Digital Pathology Podcast

98: DigiPath Digest #6 (Foundation models high level overview)

• Aleksandra Zuraw • Episode 98

Share episode

Send us a text

 Exploring Foundation Models in Digital Pathology: Insights and Tools

In today's DigiPath Digest we talk about the foundation models in pathology.
reviewing abstracts from two notable papers in Nature.

We discuss the high-level overview of these models, including Hamid Tizhoosh's insights on the vast data requirements for developing effective foundational models.

We also explore tools for literature research, comparing PubMed and Undermind.ai, and examine a useful children's book on artificial intelligence :)

The episode features audience interaction and offers updates on digital pathology trends, along with a personal anecdote on the nature of comparison based on a yoga class experience.

00:00 Introduction and Overview
00:16 Foundation Models in Pathology
00:33 Comparing Research Tools
01:03 Live Stream Interaction
01:12 Starting the Podcast
04:51 Foundation Models Explained
05:11 Research and Findings
06:34 Children's Book on AI
08:00 Deep Dive into Foundation Models
14:28 Case Studies and Examples
18:18 Discussion on Data and Models
21:00 Final Thoughts and Questions
26:24 Exploring ToxPath and Foundation Models
27:05 Introduction to Image Repositories
28:36 Using PubMed for Research
30:35 Exploring Undermined Tool
35:42 Comparing PubMed and Undermined
41:00 Final Thoughts and Recommendations


TODAY'S ABSTRACTS & RESOURCES

📄 Here are the abstracts reviewed today:

▶️ Hamid Tizhoosh's lecture:

🔧 The tool we tried today

📕 A book we discussed :)



Support the show

Become a Digital Pathology Trailblazer get the "Digital Pathology 101" FREE E-book and join us!

Welcome my digital pathology trailblazers. This podcast episode is in our recording of the digit path digest. We had the six digit path digest live online today. You do blinked in, uh, we even managed to do it on take doc. Today we discussed a high level overview of foundation models in pathology. Reviewed to nature abstract. From papers about foundational models. Then they recommended the kid's book. That actually is pretty useful for adults as well. And we compared two tools, pub med, and undermine the.ai for literature. For literature research. I got a little bit confused. Comparing those two tools. But hopefully in the next live stream, we're going to have experts. We're going to have the founders of this tool. Explain to us how to use it best. When I was reviewing the audio from the Japan digest it wasn't perfect. So I do have an edited. A little bit apologies if it's not the usual quality. And because it's a recording of the live stream, there's going to be some audience interaction. For the visuals as always go to YouTube and I'm going to leave the link to the video in the show notes as well, but for my audio, first digital pathology, trailblazers, let's dive into it. 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:

