105: Artificial intelligence in pathology Part 1 - the presentation I gave on the day I got engaged.
October 4, 2024 • 26 mins
In this episode, I celebrate another milestone of the Digital Pathology Place YouTube channel that was achieved thanks to you, my digital pathology trailblazer, reflecting on its journey since its inception in 2019.
I delve into the developments in digital pathology, focusing on the first video I ever published on YouTube about AI in pathology, highlighting trends, tools, and challenges in the field.
The video was based on a presentation I gave on the day I got engaged, so if you want to know the whole story listen in.
I explain key concepts like
- artificial intelligence,
- machine learning, and
- deep learning, and discuss
- How could AI eventually support pathology practice despite current challenges?
00:00 Welcome and AI Co-Host Feedback
00:19 YouTube Monetization Milestone
01:18 Reflecting on the First Video
02:47 Special Day and Personal Story
05:06 Introduction to AI in Pathology
07:26 AI Terminology and Concepts
13:17 Current Status of AI in Pathology
17:33 Challenges and Future of AI in Pathology
22:42 Conclusion and Call to Action
23:30 Updates and Future Plans
THIS EPISODE'S RESOURCES
- The YouTube version of "AI in Pathology" first video
- The updated "Artificial Intelligence in Pathology" video (coming soon in podcast version)
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.
(00:00):
Welcome my digital pathology,trailblazers.
How did you like the AI?
Co-hosts let me know on LinkedInor.
Or via email.
And if you haven't heard the AIcohost it's in the previous
episode, episode number 1 0 4.
At the end of the episode.
But today.
The digital pathology placeachieved an interesting
(00:22):
milestone on YouTube.
We have monetized our channel.
What does that mean?
That means that now when playads on the digital pathology
place channel.
Then they will pay us money.
Which is going to be probablymarginal But nevertheless, it is
something that happened becauseof you.
Because you, the digitalpathology trailblazer are
(00:45):
watching the YouTube videos.
So thank you so much.
And if you haven't checked theYouTube channel yet, please go
ahead, check it out andsubscribe because then you can
get even more value.
From the digital pathologyinformation that I'm trying to
spread out on the internet.
So if you have not checked thedigital pathology place channel
(01:07):
on YouTube yet, please.
Do it and subscribe.
Because then you can get evenmore as I was checking all the
settings for the channel.
I came across the very firstvideo that I recorded on the
channel.
It's about artificialintelligence in pathology.
And.
(01:28):
I looked at it.
It was.
In 2019.
It was a presentation I gave atthe conference.
Actually, my husband was invitedto.
We're actually, my husband wasinvited to present.
But I thought, oh, would theynot be interested that my
husband, isn't a clinicalpathologist and I'm trying to
(01:49):
invite him to the podcast.
To share a little bit of hisexpertise about I didn't manage
to do that yet.
So stay tuned anyways.
So he went to this conferenceand he was asked to present on
behalf of the company who waswith.
And I thought, oh, would theynot be interested in.
It was for, it was a pathologyconference.
In Hershey, Pennsylvania.
(02:10):
And.
He asked the organizers.
And they said, yes.
Sure.
So when I was preparing thepresentation, I was really
practicing.
I also recorded a YouTube videoand you can have a look at this
with.
And you're going to have a lookat this year, the video.
And see what the beginnings are.
When I look at my videos now andthe confidence.
(02:33):
And the expertise I gained.
I appreciate that there was alearning curve.
Nevertheless this This contentis still valuable.
And I found that I thought I'mgoing to share.
Glimpse from the past.
Actually this happened on apretty special day.
Because I gave.
No again.
(02:54):
So I was recording thispresentation.
The sound is not great.
Again, I was recording thispresentation at the.
Hallway of our apartment.
My son was.
Less than a year.
All Dan.
So I had to leave the apartment,go through the hallway.
I sat in front of a window and Iwas recording.
(03:14):
When you go through YouTube, thesound is not fantastic.
I of course.
Cleaned it up for you here.
So I recorded before theconference, put it on YouTube
and I am greeting all my zeroyouTube subscribers.
Now we have over 3000.
And then I think a week later Ipresent.
I was pretty stressed.
(03:35):
And that was on a very specialday.
Because.
It was the day I got engaged.
After the presentation.
We had our son there and when myhusband was presenting, I was
taking care of the son.
Then I was presenting, he wastaking care of my son.
And then we went for a hike.
We changed the from our fancyconference clothes and we went
(03:58):
for a hike.
And I was all relaxed.
All fun.
All good.
And then my husband startstaking out stuff and I'm like,
oh, that's so cool.
We're having a picnic.
I'm carrying my baby in the babycarrier.
Just hiking ahead.
Super happy that they can relaxafter this presentation that
(04:18):
caused me a little bit ofstress.
And he takes out the blanket.
And he takes out some food.
And then he takes out flowersand I'm like, are you.
Flowers.
In your backpack and Until thelast moment until the moment.
He actually kneeled on thisblanket, on the Appalachian
(04:39):
trail.
And.
Ask me.
If I'm going to marry him, gaveme the ring and asked me if I'm
going to marry him.
I like had no clue that this isthe day that I get engaged.
So this video will always remindme of this very special day.
And how digital pathology playsbegan.
So enjoy.
(04:59):
Dr Alex.
