August 12, 2025 • 39 mins
In this episode sponsored by Epredia, Dr. Anil Parwani explores the transformative journey of digital pathology from basic slide scanning to AI-driven diagnostics. He shares real-world implementation experiences and demonstrates how these technologies are addressing critical challenges in pathology practice.
- Pathology faces increasing demands amid workforce shortages and knowledge explosion
- Digital pathology provides standardization, objectivity, and automation beyond glass slides
- Ohio State University has scanned 4.2 million slides representing nearly 500,000 cases since 2016
- Current AI applications include biomarker quantification, rare event detection, and tumor classification
- Integration challenges remain the primary barrier to seamless adoption of AI tools
- Future technologies include virtual staining, 3D pathology, and large language model integration
- Artificial intelligence remains task-oriented while real intelligence is context-aware and knowledge-based
- Each institution must navigate their own "digital pathology chasm" based on specific needs
- Digital tools will augment pathologists' capabilities rather than replace human expertise
- The technology marketplace offers solutions for every stage of the digital transformation journey
This Episode's Resources
This Episode on YouTube
Coming soon!
Become a Digital Pathology Trailblazer get the "Digital Pathology 101" FREE E-book and join us!
Mark as Played
Transcript
Episode Transcript
Available transcripts are automatically generated. Complete accuracy is not guaranteed.
Speaker 1 (00:00):
What happens when academic pathology, technology
and innovation converge.
In this talk, dr Anil Parwanishares how his team is using AI,
machine learning and digitaltools to transform surgical
pathology workflows fromscanning to reporting to
education.
If you've ever wondered what ittakes to implement digital
(00:22):
pathology at scale in a hospitalsetting, or how AI is
practically used in diagnosticstoday, this presentation
delivers clear, real-worldanswers.
Speaker 2 (00:32):
Learn about the newest digital pathology trends
in science and industry.
Meet the most interestingpeople in the niche and gain
insights relevant to your ownprojects.
Here is where pathology meetscomputer science.
You are listening to theDigital Pathology Podcast with
your host, Dr Aleksandr Zhurov.
Speaker 3 (00:55):
Welcome to the morning event.
I'm going to talk to you todayabout the journey of digital
pathology.
It's different for everyone.
Some people in the audience arelooking for their first scanner
, right?
Some of us are looking at thefirst AI implementation at our
institutes.
Others are looking beyond that.
So it's really many differentjourneys and when I speak to
attendees, they share theirstories.
(01:16):
What are their pain points,what are the things that they're
looking forward to solving withtheir journey.
So how many of you are buying anew scanner?
A few of them.
How many of you have completelymoved to digital?
Several, several of you.
Excellent Greetings fromColumbus.
So I want to talk to you todayabout these journeys, the role
of digital pathology and AI, andlist some applications which we
(01:40):
can do today.
But I want to go beyond theglass slide, right?
So what are the barriers todaywhich can take us beyond the
glass slide?
So what are the barriers todaywhich can take us beyond the
glass slide and where are wetoday in that journey and what
are the potentials movingforward?
We have powerful microscopes,whole-site imaging scanning
systems, but today we havetechnologies which go beyond
(02:00):
that.
So I also want to talk aboutthat.
So it's really an exciting timeto be in medicine and pathology
.
We are generating lots of data,lots of data which is now being
converted into pixels, intobits and bytes, and we are
making decisions on that data.
We are sending the data to theelectronic medical record.
Patients are consuming thisdata in their portals.
(02:22):
They're looking at thediagnosis, they're trying to
decipher that.
What does it mean for mydiagnosis?
So it's truly an enabler ofclinical decision-making.
At the end of this, when youbuy your first scanner, when you
implement it, the goal is howdoes it transform your pathology
practice?
Right?
So that's what I want to talkto you about today.
So, as pathologists, we play acritical role in cancer
(02:44):
diagnosis.
Number of cancer cases continueto go up and the demand and
complexity of our services keepgrowing.
The workforce is continuing toshrink, there is a critical
shortage of pathologists and labstaff and we are learning more
about disease than ever before,right?
So all these combinedcombinations of shortage of
(03:07):
pathologists burnout newknowledge.
It requires some disruption, itrequires a new way of finding
things, a new way of discoveryand also a new way of making and
enabling diagnosis.
So, as pathologists, we performmany tasks, many manual tasks.
We count things, we annotateinformation, we assemble
(03:30):
information.
At the end, what is our report?
The report is our endpoint, butit's a continuum in the
patient's journey, right?
So if I look at all thechallenges in pathology today,
which is lack of standardization, subjectivity, many manual
processes, there are some labseven today which don't even have
(03:52):
barcoding.
We were at the ASDP sessionyesterday and they did a survey
which showed a lot of the labsaround the world don't even have
basic technologies.
I mean, we take things forgranted here, but globally there
is a shortage of pathologists.
Pathologists are overworked,right?
How many pathologists here arefor vacation?
(04:12):
They're here to play golf,right?
So if you ask pathologists ingeneral, they will tell you we
are overworked, right?
I see my friend Masood in theaudience there from a private
practice and he's alwayscomplaining we're looking for
pathologists, we have more workthan we can do, right?
(04:32):
So?
And there is an explosion ofmedical knowledge.
So what does digital pathologypotentially provide us?
Going beyond the glass slide,is standardization, more
objectivity, more automation,more accuracy.
Maybe you can do it faster Ifyou're first starting out.
You're signing out your firstfew cases.
You might say I'm slow on thedigital, but once you overcome
(04:55):
that, you will see thedifference In terms of why today
, why now, why in 2024?
So the current environment issuited for digital pathology and
AI models.
Ten years ago, we used to askthe question which scanner to
buy?
Right?
Many people used to ask thatquestion.
We now have cost-effective,high-performance computing,
which is cheaper, which is morereadily available.
(05:18):
Many academic centers havesupercomputers, so we now have
that environment which is neededfor growth of AI.
We also have better algorithms,right, so algorithms have
continued to evolve and theyhave become commercial grade.
There are some which are FDAapproved, and then you have more
data than ever before, right?
So if you look around the roomand if you ask each one of you
(05:42):
collectively within this room,we have several million images
scanned, maybe more than 100million images scanned globally.
Right, so we have experience.
We have data.
Is that data easily availableand shareable today?
No, but we are getting to thatpoint, right, we have several
collaborative networks andorganizations that are working
(06:03):
on it.
We have more.
We have FDA-approved algorithms, for example, for prostate
cancer.
We have clearly the cost ofmaking an H&E slide.
Digital has continued to go downand adoption has continued to
increase, right, so we havereally evolved from early
demonstration in thetelepathology in the 1980s to
(06:24):
2024, where we have several labswhich are digital and have
implemented AI, and if you walkaround the exhibit area you can
see all these products.
It amazes me.
I've been coming to thisconference for 10 years and
every year there is continuousinnovations and developments in
this area, so it's an excitingtime.
(06:45):
If you think about, where arewe, have we reached the ceiling
in terms of making a glass slidedigital?
So I think there are stillinnovations that need to happen.
There are specialized areascytopathology, for example,
hematopathology where we needdifferent types of scanning
devices.
Maybe Z stacking, maybealternative-stacking, maybe
(07:05):
alternative light sources, Idon't know.
We continue to have to continueto evolve and innovate the whole
slide scanning systems, but wecan go beyond that, right.
So today you have a suite ofscanning products out there.
If you go to the exhibit area,you can see many of these in
action and they can make slidesinto digital images.
(07:29):
They can do immunofluorescence,they can even do polarization
now.
So this is a given right.
So diagnostic quality imagesare a commodity now, right,
everybody agrees, right.
Who doesn't agree?
Who thinks we need more work inthis area?
So all the vendors are in theroom and there is room to grow
right, but on a given day.
(07:49):
So I'm on service today, I'mcovering GU and I can look at
images which are diagnosticquality and I feel confident in
my diagnosis as a pathologist.
We have achieved spatialsamplings of 0.25 microns per
pixel.
We are now able to createmultiplexing systems which allow
us to interrogate multiplebiomarkers.
(08:11):
There was a system I sawyesterday in the Wendell area
with 60 to 70 biomarkers couldbe interrogated on one slide.
We have gone beyond the glassslide to a digital image and now
we are moving forward.
So collectively, this is areally strong study.
It's a metadata andmeta-analysis of over 2,900 AI
and DP studies which looked atthe from multiple countries,
(08:35):
over 152,000 whole slide imagesrepresenting not just cancer but
many diseases.
And if you look at overall fromthese studies, 100 were
selected, 40 were drilled downinto it 96.3% sensitivity, 93.3%
specificity.
So clearly there are studieswhich have demonstrated that
(08:58):
glass slide and digitalpathology are equivalent or
digital images are non-inferioror not inferior.
These studies must continue.
We need to do more of these.
We need to do this in diversepopulations, not just very
focused on one region of thecountry, but globally, and I
think when we discuss this atthe ASDP, that is an important
(09:19):
aspect of moving these forward.
So any guesses where we are?
This is Orlando.
Who thinks it's Orlando?
This is Columbus moving theseforward.
So any guesses where we are?
This is Orlando.
Who thinks it's Orlando?
This is Columbus, ohio.
Yeah, so maybe this will bringsome recognition.
So we're excited Footballseason is on in Columbus right
now.
So I just briefly, with a fewslides, want to show you where
we are in this journey.
We started in 2016, and weimplemented.
(09:44):
We did retrospective scanninggoing back 10 years.
We started prospective scanning, we started primary diagnosis
and now we have severalpathologists who are in this
room who are completely digital.
We have David Kellogg here.
He runs the operation and thescanning site, and I think
several of our other teammembers are here as well.
But several of the pathologistsnow have come to a point where,
(10:07):
if the digital system goes down, david gets a lot of angry
emails.
So we've completely turned overright.
When we were first adopting it,there were one or two
pathologists and they were justthe lone rangers, but now many
pathologists are digital.
David just shared this datawith me yesterday.
Many pathologists are digital.
David just shared this datawith me yesterday Almost 4.2
million slides scanned, almosthalf a million cases scanned.
(10:30):
Year-to-year volume growth, youcan see, continues to grow.
And what I love about this isthe user engagement and that's
the key right.
You want to create anenvironment where users are
excited to use the system.
This is a critical piece ofchange management right.
Change management is not easy,but it'll come with user
engagement.
(10:50):
If you have two or three we hada good discussion with my
colleagues from UAB yesterdayabout this and how do you start
this journey?
It's getting a few people superengaged and get this journey
started.
So the scan slides areinstantaneously available.
They're linked to our labinformation system and we are
continuing to add more bells andwhistles to this system.
(11:12):
Right now we move from acase-level integration.
Now we have slide-levelintegration so we can
individually call out the slides.
All those 4.2 million slidesare available.
We can consult with colleagueseasily.
This, to me, has been a gamechanger.
Our neuropathologists are in adifferent building.
Our dermatopathologists arefour or five miles away.
If I have a difficult penilebiopsy, I can just press a
(11:34):
button and connect with them inreal time.
And frozen sections andeverything else is much more
easier and streamlined becausewe have a way to share images.
We have implementedPatPresenter for our
consultation work and that's nowbeing integrated into into our
system, so scan slides areinstantaneously available.
There is no waiting forfoldering.
(11:55):
It doesn't matter which orderthey are scanned in.
They show up in your queueready to be signed out, and we
are starting to explore cytologypreparations FNAs using small
scanners on the bedside.
So, in summary, digital slideshave continued to improve my
workflow.
I love the direct interface tothe LIS.
(12:16):
I love sharing cases withconsultants and colleagues and
flagging cases, looking at theprior cases, so it has improved
my turnaround time.
Because I'm not.
Our histology lab is three orfive, three miles away from the
main campus, so it's allowed usto share those images more
easily.
All right, quiz now which cityis this?
(12:36):
No, close to Columbus, though.
It's Cleveland.
Very close to it.
So if you take 71, you haveCleveland, columbus and
Cincinnati.
So 71 corridor, right.
So what do we do next?
Right, as the systems mature,as you start your own journey of
buying your own scanners,implementing it as these systems
(12:56):
mature, how do we go beyondthat right?
How do we start thinking aboutother things which we cannot do
with glass slides as adoptionincreases.
So, if you look at a typicalproduct adoption curve, you have
innovators, you have earlyadopters, but many of them hit
this chasm right.
So every one of you have theirown chasms.
You might be starting yourjourney, you might be exploring
(13:20):
how to implement digitalpathology, but the key is
finding out what is your painpoint, what is your chasm?
How do I implement digitalpathology?
But the key is finding out whatis your pain point, what is
your chasm?
How do you go beyond this chasm?
And we are facing the samething, right, as we build, as we
think about implementing AI, wehave to deal with integration,
we have to deal withinteroperability.
All these are key things.
So how do you go beyond theglass slide chasm?
(13:40):
Right?
So we have established in thisroom we can now create
diagnostic quality images.
How do we go from there?
Right?
So what can we do with digitalimages that we cannot do with
glass slides?
So it really boils down tomanaging the information,
sharing the images, connectingwith each other, connecting with
experts, but also exploitingthe pixel pipeline to build
(14:05):
algorithms or buy algorithms toidentify, quantitate, synthesize
and create knowledge pathways,create important clinical
decision-making skills, right?
This is the next part of my talk.
I'm going to focus on what aresome institutes doing and where
are we today with this chasm?
Where are we today withcomputational pathology and AI
(14:28):
for clinical decision-making?
So these are some of the thingswhich I feel are already here
today, in 2024, right, we havevery sophisticated image
analysis algorithms, aialgorithms for detection and
diagnosis.
Right, to analyze digitalimages, decipher the pixels,
(14:49):
identify features likemorphology, cell shapes, size of
nuclei, architecture andstaining patterns.
So, pattern detection, featuredetection these are things I
learned as a resident, and Ilearned that by looking at over
and over again and building analgorithm in my brain which
allowed me to look at a prostategland and say, okay, this is
(15:10):
cancer, because A, b, c, d, e,same thing, right.
So we are at a point in 2024where many institutes have
implemented AI in pathology.
Right, and they started withvery basic things counting cells
, looking at the size of nuclei,differentiating different
nuclei, differentiating positiveversus negative signals.
(15:32):
So all these focus onbiomarkers which could be
diagnostic or predictive.
Right, so these could includedetecting, classifying,
segmenting, quantifying andlocalization.
Here are the four applicationsthat we are starting to use in
our Institute.
And again, we have our ownchasm.
Our chasm is that we cannotreadily integrate many of these
(15:54):
algorithms.
So we still have to go to athird party, launch the
algorithm and use the algorithmand bring this information back
to the LIS.
In an ideal world, in my wishlist, we want this to be
completely integrated and that'swhere we are heading towards.
But quantitative digital imageanalysis for biomarkers this is
very, very routinely done.
(16:15):
Now there is a separate CPTcode for this.
You can actually get paid alittle bit more, maybe, I don't
know twenty, thirty dollars morefor a digital image analysis
system.
If you use in your lab, it'smore objective, more accurate,
more faster.
But is it easy to do?
Is it cheap?
So the answer is it is easy todo.
(16:36):
It's easier to do if youalready have digital slides in
your system.
It may not be the firstapplication you launch in your
lab, right?
If you look at this gastricneuroendocrine tumor and you can
see a 2-millimeter square areawas analyzed in 28 seconds.
So if I asked each one of youin the room to count the blue
dots and I give you threeseconds, you all will have a
(16:59):
different answer.
Right?
Everybody agree with thathypothesis.
And if I showed you this, it'sgonna be even harder, right, but
we can get this dataobjectively and it's
reproducible every time you doit, if the answer will be the
same in Cleveland or in Columbus, as long as you're using the
same algorithm and you'vevalidated it.
(17:20):
It's a locked system and so on.
These are a given.
Other type of algorithms we areusing is detecting rare events.
This is a lymph node detectionalgorithm, right, so we can
actually launch it directly fromthe viewer today.
But we still have to do a lotof manual work to get this
algorithm queued up.
But pathologists still use it.
Pathologists still want to useit.
(17:40):
So imagine a world where thisis even easier to do.
It will become much more easierto use.
So identifying metastatic foci,and the pathologist in the room
might say why do you need analgorithm?
I can just eyeball it.
This is cancer.
What if you have a few rarecells and you missed it?
But the computer didn't miss it?
These are type of things I callthem rare event detection.
(18:02):
Overall, right, it could befinding microorganisms and so on
.
Let's go beyond that, right.
So today we can find metastaticcancer in lymph nodes.
But we want to go beyond thatright.
Pathologists today doimmunostains to figure out which
cancer this is Tumor of unknown, primary, unknown origin we do
several immunostains.
I had a case last week.
(18:23):
I had to do 20 immunostains andconsult a metopathologist and
soft tissue pathologist and eventhen you know guess how it was
signed out High-grade malignancy, see comment.
So this is an ongoing issue indiagnostic pathology and today
we have algorithms which canpredict for you, just like when
you send it for next-gensequencing.
(18:44):
It can predict with greaterthan 95% certainty this is renal
cell carcinoma.
So this is an example of such aprediction model where the
computer has predicted this iscolorectal cancer.
We have GI pathologist, dr Chen, in the audience and she will
say why do you even need AI forthis?
I can look at this and say thisis colorectal cancer.
(19:04):
It has necrosis and dirtynecrosis and all the features of
colorectal cancer.
But the point is there arealgorithms out there and they
are available.
They will continue to evolveand get better and when you are
ready in your journey, you willbuy that algorithm and use it.
But before you do that, westill have to solve the
interoperability issues.
(19:25):
We still have to integrate themand some sites have, some
institutes have done better withthese, others have not.
So if you look at diagnosticsfor cancer overall, they can
help you in a pre-sign-outprocess.
They can help you duringsign-out or they can help you
post-sign-out.
In pre-sign-out setting itcould be a good screening tool.
(19:45):
I showed you the data fromhundreds of studies about
sensitivity, right?
So for prostate cancerspecifically, similar studies
have done and shown highsensitivity and high specificity
and it's one of the mostcommonly exploited cancer for
building algorithms.
Like every company I talk to,they're building their own
(20:06):
prostate cancer algorithm,pre-ordering IHC, right?
Imagine a world where you'recoming to work and the system
automatically screens the casesand chooses the ones that you
should do immunos on, anddoesn't do it automatically but
makes recommendation to youDuring sign out.
It can find other features,like intraductal cancer of the
prostate.
(20:26):
It can help you withsupervision, primary diagnosis
or without supervision.
It can also create automatedreporting templates for you and
then do a second review, likewhat about post sign out, right?
So imagine a world where youare overworked pathologist,
you've just booked a ticket togo to Las Vegas and it's six
o'clock and you're trying torush through your cases and you
(20:48):
make a mistake.
But you have the safety net,you have a gatekeeper, you have
the AI assistant.
They check your work and saywait a minute before you board
the flight.
Are you sure you want to callthis cancer?
And you'd look at it again andagain and maybe you change your
diagnosis.
So some of the studies outthere have looked at this
specifically and there are caseswhere small foci or cancer were
(21:10):
missed.
Did it make a difference?
Maybe not.
Maybe there was cancer in othercores, but what if it was the
only core and you missed it andyou wish you had the system, or
not?
So how many of you if this wasfree right, all the vendors will
give this to you for free howmany of you will use it?
And what if it wasn't free?
You'll still use it.
(21:33):
So again, we are also, as webuild and use these algorithms,
we're learning new informationabout these cancers.
So, again, these tools arebecoming, I would say, not in
routine use, but they're gettingvery close to routine use.
So you can see this work listas you are looking at your work
list for the day and the caseshave been flagged for you.
(21:56):
So I would say we're very closeto getting a routine use of
these algorithms and, again, youcan turn it on and off.
If you're in a residencyprogram, you can have the
resident use it or not.
You can use it as a way toassess their competency, right?
You can make them review thecase and then turn on the AI or
(22:16):
make it so.
Ai will only be turned on afterfive minutes of review.
How many residents in the roomYou're ready to use AI, right?
So this is an example of breastcancer, right, invasive lobular
carcinoma.
And again, the red areas arewhere the cancer is.
You can go look at it moreclosely.
You can also have algorithmswhich can distinguish the
(22:37):
subtypes of breast cancer orsubtypes of prostate cancer.
What about finding mitosis?
Who likes to count mitosis?
Residents, you like to count,you know?
I asked, I did this test.
I asked three residents tocount mitosis on the same slide
and I got three differentanswers and I would say
pathologists would also give thesame different answers.
Right?
(22:57):
But you don't have to, right,today, if you have AI, you've
implemented it.
They will find the mitosis foryou and let you check it or not.
So again, this is an example oflymphoma classification, or
finding abnormal white bloodcells in a smear.
These systems are already inplace, like many of these
systems are being used in theclinical lab.
(23:18):
What about finding acid-fastbacilli?
Who likes to do that?
These are algorithms which areavailable.
Some of the centers haveimplemented it.
What about finding H pylori?
So these systems are beingbuilt.
They're going to becomecommercially available.
There will come a time where youwill have your own personalized
(23:40):
dashboard of apps that you candownload from the App Store, and
you'll have to pay asubscription to it.
Some of them will be free, someof them will be per click, but
these are things which we didn'thave when we were talking about
whole slide imaging.
So let's go beyond that.
What else is coming?
So we were talking about wholeslide imaging, so let's go
beyond that.
What else is coming?
Right?
So I talked about the wholeslide imaging.
Is that the end of the journey?
(24:00):
Is that the end of a glassslide?
What about starting from thetissue?
Right?
So we're getting closer toradiology, and radiology is
getting closer to pathology, soit's a spectrum.
Right?
So you have synthetic data, wehave virtual staining, we have
3D pathology, integrated largelanguage models.
We have all these new toolswhich will become available in
(24:23):
your work list soon, from pixelsto diagnosis, to prognosis and
predictions.
Now we talked about thesegmentation.
But imagine a world where Iwant to go inside this segmented
area and pick one or two cellsand interrogate those cells.
What if I want to look at thisimage in three dimensions, right
(24:44):
?
What about just finding similarthings, right?
Today you can take a picture onyour phone and you're going to
the supermarket and you take apicture and find similar
products.
It gives you the prices.
What if you could do this inpathology?
When I was a resident, how did Ilearn?
I learned by opening textbooks.
(25:04):
I was in the resident room andit was 7.55.
The attending was going to comethere at 8 o'clock.
I didn't know what this was.
I wanted to scribble something,so I flipped through pages and
I put something down.
At least I tried.
But today this could be doneelectronically.
It could be done usingartificial intelligence, where,
in this case, the query imagewas a brain ependymoma and the
(25:25):
algorithm found the best matcheswhich were similar to this and
created.
At least it gave you anindication of what you're
dealing with.
And you're going to seeexamples of these different
things, like virtual staining.
So if you walk around theexhibit area, there are vendors
out there which have worked onit, right?
So this, to me, will be goingbeyond the glass slide.
(25:48):
So what about prostate biopsieswhich are unstained?
On the top panel A, you haveunstained prostate biopsies.
In panel B you have H&Es whichwere generated in the lab.
And in panel C you haveprostate biopsies which were
generated by virtual staining,and then you can actually see
(26:09):
the areas where the cancer was.
So what about staining, right?
See the areas where the cancerwas.
So what about staining, right?
So, as a prostate pathologist,we do prostate triple stain to
look for basal cells.
So here you see, on theright-hand side you have real
HNE and on the left-hand sideyou have virtual HNE.
But what is amazing to me isthe ability to now predict where
(26:30):
the stain might be and evenpredict intraductal carcinoma,
which is a difficult diagnosis,in my opinion, you know, and
there is a lot of controversyaround it.
So 3D pathology is coming.
We have commercially availablescanning systems now which can
scan in 3D, right?
So here you have a cancerprostate core and a benign
prostate core.
(26:50):
So what about helping you whileyou're signing out, right?
So these are slides from DrSingh, who's built this in Path
Presenter the ability to createa chat with the image.
So imagine a world where you'relooking at this image.
You start chatting with thechatbot what is this?
And help.
These large language models canhelp proofread your report,
(27:12):
alert pathologists if you hadmissing information.
You know this happens whenyou're in a hurry.
You're rushing through a case.
You might miss information, youmight put something like a T4
where it was actually a T2.
What if you had this way toproofread your reports and help
with billing right?
So imagine a world where youhave this image where you can
(27:33):
press a button and look at thefeatures for the residents and
the trainees.
It can highlight the featuresin this image, but it can also
suggest a diagnosis for you andsuggest some IHCs for you and
also try to help you findsimilar cases, create a
differential diagnosis, takethese images and put them into
your tumor board pile andpresent it at the tumor board.
(27:55):
So these type of tools arecoming in many areas in
pathology and again, we're notthere yet today.
But what I'm showing you todayare prototypes of things that
are coming in the pipeline.
So we've gone beyond the glassslide.
We have now demonstrated thatwe can make a good diagnosis on
digital images.
(28:15):
We have demonstrated that thereare algorithms today which can
be used clinically, and I'veshown you examples of workflows
that are coming soon.
So I just want to also separateartificial intelligence and real
intelligence, right?
So when you think aboutartificial intelligence, it's
very task-oriented.
Right?
So when you think aboutartificial intelligence, it's
very task-oriented.
Find me this feature how manynuclei are positive?
(28:40):
Biomarker quantification, lymphnode met?
I showed you all these examples, but they're very task-oriented
.
Each algorithm is composed ofsmall steps, but real
intelligence is goal-orientedMany algorithms per task.
So, as a pathologist, if youlook at your journey from
residency, from medical school,to where you are as an
experienced pathologist, it'sexperience-based, it's
context-based, it'sknowledge-based.
(29:01):
So we're not there today wherewe can say AI is equal to real
intelligence, right?
So that's a journey, just likedigital pathology is a journey.
So this is an example of a caseof prostate cancer where, on one
, one of the cores, I haveclearly established prostate
cancer, right here and then thisis which looks like prostate
(29:22):
cancer, and I actually signed itout as prostate cancer, and
that's where real intelligencecomes in.
So I signed out this case, Ireleased it into the patient's
medical record, but somethingbothered me when I reached home.
You know, that night.
I just had a thought that this,maybe this is rectal cancer.
Maybe we should think about it.
So I dug through the notes.