Good morning everyone! 6 a. m. in Pennsylvania. Good morning. Today I have my cup with the world, and I'm gonna have, oh, I think I just didn't go live on Twitter. We'll have to live with that. That's okay. I'm gonna, say hi in the chat. So tell, say hi in the chat as well. Hi, my digital pathology trail blazers. Also, I have a new microphone today. And, that means that I don't know if you guys hear me. So let me know if me. We are live. On, not on Twitter guys, on LinkedIn, on YouTube on, I'm trying to go live on TikTok, but it's getting, giving me trouble. Don't give me trouble, TikTok. Let's see, while you guys are, coming in, then, Good morning, Robert. So great to have you. Good morning, everyone. We have Masoud. Okay. And you guys hear me. This is fantastic. Let me know where you are tuning in from. Hi from Fairfield, Pennsylvania. Super excited to be here with you at 6 a. m. I know Robert is in Maine and He is live on LinkedIn. Perfect. Let me know where you are tuning in from. And, while I'm waiting for you to join, I'm gonna give you a few updates or thoughts. Okay, fantastic. Good morning. let me know where you're from because then Or where you are dialing in from because then I know like more or less. What time is it? although no, I only know what time it is in Europe, but so updates today I do yoga on Wednesdays my yoga teacher was not there and another teacher was there And they spent like at least one fourth of the class on thinking of Oh, this new teacher is not as good as the old teacher and like comparing, this other teacher. He was pretty good. Perfect. He was a great yoga teacher, but I was so used to this other teacher that I was comparing everything to him. So I'm like, okay, this is how we are, because when I started this yoga and I didn't know the first teacher, every, like for three lessons, maybe I was like, does this make sense what she teaches? Does this make sense? Are her jokes funny? and then. Once I got used to her and then, it was fine. Amazing. But what's my point with this yoga story that, we just are wired to compare, and those, self help books tell you, Oh, don't compare yourself to others because then you're going to feel either superior or inferior. Compare yourself to yourself, which is fine. But, We are wired as we are wired. So that was my little story with yoga teacher where I caught myself, just comparing the new one to the old one. And then I remembered then, that I was doing the same with the first one and today our topic is foundation models. Foundation models is like the new buzzword type in pathology. So what are they? We're going to do this today. What are those foundational models? So that when we, see those papers, see people talking about it or presenting something to us, know what these are and what I did, for the research. I did, go to YouTube, which is my, go to place to do research, because people on YouTube just take the time to synthesize stuff, and often they have, very high credentials and you can find very, good quality materials or scientifically quality materials. and I found this lecture. I found the lecture by Amit Tizoush. he is now at the Maya Clinic. this was, recently in May at the, association of pathology informatics. So everybody who's interested in pathology informatics goes there and Hamid Tissouche is a computer scientist. he now works in my clinic and hopefully I will get him on the podcast one day, sometime soon. we are talking about it and I found this lecture I marked here that only 295 people saw this lecture one month ago. It's such a good lecture. So I hope that after this one, you can go there and, look at this one. After my DigiPath digest number six, you can go and watch the full thing, but I also extracted and the stuff that is necessary for you. And then, we're gonna, check two abstracts and compare them to the definitions that we found in this very lecture. And I also want to show you something else. My new book, the book that I actually bought for my kids, ABC of artificial intelligence. I bought it for my kids and then I started reading it and I'm like. Okay. Chatbot. There's what is chatbot? but the better one was, okay. Feature no generative AI. You'd teach babies about generative AI and, other stuff like, something like this knowledge graph. I will read the definition of knowledge graph here. The generative AI definition. So generative AI is used to create new things like images, music, or text that weren't there before. They learn from existing examples and then come up with their own unique creations. This can help artists, musicians, and writers exploring new ideas and biologists and scientists. But basically like this explanation from this little book already gives you like working knowledge, what generative AI is, and. If you don't want to dive deeper, you already know how to converse about this. So guys, if you have kids, get ABC of Artificial Intelligence and read it to them, and then you will gain the knowledge as well. Or if you have somebody who has kids, then get this one as well. I'm going to link it, in the description of this live stream. But let's go back to the topic of our Digi Path Digest. Mm. So find foundation models. What are they? What are those models? What is my mouse? Okay, so when is a foundation Model F foundation model. And he goes, also, Hamid TEUs wrote this, nice article in media. In the pathology moonshot, we need 5 million host light images to train a foundation model. How did he come up with this 5 million Why 5 million? And where are we going to take 5 million from? In the lecture, he says a couple of things. I picked the ones that are most important. are good for, I thought are going to be good for like general working knowledge. 