and replace.
Yeah. And that's always the fearthat comes front and center.
But a study published lastyear in the Lancet said
mammography screeninghas been a cornerstone
(18:34):
of early detection of breastcancer since the eighties,
which you've mentioned, andamong its challenges
is a marked variabilitybetween the radiologists in
diagnostic accuracy, which leadsto unnecessary recalls and missed cancer.
- Right.
- Is there any data that you know ofthat supports the, the use
of AI in reducing recallsor missed cancer diagnosis?
(18:56):
- So there's several journalsthat are currently trying
to evaluate how AI isimpacting our callback rate
or our statistics.
How often are radiologists calling backabnormalities from the
screening mammogram?
Now, to first address yourfirst point, there is going
to be inherent variabilitybetween humans, right?
So between one radiologist to another,there may be differences in
quote unquote sensitivity.
(19:19):
Some radiologists may callback a little bit more,
versus other radiologistswho may call back less.
The goal in the UnitedStates, which is standardized,
we do have standardized numbersthat we try to attain, is
that the radiologist doesn'tcall back more than around 10
to 12% of screening screenersfor diagnostic evaluation.
We are not in the habit oftrying to scare everybody
(19:41):
and say, oh my gosh,this could be something,
this could be something,this could be something
we have to be reasonable.
Which is why getting mammogramsevery year is the standard
of care in the United States.
Because if you get a mammogram every year,it enables the radiologist
to increase their specificityand be sure that, ah,
this is a new finding.
This wasn't there 30 years ago.
(20:02):
Now, things that haveimproved over the years
that actually improve ourradiologist's sensitivity
and specificity, we nowhave fellowship programs
that are dedicated toentirely women's imaging.
So I myself underwent thattype of fellowship training.
So during that year, essentiallywhat you're doing is honing your skillset.
You're seeing mammogram after mammogram.
(20:22):
So you start to see, ah, you know,the last time I called
back this calcificationor this mass, it ended up being benign.
So what ends up happeningis that with experience
and with trainingand increased years of
education, you end up findingthat your specificity goes up
and you're not calling back.
Quite so many AI does assist with thisbecause like I said,
there's several scenarioswhere sometimes you look at a finding
(20:46):
and you could kind of go eitherway on it, you could say,
you know, I could buy that this is benign.
But part of me kind of is like, you know,it's only a finding in
this part of the rest.
Maybe I should call it backfor more diagnostic evaluation.
If the AI algorithm says,no, this is nothing,
it assists in preventingtoo many from being called
back initially.
(21:06):
So it does have a role.
It it, it does help theradiologist maybe get rid
of those in-betweener casesand downgrade it instead of over calling.
Now, when the Lancet, the reasonthat there's also some variety
in terms of how sensitiveand how specific certain AI
algorithms are is, we haveto remember that there's several
AI algorithms that exist.
And we also have to keep in mindthat different patient populations have
(21:29):
different breast densities.
So some populations of the worldand in the country have
very, very dense breasts,and other populations
have fattier breasts.
We have to keep in mindthat the AI algorithm may have
been trained on a data poolthat maybe had denser breasts.
So what happens when wetake that same AI algorithm
and we apply it to adifferent group of women,
maybe a different age group,maybe a different density
(21:50):
group, a different racial group.
So we have to be very carefulthat when we train these AI
algorithms, that we make surethat we train them on a
diversity of mammogramsand breast types, because it
can train itself over time onone specific subtype
and then maybe won't be asuseful or sensitive on another.
- And who are you talkingto when you say, we have
(22:11):
to be careful to train the AIon the manufacturers of the,
- So manufacturers of the, so right now,AI is being developed
in large part as a kindof a conglomerate between
physicians as wellas manufacturers.
So the manufacturers or thedesigners of these softwares
are largely medical engineers.
Now, medical engineersdon't just make algorithms
(22:33):
without consulting physicians.
Oftentimes physicians are onthe boards for these things
because, you know, we needto know how to apply it
and who knows better how toapply it than the physicians.
We, we know what we need for our patients,and we also demand a certain
level of, you know, sensitivityand accuracy before we
think that it's good enoughto ultimately treat a patient
or even participate inthe treatment process.
(22:54):
I will say this though, that AI algorithmsthat exist today are already
light years better thanwhat was available even
just five years ago.
So that's very promising. Soit's, it's very interesting.
It's very important tocontinue researching.
But overall, I thinkthe trajectory of AI is
that it's only gonna help our patients.
- I can see that we're allfinding, the more we use ai,
the more comfortable we get with it.
(23:16):
I've seen the reading room, it's dark,you guys are sitting in
there for hours on end,like you said, you might
have 50 mammograms to read.
How do you even keep yourselfalert to look at each image,
each different image, isn't it?
And maybe it would just beeasy to let AI take over
and oh, if AI says it's okay,I'm gonna let this one go.
(23:37):
- Yeah, no,- I mean, how is it, how
do you keep from gettingthat lackadaisical about the, the work?
- So radiologists, I thinkwe're an interesting breed.
Like if you, if you talk toa neurosurgeon, you know, I,
I always, I think thatthey're like very revered
and I, I respect them immensely.
You know, they go in for 12 hour surgeriesand I sit here thinking, oh my gosh,
how do they do a surgery for 12 hours?
And I think with radiologists,I think the common
(23:58):
misconception is, oh my gosh,you're in a reading room,
how do you stay awake?
But it's so funny withradiologists, we love our jobs.
Most of us think of ourselvesas kind of detectives.
It's fun and it's enjoyable forus to actually help patients
and be able to do itthrough medical imaging.
We're kind of the onesthat enjoy looking at every
single millimeter of that imageand participating in the medical process
(24:23):
by finding disease in that way.
So while fatigue can happenacross any medical specialty
with radiology, you know, the dimnessof the reading room rarely has anything
to do with our alertness.
We, we are trainedand specialized in interpreting
mammograms for, for hours.
And quite frankly, most ofus would be able to do this
for eight to 10 hours if we so desired.
(24:43):
And most of us don't do that, though.
We typically read screeningmammograms in several hour
increments, And the AIsimply acts as a helper tool.
But it's more so just to help us in termsof our sensitivity, making sure
that we have a double checkand a two check system.
I would say it less hasto do with like fatigue
or anything of that sort,because most radiologists
know that, you know,if my fatigue limit is two hours,
(25:06):
they're not gonna continuereading beyond that.
So most radiologists have that,and we know at this point
as part of our process,when is it a safe time to kind of sign off
and then we resume work the next day?
- Explain the two check system.
- So when you have a two check systemor a double check system, it
used to be kind of theorized,like, what would it be like
if we had two doctors look atthe same mammogram just as a
(25:26):
double check to each other?
Because we really didn'thave back in the day two set
of eyes unless it was another human.
Now the problem withthat is that the capacity
to take two doctors away fromthe patients on a given day
to only be reading, for example, screens,is not very realistic,
nor is it efficient.
We want our doctors tobe in front of patients.
(25:46):
I wanna be with the patients, making surethat I interpret imaging
with them and alongside them,and I'm also performing
biopsies with them.
So when you can actuallyhave your double check
or your two part systembe a component of ai,
it increases efficiency.
And therefore we can have one doctor gothrough a screening list
confidently with a double checkand have another doctor perhaps
(26:08):
with the patients doingdiagnostics at the same exact time.
So efficiency is criticaland AI is very, very helpful. With that,
- Are there concerns about bias,whether the technology performs
equally well across allages, races, and breast densities?
What's being done to makesure that these tools are fair
and accurate for everyone?
- I think I should gointo a little bit about
(26:29):
what does it mean when we say ai, ai,because AI is a very umbrella term
and it doesn't really discuss,what are we talking about?
Like it, it, it's obviously an algorithm,it's a machine, but what does that mean?
Artificial intelligenceis typically referring
to something called adeep learning network.
So deep learning, when wedevelop a deep learning network,
which is ai, we're trying to simulatehow the brain works in a human
(26:50):
with a computer algorithm.
So we're trying to simulatehow the brain interprets
findings, whether that's imagingor a picture or, or
speech or even language.
So when you're talking aboutdeveloping a deep learning
network or ai, there areseveral ways you can do that.
A subtype of way of doingthat is something called
(27:11):
a convolutional neuralnetwork, A CNN.
So when we talk about makingAI in radiology for CT images,
or for MRI or from mammography, oftentimeswhat we're referring to
is creating an algorithmthat simulates how the human brain works.
So when we see something on a picture,our eyeballs take it in,
(27:33):
and then we convey thatinformation slice by slice.
In some scenarios, ifwe're talking about CTS
or with X-rays, it's oftentimestwo dimensional imaging.
And we convey that to our brain,and our brain does its
magic, like it always does.
And it says, okay, I made a decision.
This is what I think it is.
When you're training an AIalgorithm to do such a thing
through these convolutionalneural networks
(27:53):
or deep learning states, whatyou're doing is you're saying,
okay, I'm gonna train thealgorithm to take slice
by slice information orvoxel information from an MRI
or maybe just a two dimensionalimage from a 2D mammogram.
And it's gonna make a decisionand it's gonna ultimately
form a conclusion.
And that's how we typicallyform the AI algorithms.
(28:15):
Now, with respect to yourquestion about how do we train it
so that it's used to seeinga diversity of patients,
whether it's dense breasted patients,fatty breasted patients,
patients who are small, large,different ethnicities, races,
even different genders.
'cause we sometimes have to usemammography for men as well.
What we do is we increase the data set.
(28:36):
So typically what mostmedical engineers will do
and doctors is we say, okay,you can't train an AI
algorithm on only 10 patients.
Typically you need thousands.
And we try to make surethat that population is
as diverse as possible.
And then what we do iswe test the algorithm
against a doctor.
So a lot of these algorithmswhen they're being developed
(28:57):
will say, okay, what didthe breast radiologist think
this this patient had?
Did they say it was a negative caseor was it a positive case?
And then they're gonnatest the AI algorithm,
and if there's ever adifference, interpretation was
between the doctor and the AI algorithm.
Doctors interpretation isultimately what's gonna be used
to train the AI algorithm.
And the more times you canhave the AI algorithm learn
(29:21):
and practice on increasingmammograms in a diversity
of patients, it gets better and better.
It's very promising, and that'show we ultimately train it
to provide good data regardlessof the diversity of patients.
- Well, it sounds a loteasier than getting in a car
and getting behind a wheel and letting thecar drive you. Yeah, I mean,
- It's very complicated,but you know what, thank,
thankfully we have such a skilled,like medical engineers nowadays
(29:44):
and scientists who are working on this,and they are obviously
including medical expertsand doctors so
that we can ultimately developthis artificial intelligence.
So it's useful for us andnot the other way around.
- Have you had a discrepancyin what you have found versus
what the AI has found andand how do you handle that
if there is a discrepancy?
- Yes, absolutely. Ithappens actually quite often,
(30:04):
but it's nothing to be feared.
I gave an example earlier ofwhen patients have a history
of surgery and the AI algorithm thinks,oh my gosh, this is a distortion.
This must be like a cancer,like a huge speculated mask.
However, the doctor looks at the imageand knows that this patient,
they've been in the practicefor years, we know that they
had perhaps a reduction onthat breast or they had a lumpectomy.
(30:25):
We ultimately know what the results areand we know why it's flagging it.
So while it's stilluseful to double check,
ultimately if there's ever discrepancyand the radiologist looks at the finding
and has a reasonablereason for the finding,
the radiologist's interpretationas usual will be the one
that is interpreted, not the ai.
The AI is again, meant to bea help, not a replacement.
(30:46):
- Even if you in inform the, the softwareor the, the ai, the patient
had a, a lumpectomy here,the patient has dense breast tissue.
- It can still happen, unfortunately.
Yeah, it, it does its best.
You know, it does its bestwith dense breasted women.
It does its best to, you know,not mark everything under the sun.
And it's, it's reallyimproved over the course
of the last five to 10 years.
(31:06):
However, it's not perfect.
Again, it's not a human.
So it's very hard to trainan algorithm to know things
that a human perceives whenthey talk to a patient.
For example, if I go talk to a patientand the patient looks very
nervous and is guardingand perhaps guarding a certain
component of their breast,they may have already undergone
mammography and ultrasound.
But when you go in and talk to the patientand the patient then reveals,
oh, I, you know, I had,I actually fell several days ago
(31:31):
and I actually have this lumpon the left side of my breast.
The physical exam has not lost.
Its, its ju if you will,like, it's still important
to follow up with the patientand ask about history,
because sometimes findings on mammographyand ultrasound mean very little without a
pertinent patient history.
So for example, a hematomacan look like a big ginormous
(31:51):
mass on a mammogram.
If I know that the patientjust had a trauma a couple days
ago, I'm gonna lean with,this is probably hematoma.
If I don't have that historyand I just leave it up to an algorithm
or anything really to justsay, oh, there's a mass,
and then call it that, it maylead to unnecessary workup
for something that's quitesimply a benign finding.
- What excites you mostabout this technology
(32:12):
and what still needs to be figured- Out?
I think what excites me most personally ishow we can apply it to
different modalities,not just mammography.
So everybody kind oftalks about mammography
because that's what we havethe most exposure to, is
that we can use AI inour screening mammograms
and our diagnostic mammogramsas we've been discussing.
But you know, I think one of the thingsthat excites me in particular is
(32:33):
how we can use it in ultrasound or MRI.
So MRI is used in breastimaging a lot when it comes
to breast cancer screening detection,especially in our high
risk genetic carriers,high risk patients in general.
There are things that AIcan pick up, especially
with contrast enhanced exams like MRI,that we simply cannot
see with the naked eyeas opposed to mammography.
(32:55):
Most of the time when we talkabout AI with mammography,
it's not so much picking up thingsthat we can't see with our naked eye.
We can see them, it's just a matter of,did you look at this area with MRI?
However, when you give someone contrastand it spreads throughout the breast
and you start to see lesions pop up,sometimes you can train AI algorithms
to pick up on informationthat you just cannot
(33:15):
see with the naked eye.
For example, if youknow someone has cancer,
and I actually did a studyon this several years ago,
and if you know someone has cancerand they get an MRI exam to stage
and to see how extensive thecancer is, where did it go?
How big is it? It's very interestingthat AI we're finding has
the ability to be trainedto tell you whether or not a
(33:35):
mass will necessarily respondto chemotherapy
before the patient evenundergoes chemotherapy.
- Wow. - Now that's incredibleand that's incredibly useful
because you still need the radiologistto interpret the imaging.
However, it adds adifferent benefit that I
otherwise would not be able to offer.
It tells me, you know, based off ofhow this mass took up
(33:56):
the contrast on a very,very micro level, looking
at the voxels in an MRI,it can tell me whether or
not it thinks they're gonnarespond to neoadjuvant chemotherapy
before ultimately undergoing surgery.
So why is that important?
If I know from the AI algorithm,if we develop an algorithm
that's so good that itcan tell me, you know,
this person's not gonna respondto neoadjuvant chemotherapy
(34:18):
'cause this cancer is pickingup the contrast in such a way
that it just doesn't looklike it's gonna be responsive,
then I'm not gonna waste mytime potentially delaying the
patient's surgery for severalmonths waiting for the cancer
to shrink on chemotherapyif there was no chance
for it to in the first place.
Hmm. So it's, it offersadditional information
that perhaps we just wouldn'thave at our fingertips just
(34:41):
by simply imaging alone.
So this is just a littlekind of splice of the pie
and just an idea of how wecan apply AI in the future.
- Exciting stuff.
- Yeah, very.
The AI is an assistance tool,but it is in no way a
replacement for the physician.
The physician still hasto do the interpretation
and still ultimately has tomake a decision based off of all
of the information at hand,whether it be the
(35:01):
patient's history, surgery,health overall,
and what, what the ultimatequestion is for each exam.
- I, I appreciate thatbecause the more patients understand about
how AI is working,putting their trust still in their doctors
and how it's helping the doctors, I thinkthat they'll be more confident in
how they'll feel about thewhole screening experience.
(35:22):
- Right. - And not shying awayfrom mammograms. Dr. Sakla, thank
with us here on DocTalk.
- Awesome. Thank you guysso much for having me.
It was an absolutepleasure. This was very fun.
And yes, just remember, don't forgetto get your screening
mammograms every year.
- For more informationabout breast imaging
and screening at MedStar Healthvisit MedStar Health.org/breast health
(35:47):
or call 2 0 2 8 7 7 2 8 0 0.
If you would like tocomment on this podcast
or recommend a topic foranother episode of DocTalk,
send an email to DocTalk@medstar.net.
- Comprehensive, relevantand insightful conversations about health
and medicine happen here.
When Health DocTalk,these are real conversations
with physician experts fromaround the largest healthcare
system in the Maryland DCregion When it comes to
detecting breast cancer,we know that finding it early
makes every difference inwhat happens next in the journey.
(00:26):
For decades, radiologistshave relied on mammograms
and their own trained eyesto find even the smallest
signs of disease.
But now a new tool is changing the gain.
Artificial intelligence fromflagging subtle patterns
invisible to the human eye,to helping radiologists read images faster
and more accurately.
(00:47):
AI is beginning to play animportant role in breast cancer
detection and diagnosisearly, perhaps even by years
to better understand what people knowand how they feel about ai.
In breast imaging,Health conducted a national survey
of more than a thousand adultsand found only one
third of women surveyed,and were aware that AI was even being used
(01:08):
to help doctors read mammograms.
In today's episode ofMedStar Health DocTalk,
we are taking a closer look athow this technology actually
works, what it meansfor accuracy and detection,
and how it can ease, notadd to patient anxiety.
Welcome Health Radiologist, Dr.
Nicole SLA to help us separatefact from fear when it comes
(01:29):
to artificial intelligenceuse in mammography.
I'm your host Schindler, Dr.
Sakla. Thanks for joining me.
- It's an absolute pleasure.
Thank you for having me.
- So you have an extensive backgroundof published medical research,
including in artificial intelligence usefor early breast cancer detection in some
of the most prestigious andhigh impact medical journals.
(01:50):
So while many of us arejust learning about chat GBT
and artificial intelligence,and we haven't really heard of this,
you were already learningabout it using it in diagnosing
breast cancer, you've beenstudying it for years.
When was the writing on the wall for you?
- So it's interesting. Thankyou again for having me.
So AI has been on the docket,some people would say since
really around the eighties,but even there's been early research done
(02:13):
even prior to that.
For me personally, I had my first exposureto artificial intelligence in
medicine when I was a residentin New Jersey, a peer
of mine who's named Dr.
Duggan Deep Singh, and he was the onewho first introduced me
to artificial intelligenceand its applications in medical imaging.
What I started to do was to researchhow can AI help the doctor,
especially the radiologist,identify pathology with respect to CT MRI
(02:40):
and mammography better and faster.
As I got further on in my trainingand I started to see, okay, wait, this,
this technology may actually enable usto catch small findings that
may potentially not be ableto be seen with the naked
eye, particularly with MRI.
When I ended up going intowomen's imaging, the application
of AI for breast cancerbecame even more evident,
(03:01):
and it has really, really astounded mehow far we've come in just
even the last five years as faras the development of
artificial intelligencealgorithms and techniques.
- Let's start with the basics.
We hear the term AI all the time.
What does it actually mean inthe context of breast imaging?
- So that's a good question.
I think when people hear theword ai, they automatically
(03:21):
think of like the Jetsons,and they think that there's
gonna be like a machine poppingout of the wall and it's gonna do the job
of the doctor for you,- Right? Exactly.
- And that is what I'm actually tryingto dissuade people from
thinking of AI in that manner.
Here's why AI is moreof an assistance tool.
It's not meant to replace the radiologist.
So when it comes to a breastradiologist specifically
(03:44):
and how we use AI to findbreast cancer, we have to kind
of divide our thinkinginto two main categories.
We have the people who come infor their screening mammograms,
and then we have the womenwho come in for what's known
as diagnostic examinations.
So on a screening mammogramlevel, when women come in
and they're asymptomatic,they're just coming in
(04:05):
for their normal annual mammogram,what typically happens on the
day-to-day basis is they comein, they get two views of
the right breast, two viewsof the left breast, they get
the squeeze, and then they goand their imaging gets sent to a list
for the radiologist to read.
When the radiologist sits down to read allof the screening mammograms from that day,
or rather the week, there'sseveral hundred oftentimes
(04:27):
screening mammograms awaiting them.
What the AI algorithm does onthe front end is it has the
capacity to flag casesthat it finds suspicious.
So ones that it thinksneed to be prioritized.
So perhaps a patient whowas screened two days ago
has a suspicious finding,but maybe there was an
additional 50 screeners also readafter that patient, it
will flag that studyand put it in a color for us
(04:51):
and it'll say, Hey, read this study firstbecause this study may
need attention faster.
Maybe there's cancer on this exam.
So when it comes to the screening, that'show we use AI on the front end.
When the radiologistultimately opens the exam,
however, they are ahundred percent the ones
who are doing the interpretationfrom front to back.
So the AI is not doing theinterpretation for them,
(05:13):
and I think that's acommon misconception that
there's this fear that the AI is goingto replace the radiologist
in terms of interpretation,and that's, that's simply not true.
The radiologist is stillinterpreting a hundred percent.
Once they're done readingthe screening mammogram
and they've made a decisionabout the findings,
we get a little kind of cheatsheet from the AI algorithm
(05:34):
and it says, did you look at this area?
Did you look at this area?
It will circle findings thatit believes on the end side
of things that it believes are important.
So you have kind of this front endand backend use of AI to act
as on the front end as a,as a, basically a flagger to say,
Hey, this is really important.
Check this case out firstbecause maybe they have cancer.
(05:57):
And on the back end, it'sused as a double check
to make absolutely sure thatthe breast radiologist didn't,
you know, accidentally skipover calcifications or a mass
or maybe a distortion,even though they've
already reviewed the case.
So that's how we use AIwith respect to screening.
Now with diagnostic patients,these are the patients
who come into the breast reading roomand they have either a symptom,
(06:20):
they were maybe sent in from a doctorbecause they feel a new lump
or they have focal pain,or perhaps it's something
that we're following up.
We saw something ontheir screening mammogram
that we wanna investigate.
Those patients are alwaysseen by a radiologist.
Live diagnostic imaging meansthat you're going to be seen
by the radiologist.
Most practices follow that.
That's pretty much the standardof care for breast imaging.
(06:44):
So when we do our investigationwith mammography on
that day, we take specialpictures on that day.
It's not your standard screening views.
When we do those specialpictures, the AI will tell us
after we're done, and wealready kind of have an idea of
what we're gonna do withthe patient, it will again,
flag any findings that itthinks might be pertinent.
And again, it acts like a double check.
(07:06):
So overall, the thing to take away isthat it's not replacing the radiologist,
but it's helping the radiologist.
Instead of just having usgo through our examinations
and make a decision withoutany double checks, this acts
as a double check that wehave trained through a number
of studies and basically throughnumerous mammograms to act
as a double check to makesure we don't miss anything.
(07:27):
- Now, is this a standardof care now in all hospitals
who have radiology andradiologists reading mammograms?
Is there always this AI set up?
- So the current AIthat typically we refer
to nowadays is not standard yetbecause we are not at the point
where every practice in thecountry has the capacity
nor the volume to necessarily be ableto have this type of technology.
(07:50):
Here at , we're very,very lucky in that we have the,
what I consider to bethe most futuristic AI
technology and access to it.
And we really do believe thatthat's best for our patients,
and I think that we'revery lucky to have it.
However, there's always beensome version of AI that's been
around in breast imagingfor the last 20 years.
To give you an example,the earlier versions
(08:12):
of quote unquote AI that wereused on mammograms from maybe
10 years ago, what it wouldsimply do is just kind
of circle a finding, but the sensitivityand the specificity was not really there.
The AI algorithms of todayare markedly different
and they're much more improved.
- So go through your dayas a radiologist, describe
how it is that you review the imagesand where AI fits into that.
(08:35):
Is it scanned first by theAI and then you look at it,
or you look at it first andthen the AI looks at it?
And what if a hospital ora system doesn't have ai?
What's the standard of care then?
- Right, so that's a good question.
So typically what happens isthe, when a patient comes in,
especially for adiagnostic exam, so we have
to remember when patients comein for a screening mammogram,
(08:57):
those patients come in areseen, they're checked in
by the front desk in most practices,and then they are seen by
a mammography specialist.
So a mam, a mammo techwho specializes in taking
mammogram pictures.
So they're not necessarily seenby a physician if they're screening,
Those are the patients wewere discussing earlier
who come in, they get twostandardized views called a CC
(09:19):
and an MLO view on each breast.
And then once those imagesare obtained, they get
to leave the centerand those images are read in
what we call not live situations.
So the radiologist can read it the nextday or the day after.
As we were discussingearlier, when they sit down
and they have a list ofscreening mammograms,
typically radiologists readthis in batches when they're
(09:39):
uninterrupted and they tend tosit there for about a couple
of hours and read severalscreening mammograms in a row.
This is the best way to honein on your interpretation
skills and to preventalso being distracted.
Now, on my typical day,what I do is I come in in the morning
and I see majority biopsypatients and diagnostic patients.
(10:00):
So as a reminder, the diagnosticpatients are the patients
who have symptoms or somethingthat needs to be evaluated.
Perhaps we saw in screeningmammogram that needs
to be interpreted and investigated,or perhaps we're following
up something that they had,we've seen it before, but we
wanna keep a close eye on it.
Or perhaps there's asymptom that they have,
like a lump or pain.
(10:20):
When I see these patients,they will get specialized
mammographic imaging mostof the time, and we use a
combination of mammographyand ultrasound.
In the many of these cases.
When we take thesespecial mammogram images,
the AI algorithm isimplemented the minute we
start taking those pictures.
So by the time theimages are sent from the,
(10:40):
what we call the machine,the mammo machine
to the reading room, and I canactually pull them up on my
screen, the AI has alreadymade a decision as to whether
or not it believes there'sanything suspicious.
So what will happen is theradiologist will typically go
through those mammo images that were sentand they're gonna make a
decision independent of the ai.
That's how most of us do it.
But before we ultimatelygo see the patient,
(11:01):
we're gonna look at that double checkand we're gonna see did
the AI flag somethingor circle something that
maybe we didn't look ator maybe we, you know, take
a double look at again.
So maybe we saw the finding,but we wanna double check.
So we always double check the AIbefore we finalize the study
and make an ultimate decision.
And that's pretty much howwe use AI in our everyday
(11:22):
practice when we're coming in-house.
Now, if we're readingscreening mammograms,
I already described to you,we use it on the front end
and on the backend totell us which studies
to prioritize firstand on the backend, again to
double check our findings.
- What's been your experience with that?
Do you find that AI often picks up thingsthat you didn't see?
- You know, to be honest with you,so I'm a a breast trained radiologist.
(11:45):
So for breast trained radiologists,I think if you ask the
majority of them today,we have extensive experience
in in reading screeningmammograms and diagnostic imaging.
You know, at manyfacilities you are doing up
to 40 diagnostics a day.
So with that level of screeningand diagnostic imaging, you kind
of become like a machine yourself.
That's actually what youultimately wanna become.
(12:06):
You wanna become the ultimatemachine that has a, a process
that you never deviate from.
Most of the time the AI is flagging thingsthat you have already seen.
However, the perfect exampleof when AI is useful is not
so much when the radiologistmisses something.
It's quite rare for the radiologistto altogether miss a finding,
if I'm being honest with you.
But when it's reallyimportant is when you're kind
(12:27):
of at a 50 50 impasse as towhat you believe a finding to be.
There are some findings that you know,you could go either way on.
Sometimes you kind of are like, you know,I wish I had a second
pair of eyes to compareand say, you know, do I
think that this is benign?
Is it probably benign, or isthis something I should biopsy?
And having an algorithmthat's ultimately trained
(12:47):
on multiple, like thousandsof additional mammograms
and is statistically backedcan help you actually make a
decision in several of those cases.
And that's what I personally findto be the most useful utility of
- Ai.
How do you make thatdetermination if, for example,
I've seen mammograms,we've all seen a mammogram,
and I'm not trained todetermine what I'm looking at,
(13:10):
or if there is a mass in all of that whitethat's showing up.
Yep. Especially when there'sheavy or dense breast tissue.
Is it most helpful then, would you say?
- So the more complicated the breast,I think the more useful the algorithm.
I agree on that. Densebreasts are notoriously harder
to read for, for radiologistsbecause they can hide small masses.
(13:31):
This is why we know that thedenser a woman's breasts are
the higher propensityand the higher risk
there is for that patientto potentially have a missed mass.
So again, having that second pairof eyes is incredibly important
because when you're goingthrough such mammograms,
especially with increased complexity,maybe the fiber glandular
tissue pattern is a little bitmore complex than someone
who has a homogeneous patternthat's kind of the same
(13:52):
tissue pattern everywhere.
AI becomes more useful asopposed to the easy cases
where you know, the tissue'svery homogeneous appearing
and maybe has more fat content.
Those cases are a little biteasier, but AI really comes
and makes a difference for caseswhere there is increased complexity.
- You mentioned, and you werevery emphatic about this as
(14:14):
as we want to be in this podcast,that AI isn't replacing a radiologist
and it's not making a diagnosis.
Right. It's just a helping tool.
- Correct. - Right. And thatdistinction is very important
because in that survey that Ireferenced earlier, we found
that while younger generationsare a little bit more aware
of AI's role, most people,especially women over 45,
(14:36):
still aren't sure how itworks or what it does, and,
and the uncertainty can fuel fears.
- Right.
- Do you ever havethat conversation with patients?
Do they even know thatyou're using this tool?
I've, how would that come out- Really?
Yeah, so I've had a mixedexperience with this.
Some patients have, are now asking,I believe this year
actually more so than any,that they're starting
(14:57):
to realize, okay, wait,I have this AI thing.
They're starting to hear aboutthis thing called artificial
intelligence that's beingused in breast imaging,
but not just in breast imaging,but in radiology in general.
And they are asking, which I appreciatebecause it gives kind of the opportunity
to disavow any fears associated with it.
But basically the thingthat I think is the most
important takeaway isthat AI is used as a helpful double check,
(15:20):
but it's never goingto, it's never intended
to replace the radiologistas far as interpreting,
and I'll give an example.
So when we talk about patientswho come in, patients go
through several differentdiagnostic histories.
Some patients have a historyof having maybe breast
reduction or surgery,or maybe they had breast
cancer in the pastand they've had a lumpectomy.
Those simple examples canresult in people having,
(15:43):
for example, scars or maybecalcifications in their breasts
because of the changes from surgery.
- And a calcification canlook like a, maybe a cancer.
- Well, so that's the thing.
Sometimes these post-surgical changes,and I'm just using this as an example,
there are several benign findings,which when you don't have
the correct history in mind,can look like cancer.
So it may be flagged, forexample, by the AI saying, whoa,
(16:05):
whoa, whoa, look at this areaof maybe weird fibro glandular
tissue or distortion.
However, we as the radiologists,we know our patients all throughout.
So holistic medicineand making sure you know
your patient, the story,the surgical history,
their pertinent medical findingsis incredibly important.
- And it, - It doesn't reallyget replaced by an algorithm
because at the end ofthe day, we may know, oh,
(16:27):
it marked an area thatwas their surgical scar.
They've had it, you know, for the last 30years, no big deal.
- Right? - So while I thinkAI is incredibly useful,
its limitation is thatit's still not a human.
It's meant to be ahelp, not a replacement.
- 36% of women in our survey aged 40and older, told us they
wouldn't feel comfortablewith AI assisting their doctor.
(16:50):
Why do you think that is?
- I think the inherentfear is that there's a,
a depersonalization.
I think we see it not justin medicine, unfortunately.