I found a note from one of theprimary care visits where this
(29:47):
patient has rectal bleeding butrefused endoscopy because it was
too expensive.
So I went in, I ordered somestains and it turned out to be
colon cancer.
So I had to admit I was wrong.
I amended my report, I calledthe urologist and the patient
actually got treated for rectalcancer and prostate cancer.
So I think that's where we needto be and we might get there in
(30:09):
20, 30 years, but we havesignificant advances in the
field even today.
So, putting it all together,everybody went to Magic Kingdom.
I was there last night Amazingfireworks, the best fireworks.
I mean, I went to Disney withmy kids many years ago, but I
went now Amazing, so I recommendit highly.
I don't have any stocks inDisney or anything.
(30:29):
So, in conclusions, how willdigital pathology and AI help
pathologists?
Right?
So we talked about the declinein number of pathologists,
increasing workloads, fewertrainees going into pathology,
and clearly what I've shown youtoday can help some of those
issues.
Help us with sharing cases,help us with connecting
(30:50):
subspecialists together,connecting spaces together.
We have increasing workloads,so are we ready for a digital
disruption?
And then augmenting yourdiagnosis, checking your work,
do some of your work and sharethe work with others right, so
you have to tame your owndigital pathology chasm, right.
Each one of you is here in thisconference because you want to
(31:12):
learn more about it.
If you've already bought thescanner, you want to learn about
AI.
If you've already implementedAI, you want to learn about 3D
pathology.
But it's a journey, right, soall of you have to solve it on
your own.
I'm not going to you know.
I'm just showing you where wewere 20 years ago, where are we
today and where are we going 20years from now.
So maybe you will have aworkstation like this in your
(31:36):
office where you will customizeit.
You'll buy your own apps.
You will create your own chasmbuilding, right, so you will use
AI to assist you to augmentyour diagnosis.
And maybe autonomous drive,right.
Yesterday, my friend drove mefrom the fireworks back and the
car was driving itself and heleft the steering wheel and I
(31:56):
was super scared and it waspretty crowded.
It was Sunday night, it wasrush hour.
I said, no, I'm not ready forthis.
So take over the steering,please.
I have a talk to give at 7.30.
So he didn't listen to me andthe car drove itself for 10
minutes and it did fine, right.
But you have to build thattrust.
You have to build that trustwith AI, right.
(32:18):
So where are we going?
Right?
The journey from glass todigital to prediction will
continue for all of us.
We are today.
We can improve our analysis.
What will come is next, we'llbe improving your diagnosis.
Maybe we will have moreintegration.
We will solve theinteroperability issues,
regulatory issues, reimbursementissues, but in the future, this
(32:39):
will really be clinicaldecision-making exercise and
integrating multiple types ofdata, and we're probably this
will be the era of AI-basedprecision medicine in its true
sense, and that's probably inpathology.
Ai could probably diagnose easycases independently, just like
the Tesla was drivingautonomously yesterday, but I
(33:02):
don't know how many pathologistsfeel comfortable about that
riding the Tesla of pathology.
So with this I'm going toconclude.
It's been an amazing journeyand I think we're going to.
I actually encourage you to talkamongst yourselves.
Go to the exhibit hall, look atall these different products
out there, and there issomething for everyone.
You know you might be justbuying your first baby scanner.
(33:25):
You might be just getting thatone big check from your
administration and you're readyto spend it.
This is the place to do it.
This is the Disney world ofdigital pathology.
So with this, I'm going to endwith a picture of football again
, which I'm excited.
Next weekend I'm going to be atthe game, and so I'm going to
stop and if you have anyquestions, I'll be glad to
answer it.
(33:45):
I'm going to be around for therest of the conference, so I
hope to interact with many ofyou.
You and I want to thankApreedia for inviting me here as
a speaker, and looks like afull house.
I'm sure it's because of thebreakfast.
That was really good.
Speaker 4 (33:59):
Thank you all thank you so much, dr Parwani, for
your lecture and I think it wasamazing because it gave this
like from the very beginning,when scanning was a problem, to
now like doing 3D pathology andactually everybody who is in
this room and who comes to thisconference, they can be at any
like single point of what youdescribed.
(34:20):
So let's start with thebeginning, and you were talking
about the chasm.
So everybody has their ownproblem, their own pain point to
solve.
But let's start with thescanners.
What scanners do you?
Speaker 3 (34:30):
have so.
So we have variety of scanners.
Uh, we have philips, we haveaperio uh laika, we have
hamamatsu and we just got thepredia.
Which one did you get the 250the fluorescent.
Yeah, yeah, so we're using itfor our kidney biopsies now is
that the one that has thepolarization?
Option as well.
Speaker 4 (34:51):
I remember in our podcast you were saying why they
are not adopting the renalpathologist, and now there is a
tool for them.
Speaker 3 (34:58):
Yes, I think that to me is an exciting part, where
you can actually see some of theusers who are turned off by
digital if they cannot do A, band C.
Today we have the capability ofproviding those options with
different types of scanners.
So I think one scanner may notaddress all the needs of a big
academic center but a smallerlab can actually do with one or
(35:18):
two scanner types.
But there are so manycomplexities in pathology and
I'm glad many of the vendors outthere, including Apreedia,
leica, are actually solvingthose problems individually or
incorporating them into yourdashboard, into their system.
You know, just like when you goto Best Buy to buy the next TV,
you have all these bells andwhistles, but they specifically
(35:39):
build those based on feedback.
You know, today I cannot do this.
Can you build this?
So I think digital pathologyscanner market is also evolving
in that direction.
To me, it's like multipletrains have left the station and
everyone has their own trainand they have their own next
stop.
The next stop might be I wantto do frozen.
The next stop might be I wantto use image analysis.
(36:00):
I want to do large languagemodels.
I want to go directly from thetissue to an image, you know.
so, all these trains, individualtrains, are users and they have
expectations and they have endand they have an endpoint.
Not an endpoint but a stop onthe way.
And they want to get off thetrain at that.
Stop, do something and then geton the train again.
Speaker 4 (36:19):
Figure out what the right course is.
Speaker 3 (36:20):
So I always think it's fascinating to me to come
to a conference like this, whereyou have different types of
users, not just pathologists,but also technicians, also
students, technologists, itvendors lab managers and they
work together to solve complexproblems.
Well, this is what I love aboutthis.
Speaker 4 (36:39):
I like it about digital pathology because
usually in medicine you don'thave a team with so many
different expertises.
You mostly are with medicalprofessionals and it's driven by
the medical professionals, andhere you have technology
specialists, you have operations, you have, obviously
pathologists, you have theadministrators and I love it.
(37:03):
It's super interactive.
Speaker 3 (37:04):
Yeah, no, I think you're exactly right.
You have the administrators whowrite the checks, you have the
end users, you have trainees whoare going to be the future of
pathology.
So they're all coming togetherfor finding their own journey,
finding their own discovery.
You know why digital pathology?
Why now?
Why me?
And I think that is animportant thing, which is
(37:26):
important for your listenersalso to know.
Speaker 4 (37:28):
And we just heard from one of the people who was
asking questions like OK, now wegot the green light, how do we
bring everybody on board?
Correct, Because I think peopleare.
Well, it's with any changemanagement you're focused first
on getting the green light andgoing somewhere, and then it
turns out you have to bringeverybody else with you.
How do you do that?
(37:48):
That's the next step, step nextstop on this journey.
Speaker 3 (37:52):
Yeah, I think, I think you have to take baby
steps.
You have to again as avisionary.
You have to look at the visionof where you want to reach right
and that's where you're here,so, but then the vision could be
.
I want to do these steps andthis will require these changes.
A good leader is also a goodchange manager you know if
you're not, you're not going tobe a good leader if you cannot
(38:13):
manage change.
The change can occur at peoplelevel.
The change could be at thetechnology level, it could be
financial, it could beregulatory, but collectively you
have to orchestrate all thosechanges to see and execute your
vision.
You have to hire the rightpeople, the right team and keep
them engaged, keep them excitedin this journey so what is your
(38:35):
chasm right now?
Speaker 4 (38:36):
what are you guys working?
Speaker 3 (38:37):
yeah.
So our chasm is, like Imentioned.
There are different taintpreparations, light preparations
, gosh frozen all those.
How do we bring all thistogether in one seamless way?
So the chasm that we're tryingto solve is integration with all
the AI tools.
Do we need to launch multipleviewers?
Do we need to createtechnologies which are not
(39:02):
compatible and try to bring themtogether?
So that is the chasm we'retrying to do.
You know we can creatediagnostic quality images.
We know we can use them.
I know we can make diagnosis onthem, but I think the next gap
is can I create a dashboardwhere I only have three AI tools
but they could be launched fromone viewer, and so on?
Speaker 4 (39:21):
For the ease of use for the pathologist, because
that's also a step in adoptionjourney.
When it's not seamless, peoplewere not already convinced and
not willing to troubleshoot.
That's gonna be a showstopperfor them.