300 million parameters. and these parameters are like everything that those deep, super deep neural nets are working with. But basically this is like the information you're going to get in the paper, like how many parameters were there? And if it's over 300 million, he considers it a foundation model. They are general purpose and they are trained on massive amounts of data. This is like no different than, okay, what is chat GPT? Oh, and I see people liking this live stream on Facebook. If you are on other platforms linked in YouTube, wherever, if you could give it a that would help a lot, get it to more people every time we have a reaction. So every time you comment, every time you like, the algorithms on social media, take it to more people who might be interested. Which, is what I want, and this is my mission to get this information to as many people as possible. And I know that sometimes it feels uncomfortable, to engage and to and whatever, and then people can see it what you like. But if this is useful, if you can give it to like, that would be fantastic. I will very much appreciate that. And it's going to take it to more people. Which then, you can also share it with more people, but that's even more effort than just liking it right now. So let's like it. Okay. He has a lot of, articles on medium. This one was, written on July 29th. last year. I'm like July 29th. That's a couple of days ago. so what he says there, he compares, like, where does he, Oh, and I see likes. Thank you so much for the likes. Oh, beautiful. Where did he come up with this 5 million, right? So there is a model. Let me just make, I can like draw. Can I draw? Come on. A clip. This clip has Specific name, like it's, of course, it's an acronym, whatever. this was developed by OpenAI, and here is the paper called learning transferable visual models from natural language supervision. they made this model that, creates, image So they paired. A lot of images. with, the description. let's say here we have a dog, we have a dog here. So it would be like a dog, orange dog, or red brown dog laying on the floor. So basically they describe, I see myself disappearing from this live stream. let me know if you can still see me. okay. So black and white image of a house with so many windows, a description of this image, so that's what they did. for this clip and how many images did they have a lot of images so proper training of clip took 400 million images, image, caption pairs, image, caption pairs. So 400 million. So he basically said, this is 400 million image pairs. He calculated, Pixels patch from an whole slide image, because there's always this concept. Whole slide image is huge, but then patches smaller. if we want to just like patch level information, he calculated all this. And he says, To do something comparable in histopathology, you would need 5 million whole slide images, not 5 million patches, not 5 million screenshots, 5 million all slide images. and not only that, their corresponding reports, because if you can see here, we had image caption pairs. So the caption Would be our report report is the caption for, for our whole slide image. So we have a pair of huge whole slide image, then corresponding report. If we have 5 million of this, then we're good to make a really generalizable foundation model Let's go through our information. So A repository, A slide repository, for pathology called TCGA. TCGA is, the cancer genome atlas. and it has barely 40. thousand mixed whole slide images. Mixed means that there's frozen sections and, FFPE, formal and fixed paraffin embedded. And there is a little bit of text description with these. the, we wanted 5 million. What we have available in public domain is 40, 000. So a lot less, and very little. So we don't have full reports of this. We just have some text description, just here to show you what this cancer genome Atlas is. it's a project from. NIH, National Cancer Institute, part of the NIH and they have this, that cancer genome TCGA. So every time somebody says, Oh, TCGA, this is the cancer genome atlas. And it's not just holes like images. It is, also like images. Genomics, radiology images. So like a multi modal data set for all the types of cancers. And I think now I can comment my comment before we go to the abstract. And my comment is, okay, we are talking about cancer. Do I look at cancer in my work as a toxicologic pathologists where I can, if I read specific studies that are designed to check if a compound causes cancer, but most of the time I don't. So I don't know how to use, the foundation models would be useful for, Me as well, because they can do a lot of stuff, meaning they can segment, they can, do like a text description, they can maybe retrieve or search. There are different things that those models can do. But basically, this is for diagnostics on human samples. So We remember that. And let's look at the abstracts that we have today. did you read any of, Those papers, like in full, I started reading the one that recently came out, from Thomas Fuchs groups and from page AI. Let me know if you read Any of them, just, comment paper, if you have read any of them. So let's go to our abstracts and see what we have here. Can I make it even bigger? Yes, I can if I remove myself, but that's okay. I will remove myself for a second. the, if it fits to the description that we just learned. So the first one is a visual language foundation. for computational pathology. So we have this visual and language. This was published, in nature medicine on March 19th, 2024. And, what happened here? Of course, the first sentence is how important digital pathology is and how difficult it is to get data from it. And so few pathologists, very few pathologists. So they decided that they need to make a model. And most models in histopathology. only image date they did. So by day, Ming et al and the corresponding author is Faisal Mahmoud. He's also a renowned researcher in this space and they decided, okay, we need a visual language model. We need a visual language model. And so they did. We need something else. We need conch, contrastive learning for captions for histopathology, contrasting learning for captions for histopathology. And I love how they just take, whatever you want from those words to make a cool acronym. I will not, stop being amazed. Oh, so Monica actually read the paper, the full paper, from page. Perfect. Yeah, I started, it's a week long study what they did, but I have a couple of images for you. But this conch is a visual language model. A visual language foundation model developed using diverse source of histopathology images, keyword diverse source. We're going to talk about it in a second. biomedical text and notably, 1. 