From.
Over five years ago.
I let me know what you think.
Learn about the newest digitalpathology trends in science and
industry.
Meet the most interesting peoplein the niche, and gain insights
relevant to your own projects.
Here is where pathology meetscomputer science.
(05:22):
You are listening to the DigitalPathology Podcast with your
host, Dr.
Aleksandr Zhurav.
Aleksandra (05:32):
Hi, my name is Aleksandra Żura.
I'm a veterinary pathologist andI'm passionate about digital
pathology.
I had the chance to speak at theannual meeting of the
Association of ClinicalScientists in Hershey PA and I
wanted to share the presentationI held there with all my YouTube
subscribers, which is currentlyno subscribers, zero people
(05:54):
because that's my first video,but with my future subscribers.
Me tell you about artificialintelligence in pathology
practice and how advanced wereally are.
We know that there is a shortageof pathologists in the world and
AI would possibly save time ofpathologists and they would be
able to help more patients.
(06:16):
So where do we stay?
I'm gonna introduce you to thesubject.
We're gonna talk about pathologyinformatics terminology.
What is the current status of AIin pathology?
What are challenges inbiomedical image analysis?
What is the chameleon challengein that?
And what's a challenge design?
Chameleon challenges is the mostfamous one in pathology.
(06:39):
What tools are available, whichare not?
And what are the hurdles forapplications in digital
pathology?
So let me start with someintroduction.
We are bombarded with artificialintelligence and digitalization
of pathology on pathologyconferences.
(06:59):
The accuracy of the A.
I.
Models often comes very closeand sometimes exceeds interim
intra pathologist variabilityand the terms such as machine
learning, deep learning neuralnetworks.
They increasingly are being usedin pathology context.
So the question is, are weapproaching a new era where I
(07:21):
will dominate pathology or is itstill a long way off?
Let me start with theterminology to understand it
better.
What is artificial intelligence?
It is a branch of computerscience about systems which can
learn from data.
So we encounter AI every time wetouch any of our mobile devices.
(07:42):
And examples here would beGoogle predictive searches.
When we start typing somethingand to look it up in Google, we
get a suggestion.
And usually these suggestionsare correct.
And there, this is a AI productrecommendations on Amazon.
This is also artificialintelligence.
(08:02):
When we buy something and we getsuggestions of similar products,
or for example, we get thismessage customers who bought
this also bought that.
And it suggests us differentproducts, music recommendations
by Spotify, based on the songsthat we already liked.
And Google Maps fastest routethat is based on the current
(08:25):
available traffic informationwhen we are using it as a
navigation system already.
And based on the current trafficinformation, Google suggests a
faster route.
What is machine learning?
This is a branch of AI that isbased on the idea that computers
can learn from data as humanslearn from experience.
And Then they can make decisionsabout data without human
(08:48):
interactions.
We have three types of machinelearning supervised learning
based on labeled data,unsupervised learning based on
grouping unlabeled data pointswith similar features together,
and reinforcement learning,where the user of the algorithm
or the model gives the modelfeedback on what was labeled
(09:08):
incorrectly.
What is random forest?
Random forest is one of the mostpopular supervised machine
learning algorithms.
It's a forest of decision trees.
These do not tend to perform sofantastic in pathology.
What is deep learning?
This is a subfield of machinelearning in which the models
(09:29):
resemble neural networks of thehuman brain.
And those models within deeplearning are called artificial
neural networks.
And an example or a type ofartificial neural networks are
convolutional neural networks.
And these have dominated thedeep learning space.
Computer vision is a field ofcomputer science trying to mimic
(09:50):
human vision.
And patching is dividing largepathology images into small
patches.
And to better picture this, weneed to know how big such an
image is.
And so a whole slide image is 15gigabytes, more or less.
What is 15 gigabytes?
It's three, three hour long,high quality Netflix movies.
(10:15):
Just one slide, one digitalizedpathology slide.
They need to be divided insmaller squares and these
squares are called patches andthe process is called patching.
Graphical processing units arenecessary for quick processing
of pathology images.
These are chips on thecomputer's graphic cards
designed to rapidly processgraphics.
(10:37):
They're used in gaming, videogaming computers and are
indispensable for fastprocessing of whole flight
images.
Computer aided diagnosis is whenclinicians use computer defined
regions to assist them in makingthe diagnosis.
This is already in routine usein radiology.
For the detection of breastcancer foci in mammograms and
(11:00):
the potential use in pathologywould be pre screening of lymph
node sections in cancermetastasis.
This is a tedious task, takes alot of time, there are many
sections of the lymph nodes andpathologists have to screen them
under the microscope.
It would significantly, savetime if they could have them
suggested by the computer andthen just say yes, no.
(11:20):
What is data augmentation?
This is a way of getting moredata when we don't have enough
data.
And on this example, on the leftside of the screen, we have four
mitotic figures.
They have been slightly alteredshifted a little bit, and we get
this.
From those four mitotic figures,we get 16 mitotic figures.
(11:43):
We increased our data set fromfour points to 16 data points.
And these are seen as by thecomputer as separate data
points.
So we don't have enough data.
We can augment them to have morefor the training of the system.
What are probability heat maps?
This is a color coded way ofvisualizing the classification
(12:05):
results of the deep learningmodel.
And on one end of the scale iscorresponding to a high 100
percent probability of a featurein this case of the tissue being
a tumor and the other end of thescale, the blue one in this case
is corresponding to zeroprobability.
So we can appreciate that thoseall red regions, is.
(12:26):
Probably tumor and regions inbetween, and there is lower
probability of tumor.
To visualize how those termsrelate to each other, the
broadest term is artificialintelligence, and the machine
learning is a part of artificialintelligence, where we have the
patching, random forest, anddeep learning is part of machine
(12:49):
learning.
We have Artificial neuralnetworks, convolutional neural
networks, as an example of them,we have probability heat maps
and data augmentation.
Computer vision is thediscipline that uses all those
models.
GPUs, graphical processing unitsenable us to process whole slide
(13:09):
images fast.
And in the end, the algorithmcan help us with making the
diagnosis, and we have computeraided diagnosis.
On conferences and inpublications on LinkedIn as
well, we are surrounded with allthe AI news, but what is the
current status of artificialintelligence in pathology?
(13:30):
As of today, May 18, 2019, thereis no FDA approved AI solution
for pathology on the market.
There are initiatives supportingand accelerating the development
of AI applications, like the FDADigital Health Innovation Action
Plan, Digital Health SoftwarePre Certification Program, Pre
Cert, and in differentdisciplines, AI already.
(13:54):
Made it officially with an FDAclearance or approval like
diabetes research incardiovascular and brain disease
treatment in radiology, but notin pathology yet and in
pathology publications about A.
I.
All those publications come fromR.
N.
D.
Departments, either academia orindustry, different companies,
(14:17):
but it's research.
So to foster AI and to encouragepeople to develop a solutions
and challenges in biomedicalimage analysis have been
established.
And what are these?
These are competitions aiming tocompare new and existing
algorithms in biomedical imageanalysis.
(14:39):
Thank you.
And researchers are verystrongly encouraged to take part
in those challenges in order topromote and contribute to AI
development.
The participants try to solve astated problem on a common data
set, and they can use a solutionof choice and are required to
(14:59):
publish their results.
So the most famous challenge inpathology was chameleon
challenge organized in 2016 and2017.
And this and other challengesare gathered on the comic
platform standing for consortiumof open medical image computing.
(15:20):
And they are called grandchallenges.
They are the most famous inbiomedical image analysis, but
there are also other platformsthat are hosting them.
And these include Coda, labCoval, or Virtual Skeleton.
And let me tell you about thecommunal challenge.
What it was about it, as I said,was one of the most famous
(15:41):
challenges in pathology.
Maybe not in biomedical imageanalysis or different ones that
I'm not aware of, but inpathology.
This one was the most famousone.
It was hosted by in theNetherlands by, the diagnostic
image analysis group anddepartment of pathology offered
the Radboud University MedicalCenter in Nijmegen.
(16:01):
And as I said, there were twoadditions, Chameleon 16 and
Chameleon 17 in 2016 and 2017.
And almost 100 submissions forboth of these challenges were
done both from academia and fromcompanies.
And the problem here wasdetection of breast cancer,
metastasis, and whole slideimages of lymph nodes.
The participants were workingwith the same training and test
(16:25):
set of 1, 399 hematoxylin andeosin stained lymph node
sections.
And they were free to use anyimage analysis method to best
solve this problem, so theydidn't have to use deep
learning.
But it just so happened that thewinners used deep learning.
So this performed the best.
This is how the challenge isdesigned.
(16:47):
Not only the pathology ones, butthe other ones as well.
A meaningful task needs to bedefined.
In this case, that was breastcancer metastasis detection in
lymph nodes.
Representative data needs to begathered.
In this case, it was labeledwhole slide images.
A reference standard needs to bedefined.
That was the diagnosis of thepathologist and their
(17:08):
annotations.
and a discriminative evaluationmetric needs to be determined.
And in this case, the consensusof the algorithms with the
pathologist input was measuredby Cohen's kappa and the
participants are required towrite a peer reviewed pathway.
(17:29):
So what are the tools that areavailable and what are not
available?
There's no FDA approved AAsolutions for pathology, like I
said, but Why is there not?
We will see later.
But for research use only, imageanalysis software companies are
starting to incorporate AImodules in their solutions.
where a pathologist or a usercan annotate target structures
(17:52):
to train the model and later themodel detects those structures
automatically.
For example, the pathologist orthe user annotates glomeruli on
one slide or in one part of theslide and then the algorithm
detects the rest of theglomeruli on the other part of
the slide or on the rest of theslide.
These solutions are currentlyavailable from different
(18:15):
companies like Indica Labs,Viziopharm, and Aphoria.
And these tools incorporaterandom forest or deep learning
modules for image analysis, butas I said, they're for research
use only.
Nothing for clinical use so far.
And why not?
Why is there something AI basedin other disciplines and not in
(18:38):
pathology yet?
So there are some hurdles thatare pathology specific, and
these include lack of validatedtools, as we already said, lack
of labeled data.
We said that the machinelearning need supervised machine
learning needs a large amount oflabeled data.
(18:59):
This would mean that thepathologists instead of looking
at sites would have to go andstart annotating whatever
structures they want to havedetected.
And it's not going to happenbecause they are supposed to
diagnose patients and they useslides for that, the non digital
slides.
So lack of labeled data, highcomplexity of histological
images.
(19:19):
In comparison to radiology orcardiology, the complexity of
histological images is much,much higher.
We have different colors,textures, tissues, organs, and
deaths multiplied by each othercontributes to this high
complexity of histologicalimages.
(19:40):
The dimensionality of pathologydiagnostic problems is also
high.
And the example that we had thedetection of neoplastic foci in
a lymph node section is just onebinary problem that was used for
the competition.
But the dimensionality ofPrimary diagnosis in pathology
(20:02):
is a lot higher.
There is a different patternrecognition involved and
incorporation of many differentprocesses visible on the slide
and different data availableabout the patient.
Another problem or a hurdle isthat pathologists are the gold
standard or they are generatingthe ground truth.
(20:22):
So the algorithm can only be asgood as a pathologist and also
pathologists tend to differ.
Algorithm is gonna be close towhat one pathologist says, does
it mean that it's a good one ifanother pathologist says
something else?
So that's a hurdle.
(20:42):
And the affordability ofcomputational power and storage
space is a problem.
Those GPUs that are required forprocessing these huge slides
cost a lot more than a normalPC, normal computer.
That's a hurdle.
And we need to store somewherethose digital slides, enormous
amounts of data.
(21:03):
So that's another hurdle.
So to sum up, AI is affectingour lives every time we touch a
mobile device, phone, whatever,computer.
It has officially enteredvarious medical fields, but in
pathology it still remainsrestricted to researchers only.
There are many pathologyspecific hurdles to overcome,
(21:23):
and the most important beingenormous size of pathology
images and complexity ofmorphological patterns.
The regulatory pathways havebeen already paved by other
disciplines, which is good.
And the FDA officially supportsthe digital health initiative
with their digital healthinnovation action plan and
digital health software precertification program.
(21:45):
So to answer the question thatwe ask at the beginning.
Is it gonna happen?
Yes, it's gonna happen soon.
A.
I.
Is coming soon.
It's we're not quite there yet,but it's coming soon.
So pathologist should prepare.
They can prepare by goingdigital in their labs by
recognizing where A.
I.
Will add value.
So it may not add value to everystep of the pathology workflow.
(22:12):
It's crucial to recognize whereit would and apply it there.
And they can prepare byeducating themselves.
There are courses available, onecourse for example, from digital
pathology association, an onlinecourse that pathologists can
take and they should proactivelyengage with a initiatives and
beat the challenges like thechameleon challenge or
(22:36):
collaborations, cooperationswith academic partners, with
industry with differentcompanies.
Thank you very much forlistening to me.
And I hope you liked it.
If you liked it, please clicksubscribe to do my channel.
And if you have any questions,don't hesitate to write an email
(22:57):
and go and check out my blog anddigital pathology consulting dot
com, where there is an articleabout this presentation as well.
And I will also.
Put this presentation fordownload if you would like the
slide.
Thank you.
Bye Here the references in caseyou want to go back to the
primary literature or and Followthe links for the things that I
(23:22):
mentioned.
Thank you so much for listeningto the end.
That tells me you are a realdigital pathology trailblazer.
This video, was posted fiveyears ago and it has 3000 views
on YouTube.
This is one of the most viewedvideos.
Of course it has been there forthe longest.
(23:42):
And there was a whole chunk ofwhat happened in the last five
years missing, like all thegenerative AI foundation models,
chat GPT.
And all that good stuff.
So of course there is an updatevideo that I posted five months
ago.
And this was has almost 2000views already.
So I'm going to link to both theolder version and the new
(24:05):
version if you feel likecomparing.
How I was presenting them versushow I'm presenting now.
In those over five years.
My YouTube channel is telling methat there are 510 videos.
It's a little bit.
Elevated, because at some pointI connected the digital
pathology podcast.
So all the audio podcasts are onYouTube as well.
(24:27):
When Google decided to mergeGoogle podcasts with YouTube, we
migrated to YouTube.
So it doubled the number ofvideos, but definitely over 370.
Native YouTube videos.
So I decided to create astructured curriculum out of
those videos, they were beingcreated more or less on demand
(24:48):
when I was hearing from you.
Oh, I would like to have a videoon that.
I would have liked to have apresentation or if I would give
a presentation, there was anoption to record.
You're going to find things likethat as well.
Recently, I started addingconference vlogs to give you a
glimpse of what's happening atthe conferences and a life live
commentary on what's happening.
(25:10):
So I decided to create astructured curriculum.
And create.
Course based on.
Selected and curated YouTubevideos.
I am still working on putting ittogether.
But if you're interested inchecking it out, In exchange for
feedback and testimonials, Ihave a waiting list, for 100
(25:33):
digital pathology trailblazerswho will get access to this
course for free.
So if you're interested in testdriving the course, giving me
your feedback, whatever youthink about it.
You just let me know.
And if you love it, if you wouldgive me a testimonial, then this
course is going to be for youfor free.
I already shared it with thedigital pathology trailblazers
(25:56):
from my mailing list.
And several of them signed up.
There are still a few spotsleft.
For you, the podcast listenerbut before the.
free.
access waiting list, Gets sharedon social media.
I decided to share it on thepodcast.
So for you who listens to theend.
(26:16):
This is an opportunity to getaccess to this course for free.
And let me know what you think.
Check out the link in the shownotes.
YouTube course Waiting list,sign up and whenever I have
ready, it's going to end up inyour inbox.
And I talk to you and the nextepisode.
Welcome my digital pathology,trailblazers.
How did you like the AI?
Co-hosts let me know on LinkedInor.
Or via email.
And if you haven't heard the AIcohost it's in the previous
episode, episode number 1 0 4.
At the end of the episode.
But today.
The digital pathology placeachieved an interesting
(00:22):
milestone on YouTube.
We have monetized our channel.
What does that mean?
That means that now when playads on the digital pathology
place channel.
Then they will pay us money.
Which is going to be probablymarginal But nevertheless, it is
something that happened becauseof you.
Because you, the digitalpathology trailblazer are
(00:45):
watching the YouTube videos.
So thank you so much.
And if you haven't checked theYouTube channel yet, please go
ahead, check it out andsubscribe because then you can
get even more value.
From the digital pathologyinformation that I'm trying to
spread out on the internet.
So if you have not checked thedigital pathology place channel
(01:07):
on YouTube yet, please.
Do it and subscribe.
Because then you can get evenmore as I was checking all the
settings for the channel.
I came across the very firstvideo that I recorded on the
channel.
It's about artificialintelligence in pathology.
And.
(01:28):
I looked at it.
It was.
In 2019.
It was a presentation I gave atthe conference.
Actually, my husband was invitedto.
We're actually, my husband wasinvited to present.
But I thought, oh, would theynot be interested that my
husband, isn't a clinicalpathologist and I'm trying to
(01:49):
invite him to the podcast.
To share a little bit of hisexpertise about I didn't manage
to do that yet.
So stay tuned anyways.
So he went to this conferenceand he was asked to present on
behalf of the company who waswith.
And I thought, oh, would theynot be interested in.
It was for, it was a pathologyconference.
In Hershey, Pennsylvania.
(02:10):
And.
He asked the organizers.
And they said, yes.
Sure.
So when I was preparing thepresentation, I was really
practicing.
I also recorded a YouTube videoand you can have a look at this
with.
And you're going to have a lookat this year, the video.
And see what the beginnings are.
When I look at my videos now andthe confidence.
(02:33):
And the expertise I gained.
I appreciate that there was alearning curve.
Nevertheless this This contentis still valuable.
And I found that I thought I'mgoing to share.
Glimpse from the past.
Actually this happened on apretty special day.
Because I gave.
No again.
(02:54):
So I was recording thispresentation.
The sound is not great.
Again, I was recording thispresentation at the.
Hallway of our apartment.
My son was.
Less than a year.
All Dan.
So I had to leave the apartment,go through the hallway.
I sat in front of a window and Iwas recording.
(03:14):
When you go through YouTube, thesound is not fantastic.
I of course.
Cleaned it up for you here.
So I recorded before theconference, put it on YouTube
and I am greeting all my zeroyouTube subscribers.
Now we have over 3000.
And then I think a week later Ipresent.
I was pretty stressed.
(03:35):
And that was on a very specialday.
Because.
It was the day I got engaged.
After the presentation.
We had our son there and when myhusband was presenting, I was
taking care of the son.
Then I was presenting, he wastaking care of my son.
And then we went for a hike.
We changed the from our fancyconference clothes and we went
(03:58):
for a hike.
And I was all relaxed.
All fun.
All good.
And then my husband startstaking out stuff and I'm like,
oh, that's so cool.
We're having a picnic.
I'm carrying my baby in the babycarrier.
Just hiking ahead.
Super happy that they can relaxafter this presentation that
(04:18):
caused me a little bit ofstress.
And he takes out the blanket.
And he takes out some food.
And then he takes out flowersand I'm like, are you.
Flowers.
In your backpack and Until thelast moment until the moment.
He actually kneeled on thisblanket, on the Appalachian
(04:39):
trail.
And.
Ask me.
If I'm going to marry him, gaveme the ring and asked me if I'm
going to marry him.
I like had no clue that this isthe day that I get engaged.
So this video will always remindme of this very special day.
And how digital pathology playsbegan.
So enjoy.
(04:59):
Dr Alex.
From.
Over five years ago.
I let me know what you think.
Learn about the newest digitalpathology trends in science and
industry.
Meet the most interesting peoplein the niche, and gain insights
relevant to your own projects.
Here is where pathology meetscomputer science.
(05:22):
You are listening to the DigitalPathology Podcast with your
host, Dr.
Aleksandr Zhurav.
Aleksandra (05:32):
Hi, my name is Aleksandra Żura.
I'm a veterinary pathologist andI'm passionate about digital
pathology.
I had the chance to speak at theannual meeting of the
Association of ClinicalScientists in Hershey PA and I
wanted to share the presentationI held there with all my YouTube
subscribers, which is currentlyno subscribers, zero people
(05:54):
because that's my first video,but with my future subscribers.
Me tell you about artificialintelligence in pathology
practice and how advanced wereally are.
We know that there is a shortageof pathologists in the world and
AI would possibly save time ofpathologists and they would be
able to help more patients.
(06:16):
So where do we stay?
I'm gonna introduce you to thesubject.
We're gonna talk about pathologyinformatics terminology.
What is the current status of AIin pathology?
What are challenges inbiomedical image analysis?
What is the chameleon challengein that?
And what's a challenge design?
Chameleon challenges is the mostfamous one in pathology.
(06:39):
What tools are available, whichare not?
And what are the hurdles forapplications in digital
pathology?
So let me start with someintroduction.
We are bombarded with artificialintelligence and digitalization
of pathology on pathologyconferences.
(06:59):
The accuracy of the A.
I.
Models often comes very closeand sometimes exceeds interim
intra pathologist variabilityand the terms such as machine
learning, deep learning neuralnetworks.
They increasingly are being usedin pathology context.
So the question is, are weapproaching a new era where I
(07:21):
will dominate pathology or is itstill a long way off?
Let me start with theterminology to understand it
better.
What is artificial intelligence?
It is a branch of computerscience about systems which can
learn from data.
So we encounter AI every time wetouch any of our mobile devices.
(07:42):
And examples here would beGoogle predictive searches.
When we start typing somethingand to look it up in Google, we
get a suggestion.
And usually these suggestionsare correct.
And there, this is a AI productrecommendations on Amazon.
This is also artificialintelligence.
(08:02):
When we buy something and we getsuggestions of similar products,
or for example, we get thismessage customers who bought
this also bought that.
And it suggests us differentproducts, music recommendations
by Spotify, based on the songsthat we already liked.
And Google Maps fastest routethat is based on the current
(08:25):
available traffic informationwhen we are using it as a
navigation system already.
And based on the current trafficinformation, Google suggests a
faster route.
What is machine learning?
This is a branch of AI that isbased on the idea that computers
can learn from data as humanslearn from experience.
And Then they can make decisionsabout data without human
(08:48):
interactions.
We have three types of machinelearning supervised learning
based on labeled data,unsupervised learning based on
grouping unlabeled data pointswith similar features together,
and reinforcement learning,where the user of the algorithm
or the model gives the modelfeedback on what was labeled
(09:08):
incorrectly.
What is random forest?
Random forest is one of the mostpopular supervised machine
learning algorithms.
It's a forest of decision trees.
These do not tend to perform sofantastic in pathology.
What is deep learning?
This is a subfield of machinelearning in which the models
(09:29):
resemble neural networks of thehuman brain.
And those models within deeplearning are called artificial
neural networks.
And an example or a type ofartificial neural networks are
convolutional neural networks.
And these have dominated thedeep learning space.
Computer vision is a field ofcomputer science trying to mimic
(09:50):
human vision.
And patching is dividing largepathology images into small
patches.
And to better picture this, weneed to know how big such an
image is.
And so a whole slide image is 15gigabytes, more or less.
What is 15 gigabytes?
It's three, three hour long,high quality Netflix movies.
(10:15):
Just one slide, one digitalizedpathology slide.
They need to be divided insmaller squares and these
squares are called patches andthe process is called patching.
Graphical processing units arenecessary for quick processing
of pathology images.
These are chips on thecomputer's graphic cards
designed to rapidly processgraphics.
(10:37):
They're used in gaming, videogaming computers and are
indispensable for fastprocessing of whole flight
images.
Computer aided diagnosis is whenclinicians use computer defined
regions to assist them in makingthe diagnosis.
This is already in routine usein radiology.
For the detection of breastcancer foci in mammograms and
(11:00):
the potential use in pathologywould be pre screening of lymph
node sections in cancermetastasis.
This is a tedious task, takes alot of time, there are many
sections of the lymph nodes andpathologists have to screen them
under the microscope.
It would significantly, savetime if they could have them
suggested by the computer andthen just say yes, no.
(11:20):
What is data augmentation?
This is a way of getting moredata when we don't have enough
data.
And on this example, on the leftside of the screen, we have four
mitotic figures.
They have been slightly alteredshifted a little bit, and we get
this.
From those four mitotic figures,we get 16 mitotic figures.
(11:43):
We increased our data set fromfour points to 16 data points.
And these are seen as by thecomputer as separate data
points.
So we don't have enough data.
We can augment them to have morefor the training of the system.
What are probability heat maps?
This is a color coded way ofvisualizing the classification
(12:05):
results of the deep learningmodel.
And on one end of the scale iscorresponding to a high 100
percent probability of a featurein this case of the tissue being
a tumor and the other end of thescale, the blue one in this case
is corresponding to zeroprobability.
So we can appreciate that thoseall red regions, is.
(12:26):
Probably tumor and regions inbetween, and there is lower
probability of tumor.
To visualize how those termsrelate to each other, the
broadest term is artificialintelligence, and the machine
learning is a part of artificialintelligence, where we have the
patching, random forest, anddeep learning is part of machine
(12:49):
learning.
We have Artificial neuralnetworks, convolutional neural
networks, as an example of them,we have probability heat maps
and data augmentation.
Computer vision is thediscipline that uses all those
models.
GPUs, graphical processing unitsenable us to process whole slide
(13:09):
images fast.
And in the end, the algorithmcan help us with making the
diagnosis, and we have computeraided diagnosis.
On conferences and inpublications on LinkedIn as
well, we are surrounded with allthe AI news, but what is the
current status of artificialintelligence in pathology?
(13:30):
As of today, May 18, 2019, thereis no FDA approved AI solution
for pathology on the market.
There are initiatives supportingand accelerating the development
of AI applications, like the FDADigital Health Innovation Action
Plan, Digital Health SoftwarePre Certification Program, Pre
Cert, and in differentdisciplines, AI already.
(13:54):
Made it officially with an FDAclearance or approval like
diabetes research incardiovascular and brain disease
treatment in radiology, but notin pathology yet and in
pathology publications about A.
I.
All those publications come fromR.
N.
D.
Departments, either academia orindustry, different companies,
(14:17):
but it's research.
So to foster AI and to encouragepeople to develop a solutions
and challenges in biomedicalimage analysis have been
established.
And what are these?
These are competitions aiming tocompare new and existing
algorithms in biomedical imageanalysis.
(14:39):
Thank you.
And researchers are verystrongly encouraged to take part
in those challenges in order topromote and contribute to AI
development.
The participants try to solve astated problem on a common data
set, and they can use a solutionof choice and are required to
(14:59):
publish their results.
So the most famous challenge inpathology was chameleon
challenge organized in 2016 and2017.
And this and other challengesare gathered on the comic
platform standing for consortiumof open medical image computing.
(15:20):
And they are called grandchallenges.
They are the most famous inbiomedical image analysis, but
there are also other platformsthat are hosting them.
And these include Coda, labCoval, or Virtual Skeleton.
And let me tell you about thecommunal challenge.
What it was about it, as I said,was one of the most famous
(15:41):
challenges in pathology.
Maybe not in biomedical imageanalysis or different ones that
I'm not aware of, but inpathology.
This one was the most famousone.
It was hosted by in theNetherlands by, the diagnostic
image analysis group anddepartment of pathology offered
the Radboud University MedicalCenter in Nijmegen.
(16:01):
And as I said, there were twoadditions, Chameleon 16 and
Chameleon 17 in 2016 and 2017.
And almost 100 submissions forboth of these challenges were
done both from academia and fromcompanies.
And the problem here wasdetection of breast cancer,
metastasis, and whole slideimages of lymph nodes.
The participants were workingwith the same training and test
(16:25):
set of 1, 399 hematoxylin andeosin stained lymph node
sections.
And they were free to use anyimage analysis method to best
solve this problem, so theydidn't have to use deep
learning.
But it just so happened that thewinners used deep learning.
So this performed the best.
This is how the challenge isdesigned.
(16:47):
Not only the pathology ones, butthe other ones as well.
A meaningful task needs to bedefined.
In this case, that was breastcancer metastasis detection in
lymph nodes.
Representative data needs to begathered.
In this case, it was labeledwhole slide images.
A reference standard needs to bedefined.
That was the diagnosis of thepathologist and their
(17:08):
annotations.
and a discriminative evaluationmetric needs to be determined.
And in this case, the consensusof the algorithms with the
pathologist input was measuredby Cohen's kappa and the
participants are required towrite a peer reviewed pathway.
(17:29):
So what are the tools that areavailable and what are not
available?
There's no FDA approved AAsolutions for pathology, like I
said, but Why is there not?
We will see later.
But for research use only, imageanalysis software companies are
starting to incorporate AImodules in their solutions.
where a pathologist or a usercan annotate target structures
(17:52):
to train the model and later themodel detects those structures
automatically.
For example, the pathologist orthe user annotates glomeruli on
one slide or in one part of theslide and then the algorithm
detects the rest of theglomeruli on the other part of
the slide or on the rest of theslide.
These solutions are currentlyavailable from different
(18:15):
companies like Indica Labs,Viziopharm, and Aphoria.
And these tools incorporaterandom forest or deep learning
modules for image analysis, butas I said, they're for research
use only.
Nothing for clinical use so far.
And why not?
Why is there something AI basedin other disciplines and not in
(18:38):
pathology yet?
So there are some hurdles thatare pathology specific, and
these include lack of validatedtools, as we already said, lack
of labeled data.
We said that the machinelearning need supervised machine
learning needs a large amount oflabeled data.
(18:59):
This would mean that thepathologists instead of looking
at sites would have to go andstart annotating whatever
structures they want to havedetected.
And it's not going to happenbecause they are supposed to
diagnose patients and they useslides for that, the non digital
slides.
So lack of labeled data, highcomplexity of histological
images.
(19:19):
In comparison to radiology orcardiology, the complexity of
histological images is much,much higher.
We have different colors,textures, tissues, organs, and
deaths multiplied by each othercontributes to this high
complexity of histologicalimages.
(19:40):
The dimensionality of pathologydiagnostic problems is also
high.
And the example that we had thedetection of neoplastic foci in
a lymph node section is just onebinary problem that was used for
the competition.
But the dimensionality ofPrimary diagnosis in pathology
(20:02):
is a lot higher.
There is a different patternrecognition involved and
incorporation of many differentprocesses visible on the slide
and different data availableabout the patient.
Another problem or a hurdle isthat pathologists are the gold
standard or they are generatingthe ground truth.
(20:22):
So the algorithm can only be asgood as a pathologist and also
pathologists tend to differ.
Algorithm is gonna be close towhat one pathologist says, does
it mean that it's a good one ifanother pathologist says
something else?
So that's a hurdle.
(20:42):
And the affordability ofcomputational power and storage
space is a problem.
Those GPUs that are required forprocessing these huge slides
cost a lot more than a normalPC, normal computer.
That's a hurdle.
And we need to store somewherethose digital slides, enormous
amounts of data.
(21:03):
So that's another hurdle.
So to sum up, AI is affectingour lives every time we touch a
mobile device, phone, whatever,computer.
It has officially enteredvarious medical fields, but in
pathology it still remainsrestricted to researchers only.
There are many pathologyspecific hurdles to overcome,
(21:23):
and the most important beingenormous size of pathology
images and complexity ofmorphological patterns.
The regulatory pathways havebeen already paved by other
disciplines, which is good.
And the FDA officially supportsthe digital health initiative
with their digital healthinnovation action plan and
digital health software precertification program.
(21:45):
So to answer the question thatwe ask at the beginning.
Is it gonna happen?
Yes, it's gonna happen soon.
A.
I.
Is coming soon.
It's we're not quite there yet,but it's coming soon.
So pathologist should prepare.
They can prepare by goingdigital in their labs by
recognizing where A.
I.
Will add value.
So it may not add value to everystep of the pathology workflow.
(22:12):
It's crucial to recognize whereit would and apply it there.
And they can prepare byeducating themselves.
There are courses available, onecourse for example, from digital
pathology association, an onlinecourse that pathologists can
take and they should proactivelyengage with a initiatives and
beat the challenges like thechameleon challenge or
(22:36):
collaborations, cooperationswith academic partners, with
industry with differentcompanies.
Thank you very much forlistening to me.
And I hope you liked it.
If you liked it, please clicksubscribe to do my channel.
And if you have any questions,don't hesitate to write an email
(22:57):
and go and check out my blog anddigital pathology consulting dot
com, where there is an articleabout this presentation as well.
And I will also.
Put this presentation fordownload if you would like the
slide.
Thank you.
Bye Here the references in caseyou want to go back to the
primary literature or and Followthe links for the things that I
(23:22):
mentioned.
Thank you so much for listeningto the end.
That tells me you are a realdigital pathology trailblazer.
This video, was posted fiveyears ago and it has 3000 views
on YouTube.
This is one of the most viewedvideos.
Of course it has been there forthe longest.
(23:42):
And there was a whole chunk ofwhat happened in the last five
years missing, like all thegenerative AI foundation models,
chat GPT.
And all that good stuff.
So of course there is an updatevideo that I posted five months
ago.
And this was has almost 2000views already.
So I'm going to link to both theolder version and the new
(24:05):
version if you feel likecomparing.
How I was presenting them versushow I'm presenting now.
In those over five years.
My YouTube channel is telling methat there are 510 videos.
It's a little bit.
Elevated, because at some pointI connected the digital
pathology podcast.
So all the audio podcasts are onYouTube as well.
(24:27):
When Google decided to mergeGoogle podcasts with YouTube, we
migrated to YouTube.
So it doubled the number ofvideos, but definitely over 370.
Native YouTube videos.
So I decided to create astructured curriculum out of
those videos, they were beingcreated more or less on demand
(24:48):
when I was hearing from you.
Oh, I would like to have a videoon that.
I would have liked to have apresentation or if I would give
a presentation, there was anoption to record.
You're going to find things likethat as well.
Recently, I started addingconference vlogs to give you a
glimpse of what's happening atthe conferences and a life live
commentary on what's happening.
(25:10):
So I decided to create astructured curriculum.
And create.
Course based on.
Selected and curated YouTubevideos.
I am still working on putting ittogether.
But if you're interested inchecking it out, In exchange for
feedback and testimonials, Ihave a waiting list, for 100
(25:33):
digital pathology trailblazerswho will get access to this
course for free.
So if you're interested in testdriving the course, giving me
your feedback, whatever youthink about it.
You just let me know.
And if you love it, if you wouldgive me a testimonial, then this
course is going to be for youfor free.
I already shared it with thedigital pathology trailblazers
(25:56):
from my mailing list.
And several of them signed up.
There are still a few spotsleft.
For you, the podcast listenerbut before the.
free.
access waiting list, Gets sharedon social media.
I decided to share it on thepodcast.
So for you who listens to theend.
(26:16):
This is an opportunity to getaccess to this course for free.
And let me know what you think.
Check out the link in the shownotes.
YouTube course Waiting list,sign up and whenever I have
ready, it's going to end up inyour inbox.
And I talk to you and the nextepisode.
Popular Podcasts
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
Amy Robach & T.J. Holmes present: Aubrey O’Day, Covering the Diddy Trial
Introducing… Aubrey O’Day Diddy’s former protege, television personality, platinum selling music artist, Danity Kane alum Aubrey O’Day joins veteran journalists Amy Robach and TJ Holmes to provide a unique perspective on the trial that has captivated the attention of the nation. Join them throughout the trial as they discuss, debate, and dissect every detail, every aspect of the proceedings. Aubrey will offer her opinions and expertise, as only she is qualified to do given her first-hand knowledge. From her days on Making the Band, as she emerged as the breakout star, the truth of the situation would be the opposite of the glitz and glamour. Listen throughout every minute of the trial, for this exclusive coverage. Amy Robach and TJ Holmes present Aubrey O’Day, Covering the Diddy Trial, an iHeartRadio podcast.
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.