Speaker 3 (39:33):
Yes.
Speaker 4 (39:34):
I love your stories.
Thank you so much for thisfantastic presentation.
What happens when academic pathology, technology
and innovation converge.
In this talk, dr Anil Parwanishares how his team is using AI,
machine learning and digitaltools to transform surgical
pathology workflows fromscanning to reporting to
education.
If you've ever wondered what ittakes to implement digital
(00:22):
pathology at scale in a hospitalsetting, or how AI is
practically used in diagnosticstoday, this presentation
delivers clear, real-worldanswers.
Speaker 2 (00:32):
Learn about the newest digital pathology trends
in science and industry.
Meet the most interestingpeople in the niche and gain
insights relevant to your ownprojects.
Here is where pathology meetscomputer science.
You are listening to theDigital Pathology Podcast with
your host, Dr Aleksandr Zhurov.
Speaker 3 (00:55):
Welcome to the morning event.
I'm going to talk to you todayabout the journey of digital
pathology.
It's different for everyone.
Some people in the audience arelooking for their first scanner
, right?
Some of us are looking at thefirst AI implementation at our
institutes.
Others are looking beyond that.
So it's really many differentjourneys and when I speak to
attendees, they share theirstories.
(01:16):
What are their pain points,what are the things that they're
looking forward to solving withtheir journey.
So how many of you are buying anew scanner?
A few of them.
How many of you have completelymoved to digital?
Several, several of you.
Excellent Greetings fromColumbus.
So I want to talk to you todayabout these journeys, the role
of digital pathology and AI, andlist some applications which we
(01:40):
can do today.
But I want to go beyond theglass slide, right?
So what are the barriers todaywhich can take us beyond the
glass slide?
So what are the barriers todaywhich can take us beyond the
glass slide and where are wetoday in that journey and what
are the potentials movingforward?
We have powerful microscopes,whole-site imaging scanning
systems, but today we havetechnologies which go beyond
(02:00):
that.
So I also want to talk aboutthat.
So it's really an exciting timeto be in medicine and pathology
.
We are generating lots of data,lots of data which is now being
converted into pixels, intobits and bytes, and we are
making decisions on that data.
We are sending the data to theelectronic medical record.
Patients are consuming thisdata in their portals.
(02:22):
They're looking at thediagnosis, they're trying to
decipher that.
What does it mean for mydiagnosis?
So it's truly an enabler ofclinical decision-making.
At the end of this, when youbuy your first scanner, when you
implement it, the goal is howdoes it transform your pathology
practice?
Right?
So that's what I want to talkto you about today.
So, as pathologists, we play acritical role in cancer
(02:44):
diagnosis.
Number of cancer cases continueto go up and the demand and
complexity of our services keepgrowing.
The workforce is continuing toshrink, there is a critical
shortage of pathologists and labstaff and we are learning more
about disease than ever before,right?
So all these combinedcombinations of shortage of
(03:07):
pathologists burnout newknowledge.
It requires some disruption, itrequires a new way of finding
things, a new way of discoveryand also a new way of making and
enabling diagnosis.
So, as pathologists, we performmany tasks, many manual tasks.
We count things, we annotateinformation, we assemble
(03:30):
information.
At the end, what is our report?
The report is our endpoint, butit's a continuum in the
patient's journey, right?
So if I look at all thechallenges in pathology today,
which is lack of standardization, subjectivity, many manual
processes, there are some labseven today which don't even have
(03:52):
barcoding.
We were at the ASDP sessionyesterday and they did a survey
which showed a lot of the labsaround the world don't even have
basic technologies.
I mean, we take things forgranted here, but globally there
is a shortage of pathologists.
Pathologists are overworked,right?
How many pathologists here arefor vacation?
(04:12):
They're here to play golf,right?
So if you ask pathologists ingeneral, they will tell you we
are overworked, right?
I see my friend Masood in theaudience there from a private
practice and he's alwayscomplaining we're looking for
pathologists, we have more workthan we can do, right?
(04:32):
So?
And there is an explosion ofmedical knowledge.
So what does digital pathologypotentially provide us?
Going beyond the glass slide,is standardization, more
objectivity, more automation,more accuracy.
Maybe you can do it faster Ifyou're first starting out.
You're signing out your firstfew cases.
You might say I'm slow on thedigital, but once you overcome
(04:55):
that, you will see thedifference In terms of why today
, why now, why in 2024?
So the current environment issuited for digital pathology and
AI models.
Ten years ago, we used to askthe question which scanner to
buy?
Right?
Many people used to ask thatquestion.
We now have cost-effective,high-performance computing,
which is cheaper, which is morereadily available.
(05:18):
Many academic centers havesupercomputers, so we now have
that environment which is neededfor growth of AI.
We also have better algorithms,right, so algorithms have
continued to evolve and theyhave become commercial grade.
There are some which are FDAapproved, and then you have more
data than ever before, right?
So if you look around the roomand if you ask each one of you
(05:42):
collectively within this room,we have several million images
scanned, maybe more than 100million images scanned globally.
Right, so we have experience.
We have data.
Is that data easily availableand shareable today?
No, but we are getting to thatpoint, right, we have several
collaborative networks andorganizations that are working
(06:03):
on it.
We have more.
We have FDA-approved algorithms, for example, for prostate
cancer.
We have clearly the cost ofmaking an H&E slide.
Digital has continued to go downand adoption has continued to
increase, right, so we havereally evolved from early
demonstration in thetelepathology in the 1980s to
(06:24):
2024, where we have several labswhich are digital and have
implemented AI, and if you walkaround the exhibit area you can
see all these products.
It amazes me.
I've been coming to thisconference for 10 years and
every year there is continuousinnovations and developments in
this area, so it's an excitingtime.
(06:45):
If you think about, where arewe, have we reached the ceiling
in terms of making a glass slidedigital?
So I think there are stillinnovations that need to happen.
There are specialized areascytopathology, for example,
hematopathology where we needdifferent types of scanning
devices.
Maybe Z stacking, maybealternative-stacking, maybe
(07:05):
alternative light sources, Idon't know.
We continue to have to continueto evolve and innovate the whole
slide scanning systems, but wecan go beyond that, right.
So today you have a suite ofscanning products out there.
If you go to the exhibit area,you can see many of these in
action and they can make slidesinto digital images.
(07:29):
They can do immunofluorescence,they can even do polarization
now.
So this is a given right.
So diagnostic quality imagesare a commodity now, right,
everybody agrees, right.
Who doesn't agree?
Who thinks we need more work inthis area?
So all the vendors are in theroom and there is room to grow
right, but on a given day.
(07:49):
So I'm on service today, I'mcovering GU and I can look at
images which are diagnosticquality and I feel confident in
my diagnosis as a pathologist.
We have achieved spatialsamplings of 0.25 microns per
pixel.
We are now able to createmultiplexing systems which allow
us to interrogate multiplebiomarkers.
(08:11):
There was a system I sawyesterday in the Wendell area
with 60 to 70 biomarkers couldbe interrogated on one slide.
We have gone beyond the glassslide to a digital image and now
we are moving forward.
So collectively, this is areally strong study.
It's a metadata andmeta-analysis of over 2,900 AI
and DP studies which looked atthe from multiple countries,
(08:35):
over 152,000 whole slide imagesrepresenting not just cancer but
many diseases.
And if you look at overall fromthese studies, 100 were
selected, 40 were drilled downinto it 96.3% sensitivity, 93.3%
specificity.
So clearly there are studieswhich have demonstrated that
(08:58):
glass slide and digitalpathology are equivalent or
digital images are non-inferioror not inferior.
These studies must continue.
We need to do more of these.
We need to do this in diversepopulations, not just very
focused on one region of thecountry, but globally, and I
think when we discuss this atthe ASDP, that is an important
(09:19):
aspect of moving these forward.
So any guesses where we are?
This is Orlando.
Who thinks it's Orlando?
This is Columbus moving theseforward.
So any guesses where we are?
This is Orlando.
Who thinks it's Orlando?
This is Columbus, ohio.
Yeah, so maybe this will bringsome recognition.
So we're excited Footballseason is on in Columbus right
now.
So I just briefly, with a fewslides, want to show you where
we are in this journey.
We started in 2016, and weimplemented.
(09:44):
We did retrospective scanninggoing back 10 years.
We started prospective scanning, we started primary diagnosis
and now we have severalpathologists who are in this
room who are completely digital.
We have David Kellogg here.
He runs the operation and thescanning site, and I think
several of our other teammembers are here as well.
But several of the pathologistsnow have come to a point where,
(10:07):
if the digital system goes down, david gets a lot of angry
emails.
So we've completely turned overright.
When we were first adopting it,there were one or two
pathologists and they were justthe lone rangers, but now many
pathologists are digital.
David just shared this datawith me yesterday.
Many pathologists are digital.
David just shared this datawith me yesterday Almost 4.2
million slides scanned, almosthalf a million cases scanned.
(10:30):
Year-to-year volume growth, youcan see, continues to grow.
And what I love about this isthe user engagement and that's
the key right.
You want to create anenvironment where users are
excited to use the system.
This is a critical piece ofchange management right.
Change management is not easy,but it'll come with user
engagement.
(10:50):
If you have two or three we hada good discussion with my
colleagues from UAB yesterdayabout this and how do you start
this journey?
It's getting a few people superengaged and get this journey
started.
So the scan slides areinstantaneously available.
They're linked to our labinformation system and we are
continuing to add more bells andwhistles to this system.
(11:12):
Right now we move from acase-level integration.
Now we have slide-levelintegration so we can
individually call out the slides.
All those 4.2 million slidesare available.
We can consult with colleagueseasily.
This, to me, has been a gamechanger.
Our neuropathologists are in adifferent building.
Our dermatopathologists arefour or five miles away.
If I have a difficult penilebiopsy, I can just press a
(11:34):
button and connect with them inreal time.
And frozen sections andeverything else is much more
easier and streamlined becausewe have a way to share images.
We have implementedPatPresenter for our
consultation work and that's nowbeing integrated into into our
system, so scan slides areinstantaneously available.
There is no waiting forfoldering.
(11:55):
It doesn't matter which orderthey are scanned in.
They show up in your queueready to be signed out, and we
are starting to explore cytologypreparations FNAs using small
scanners on the bedside.
So, in summary, digital slideshave continued to improve my
workflow.
I love the direct interface tothe LIS.
(12:16):
I love sharing cases withconsultants and colleagues and
flagging cases, looking at theprior cases, so it has improved
my turnaround time.
Because I'm not.
Our histology lab is three orfive, three miles away from the
main campus, so it's allowed usto share those images more
easily.
All right, quiz now which cityis this?
(12:36):
No, close to Columbus, though.
It's Cleveland.
Very close to it.
So if you take 71, you haveCleveland, columbus and
Cincinnati.
So 71 corridor, right.
So what do we do next?
Right, as the systems mature,as you start your own journey of
buying your own scanners,implementing it as these systems
(12:56):
mature, how do we go beyondthat right?
How do we start thinking aboutother things which we cannot do
with glass slides as adoptionincreases.
So, if you look at a typicalproduct adoption curve, you have
innovators, you have earlyadopters, but many of them hit
this chasm right.
So every one of you have theirown chasms.
You might be starting yourjourney, you might be exploring
(13:20):
how to implement digitalpathology, but the key is
finding out what is your painpoint, what is your chasm?
How do I implement digitalpathology?
But the key is finding out whatis your pain point, what is
your chasm?
How do you go beyond this chasm?
And we are facing the samething, right, as we build, as we
think about implementing AI, wehave to deal with integration,
we have to deal withinteroperability.
All these are key things.
So how do you go beyond theglass slide chasm?
(13:40):
Right?
So we have established in thisroom we can now create
diagnostic quality images.
How do we go from there?
Right?
So what can we do with digitalimages that we cannot do with
glass slides?
So it really boils down tomanaging the information,
sharing the images, connectingwith each other, connecting with
experts, but also exploitingthe pixel pipeline to build
(14:05):
algorithms or buy algorithms toidentify, quantitate, synthesize
and create knowledge pathways,create important clinical
decision-making skills, right?
This is the next part of my talk.
I'm going to focus on what aresome institutes doing and where
are we today with this chasm?
Where are we today withcomputational pathology and AI
(14:28):
for clinical decision-making?
So these are some of the thingswhich I feel are already here
today, in 2024, right, we havevery sophisticated image
analysis algorithms, aialgorithms for detection and
diagnosis.
Right, to analyze digitalimages, decipher the pixels,
(14:49):
identify features likemorphology, cell shapes, size of
nuclei, architecture andstaining patterns.
So, pattern detection, featuredetection these are things I
learned as a resident, and Ilearned that by looking at over
and over again and building analgorithm in my brain which
allowed me to look at a prostategland and say, okay, this is
(15:10):
cancer, because A, b, c, d, e,same thing, right.
So we are at a point in 2024where many institutes have
implemented AI in pathology.
Right, and they started withvery basic things counting cells
, looking at the size of nuclei,differentiating different
nuclei, differentiating positiveversus negative signals.
(15:32):
So all these focus onbiomarkers which could be
diagnostic or predictive.
Right, so these could includedetecting, classifying,
segmenting, quantifying andlocalization.
Here are the four applicationsthat we are starting to use in
our Institute.
And again, we have our ownchasm.
Our chasm is that we cannotreadily integrate many of these
(15:54):
algorithms.
So we still have to go to athird party, launch the
algorithm and use the algorithmand bring this information back
to the LIS.
In an ideal world, in my wishlist, we want this to be
completely integrated and that'swhere we are heading towards.
But quantitative digital imageanalysis for biomarkers this is
very, very routinely done.
(16:15):
Now there is a separate CPTcode for this.
You can actually get paid alittle bit more, maybe, I don't
know twenty, thirty dollars morefor a digital image analysis
system.
If you use in your lab, it'smore objective, more accurate,
more faster.
But is it easy to do?
Is it cheap?
So the answer is it is easy todo.
(16:36):
It's easier to do if youalready have digital slides in
your system.
It may not be the firstapplication you launch in your
lab, right?
If you look at this gastricneuroendocrine tumor and you can
see a 2-millimeter square areawas analyzed in 28 seconds.
So if I asked each one of youin the room to count the blue
dots and I give you threeseconds, you all will have a
(16:59):
different answer.
Right?
Everybody agree with thathypothesis.
And if I showed you this, it'sgonna be even harder, right, but
we can get this dataobjectively and it's
reproducible every time you doit, if the answer will be the
same in Cleveland or in Columbus, as long as you're using the
same algorithm and you'vevalidated it.
(17:20):
It's a locked system and so on.
These are a given.
Other type of algorithms we areusing is detecting rare events.
This is a lymph node detectionalgorithm, right, so we can
actually launch it directly fromthe viewer today.
But we still have to do a lotof manual work to get this
algorithm queued up.
But pathologists still use it.
Pathologists still want to useit.
(17:40):
So imagine a world where thisis even easier to do.
It will become much more easierto use.
So identifying metastatic foci,and the pathologist in the room
might say why do you need analgorithm?
I can just eyeball it.
This is cancer.
What if you have a few rarecells and you missed it?
But the computer didn't miss it?
These are type of things I callthem rare event detection.
(18:02):
Overall, right, it could befinding microorganisms and so on
.
Let's go beyond that, right.
So today we can find metastaticcancer in lymph nodes.
But we want to go beyond thatright.
Pathologists today doimmunostains to figure out which
cancer this is Tumor of unknown, primary, unknown origin we do
several immunostains.
I had a case last week.
(18:23):
I had to do 20 immunostains andconsult a metopathologist and
soft tissue pathologist and eventhen you know guess how it was
signed out High-grade malignancy, see comment.
So this is an ongoing issue indiagnostic pathology and today
we have algorithms which canpredict for you, just like when
you send it for next-gensequencing.
(18:44):
It can predict with greaterthan 95% certainty this is renal
cell carcinoma.
So this is an example of such aprediction model where the
computer has predicted this iscolorectal cancer.
We have GI pathologist, dr Chen, in the audience and she will
say why do you even need AI forthis?
I can look at this and say thisis colorectal cancer.
(19:04):
It has necrosis and dirtynecrosis and all the features of
colorectal cancer.
But the point is there arealgorithms out there and they
are available.
They will continue to evolveand get better and when you are
ready in your journey, you willbuy that algorithm and use it.
But before you do that, westill have to solve the
interoperability issues.
(19:25):
We still have to integrate themand some sites have, some
institutes have done better withthese, others have not.
So if you look at diagnosticsfor cancer overall, they can
help you in a pre-sign-outprocess.
They can help you duringsign-out or they can help you
post-sign-out.
In pre-sign-out setting itcould be a good screening tool.
(19:45):
I showed you the data fromhundreds of studies about
sensitivity, right?
So for prostate cancerspecifically, similar studies
have done and shown highsensitivity and high specificity
and it's one of the mostcommonly exploited cancer for
building algorithms.
Like every company I talk to,they're building their own
(20:06):
prostate cancer algorithm,pre-ordering IHC, right?
Imagine a world where you'recoming to work and the system
automatically screens the casesand chooses the ones that you
should do immunos on, anddoesn't do it automatically but
makes recommendation to youDuring sign out.
It can find other features,like intraductal cancer of the
prostate.
(20:26):
It can help you withsupervision, primary diagnosis
or without supervision.
It can also create automatedreporting templates for you and
then do a second review, likewhat about post sign out, right?
So imagine a world where youare overworked pathologist,
you've just booked a ticket togo to Las Vegas and it's six
o'clock and you're trying torush through your cases and you
(20:48):
make a mistake.