17 million image caption pairs. Amazing. It's a little less than 5 million, but it's already in million. It's not like hundreds, thousands, it's millions. that is a lot, It's focused on this phrase, diverse source of histopathology images. That means they took those images from different sources. So not only TCGA, but also PubMed and some, I don't know if it was screenshots, but basically from, different places from literature. So it was not only whole slide images and that were, Oh, and we have, we have somebody from Kosovo. Those diverse source of images was from, let's say, not only host light images, but also, PubMed. So like online images, which like for a human observer, fantastic, this is how pathologists learn. They look, in, in the book and then they check, Oh, this is the same in what I'm looking for. if it is, then you'll have your diagnosis, right? So I do not dismiss this, but I am, agreeing with not I am agreeing with Hamid Tizoush because he says, that this is not the data that we wanted to perform. This is good information, but we wanted to work on whole slide images. So we need to train with whole slide images. So how many whole slide images were, in this 1 million? Image captured pairs, we would need to go to the full paper. but not all of them. So that's one thing. So then what happens then is we are in August now, and in August a new paper is being published, a foundation. Shouldn't model for clinical grade computational pathology in the rare cancer detection. And that is the one that was published on July 22nd received I looked at and it's a long process to publish a paper. We know that, right? So what happened in this beautiful paper, that's the Thomas Fuchs group. And this paper is by Vora Notsov et al. Vora Notsov, Eugene Vora Notsov. And this is the group of Thomas Fuchs. What did they do here? Let me remove myself from the live stream so that you can see the paper and they called this Virko. so now we have the largest foundation model for computational pathology to date. So I guess every time you have the largest, you can go to nature medicine and say, Hey, this one is larger than the one you published last time. So let's publish this one. Anyway they made this Virho and it's the largest for computational pathology to date. It can do even biomarker prediction and cell identification. it does pan cancer detection has a very nice 0. 95 specimen level area under the receiver operating characteristics curve across nine common and seven rare cancers. Is this impressive? Yes, it is. Nine common and seven rare cancers. Are these all cancers in the world? There are more cancer, cancers. Are these all pathologies in the world? No, there is nothing like inflammation. There is, it's just neoplastic. It's one part of pathological processes. So why am I saying that? Not to say that this is not enough. It's so much the hugest thing we have now in pathology. but I just want to. tell you that this is a fraction of pathology. This is just part of what pathologists deal with. And obviously in the clinical human medicine diagnostics, this is a lot of, work. This is also a very, the disease that is, super important, super dangerous and disease. I say cancer is disease. It's like a lot of diseases. and they address nine common and seven rare. So amazing. Let's look at, the data set. I'm gonna make this bigger and remove myself. And by the way, if you have questions, give me questions in the chat, because that's our last, abstract. So give me questions, give me comments, especially if you read the paper, give me comments and give me additional info. I started studying, this. So what do we have? How many patients? A hundred, 19,629 patients, unique individuals represented in the data. So over 120,000 people were represented in this data. And this was, twice as many cases, almost 208, 000 cases that gives us almost twice as many specimen tissue samples. So almost 400, 000 specimen. And that gives us, over a million blocks in that gives us almost one and a half million and a half. H and E slides. So here we are not talking about PubMed images. We are talking about H and E slides, diagnostic samples, and tens of thousands of square pixels after digitization. Tens of thousands of square pixels, just tens of thousands per slide, probably. So times a million and a half. So a lot of data. What Hamid Tissouche says in his, in his talk is that we basically don't have repositories that are as big. we don't have repositories that have 5 million, 5 million images. At the end of the talk, if you go to YouTube, I'm going to link it in the description of this. somebody asked him, Hey, so are you basically like trying to make a repository of 5 million or looking for a repository of 5 million? And he says, yes, that's what I'm looking for. I'm like, wow. Okay. and we were also saying that, remember when at the beginning we said that, model like this is supposed to have over, over 300 million parameters. Are we hitting? Here we have 632 million parameters, trained this Dino V version 2 framework. Dino, what is Dino? some, another acronym of theirs, but basically a lot of data. this is what those foundation models are. They are huge. So chat GPT, those language models are foundation models. They can take, like they can talk to you, right? They were trained on massive amounts of data and they can now talk to you, can ask it questions. and we are trying, or there are groups trying to do that in pathology as well. what I'm thinking is okay, do we still need to invest so much? Effort and so much time into training those like one purpose model just for prostate cancer, just for, let's take an example in my space and talks, but just for the talks about five or seven or 15 talks, but the lesions, or should we wait till these models are somehow adapted to real life? And we can take this instead of just a neural network architecture. I don't know, but I'm thinking, oh, maybe we should Wait a little bit, or figure out a way how to get this, to work. It's I know there's always a balance. It may not be feasible, but there's always a balance between, Oh, the new technology that can do so much, is so amazing. And then you feel like using the old one, you're like, so behind. But then on the other hand, if you don't have access to the new one, using the old one, still gives you some leverage, right? So I think it's going to be a balance. I'm also interested about the models for life sciences, more for preclinical pathology. And let me tell you about some updates today. So do you, if somebody asks you about what is this foundation model, can it do everything? Will you be able to, give, an educated answer? Give me a yes. If you feel like you will be able to give at least this ABC of artificial intelligence level answer to this question, what is a foundation model? give me a yes in the chat if you feel confident after this, live stream about it, because now I feel comfortable at least engaging in the conversation about this. I know that the ones that we have out there, different types of data are being used. If it's data that is, