Artificial Intelligence & Machine Learning: Transforming Pathology | Webinar

If any of the statements applies:

➡️ You know AI and Machine Learning are already part of the pathology workflow, but maybe you are not exactly sure which part of the workflow?

➡️ “AI” is still a bit of overhyped, fuzzy buzzword for you?

➡️ You would like to learn about how it can help pathologist and labs work smarter and patients get better care.

Then this webinar is for you!

This is the second part of the “Digital Pathology 101” webinar series, based on the “Digital Pathology 101” book, where Dr. Aleks Zuraw explains digital pathology and AI concepts.

This journey through Chapter 3 illuminates how image analysis, AI, and machine learning not only complement traditional pathology but propel it into new realms of precision and insight.

As we delve into the essence of tissue image analysis and the transformative role of AI and machine learning in modern pathology, you'll discover how these technologies augment diagnostic methods, enhance research, and redefine what's possible in our field.

From the basics of tissue image analysis to the advanced realms of computer vision and the pivotal role of quality control, this webinar bridges the gap between high-level computational domains and daily pathology practice.

What You'll Explore:

The foundational principles of image analysis, AI, and machine learning in pathology.
The crucial balance between classical and AI-based approaches to tissue image analysis and their applications in both regulated and non-regulated environments.
The importance of quality control in ensuring accurate, reliable results from AI-assisted analyses.
An introduction to the key terminology of pathology informatics, demystifying the language that underpins digital pathology and AI.
Who Should Attend:
This webinar is tailored for:

🔴 pathologists,

🔴 researchers, and

🔴 healthcare professionals

who are eager to learn about and/or integrate AI and machine learning into their work.

Whether you're just starting or looking to deepen your expertise in digital pathology, this series offers invaluable insights into leveraging technology for enhanced diagnostic precision and patient care.

Date and Time:
April 11, 2024 at 9:00 - 10:30 a.m., EST

Location:
Online

AI is here, so let’s learn what it means for pathology

how can you leverage it for your work?

and how to navigate this new technology responsibly.


Looking forward to seeing you on the inside!

Transcript

[00:00:00] My digital pathology, trailblazers. This is a recording of a webinar I hosted today. And thank you so much for showing live. So many of you have shown live. 

But I know that not everybody was able to. And I know that my podcast listening, digital pathology, trailblazers, I'm not always the same one that are looking at the videos. 

So that's why I wanted to give you the opportunity to listen to this as well. So let's dive into it.

Intro: Learn about the newest digital pathology trends in science and industry. Meet the most interesting people in the niche, and gain insights relevant to your own projects. Here is where pathology meets computer science. You are listening to the Digital Pathology Podcast with your host, Dr. Aleksandr Zhurav.

Aleksandra: Welcome, welcome, welcome. Welcome, everyone. Thank you so much for joining me live. I see already people joining. So whenever you join the [00:01:00] live stream, the live webinar, let me know in the comments, where are you tuning in from? We are a little bit early, maybe one minute early. So there's gonna be enough time for everybody to join.

And I'm looking forward to hearing, where are you dialing in from? I'm in, , no, I'm in Fairfield, Pennsylvania, population of 500, , talking to you about AI and machine learning in pathology today. Hi, Christiane, great to have you. Let me know where you're dialing in from. This topic is gonna, I hope it's gonna spark some discussion because, , It is very, it isn't really, but it can be controversial.

And we have Fabio from Brazil. Welcome, Fabio. Great to have you here again. Amazing from Brazil. What time is it in Brazil? I assume. So it's 9 a. m. in Pennsylvania. Let me know what time it is, , where you guys are, but great to have you here. So this, , let me tell you. Oh, I have some, [00:02:00] I. Welcome France, Ali, hi.

Hello, Germany. We have Cape Town. Hi, Brendan, how are you? Okay, we have, , it's 10 a. m. in Fortaleza. Thank you, Fabio. We have Atlanta, Georgia. Amazing. We have Johannesburg, South Africa. This is so cool. It's great to have you here. Everyone. Hi Christophe and good afternoon to LinkedIn. So we are live on LinkedIn, on Facebook and on YouTube for sure.

So also let me know which platform are you using? Is it YouTube? Do do a Y for YouTube and L for LinkedIn, or the first letter of, , whatever it is that you are watching at. Good morning, Fernando. Good morning Stacy. Amazing to have. to have you here. USA! We have Sweden, Mari, welcome. Okay, we have YouTube, we have LinkedIn, perfect.

And some of you maybe just, , dialing in from a direct stream, stream [00:03:00] yard link, , whatever you're using and whatever works for you is amazing. Hello, Denmark. Why do I say that this, , topic actually is so why is this topic controversial? Why would it be controversial? So because I think it's very polarizing and all the polarizing topics are controversial.

So when I published my last video that you might have Gotten in the mail, , AI, , this was AI in pathology and this is the thumbnail of the video on YouTube. , and, , surprisingly enough, not too many people have, , watched this. I think we have, we're going to have more people watching this live stream than the video, , but that's okay.

, I just wanted it to be out there and I got some emails. I got some comments, , for example, , One comment was, AI can be used for prostate and breast core biopsy. It's helping, but it's not absolute. And then the other comment was, what is the role of pathologists if AI is being used? Like, there is this role of pathologists is being questioned [00:04:00] when in the context of AI being, , used.

Helping pathologists, replacing pathologists, where do we stand on that, right? Other comments that I've gotten was that predicting molecular status is very random. The person didn't trust predicting molecular stuff. And then, , there was another one that the real utility of AI is just to filter out the negative cases.

So, , a couple of comments, a couple of opinions. If you have your AI. Opinions, comments, things that you think right now about AI. Let me know in the chat, like, the first thing that comes to your mind, AI and pathology. Why did you join? I assume you're enthusiastic about it. That's why you joined. But maybe you have some concerns.

Maybe there are ethical concerns. Maybe there is. , maybe, , job security, maybe the reliability of this, or maybe you're super excited that it's going to streamline, , work, and we are going to touch on all of these, , and my first reaction to, to the comments [00:05:00] was, let me tell you, it was defensive. I'm like, How do you mean it's only useful for, for negative cases?

There are so many other applications that it can be used for. Why do you think this predicting of molecular status is random? When is it random? Have you not seen this publication, that publication? , and then I kind of took a step back. , basically they are, they, depending how you look at it is they can all be true and they can all be false.

, Why? Because, now let me show you a few more, a few more, for example, it's all about validation, actually only as good as the validator, so it cannot be better than the validator. And the validator is a person, so it's actually fallible, like makes mistakes because we make mistakes. And then, , okay, are we really there yet?