But you have the safety net,you have a gatekeeper, you have
the AI assistant.
They check your work and saywait a minute before you board
the flight.
Are you sure you want to callthis cancer?
And you'd look at it again andagain and maybe you change your
diagnosis.
So some of the studies outthere have looked at this
specifically and there are caseswhere small foci or cancer were
(21:10):
missed.
Did it make a difference?
Maybe not.
Maybe there was cancer in othercores, but what if it was the
only core and you missed it andyou wish you had the system, or
not?
So how many of you if this wasfree right, all the vendors will
give this to you for free howmany of you will use it?
And what if it wasn't free?
You'll still use it.
(21:33):
So again, we are also, as webuild and use these algorithms,
we're learning new informationabout these cancers.
So, again, these tools arebecoming, I would say, not in
routine use, but they're gettingvery close to routine use.
So you can see this work listas you are looking at your work
list for the day and the caseshave been flagged for you.
(21:56):
So I would say we're very closeto getting a routine use of
these algorithms and, again, youcan turn it on and off.
If you're in a residencyprogram, you can have the
resident use it or not.
You can use it as a way toassess their competency, right?
You can make them review thecase and then turn on the AI or
(22:16):
make it so.
Ai will only be turned on afterfive minutes of review.
How many residents in the roomYou're ready to use AI, right?
So this is an example of breastcancer, right, invasive lobular
carcinoma.
And again, the red areas arewhere the cancer is.
You can go look at it moreclosely.
You can also have algorithmswhich can distinguish the
(22:37):
subtypes of breast cancer orsubtypes of prostate cancer.
What about finding mitosis?
Who likes to count mitosis?
Residents, you like to count,you know?
I asked, I did this test.
I asked three residents tocount mitosis on the same slide
and I got three differentanswers and I would say
pathologists would also give thesame different answers.
Right?
(22:57):
But you don't have to, right,today, if you have AI, you've
implemented it.
They will find the mitosis foryou and let you check it or not.
So again, this is an example oflymphoma classification, or
finding abnormal white bloodcells in a smear.
These systems are already inplace, like many of these
systems are being used in theclinical lab.
(23:18):
What about finding acid-fastbacilli?
Who likes to do that?
These are algorithms which areavailable.
Some of the centers haveimplemented it.
What about finding H pylori?
So these systems are beingbuilt.
They're going to becomecommercially available.
There will come a time where youwill have your own personalized
(23:40):
dashboard of apps that you candownload from the App Store, and
you'll have to pay asubscription to it.
Some of them will be free, someof them will be per click, but
these are things which we didn'thave when we were talking about
whole slide imaging.
So let's go beyond that.
What else is coming?
So we were talking about wholeslide imaging, so let's go
beyond that.
What else is coming?
Right?
So I talked about the wholeslide imaging.
Is that the end of the journey?
(24:00):
Is that the end of a glassslide?
What about starting from thetissue?
Right?
So we're getting closer toradiology, and radiology is
getting closer to pathology, soit's a spectrum.
Right?
So you have synthetic data, wehave virtual staining, we have
3D pathology, integrated largelanguage models.
We have all these new toolswhich will become available in
(24:23):
your work list soon, from pixelsto diagnosis, to prognosis and
predictions.
Now we talked about thesegmentation.
But imagine a world where Iwant to go inside this segmented
area and pick one or two cellsand interrogate those cells.
What if I want to look at thisimage in three dimensions, right
(24:44):
?
What about just finding similarthings, right?
Today you can take a picture onyour phone and you're going to
the supermarket and you take apicture and find similar
products.
It gives you the prices.
What if you could do this inpathology?
When I was a resident, how did Ilearn?
I learned by opening textbooks.
(25:04):
I was in the resident room andit was 7.55.
The attending was going to comethere at 8 o'clock.
I didn't know what this was.
I wanted to scribble something,so I flipped through pages and
I put something down.
At least I tried.
But today this could be doneelectronically.
It could be done usingartificial intelligence, where,
in this case, the query imagewas a brain ependymoma and the
(25:25):
algorithm found the best matcheswhich were similar to this and
created.
At least it gave you anindication of what you're
dealing with.
And you're going to seeexamples of these different
things, like virtual staining.
So if you walk around theexhibit area, there are vendors
out there which have worked onit, right?
So this, to me, will be goingbeyond the glass slide.
(25:48):
So what about prostate biopsieswhich are unstained?
On the top panel A, you haveunstained prostate biopsies.
In panel B you have H&Es whichwere generated in the lab.
And in panel C you haveprostate biopsies which were
generated by virtual staining,and then you can actually see
(26:09):
the areas where the cancer was.
So what about staining, right?
See the areas where the cancerwas.
So what about staining, right?
So, as a prostate pathologist,we do prostate triple stain to
look for basal cells.
So here you see, on theright-hand side you have real
HNE and on the left-hand sideyou have virtual HNE.
But what is amazing to me isthe ability to now predict where
(26:30):
the stain might be and evenpredict intraductal carcinoma,
which is a difficult diagnosis,in my opinion, you know, and
there is a lot of controversyaround it.
So 3D pathology is coming.
We have commercially availablescanning systems now which can
scan in 3D, right?
So here you have a cancerprostate core and a benign
prostate core.
(26:50):
So what about helping you whileyou're signing out, right?
So these are slides from DrSingh, who's built this in Path
Presenter the ability to createa chat with the image.
So imagine a world where you'relooking at this image.
You start chatting with thechatbot what is this?
And help.
These large language models canhelp proofread your report,
(27:12):
alert pathologists if you hadmissing information.
You know this happens whenyou're in a hurry.
You're rushing through a case.
You might miss information, youmight put something like a T4
where it was actually a T2.
What if you had this way toproofread your reports and help
with billing right?
So imagine a world where youhave this image where you can
(27:33):
press a button and look at thefeatures for the residents and
the trainees.
It can highlight the featuresin this image, but it can also
suggest a diagnosis for you andsuggest some IHCs for you and
also try to help you findsimilar cases, create a
differential diagnosis, takethese images and put them into
your tumor board pile andpresent it at the tumor board.
(27:55):
So these type of tools arecoming in many areas in
pathology and again, we're notthere yet today.
But what I'm showing you todayare prototypes of things that
are coming in the pipeline.
So we've gone beyond the glassslide.
We have now demonstrated thatwe can make a good diagnosis on
digital images.
(28:15):
We have demonstrated that thereare algorithms today which can
be used clinically, and I'veshown you examples of workflows
that are coming soon.
So I just want to also separateartificial intelligence and real
intelligence, right?
So when you think aboutartificial intelligence, it's
very task-oriented.
Right?
So when you think aboutartificial intelligence, it's
very task-oriented.
Find me this feature how manynuclei are positive?
(28:40):
Biomarker quantification, lymphnode met?
I showed you all these examples, but they're very task-oriented
.
Each algorithm is composed ofsmall steps, but real
intelligence is goal-orientedMany algorithms per task.
So, as a pathologist, if youlook at your journey from
residency, from medical school,to where you are as an
experienced pathologist, it'sexperience-based, it's
context-based, it'sknowledge-based.
(29:01):
So we're not there today wherewe can say AI is equal to real
intelligence, right?
So that's a journey, just likedigital pathology is a journey.
So this is an example of a caseof prostate cancer where, on one
, one of the cores, I haveclearly established prostate
cancer, right here and then thisis which looks like prostate
(29:22):
cancer, and I actually signed itout as prostate cancer, and
that's where real intelligencecomes in.
So I signed out this case, Ireleased it into the patient's
medical record, but somethingbothered me when I reached home.
You know, that night.
I just had a thought that this,maybe this is rectal cancer.
Maybe we should think about it.
So I dug through the notes.
I found a note from one of theprimary care visits where this
(29:47):
patient has rectal bleeding butrefused endoscopy because it was
too expensive.
So I went in, I ordered somestains and it turned out to be
colon cancer.
So I had to admit I was wrong.
I amended my report, I calledthe urologist and the patient
actually got treated for rectalcancer and prostate cancer.
So I think that's where we needto be and we might get there in
(30:09):
20, 30 years, but we havesignificant advances in the
field even today.
So, putting it all together,everybody went to Magic Kingdom.
I was there last night Amazingfireworks, the best fireworks.
I mean, I went to Disney withmy kids many years ago, but I
went now Amazing, so I recommendit highly.
I don't have any stocks inDisney or anything.
(30:29):
So, in conclusions, how willdigital pathology and AI help
pathologists?
Right?
So we talked about the declinein number of pathologists,
increasing workloads, fewertrainees going into pathology,
and clearly what I've shown youtoday can help some of those
issues.
Help us with sharing cases,help us with connecting
(30:50):
subspecialists together,connecting spaces together.
We have increasing workloads,so are we ready for a digital
disruption?
And then augmenting yourdiagnosis, checking your work,
do some of your work and sharethe work with others right, so
you have to tame your owndigital pathology chasm, right.