You know, we all know whatit's like when you're trying
to like return somethingto an organization
and like you get like a voicemailor you get like a voice machine.
I think most patients want to knowthat there's still a doctor there,
that we still have a support system,especially when we are
dealing with cancer.
(17:11):
And I think my most important takeaway,and what I hope most people
take away from this podcast isthat we are never deserting our patients.
The interpreting physician isa hundred percent interpreting
their mammogram from start to finish.
That will never change.
This algorithm simply helps usto make less mistakes in the event
that we are maybe on our50th study of that day.
(17:31):
And we're looking at a scanthat's maybe particularly
complex, it acts as double check.
And I think the fear comes from a worrythat the doctor's gonna
be removed one day,and that now I'm gonna
have a machine lookingand maybe the machine's
gonna miss something,or maybe the machine won't know
that I had a family historyof breast cancer, or that
I felt a lump last week.
(17:53):
So my hope is that I can reassure patientsthat we are not trying to
replace the doctor in any way.
I love my patients, I love my job.
I would never want to giveaway the opportunity to act,
to participate in apatient's care to a machine,
because we're just notthere in terms of the,
at the sensitivity specificity.
(18:13):
But not only that, history is importantand physical exam is still important,
and talking to patientsis still important.
So it's more of a help,not a replacement. Right.
- I think when anyone talksabout ai, it's always that,
that it's gonna replacehumans and replace.
Yeah. And that's always the fearthat comes front and center.
But a study published lastyear in the Lancet said
mammography screeninghas been a cornerstone
(18:34):
of early detection of breastcancer since the eighties,
which you've mentioned, andamong its challenges
is a marked variabilitybetween the radiologists in
diagnostic accuracy, which leadsto unnecessary recalls and missed cancer.
- Right.
- Is there any data that you know ofthat supports the, the use
of AI in reducing recallsor missed cancer diagnosis?
(18:56):
- So there's several journalsthat are currently trying
to evaluate how AI isimpacting our callback rate
or our statistics.
How often are radiologists calling backabnormalities from the
screening mammogram?
Now, to first address yourfirst point, there is going
to be inherent variabilitybetween humans, right?
So between one radiologist to another,there may be differences in
quote unquote sensitivity.
(19:19):
Some radiologists may callback a little bit more,
versus other radiologistswho may call back less.
The goal in the UnitedStates, which is standardized,
we do have standardized numbersthat we try to attain, is
that the radiologist doesn'tcall back more than around 10
to 12% of screening screenersfor diagnostic evaluation.
We are not in the habit oftrying to scare everybody
(19:41):
and say, oh my gosh,this could be something,
this could be something,this could be something
we have to be reasonable.
Which is why getting mammogramsevery year is the standard
of care in the United States.
Because if you get a mammogram every year,it enables the radiologist
to increase their specificityand be sure that, ah,
this is a new finding.
This wasn't there 30 years ago.
(20:02):
Now, things that haveimproved over the years
that actually improve ourradiologist's sensitivity
and specificity, we nowhave fellowship programs
that are dedicated toentirely women's imaging.
So I myself underwent thattype of fellowship training.
So during that year, essentiallywhat you're doing is honing your skillset.
You're seeing mammogram after mammogram.
(20:22):
So you start to see, ah, you know,the last time I called
back this calcificationor this mass, it ended up being benign.
So what ends up happeningis that with experience
and with trainingand increased years of
education, you end up findingthat your specificity goes up
and you're not calling back.
Quite so many AI does assist with thisbecause like I said,
there's several scenarioswhere sometimes you look at a finding
(20:46):
and you could kind of go eitherway on it, you could say,
you know, I could buy that this is benign.
But part of me kind of is like, you know,it's only a finding in
this part of the rest.
Maybe I should call it backfor more diagnostic evaluation.
If the AI algorithm says,no, this is nothing,
it assists in preventingtoo many from being called
back initially.
(21:06):
So it does have a role.
It it, it does help theradiologist maybe get rid
of those in-betweener casesand downgrade it instead of over calling.
Now, when the Lancet, the reasonthat there's also some variety
in terms of how sensitiveand how specific certain AI
algorithms are is, we haveto remember that there's several
AI algorithms that exist.
And we also have to keep in mindthat different patient populations have
(21:29):
different breast densities.
So some populations of the worldand in the country have
very, very dense breasts,and other populations
have fattier breasts.
We have to keep in mindthat the AI algorithm may have
been trained on a data poolthat maybe had denser breasts.
So what happens when wetake that same AI algorithm
and we apply it to adifferent group of women,
maybe a different age group,maybe a different density
(21:50):
group, a different racial group.
So we have to be very carefulthat when we train these AI
algorithms, that we make surethat we train them on a
diversity of mammogramsand breast types, because it
can train itself over time onone specific subtype
and then maybe won't be asuseful or sensitive on another.
- And who are you talkingto when you say, we have
(22:11):
to be careful to train the AIon the manufacturers of the,
- So manufacturers of the, so right now,AI is being developed
in large part as a kindof a conglomerate between
physicians as wellas manufacturers.
So the manufacturers or thedesigners of these softwares
are largely medical engineers.
Now, medical engineersdon't just make algorithms
(22:33):
without consulting physicians.
Oftentimes physicians are onthe boards for these things
because, you know, we needto know how to apply it
and who knows better how toapply it than the physicians.
We, we know what we need for our patients,and we also demand a certain
level of, you know, sensitivityand accuracy before we
think that it's good enoughto ultimately treat a patient
or even participate inthe treatment process.
(22:54):
I will say this though, that AI algorithmsthat exist today are already
light years better thanwhat was available even
just five years ago.
So that's very promising. Soit's, it's very interesting.
It's very important tocontinue researching.
But overall, I thinkthe trajectory of AI is
that it's only gonna help our patients.
- I can see that we're allfinding, the more we use ai,
the more comfortable we get with it.
(23:16):
I've seen the reading room, it's dark,you guys are sitting in
there for hours on end,like you said, you might
have 50 mammograms to read.
How do you even keep yourselfalert to look at each image,
each different image, isn't it?
And maybe it would just beeasy to let AI take over
and oh, if AI says it's okay,I'm gonna let this one go.
(23:37):
- Yeah, no,- I mean, how is it, how
do you keep from gettingthat lackadaisical about the, the work?
- So radiologists, I thinkwe're an interesting breed.
Like if you, if you talk toa neurosurgeon, you know, I,
I always, I think thatthey're like very revered
and I, I respect them immensely.