Are we really, can we use it? Are we like, no, let's not use it. So this kind of comment, don't leave me. If I accidentally end this Trim, then no, I didn't end the stream and just say, , logistics here. , whenever something goes wrong, just wait for me. I'm going to try to troubleshoot. And if my internet goes down, I have my phone.[00:06:00] 

I can use my phone data. So don't leave me. We're just starting. And, , I just want to welcome a welcome Belgium. My suit. Hello. We have Nigeria and we have London. Amazing. Thank you so much for joining. So, , we're going to start with what AI is because as I told you, I have my super, my super, , thing to write.

Let's see if I'm better than last time. What is AI? , because, okay, I am getting ahead of myself. Let me go through the agenda for you. , what is machine learning? We're going to talk about narrow versus general versus super AI. Super AI. Who wouldn't want super AI? Well, not everybody would like super AI.

Then we're going to talk about some AI applications in pathology Divided by AI branch. You're gonna you're gonna know what that is in a second and then application divided by area of use and then we're gonna , talk about AI to computer vision translation. Not because I want you to learn AI to computer vision translation, [00:07:00] but this is kind of an example of, , a translation that we need to learn to be able to leverage those tools or discard those tools or do whatever we want with these tools.

So let's start with what is AI? What is AI? What is AI? It is a term used to describe Machines that can perform tasks that normally require human intelligence. , so what can we use it for? We can use it for visual perception. , you know, we have cameras that recognize faces, everything. We can use it for speech recognition, decision making, , whether that's, you know, Totally ethical or not.

That's a topic for another discussion. And, , by the way, I did give a presentation on ethics in AI, , at some point. So if you would like to, , have me present this to you and talk about this in one of our future webinars, let me know, , in the comments. Just write ethics, and I will know that, , , That you wanted and we're gonna have this conversation as well And i'm gonna make sure that we have enough time for discussion because this is also like a non black and [00:08:00] white topic That will require discussion language translation.

I love languages I speak like six or seven But you know what if I didn't I could get by everywhere because now with my phone I can just basically take google translate or any other translate chat gpt translate and go and be happy Across the world. So that's what we're going to be talking about today ai.

So, , also another comment to what is ai, because, , I often hear, and I kind of like teased it in the email about the video, that, oh, I would love to use ai, , in my pathology work. So, , maybe at the, I don't know in which year, but maybe it was just connected to one thing when I was starting, it was just image analysis.

, when somebody was, , talking about using ai, , they wanted to do image analysis. Now, when, , I hear this statement, it's like, I would like to use. AI in I don't know in anything to me. It's like I would like to use a machine, but what kind of machine? What do you want to use it for? Well, I would like to use a machine in my kitchen Okay, that's still not specific enough.

So this is We need to be more [00:09:00] specific or I mean We don't need to but if you want we want to communicate and we want to engage in the discussion and not generalize and not discard or , What how do you say that? , not not , You In Polish, you, you, you, , like spill the kid with the water. Write me in the comments what's the correct English version of this idiom, , to, to like get rid of the baby with the bathwater.

, right. Because if we just have this categorical, , opinion or, or preconception about AI, , I mean, we might be as right as we might be wrong. And there's a lot of nuances, , to this. So that's kind of my general. goal with this presentation today. So, the next question that we would like to answer is, what is machine learning?

You can also hear, I'd like to use machine learning for my pathology work. Yeah, I'd like to use machine learning. And to me, whenever I, like the first thing, , I, I think about when I hear this is image analysis. Because I'm biased, because I started in the image analysis field. To [00:10:00] me, always AI is image analysis, but now we know it's a lot more for a lot more things than just image analysis.

But, Nevertheless, every time I hear this AI and machine learning, when I hear machine learning, I'm thinking random forest, and I'm thinking, don't use it. It doesn't work for pathology, which is exactly a statement that is as incorrect as correct. I've seen enough images, , you were enough images analyzed with random forest that didn't match the tissue.

, but there were others that were good and random forest. Is good for other stuff, maybe, and not necessarily for, , tissue image analysis. Or, it's maybe good for, , tumor stroma separation, but not, , cellular analysis and things like that. So, me saying, oh, no, scrap, , scratch this, , sorry, random forest, , go and do everything with deep learning.

It's like, well, yes and no. You're right, but you're not really right. So anyway, going back to our presentation is, and I see, , some of you wanting the ethics presentation. Fantastic. Luca, thank you. Whoever wants the ethics [00:11:00] presentation, let me know. , if there is, , some of you that want that, we're gonna, I'm gonna organize that.

Basically, I'm going to give you the presentation updated because every time I prepare a presentation, I think that I already have everything. I don't have it because I need to update. So, , go ahead. Crucial thing is here. Subfield of ai, what did I use right now? I want this one, , subfield. , so what can it be used for?

, and, and the subfield. , you can create models that let machines perform tasks that would otherwise be possible for, , we only be possible for humans. It's like, okay, Alex, are you repeating the same definition here? It was also normally would require. Require human intelligence. Well, because it's subfield and it's pretty broad subfield, , it can also be used for categorizing images, social media optimization.

So the feed that is being shown, for example, on LinkedIn, it's based on your preferences. It's updated by AI. While you're scrolling and it's learning what you like. So in my feed I have a 99. 9 percent digital pathology content or [00:12:00] Like health digital health content mostly digital pathology. So linkedin knows my preferences mobile voice to text and predictive text price fluctuations I mean all kinds of predictions facial recognition and product recommendations.

So, , this is machine learning and Basically, when it's being used in the context of AI or pathology, machine learning is such a huge subfield of AI that it kind of is being used interchangeably. In theory, and like to be very correct, when we, when we look at this diagram, AI is the biggest circle of the computer science.

Then inside of AI, we have machine learning, and then even more inside is deep learning and deep learning. This is from the book. By the way, we are talking about chapter three of the book, and if you don't have the book yet, at the end of the presentation, I'm gonna have a QR code so that you can download the book.

But basically, now, deep learning is [00:13:00] pretty popular, but it's not exclusively what's being used. I have seen, and I'm seeing, a combination of deep learning and classical, , computer vision and classical, , Yeah, image analysis, image analysis. I'm going back to image analysis. It's not only image analysis.

Anyway, , deep learning is, , like a well performing way of doing, , AI. And, , spoiler alert, we're not going to be talking exactly like about the, , building blocks of AI. We're going to be talking about concepts today. So anyway, , just to summarize, , AI versus machine learning versus, , deep learning.

Now, very often, they're used interchangeably. correctly and correctly. I think now everybody more or less knows and assumes that when we're talking about, about AI, we're talking deep learning or a combination of deep learning with something else. So that's what this is. Yes. And I already am getting, , good comments about this.

generative AI that can have many applications. We're going to be talking about applications, and it's not exhaustive what I'm going to [00:14:00] be telling you about. There are probably more applications than we're going to be talking about, but we're going to talk about categories that will help you navigate what AI can be used, and I hope that's going to be useful.

So let's talk about narrow AI versus general AI versus super AI. So narrow AI is also known weak AI. My friends, it's very weak. This AI is supposed to be so powerful and is weak. , it is the most common type of AI. I want you to remember this, because this kind of, I can't even, the fact that this is the most common type of AI, and some people, including truly yours, Dr.

Alexandra Zhurav, , would say this is the only type of AI that we have. It's only doing, some limited, , specific tasks. Or a range of tasks, a specific task or a range of tasks. So, facial recognition is a specific task, right? Then, , speech recognition is a specific task. We have image recognition. We have natural [00:15:00] language processing, which, , these two are related.

Speech recognition and natural language processing. And then recommendation systems. Whenever I go on Amazon, My Amazon miraculously knows what is the next thing I want to buy. It's, you know, whatever, workout equipment or clothes or whatever. It suggests me those super cool things and this is based on AI.

So, NarrowAI, right? Let's remember NarrowAI is the most common type. What do we have? Next, we have General AI, make it big, General AI. This one would be strong, and you already hear me saying it would be. It's designed to perform any intellectual task that a human could do. So if we use, we go here, we have the narrow that can do facial recognition, then the narrow one can do speech recognition, then another image recognition.

Like, I can do it all at once, right? So, , the general, if there would be a general one, it could do it all at once. And, Here, the thing here is it is a theoretical form of AI that is not yet possible [00:16:00] to achieve. It's a theoretical form of AI that is not yet possible to achieve. This is like the scientific classification of, , these, , AI capabilities.

, and you know, you can Google it and there's plenty of references. The reference I used was an article, , on LinkedIn by, by Professor Ahmed Banafa. And, , Yeah, so what would this AI be able to do? It would be able to reason, to learn. It would be able to understand complex concepts, just like us. It would be our AI robot friend.

Wouldn't that be great? Well, maybe yes, maybe not, right? So, this is general AI, and this is AI. a theoretical concept. And now, I am surprised I'm not getting any comments telling me that, hey Alex, what about chat GPT? Is that not general AI? They say it's general, right? It can understand, like, all this text and you can talk to it and it talks back and all the other large language models.

Well, it's not. Let me tell you, it can understand text, but it's, , just a chatbot. It uses [00:17:00] natural language processing and generative AI. And we had a comment about generative AI. So generative AI is basically AI that can create something from scratch. Based on some training, it not only predicts something, but it actually creates.

So this one creates text and, , you know, there are other creations that AI can do. , but it's just for conversations, right? Okay. So it can only do one thing. It's very, very good at this one thing. So to me, this is still narrow AI. And I totally agree with this classification. We are in the world, in the reality of narrow AI.

We can stack several narrow AIs to have a pretty powerful tool. But it has these narrow AI components. There's always going to be something that's missing. Like, , I don't know. Can it recognize the flavor of my tea? Is there a model for that? Would you even want a model for that? I don't think so. , so chat GPT, narrow AI.

Alex, what about DALI? DALI can make images, but you already know where I'm getting at, right? , DALI can make images, but, , well, it actually, look, it's, it's powerful. It actually does text to image. So text to image, it can, , [00:18:00] generate, sorry, generate images, , from A prompt from text, and I'm going to show you something funny later, because this tool that I'm using for these presentations, it is AI powered, AI enabled, so what does it do?

It can, it has a language model that generates some text, so if I put some concepts, , it's gonna, , put it in nice words, it can actually generate images. So several of those images, , are being generated. If you don't have a citation, , or if it's not from a paper that's cited, assume that this is an AI generated.

, and, , then, yeah, it's not that, , as you would think. I'll show you. And, , hello to bad candy. , so anyway, Dali. It can do text and image, text to image, it doesn't do image to text, but now you can combine these things, and we're going to be talking about this, and, and you have more powerful use cases, but still narrow AI, and, , and Berlin is here, hello, and let me know, , if you want the ethics presentation, I am getting some more ethics, , yeses, and also, if you like, if something comes up for you when I'm saying, oh, this is this way, this is that [00:19:00] way, drop me a counter argument or a counter example, in the comments.

I'm going to be starring these so that we can, , talk about this in the Q& A. So here, I hope I convinced you that, , we are in the era of narrow AI. But still, when we talk AI, it's, , talk about AI. I want to use AI. It's like, I want to use a machine. Okay. What for? What kind of AI? What about super AI?

Super AI. It's super. It's better than us. It does everything. It's capable of surpassing human intelligence in all areas, in all areas. Basically, it can solve climate change, eliminate disease and poverty, and the examples are Terminator. Literally, there was a reference, because, so let's go back to, , to something that I skipped.

It is a hypothetical form of AI. This one was theoretical, so I assume theoretical is closer to reality than hypothetical, , so that's why Terminator. There was another, , Iron Man and some other examples in, , , giving this example. So, [00:20:00] but actually, the development of super AI is the goal, is the ultimate goal of AI research.

Do we want that? I, I'm gonna confess something. I actually never watched Terminator. , there are a couple of movies somebody recommended, and if you have any recommendations, let me know in the comments about AI. There were, there was one positive one, and one, Terminator was not so positive, although I didn't watch it.

I will watch it next time, and, , I report if I did. , But yeah, from common knowledge, I assume Terminator was a robot that went wrong, but then he was good. Anyway, it doesn't matter, digressing. So, super AI, not there, hypothetical form of AI that's going to be even better than human. But we're not even close, right?

Which is fine. So, thank you, Professor Ahmed, for those nice definitions that I used in our presentation. And let's talk about AI. applications, AI applications in pathology. We're gonna start with AI by branch. And by branch, I mean, , by like, like the area of [00:21:00] computer science is being used. So as you all know, and we'll hear several times, , I started in the image analysis field.

This is what, for the longest time, AI in pathology was used, before the language models were even considered to be incorporated. Oh, I have a good recommendation. , I should watch Moonfall. Okay, I'm gonna, I'm starring this recommendation, Dr. Kim. So, , yeah, by branch image analysis, then natural language processing, we already started talking about this, and we can combine the two right now.

Right. And at the moment, this is like the state of knowledge that I have, state of research that I have done for you. It doesn't mean that, you know, next month there's not going to be a new application. So, , obviously if you're interested in that, , LinkedIn is going to give you the correct feed if you're liking and engaging with the posts about, , these topics, about, , this subject.

, But yeah, so it doesn't really matter how current this information is. It gives [00:22:00] a framework to see what kind of AI is out there and how can we use it and what the applications are. So, A, for image analysis, it also has several applications. Oh, and I have a great recommendation, , another one. I have heard about this movie, Her, but I haven't watched it.

So, , homework for me, watch the AI related movies to be able to, , reference some pop culture references and not just papers. My sister is so good at this. I am not good at pop culture. Anyway, so, image analysis can be used for different things as well. And the beginning was, of course, quantification. We were able to quantify.

So these are Alzheimer plaques, and these are my personal annotations in a specific field of view where I wanted to, , annotate them. , and they are both counted, and they are delineated. So, , Here is the delineation. The dots, , are for counting. Basically, they are quantified in one way or another.

Here, this tumor is being delineated as well. Once it's being delineated, you can, , calculate the size. You can do different [00:23:00] things with it. And if you want to be even more granular, here I was not that granular, that you can separate this stroma from the tumor epithelium. And here you have this, , that is exactly delineating.

, so if, for example, , the amount of stroma, which I know in pancreatic cancer, that is the case, is a kind of biomarker, then you can exactly quantify and do your calculations with these quantifications, with these numbers. We can use image analysis for computer aided diagnosis. And go ahead, if something comes up, , for you, like, yes, no, I'm afraid, let me know in the comments.

Just let's, , let's have a base for the discussion later. So computer aided diagnosis, I have noticed in around 2019 that there has been a shift, the concept, especially in the, in the image analysis world and AI for pathology world, which at that time was focusing on image analysis, that there was a shift from this concept.

Oh, AI is going to replace pathologists, which I am not saying that this concept is wrong. , [00:24:00] eradicated, , because this is, , something that comes up very often when people start learning about this, , and do not have yet granular information, nor do they have information how, , these tools are being developed.

So, before 2019, and this is my arbitrary, , , date where, when I went to conferences, the, the paradigm started shifting. Then before it was AI is going to replace pathologists and you will not have to look at things, it's going to diagnose for you. , I assume people actually hope for that level of performance, regardless whether, you know, the ethics questions and if we want that and, , if it's legal.

I think they were hoping for that level of performance. And very quickly with deep learning, when deep learning entered the scene and Alex Net was outperforming everything in image classification and competitions, they thought, okay, everything that's image based will no longer needed to be done by a pathologist, by doctor, right?

And very quickly [00:25:00] while working with pathology images, , the scientists realized, well, it's not like, One thing takes care of all. It's still narrow. And when we, when it comes to diagnostics, it's even more narrow because for every entity, you actually need an algorithm if you want it to go the diagnostic route and never.

So, so the paradigm shifted to something called computer aided diagnostics, not computer diagnostics, but computer aided diagnostics where a pathologist, , is using AI as help. In 2021, we had the first FDA cleared solution for, , for this computer aided diagnostics based on image analysis, deep learning, September 21, 2021.

And who did that? I bet we all know who did that. That was Paige AI and the, , The, the algorithm was called PageProstate, so they went to the FDA with their solution, saying and showing data, backing up the, , the, this is a good hub for pathologists. From what I remember, [00:26:00] it was the pathologists with AI were 7 percent better in whatever their metrics were and, , you know, you would have to go back to the paper, , which I don't have cited here.

I'm going to do this after the presentation. But, , basically they were seven percentage point better than pathologists alone and then AI alone. So the combination was better in detection. And, , here also keyword detection. of metastatic, , not metastatic, , malignant, malignant foci, and this is the task of this detection of foci that are suspicious of malignancy, and they are being highlighted to the pathologist.

That's the only thing that this algorithm can do. This particular one that got the clearance. Can you go even more narrow than that? Maybe you can. But this is not specific to this one, right? That this, I don't mean it in a negative way. This is what we're dealing with in pathology. So then also, like, prioritization of this algorithm development is going to be, okay, by the number of cases per year, [00:27:00] by the number of slides pathologists have to review per year.

, so, yeah. It's not going to be, oh, let's now, although there's a concept called foundational models, where, where they are developing models that can recognize more general things than just particular cancer, , like in this case, but that's outside of the scope of this presentation. So that was 2021 and we had computer aided diagnostic application.

So that is the only algorithm, only model that was actually authorized by the FDA. Does that mean that that's the only one out there? Not at all. Not at all. There is a publication. It's actually preprint from this, , from, from the, where is the publication? Here. The, the reference is here. Anyway, it's a preprint.

, but, , there is, , public evidence on AI products for digital pathology. It's a very interesting, , publication that, , let me show you everything I have from it and then I'm going to comment on that. Hmm. Very interesting publication showing that, , All the different things that [00:28:00] are being used, and there is, , I found this figure.

It's, it's part C of figure 2, actually. So we have our page prostate, right? And when you look at these arrows, date of first publication. Prostate published here, and then got approved here. This is like a stellar example. They published first, and then they, their data was so strong, , Then that they, , got this, , this approval, right?

And here, blue is publication priority approval. The bio ex prostate. There is approval. Where's the publication? I don't see any publication. Ibex Galen prostate. We had the approval. Where's the publication? The publication is post approval. , and then there is pro page prostate grade and quantity. So they are adding on top of the page that, , page prostate that only detects, they are, , adding the grade and quantity.

, approval, here is the publication. And it doesn't mean approval by, , FDA, because these are also, , European, , , European companies, , or different country companies. And I'm sorry, [00:29:00] Because I didn't put it big. I'm sorry. , this should be big. Let me know if, if I'm, , doing the, , wrong view, please let me know so that you can actually see what I'm drawing on this, , on this drawing board.

But anyway, this is from across the world. , this is from 2019 to 2023. And this paper, the, the Public Evidence on AI Products of Digital Pathology. This paper, , gives. Like a list and breaks them down. There is a lot of, , those, , used tools that are, never get any, never had any, , approval or never had any publication.

They were developed, I assume somebody did some kind of validation in the lab, , where they are using this tool. , I definitely would, , but over a hundred publications, so here they're only showing, , 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 12. 15, right? 15 publications, sorry, 15 tools here, , that actually got some kind of approval.

So they all have a dot. Dot is the approval. And whether the publication [00:30:00] was before or after, I think there's only one that doesn't have a publication. All the other ones have publications, which is great. You know, some were after them before. , anyway, there is enough proof for regulatory bodies for these particular tools to go validate on your own as well and, because that's probably a requirement of every lab, and and use them.

But this is just 15 and in the publication there is at least there's over a hundred, at least a hundred more than a hundred. I'm I'm Guessing 123. , who's using that? I guess you can just use it if you want to. I know one particular, , practice that is very much using it and has been using it for several years.

This is, , a practice in Puerto Rico, CORE CORE , is from, what I wanted to show you is all the partners that they have. That they are working with ALP and Glow Biosciences. , , here is Tech Site. This is a webinar they had with Tech Site, , core Plus Partners with ALP and Glow. And , yeah, this is Mariano de Sakara, the CEO of this , practice.

And by the way, [00:31:00] I will have him on the podcast I actually already recorded, so stay tuned for this. It's really interesting. , but you get the point they have partnerships with different companies They're very very early adopters and they have been using it. They have been using it for several years They have been doing their own validations and they have been , like gaining confidence how good is this thing, is this particular algorithm that we're using for the use case that we have.

So, , they did it in the CLIA regulatory framework as a lab developed test for several of these tests, including prostate, , not from PAGE, , it's from another company. Why can't I change my slide? Hmm. Okay. So, AEI 4, molecular predictions. It's still, we're still talking image analysis. And, , here. There is skepticism.

And let me tell you, , until I have seen this publication that I already talked about last time, , and I have more, more images from this publication. I talked about this artificial, , intelligence, molecular analysis tool assists in rapid treatment decision in lung [00:32:00] cancer, a case report. Until I actually have seen this case report by Dr.

Barry Leeds Weissengren, I was skeptical. I thought it was a hype. I thought, well, okay, nice, like, academic work, but when is it gonna make any impact? Because the claims were, well, , this can then help, , those who don't have access to molecular tools, because, , it's, , in the under resourced countries, there is no access to that, right?

So if we have those image based, , diagnostics, then we kind of, like, solve the problem. So, , Well, I was skeptical. And for some application of these molecular predictions, , like for example, virtual immunohistochemistry stains, depending on the stain, , I am still skeptical. But when I saw this publication, I was very, very happy.

They had a way of actually checking if this prediction is okay. So, , what happened here? They basically did a histopathological, , evaluation of a sample, and then they run, , this, , algorithm to detect, , the algorithm predicted the [00:33:00] EGFR mutation, and here is, like, this is the area where it's, and then the pathologist looked at it and said, yeah, that makes sense, this is, , this is cancer, it can have the mutation there, and, , they didn't blindly just move With treatment, they confirmed it with another method, which was they just did they had this direct essay.

This was a rapid PCR direct essay for EGFR. And they checked. Is this really EGFR in that particular spot? And the, , the answer was yes. And here you can see, so this is routine practice. What would they have done? , they would have started the radiation therapy already. Then they would have waited 7 to 14 days for a next gene sequencing report.

And then only after those, , for lucky a week, and if we're not that lucky after two weeks, they would start a targeted treatment. Whereas here, they already knew It's an EGFR we can start treating. And this was based on, I want to, I want to give a shout out to this company. Let me see if I have it on the previous slide.

I think it [00:34:00] was Imagine you can confirm that in this publication and it was a commercially available algorithm. It was a commercially available algorithm validated for that particular lab in Israel. And it was. It is a case report, right? So we are, , all have read enough scientific papers that we know that the case report is the, like, lowest level of evidence and it's not a randomized clinical trial or whatever the highest level of evidence is.

And so with all that in mind, I'm still excited about the potential of molecular predictions. Can I, , give a blanket statement that molecular predictions work? No, it's a prediction for EGFR mutation in mice. , lung cancer. That's even more narrow than the prostate. Yeah, because now you go molecular, so you, you predict a particular, , mutation.

There it was malignant cells in this cancer. So anyway, I'm excited. Let me know what you think about that. This is image analysis. What else do we have? Oh, we do have applications of Natural language processing and natural language processing and this [00:35:00] beautiful AI generated image and wait for the semi joke that they have for you.

It's going to come soon. , but basically this beautiful image is, , was generated by this presentation software. When I said somebody talking to the microphone, they have like a hybrid microphone and interesting glasses as well. But the point of the slide is we, , can now with a natural language processing, we can leverage that capability for.

, Speech to text for, for dictation that is going to be transcribed immediately. So, , now several of these transcription, , software that are powered by AI, they can be trained, , to the particular pathologist's voice, , several of the, , digital pathology, or even a pathology, , lab, or pathology, softwares that you use for reporting.

They have this integrated and, , it's got better. It gets better and better. When I was doing my residency, , my, , supervisor was dictating everything. And then the, , there was a, , transcriptionist transcribing all this and he was [00:36:00] very He spoke a lot. And then sometimes when, , so he was like very descriptive in all this and she was always so tired when, , he was diagnosing and then sometimes, , another, , person was, , when he was not available, another person was doing this and this other person was so much shorter and so much shorter.

concise. And I remember this transcriptionist going to, , going home like an hour earlier because, or a lot, you know, a lot earlier, , because she didn't have this, , this first person who was very descriptive in everything. So now you don't need a transcriptionist. Speech text is good enough, or at least There are some instances of speech to text programs that are good enough that you can eliminate that transcriptionism.

Now, look at this. So, , the example here is, we're still talking about natural language processing, structured reports, right? So, , Just, , look at this beautiful structured report. The prompt that I gave is text report. I am wondering about AI myself. [00:37:00] So this is, , the other prompt I get was pathology report with text and I got some car.

So you know what? We need to verify sometimes. , I just wanted to show you that, , this was the output of the, of the report. of the AI today, or whenever I was preparing this presentation, but structure reports. So building on the story about the transcriptionist going home earlier, , when the other person was dictating, the more concise, how about, , we don't have the differences between people, , people actually, , dictating a lot or dictating not that much, , because we can, , simplify it by structured reports.

And you know what? You can even simplify it more by having AI fill in the blanks. The report. So basically you can dictate however you would dictate. If you are a person of many words or a few words, it doesn't matter. Your expertise is being conveyed in those words. And then you have AI, , working on a structured report.

, I do not have a concrete example here. I don't know a company that actually does it. , but I know the technology is [00:38:00] there and I know , that at least one company is working with a sound system company to actually enable the, the speech to text and, , voice navigation of the whole software. That's the natural language processing.

, smart search, smart search based, , on retrieval augmented generation. What is that? So sometimes, We, , are also pathologists. Look at that poster. It's about pathologists and chatbots, , by Andrew Bychkov. , go to chatgpt or any language model and they, , ask questions. They kind of Google, , the chatbot instead of going, , to a search engine.

And, and we all know that not always those, , answers are reliable. , like I saw with my flowers. They're not always reliable. How can we make them more reliable? How can we leverage this chatbot for search? Well, the technology or the method is called RAG, Retrieval Augmented Generation. What is that?

Let's start with Smart Search. Smart Search is when you, for example, have a database and you search in this [00:39:00] database. , for a healthcare institution, in a pharmaceutical company, in a biotech company, wherever, right? But you have this search, it's powered by generative AI, that's why it's called generation.

So it's not just like, , it's, it's not this word document search where it highlights the words, but it actually synthesizes, , An answer for you based on search in this particular database. So we kind of mitigate this or minimize or address the risk of the hallucinations of those models. Because we restrict the generation of new content to a certain database.

, that. is within an institution. It is, , hopefully curated and it's, it's controlled and prepared. And then we also make the model reference where it took the information from. Am I going to show you something here? I go to my other slides at the beginning. Here, let me show you something. , I don't know if I have it, but, , there is an option in this software actually, , like find a reference for what I'm telling you [00:40:00] about.

So, , it, it has this, let's see if I can show you. No, because I'm in a, , but basically, it's out of, I write something, let's say, narrow AI is only used for one particular task, and then I have an option to click and find reference, and it's gonna find me a reference, and it's gonna, , show this reference at the particular citation where it actually claims what I wrote.

So, how cool is that for any type of referenced material preparation? , I like it very much. If we think of, , the combination of image analysis and natural language processing, Then, , I found a paper from, , 2023, , that, , used medical Twitter. So what happened in this, , in this publication, , there, there was a, , they, , used crowd platforms to curate open path and open path.

Let me make it big. Open path, , is a data set of two. 200, 000 over 200, 000 pathology images and paired with their natural language descriptions. If you're part of the Path Twitter or Med Twitter hashtag, I assume that that's the hashtag, , [00:41:00] then you will see that there is a lot of diagnostic discussions going on there.

So they basically leveraged this database and created LIP, Pathology Language Image Pre training Multimodal Artificial Intelligence. So yeah, the combination of image analysis and natural language. Language processing is happening already, and let's look at it, , at our AI from a slightly different perspective.

This was by method, and now it's going to be by area of use. And I'm aware that not everybody can stay, , for the full one hour and a half. I'm almost done with, , this part five out of seven, , topics covered, , at , 10, which is in two minutes. So maybe I will go over time, five minutes. We're gonna have q and a session, , regarding by area of use.

We can use it for workflow optimization. We can use it for research and we can use it for diagnostics. And I think this application causes. the most controversy because this basically kind of implies that, oh, now we have an automatic tool instead of a doctor [00:42:00] making medical decisions or helping with medical decisions.

What's, , with confirmation bias. So even if this tool is so great, but at some point it's not going to perform that well, the pathologist is going to look for confirmation. This is how we're wired. So this is, this is the kind of controversial part, but I would love to have AI for workflow optimization. I would love to leverage it in the research.

Assuming it performs relatively well. And for workflow optimization, what can we use it for? Automated slide QC, speech to text, we already mentioned that. The research rag is something that could be leveraged very much, especially in the pharmaceutical industry. Let me make it big again. In the pharmaceutical industry where a lot of data is siloed and the search of these databases, , , It's very difficult, you have to search multiple databases at the time, , to, to reach, , like a full set of information.

, and I know there are efforts being made at several, , pharmaceutical companies that are implementing this rag strategy, , with multi modal data searching [00:43:00] and insight generation. And, , We can, of course, do biomarker quantification and automatic lesion detection. And when we talk about diagnostics, , we can have something that actually is not, doesn't seem that, , control automated case prioritization.

Let's say an algorithm runs on, , a set of slides and prioritizes the ones that, , Need to go out earlier or that are positive or I mean that's also an interesting thing because like why would you? Prioritize the positive ones if the people who have negative ones also want to know very fast if it's negative But the answer to this one was okay.

We have Usually, more cognitive power in the morning, so maybe more difficult cases, which usually are the positive cases, should be prioritized. And then, you know, you still do everything within one day and everybody gets their results at the same time. But you are working on the most difficult things when your mind has the prime hours of functioning.

Computer aided diagnostics and prediction of molecular tissue [00:44:00] properties in the images. We covered that in the previous part already. Bye! The last, last thing before we do a Q& A, six out of seven concepts, , and this is just to illustrate something. So, image analysis is also a blanket term. You basically analyze images.

How specific is that? Not too specific, right? , and I have a very good Start comment, , from Anisa that, , she is doing instant segmentation to achieve precise quantification breast cells, right? Super specific. This is what I do for my image analysis, and this is the level of specificity we kind of need to communicate clearly.

Let me show you this example with my gir. These are drafts that I photographed myself. in Disney in Florida in Animal Kingdom. So, , when you look at those giraffes on this first picture, you see like, okay, maybe it's not big enough, but it's from the book. So if you have the book, you will see a better version of this picture, but basically you can appreciate that there, there's one giraffe, there's another one, and there's another one.

And you know, this one is [00:45:00] very far, and this one is not that far, and this one is super close. And you have seen it immediately without any problems. But! And the computer is divided into, okay, I want to just detect these giraffes. And then we put a binding, bounding, bounding box, a box around the giraffe. So, , this one is in a box, right?

And there's another bounding box around the small one and another one around the middle one. And here it was, it has been detected. If you just need to detect cells. You can use this approach. Why wouldn't you go for a more fancy approach? If you only need to detect these cells and, , count them, this is a computationally cheaper approach.

It's gonna be faster, it's gonna use less, , of your computational power, and you don't have to overdo it. Even though natural thing for people is, I see it and I know what it is, right? Let's say we needed to, to quantify how much area on the spectra those giraffes occupy. , do I know, do I need to know, where each of those drafts is.

No, I could just make like a, , red on red is not optimal, but I can basically like draw a [00:46:00] boundary around all of them. And let's, Let's assume I actually did it around the legs as well. And all this area is gonna be my giraffe area, right? So if I need an area of something, and I don't need to count it, and the only metric is area, then I'm gonna go for semantic segmentation.

This is how it's called. And if I want everything, if I want to detect the objects, and if I want to segment them, each of them, like giraffe one, two, and three, the red one, yellow, and the pink one. Then I'm going to go for instance segmentation. And these are the computer vision terms associated with solving particular pathology problem with detecting cells.

And this example here, those cells are actually not cells. These are Alzheimer plaques in a brain sample stained with IHC. And so you can count them, but maybe one is bigger and the other is smaller. So maybe the counting is not really giving you enough information. You can, , You can also, , just delineate their area, , and maybe that's enough information, maybe it has been shown that the area [00:47:00] corresponds to something.

But maybe, , you wanna have everything, and maybe you wanna know the area, the exact number, and then check, like, how many big ones are there, and how many little ones, and do the little ones, , matter or not. And things like that, then you have to go for instance segmentation. This is like a micro example.

What I want you to take away for AI. So is image analysis, , good enough or not? It's fantastic. No, it's terrible. Both statements are correct. It depends. What are you using it for? What kind of method are you using? And then only how well it performs. We kind of need to be super specific. Be able to judge, which we have to judge.

We have to judge, , , and I'm going to be staring, I already see some good comments. Janet, I have you. So, , yeah, we, we have to know the use case, know the tool that we're supposed to use it, because if we're going to be doing semantic segmentation, then we start counting those plaques. These are not good numbers because you have merging of plaques.

So that's the wrong tool. It's like, I don't know, using your forks for the soup. [00:48:00] With that, I officially start Q& A. We have less than half an hour for Q& A. And while you guys are asking questions, if somebody does not have the book yet, , put it big, let me know if this code is actually working. If somebody can scan with their phone, and you should be able to download the book through this code.