Each one of you is here in thisconference because you want to
(31:12):
learn more about it.
If you've already bought thescanner, you want to learn about
AI.
If you've already implementedAI, you want to learn about 3D
pathology.
But it's a journey, right, soall of you have to solve it on
your own.
I'm not going to you know.
I'm just showing you where wewere 20 years ago, where are we
today and where are we going 20years from now.
So maybe you will have aworkstation like this in your
(31:36):
office where you will customizeit.
You'll buy your own apps.
You will create your own chasmbuilding, right, so you will use
AI to assist you to augmentyour diagnosis.
And maybe autonomous drive,right.
Yesterday, my friend drove mefrom the fireworks back and the
car was driving itself and heleft the steering wheel and I
(31:56):
was super scared and it waspretty crowded.
It was Sunday night, it wasrush hour.
I said, no, I'm not ready forthis.
So take over the steering,please.
I have a talk to give at 7.30.
So he didn't listen to me andthe car drove itself for 10
minutes and it did fine, right.
But you have to build thattrust.
You have to build that trustwith AI, right.
(32:18):
So where are we going?
Right?
The journey from glass todigital to prediction will
continue for all of us.
We are today.
We can improve our analysis.
What will come is next, we'llbe improving your diagnosis.
Maybe we will have moreintegration.
We will solve theinteroperability issues,
regulatory issues, reimbursementissues, but in the future, this
(32:39):
will really be clinicaldecision-making exercise and
integrating multiple types ofdata, and we're probably this
will be the era of AI-basedprecision medicine in its true
sense, and that's probably inpathology.
Ai could probably diagnose easycases independently, just like
the Tesla was drivingautonomously yesterday, but I
(33:02):
don't know how many pathologistsfeel comfortable about that
riding the Tesla of pathology.
So with this I'm going toconclude.
It's been an amazing journeyand I think we're going to.
I actually encourage you to talkamongst yourselves.
Go to the exhibit hall, look atall these different products
out there, and there issomething for everyone.
You know you might be justbuying your first baby scanner.
(33:25):
You might be just getting thatone big check from your
administration and you're readyto spend it.
This is the place to do it.
This is the Disney world ofdigital pathology.
So with this, I'm going to endwith a picture of football again
, which I'm excited.
Next weekend I'm going to be atthe game, and so I'm going to
stop and if you have anyquestions, I'll be glad to
answer it.
(33:45):
I'm going to be around for therest of the conference, so I
hope to interact with many ofyou.
You and I want to thankApreedia for inviting me here as
a speaker, and looks like afull house.
I'm sure it's because of thebreakfast.
That was really good.
Speaker 4 (33:59):
Thank you all thank you so much, dr Parwani, for
your lecture and I think it wasamazing because it gave this
like from the very beginning,when scanning was a problem, to
now like doing 3D pathology andactually everybody who is in
this room and who comes to thisconference, they can be at any
like single point of what youdescribed.
(34:20):
So let's start with thebeginning, and you were talking
about the chasm.
So everybody has their ownproblem, their own pain point to
solve.
But let's start with thescanners.
What scanners do you?
Speaker 3 (34:30):
have so.
So we have variety of scanners.
Uh, we have philips, we haveaperio uh laika, we have
hamamatsu and we just got thepredia.
Which one did you get the 250the fluorescent.
Yeah, yeah, so we're using itfor our kidney biopsies now is
that the one that has thepolarization?
Option as well.
Speaker 4 (34:51):
I remember in our podcast you were saying why they
are not adopting the renalpathologist, and now there is a
tool for them.
Speaker 3 (34:58):
Yes, I think that to me is an exciting part, where
you can actually see some of theusers who are turned off by
digital if they cannot do A, band C.
Today we have the capability ofproviding those options with
different types of scanners.
So I think one scanner may notaddress all the needs of a big
academic center but a smallerlab can actually do with one or
(35:18):
two scanner types.
But there are so manycomplexities in pathology and
I'm glad many of the vendors outthere, including Apreedia,
leica, are actually solvingthose problems individually or
incorporating them into yourdashboard, into their system.
You know, just like when you goto Best Buy to buy the next TV,
you have all these bells andwhistles, but they specifically
(35:39):
build those based on feedback.
You know, today I cannot do this.
Can you build this?
So I think digital pathologyscanner market is also evolving
in that direction.
To me, it's like multipletrains have left the station and
everyone has their own trainand they have their own next
stop.
The next stop might be I wantto do frozen.
The next stop might be I wantto use image analysis.
(36:00):
I want to do large languagemodels.
I want to go directly from thetissue to an image, you know.
so, all these trains, individualtrains, are users and they have
expectations and they have endand they have an endpoint.
Not an endpoint but a stop onthe way.
And they want to get off thetrain at that.
Stop, do something and then geton the train again.
Speaker 4 (36:19):
Figure out what the right course is.
Speaker 3 (36:20):
So I always think it's fascinating to me to come
to a conference like this, whereyou have different types of
users, not just pathologists,but also technicians, also
students, technologists, itvendors lab managers and they
work together to solve complexproblems.
Well, this is what I love aboutthis.
Speaker 4 (36:39):
I like it about digital pathology because
usually in medicine you don'thave a team with so many
different expertises.
You mostly are with medicalprofessionals and it's driven by
the medical professionals, andhere you have technology
specialists, you have operations, you have, obviously
pathologists, you have theadministrators and I love it.
(37:03):
It's super interactive.
Speaker 3 (37:04):
Yeah, no, I think you're exactly right.
You have the administrators whowrite the checks, you have the
end users, you have trainees whoare going to be the future of
pathology.
So they're all coming togetherfor finding their own journey,
finding their own discovery.
You know why digital pathology?
Why now?
Why me?
And I think that is animportant thing, which is
(37:26):
important for your listenersalso to know.
Speaker 4 (37:28):
And we just heard from one of the people who was
asking questions like OK, now wegot the green light, how do we
bring everybody on board?
Correct, Because I think peopleare.
Well, it's with any changemanagement you're focused first
on getting the green light andgoing somewhere, and then it
turns out you have to bringeverybody else with you.
How do you do that?
(37:48):
That's the next step, step nextstop on this journey.
Speaker 3 (37:52):
Yeah, I think, I think you have to take baby
steps.
You have to again as avisionary.
You have to look at the visionof where you want to reach right
and that's where you're here,so, but then the vision could be
.
I want to do these steps andthis will require these changes.
A good leader is also a goodchange manager you know if
you're not, you're not going tobe a good leader if you cannot
(38:13):
manage change.
The change can occur at peoplelevel.
The change could be at thetechnology level, it could be
financial, it could beregulatory, but collectively you
have to orchestrate all thosechanges to see and execute your
vision.
You have to hire the rightpeople, the right team and keep
them engaged, keep them excitedin this journey so what is your
(38:35):
chasm right now?
Speaker 4 (38:36):
what are you guys working?
Speaker 3 (38:37):
yeah.
So our chasm is, like Imentioned.
There are different taintpreparations, light preparations
, gosh frozen all those.
How do we bring all thistogether in one seamless way?
So the chasm that we're tryingto solve is integration with all
the AI tools.
Do we need to launch multipleviewers?
Do we need to createtechnologies which are not
(39:02):
compatible and try to bring themtogether?
So that is the chasm we'retrying to do.
You know we can creatediagnostic quality images.
We know we can use them.
I know we can make diagnosis onthem, but I think the next gap
is can I create a dashboardwhere I only have three AI tools
but they could be launched fromone viewer, and so on?
Speaker 4 (39:21):
For the ease of use for the pathologist, because
that's also a step in adoptionjourney.
When it's not seamless, peoplewere not already convinced and
not willing to troubleshoot.
That's gonna be a showstopperfor them.
Speaker 3 (39:33):
Yes.
Speaker 4 (39:34):
I love your stories.
Thank you so much for thisfantastic presentation.
Popular Podcasts
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.
On Purpose with Jay Shetty
I’m Jay Shetty host of On Purpose the worlds #1 Mental Health podcast and I’m so grateful you found us. I started this podcast 5 years ago to invite you into conversations and workshops that are designed to help make you happier, healthier and more healed. I believe that when you (yes you) feel seen, heard and understood you’re able to deal with relationship struggles, work challenges and life’s ups and downs with more ease and grace. I interview experts, celebrities, thought leaders and athletes so that we can grow our mindset, build better habits and uncover a side of them we’ve never seen before. New episodes every Monday and Friday. Your support means the world to me and I don’t take it for granted — click the follow button and leave a review to help us spread the love with On Purpose. I can’t wait for you to listen to your first or 500th episode!
Dateline NBC
Current and classic episodes, featuring compelling true-crime mysteries, powerful documentaries and in-depth investigations. Follow now to get the latest episodes of Dateline NBC completely free, or subscribe to Dateline Premium for ad-free listening and exclusive bonus content: DatelinePremium.com