You know, they go in for 12 hour surgeriesand I sit here thinking, oh my gosh,
how do they do a surgery for 12 hours?
And I think with radiologists,I think the common
(23:58):
misconception is, oh my gosh,you're in a reading room,
how do you stay awake?
But it's so funny withradiologists, we love our jobs.
Most of us think of ourselvesas kind of detectives.
It's fun and it's enjoyable forus to actually help patients
and be able to do itthrough medical imaging.
We're kind of the onesthat enjoy looking at every
single millimeter of that imageand participating in the medical process
(24:23):
by finding disease in that way.
So while fatigue can happenacross any medical specialty
with radiology, you know, the dimnessof the reading room rarely has anything
to do with our alertness.
We, we are trainedand specialized in interpreting
mammograms for, for hours.
And quite frankly, most ofus would be able to do this
for eight to 10 hours if we so desired.
(24:43):
And most of us don't do that, though.
We typically read screeningmammograms in several hour
increments, And the AIsimply acts as a helper tool.
But it's more so just to help us in termsof our sensitivity, making sure
that we have a double checkand a two check system.
I would say it less hasto do with like fatigue
or anything of that sort,because most radiologists
know that, you know,if my fatigue limit is two hours,
(25:06):
they're not gonna continuereading beyond that.
So most radiologists have that,and we know at this point
as part of our process,when is it a safe time to kind of sign off
and then we resume work the next day?
- Explain the two check system.
- So when you have a two check systemor a double check system, it
used to be kind of theorized,like, what would it be like
if we had two doctors look atthe same mammogram just as a
(25:26):
double check to each other?
Because we really didn'thave back in the day two set
of eyes unless it was another human.
Now the problem withthat is that the capacity
to take two doctors away fromthe patients on a given day
to only be reading, for example, screens,is not very realistic,
nor is it efficient.
We want our doctors tobe in front of patients.
(25:46):
I wanna be with the patients, making surethat I interpret imaging
with them and alongside them,and I'm also performing
biopsies with them.
So when you can actuallyhave your double check
or your two part systembe a component of ai,
it increases efficiency.
And therefore we can have one doctor gothrough a screening list
confidently with a double checkand have another doctor perhaps
(26:08):
with the patients doingdiagnostics at the same exact time.
So efficiency is criticaland AI is very, very helpful. With that,
- Are there concerns about bias,whether the technology performs
equally well across allages, races, and breast densities?
What's being done to makesure that these tools are fair
and accurate for everyone?
- I think I should gointo a little bit about
(26:29):
what does it mean when we say ai, ai,because AI is a very umbrella term
and it doesn't really discuss,what are we talking about?
Like it, it, it's obviously an algorithm,it's a machine, but what does that mean?
Artificial intelligenceis typically referring
to something called adeep learning network.
So deep learning, when wedevelop a deep learning network,
which is ai, we're trying to simulatehow the brain works in a human
(26:50):
with a computer algorithm.
So we're trying to simulatehow the brain interprets
findings, whether that's imagingor a picture or, or
speech or even language.
So when you're talking aboutdeveloping a deep learning
network or ai, there areseveral ways you can do that.
A subtype of way of doingthat is something called
(27:11):
a convolutional neuralnetwork, A CNN.
So when we talk about makingAI in radiology for CT images,
or for MRI or from mammography, oftentimeswhat we're referring to
is creating an algorithmthat simulates how the human brain works.
So when we see something on a picture,our eyeballs take it in,
(27:33):
and then we convey thatinformation slice by slice.
In some scenarios, ifwe're talking about CTS
or with X-rays, it's oftentimestwo dimensional imaging.
And we convey that to our brain,and our brain does its
magic, like it always does.
And it says, okay, I made a decision.
This is what I think it is.
When you're training an AIalgorithm to do such a thing
through these convolutionalneural networks
(27:53):
or deep learning states, whatyou're doing is you're saying,
okay, I'm gonna train thealgorithm to take slice
by slice information orvoxel information from an MRI
or maybe just a two dimensionalimage from a 2D mammogram.
And it's gonna make a decisionand it's gonna ultimately
form a conclusion.
And that's how we typicallyform the AI algorithms.
(28:15):
Now, with respect to yourquestion about how do we train it
so that it's used to seeinga diversity of patients,
whether it's dense breasted patients,fatty breasted patients,
patients who are small, large,different ethnicities, races,
even different genders.
'cause we sometimes have to usemammography for men as well.
What we do is we increase the data set.
(28:36):
So typically what mostmedical engineers will do
and doctors is we say, okay,you can't train an AI
algorithm on only 10 patients.
Typically you need thousands.
And we try to make surethat that population is
as diverse as possible.
And then what we do iswe test the algorithm
against a doctor.
So a lot of these algorithmswhen they're being developed
(28:57):
will say, okay, what didthe breast radiologist think
this this patient had?
Did they say it was a negative caseor was it a positive case?
And then they're gonnatest the AI algorithm,
and if there's ever adifference, interpretation was
between the doctor and the AI algorithm.
Doctors interpretation isultimately what's gonna be used
to train the AI algorithm.
And the more times you canhave the AI algorithm learn
(29:21):
and practice on increasingmammograms in a diversity
of patients, it gets better and better.
It's very promising, and that'show we ultimately train it
to provide good data regardlessof the diversity of patients.
- Well, it sounds a loteasier than getting in a car
and getting behind a wheel and letting thecar drive you. Yeah, I mean,
- It's very complicated,but you know what, thank,
thankfully we have such a skilled,like medical engineers nowadays
(29:44):
and scientists who are working on this,and they are obviously
including medical expertsand doctors so
that we can ultimately developthis artificial intelligence.
So it's useful for us andnot the other way around.
- Have you had a discrepancyin what you have found versus
what the AI has found andand how do you handle that
if there is a discrepancy?
- Yes, absolutely. Ithappens actually quite often,
(30:04):
but it's nothing to be feared.
I gave an example earlier ofwhen patients have a history
of surgery and the AI algorithm thinks,oh my gosh, this is a distortion.
This must be like a cancer,like a huge speculated mask.
However, the doctor looks at the imageand knows that this patient,
they've been in the practicefor years, we know that they
had perhaps a reduction onthat breast or they had a lumpectomy.
(30:25):
We ultimately know what the results areand we know why it's flagging it.
So while it's stilluseful to double check,
ultimately if there's ever discrepancyand the radiologist looks at the finding
and has a reasonablereason for the finding,
the radiologist's interpretationas usual will be the one
that is interpreted, not the ai.