So just, , if anybody does that right now, let me know by code working. And then I will know. So let's start. We could start by optimizing the workflow with AI, but it's still difficult for us to have the necessary technology. Yes. So that's another question. I agree. And yeah, if I understand that not everybody can stay all the time, thank you so much for joining, Brittany.

Going back to this question, Janet's question. Yes, not everybody has that, , necessary technology. And I don't have a solution to that. But there are different And I've had somebody reach out to me saying that, Oh, no, in my region, this is still SOFRA fetched. We don't have whole slide images. We don't have something else.[00:49:00] 

And yes, there are you know, places. I mean, it doesn't have to be a different country. It can be in the U. S. or wherever, right? You have one practice that has gone fully digital and the other one, , didn't do anything digitally. , so, , one thing that is kind of an example of let's try to figure out workarounds and let's try to figure out leverages.

even when we don't have access to the full pipeline is, , for example, image analysis on static images. You don't need, , host led images to do image analysis. There are, , solutions that, , there are models available that run on static images that you can take from a microscope camera. You have, , companies that have microscope cameras that partner with AI companies that can, , deploy it while you're looking at it under your microscope and then , on the screen next to you, there is this algorithm running.

So for example, smart in media, , I recently saw their Kai 67 counting, , capability, and then there's also another company Augmentix that I know that they were working on something. This is like a module, , in between your eyepieces [00:50:00] and your objectives that also kind of does it real time. , and you don't have to count.

things manually. So I hope that answers. Let me know if you have, let me, I'm going to be scrolling and checking, but yeah, give me questions. Give me questions. Give me things that were not clear yet that maybe, you know, you don't believe, but somebody told you and you didn't know how to answer. And I can maybe suggest, , a different point of view.

And if you are still here and if you can give this a live stream, a like on whichever platform you are dialing in from, that would mean a lot. And a lot to me because then more people, , , are going to be seeing it. Okay, so I'm going to show something that is not a question, but Dr. Kim is struggling to find internship or volunteer opportunity to work with companies who are currently involved in training of AI or machine learning models to improve diagnostics and pathology.

Do you know any, , forms where I could look for such companies to apply, please? So, if anybody from such company is interested, , feel free to reach out on LinkedIn and, [00:51:00] and, , Dr. Kim, , I'm gonna get back to you, , about anything else that I know , about getting involved, , later. And there's another question.

You can give me all the questions. , it doesn't need to be strictly related to the talk. There's another question. I'm looking for a database with high resolution images of toxicologic pathology studies, but haven't found it yet. Well, have not found it yet. So, I'm a toxicologic pathologist as well. I'm sorry not to be able to, , answer your questions.

So you probably are aware of the partnership that Charles River has with the Cypex. They have been doing this internally, leveraging their, , they have been developing models, but this was an internal database. I am not aware of any ToxPath available database, but there is a an initiative ongoing. I was mentioning this last time, big picture, but this is not going to be something that is for free and, , you know, openly available.

This is something where you will be able to apply and they are in their year three or four out of six years of generating this database and TalkSpot database of images [00:52:00] is part of that. Okay. I am having here some, , questions. So how can we start to optimize the workflow with AI? So for this, to answer this question, I need a more More specifics.

So examples that I gave, and feel free, Ray, to drop consecutive comments as I speak. So the examples to optimize workflow is the prioritization of cases. So there would be an algorithm running on scanned images that prioritizes cases. Then there is an automatic quality control of cases. images. , there can be, , things happening while the images are being scanned.

So this quality control for digital images can actually happen while images are being scanned, , in the scanner. Other things is maybe automatic, , assigning of cases to a specialist, depending which part of the workflow you want to address with AI. Let me know which part of the workflow, maybe I'll have some [00:53:00] idea.

And then, , Christina is saying, one problem is that the companies will not take over the responsibility for the diagnosis. So, if it is still the pathologist's responsibility, they will, of course, want to review the slide. Yes, Which they should. And, , so what kind of application, , are you talking about?

Because, , I'm not sure of any applications that, , kind of allow the pathologist not to look at the slide. Exactly for this very reason. Because, , , they are signing, the pathologist is signing off on the case, the pathologist is signing off on the study, so the company is giving you a tool, and you see if this tool is good enough, so like, a company that, , produces hammers is not really responsible for you doing damage to your finger with this tool.

Hammer, right? You need to know how to use it. And that's how I see it. And that's why I like to know how the tools work. I like to see the performance. I like to see the performance over time so that I can be confident that this is actually , helping me and not, , not helping me. I hope that makes sense.

[00:54:00] And, , feel free to, , give, give, , follow up comments. I see another one. , Christina, I see the space, , for digital pathology in triage and counting heavy stains, but not for diagnosis. Why, why not for diagnosis? So you don't like the computer aided diagnosis approach? I mean, You can not like it.

That's okay. So I don't believe in the AI replacing pathologists. This is not something I subscribe to. I do subscribe to AI helping pathologists. , triage is gonna help. Highlighting, , foci if the algorithm performs well enough is gonna help as well. It directs you to things that you can then, , evaluate, , In that case, I do believe in diagnostic applications, but not, , something, like it has to be an aid where there is still enough cognitive power of the pathologist, , to actually sign off of it, on it, and not be biased.

And I totally agree that this can also introduce bias. From what I have seen from the publications, the level of like the, the, , things that the AI is helping with is, , pretty [00:55:00] binary. Is there something or is there not? It's counting. So for example, a diagnostic application would be counting chi 67 currently, and I don't know in how many fields of view, but for breast cancer, , diagnostics, you need to count at least 500 cells in a field of view.

If I can, Not count them and have an AI count them and then see if the markups match. That's a good diagnostic help for me. So, , yeah, Christina, you're asking what's the benefit. So, 7 percent more accuracy in the page AI study. That was, for example, the benefit. If that's good enough or not, that's, you know, depending.

If that's enough for you, maybe it's not. Now there is a comment from Christian. We need to see remote biopsies to make money. That is the way in which pathologists today can take advantage of digital. Do you know anything about it or any international company that needs diagnosis from pathologists? , there are, I don't know about international companies.

I know about applications. , thank you so much, Christina, also for the good discussion. I love it. , but going back to Christian's, , I know, , applications abroad diagnostics. So for [00:56:00] example, remote consultation, , you can consult. from abroad on digital slide. , I know applications of companies hiring remote pathologists, but, , so whenever it's, , coming to hiring somebody from a different country, regardless of the country, and USA is the sixth country I've lived in, , regardless whether you like, , work from abroad, there are regulations involved.

So it's not like, Oh, now we can, , other than European Union, but, , for me, at least, , as a European Union citizen, it was easy to go and work in Germany, , and in Spain. But, basically, there's immigration involved. So, there are companies, and they've actually seen this being advertised on LinkedIn, , but I don't think it was, , for diagnostic applications.

This was for , Annotation. There is like a part time, and I think, , Dr. Kim was asking about that. So there was a part time position for annotating, , pathology images, and the requirement was an MD pathologist. So there are options. Anything else? , don't forget to let me know if you want the ethics presentation.

I don't see anybody who [00:57:00] tried the code, so I don't know. I hope it's working. I can try it on my own. I think it is. Yes, it's working. Okay, perfect. So, , Yes, I'm here for you. We still have 10 minutes. Let's, let's discuss ethics. Okay, good. People want the ethics. , and I see Kristen, you're interested in moving to New Zealand.

I know they have a program for health professionals, where it's actually from the immigration standpoint, not that difficult to go there. Let me know if you have any more questions regarding anything, digital pathology, if there is, if there are any other topics. So I, , Okay. I wanted to ask you something, so I'm thinking of exploring something that I found on LinkedIn.

It's called audio events. , why am I trying to explore this? This is something where I can go live and talk to you and we can actually have a conversation. It's audio only, and I see that the code is working. Perfect. Thanks so much, Samuel. And there, there is, so instead of me giving you a presentation and you having to watch it on video, it's something that you can just join.

[00:58:00] As audio only, would you be interested in me hosting an audio event? Give me an audio in the comments if you're still listening to it. If you would like to join an audio event. For me, I want to start doing this because it's a lower, , Like easier for me to just show up for you more often and to have these discussions , but let me know if you would like that where you would like audio give me audio in the comments LinkedIn audio events, then you would get a notification probably per email if you're on my mailing list and Have a good question.

, so and then a invitation from linkedin. Let me know if you would like audio event just a listen in thing where you can actually, like, be driving and, , not watching me. Question. Brief insights about data protection regarding use of AI solutions cloud based. So, do you mean if data is protected in the cloud?

Is that a question? , so, If you want to provide cloud solutions for medical purposes here in the U. S., you need to [00:59:00] be HIPAA compliant. So you need to comply with this regulation of not disclosing any identifiable data. And I know that the companies that are actually having their images and deploying algorithms in the cloud, you need to be HIPAA compliant.

, do, do comply with, , if it's for, , for medical purposes, they need to comply. Otherwise it cannot be used for medical purposes. , so I hope that answers the question. If not, let me know in a subsequent comment. So basically, whether regulators have something being done in the cloud or not, it isn't our question.

So for example, , in Europe, , It's not that in Europe, the GDPR regulation and there is a new AI risk stratification regulation, these things, , kind of have to be, , have to be complied with. And in Europe, for example, , for different regulatory bodies, it is necessary to be able to, , audit the location where a server with the data is located.

So if you have cloud [01:00:00] based computing and the data is in the cloud, the data can actually be in a server that's outside of the country, and that is not acceptable for your regulators. So it depends on the country. And if that's not acceptable for your regulators, then you cannot do it. Then you have to, for example, have your own.

server, or like you can do cloud, but you have to have a dedicated server that you're renting in a particular space when a regulator wants to inspect, you have to go to that particular server. Okay. And I see people are interested in audio events. Let me know if you like the audio events. Tell me, tell me what you didn't like about this presentation.

What do you think is like, totally useless. And for those who need to leave, thank you so much for joining. I appreciate you staying till the end, , this means a lot to me. I know that we are bombarded with different types of information, and I very much appreciate you joining me here on those presentations.

I'm planning to do one more next, , next week, , might be the ethics one, so I'm going to kind of deviate, because several people wanted the ethics one, , from the digital pathology [01:01:00] one on one book series, , or Any of the remaining chapters, , one is very toxpath specific. So I might Leave this one till the end and just , I mean i'm gonna let you all know if you want to learn more about toxpath That's going to be a good presentation It's very underrepresented and under talk about a specialty of pathology And then there's one about the history of digital pathology and general digitization concepts and one more chapter, that i'm No.

Or the ethics. You can let me know if you're watching the replay, , then you can let me know which is your preference, either book or ethics for next week. , I'm gonna still stay here for the couple of minutes that are, , left for our Q& A. Maybe some questions are gonna, , sneak in. And yeah, you're welcome, Suman.

I'm glad this was a helpful answer. Ah, what else I'm going to do? There's going to be a replay, so everybody who needs to leave, you will get a replay. And if you would want to share it with somebody, I will have this replay [01:02:00] available for everyone on YouTube. So if there's anybody you think would benefit from this content, and you can share it with them, I would very, very much appreciate it.

, and I'm also going to make audio content out of this like I did from the last webinar So if you want to have a quick audio recap of what we were talking about With maybe a little bit of editing Cutting out my ums and ums and the pauses that I'm making Then go ahead and visit the digital pathology Podcast page and thank you so much for those Very, very nice words that the presentation is great and good introduction to AI in pathology, AI in pathology.

Were you surprised with anything? , did you already know, , about all these things? Did it, , make you look at it from a different angle or was it just a recap? Hmm. Okay. I'm going to do the ethics seminar, , definitely either next week or the week after. After next week, I'm going to be in Poland. So I'm not sure how, , if I'm going to be able to do the.

 Live streams, but i'm definitely I already have [01:03:00] videos prepared for you on youtube So definitely you can see me there and interact there in the comments And if we don't have any questions, then I thank you so much for staying till the end It was a long presentation if anything comes to your mind after this presentation something pops up in your head You can go back to the recording that i'm going to be sharing with you and You Put a comment there.

And, , okay. There's a request for natural. Language, please. And there actually, I did give a presentation about chat GPT, where I was explaining natural language in it. So, , that's a new, but I can do an updated one because that was like, I don't know, over half a year ago. And so many has changed in that space.

And then I can also incorporate the data from Andrew's podcast. The literature has, the literature has been more and more compelling, though it remains to be seen what the interface is between pathology training and AI deployment as a whole, definitely. Yeah, with every single application. I like, there's no blanket statement, yes AI, [01:04:00] no AI.

It's the same, like, scrutiny for every single application. It's like every, , let's think about it as a, , each of these is a separate medical device. And for every medical device, you need to, , show the supporting data to get your clearance or to get your approval. So it's the same with AI. It's not like all AI, it's this particular application of AI that has enough supporting data.

Thank you so much. And now I'm going to be jumping to the private Q& A with the Digital Pathology Club members. So if next time you want to join the club, , you will be able to find the link when you get the book. This is, , Like a more closed group, , a paid membership that I'm hosting. And, , I hope to see you there one day.

And if not, then see you on YouTube and all the other platforms. Thank you so much.

Obviously on the podcast, there is no place for comments, but feel free to send me messages on LinkedIn. If for my email list, go ahead and shoot me an [01:05:00] email or leave comments on any social media platform that you're following and I'm going to be collecting all those questions. And the, maybe that's going to be a fantastic fuel for the audio events. 

Stay tuned for the next audio event and join me there. And I talk to you in the next episode.