The AI is again, meant to bea help, not a replacement.
(30:46):
- Even if you in inform the, the softwareor the, the ai, the patient
had a, a lumpectomy here,the patient has dense breast tissue.
- It can still happen, unfortunately.
Yeah, it, it does its best.
You know, it does its bestwith dense breasted women.
It does its best to, you know,not mark everything under the sun.
And it's, it's reallyimproved over the course
of the last five to 10 years.
(31:06):
However, it's not perfect.
Again, it's not a human.
So it's very hard to trainan algorithm to know things
that a human perceives whenthey talk to a patient.
For example, if I go talk to a patientand the patient looks very
nervous and is guardingand perhaps guarding a certain
component of their breast,they may have already undergone
mammography and ultrasound.
But when you go in and talk to the patientand the patient then reveals,
oh, I, you know, I had,I actually fell several days ago
(31:31):
and I actually have this lumpon the left side of my breast.
The physical exam has not lost.
Its, its ju if you will,like, it's still important
to follow up with the patientand ask about history,
because sometimes findings on mammographyand ultrasound mean very little without a
pertinent patient history.
So for example, a hematomacan look like a big ginormous
(31:51):
mass on a mammogram.
If I know that the patientjust had a trauma a couple days
ago, I'm gonna lean with,this is probably hematoma.
If I don't have that historyand I just leave it up to an algorithm
or anything really to justsay, oh, there's a mass,
and then call it that, it maylead to unnecessary workup
for something that's quitesimply a benign finding.
- What excites you mostabout this technology
(32:12):
and what still needs to be figured- Out?
I think what excites me most personally ishow we can apply it to
different modalities,not just mammography.
So everybody kind oftalks about mammography
because that's what we havethe most exposure to, is
that we can use AI inour screening mammograms
and our diagnostic mammogramsas we've been discussing.
But you know, I think one of the thingsthat excites me in particular is
(32:33):
how we can use it in ultrasound or MRI.
So MRI is used in breastimaging a lot when it comes
to breast cancer screening detection,especially in our high
risk genetic carriers,high risk patients in general.
There are things that AIcan pick up, especially
with contrast enhanced exams like MRI,that we simply cannot
see with the naked eyeas opposed to mammography.
(32:55):
Most of the time when we talkabout AI with mammography,
it's not so much picking up thingsthat we can't see with our naked eye.
We can see them, it's just a matter of,did you look at this area with MRI?
However, when you give someone contrastand it spreads throughout the breast
and you start to see lesions pop up,sometimes you can train AI algorithms
to pick up on informationthat you just cannot
(33:15):
see with the naked eye.
For example, if youknow someone has cancer,
and I actually did a studyon this several years ago,
and if you know someone has cancerand they get an MRI exam to stage
and to see how extensive thecancer is, where did it go?
How big is it? It's very interestingthat AI we're finding has
the ability to be trainedto tell you whether or not a
(33:35):
mass will necessarily respondto chemotherapy
before the patient evenundergoes chemotherapy.
- Wow. - Now that's incredibleand that's incredibly useful
because you still need the radiologistto interpret the imaging.
However, it adds adifferent benefit that I
otherwise would not be able to offer.
It tells me, you know, based off ofhow this mass took up
(33:56):
the contrast on a very,very micro level, looking
at the voxels in an MRI,it can tell me whether or
not it thinks they're gonnarespond to neoadjuvant chemotherapy
before ultimately undergoing surgery.
So why is that important?
If I know from the AI algorithm,if we develop an algorithm
that's so good that itcan tell me, you know,
this person's not gonna respondto neoadjuvant chemotherapy
(34:18):
'cause this cancer is pickingup the contrast in such a way
that it just doesn't looklike it's gonna be responsive,
then I'm not gonna waste mytime potentially delaying the
patient's surgery for severalmonths waiting for the cancer
to shrink on chemotherapyif there was no chance
for it to in the first place.
Hmm. So it's, it offersadditional information
that perhaps we just wouldn'thave at our fingertips just
(34:41):
by simply imaging alone.
So this is just a littlekind of splice of the pie
and just an idea of how wecan apply AI in the future.
- Exciting stuff.
- Yeah, very.
The AI is an assistance tool,but it is in no way a
replacement for the physician.
The physician still hasto do the interpretation
and still ultimately has tomake a decision based off of all
of the information at hand,whether it be the
(35:01):
patient's history, surgery,health overall,
and what, what the ultimatequestion is for each exam.
- I, I appreciate thatbecause the more patients understand about
how AI is working,putting their trust still in their doctors
and how it's helping the doctors, I thinkthat they'll be more confident in
how they'll feel about thewhole screening experience.
(35:22):
- Right. - And not shying awayfrom mammograms. Dr. Sakla, thank
with us here on DocTalk.
- Awesome. Thank you guysso much for having me.
It was an absolutepleasure. This was very fun.
And yes, just remember, don't forgetto get your screening
mammograms every year.
- For more informationabout breast imaging
and screening at MedStar Healthvisit MedStar Health.org/breast health
(35:47):
or call 2 0 2 8 7 7 2 8 0 0.
If you would like tocomment on this podcast
or recommend a topic foranother episode of DocTalk,
send an email to DocTalk@medstar.net.
Popular Podcasts
Stuff You Should Know
If you've ever wanted to know about champagne, satanism, the Stonewall Uprising, chaos theory, LSD, El Nino, true crime and Rosa Parks, then look no further. Josh and Chuck have you covered.
Crime Junkie
Does hearing about a true crime case always leave you scouring the internet for the truth behind the story? Dive into your next mystery with Crime Junkie. Every Monday, join your host Ashley Flowers as she unravels all the details of infamous and underreported true crime cases with her best friend Brit Prawat. From cold cases to missing persons and heroes in our community who seek justice, Crime Junkie is your destination for theories and stories you won’t hear anywhere else. Whether you're a seasoned true crime enthusiast or new to the genre, you'll find yourself on the edge of your seat awaiting a new episode every Monday. If you can never get enough true crime... Congratulations, you’ve found your people. Follow to join a community of Crime Junkies! Crime Junkie is presented by Audiochuck Media Company.
NFL Daily with Gregg Rosenthal
Gregg Rosenthal and a rotating crew of elite NFL Media co-hosts, including Patrick Claybon, Colleen Wolfe, Steve Wyche, Nick Shook and Jourdan Rodrigue of The Athletic get you caught up daily on all the NFL news and analysis you need to be smarter and funnier than your friends.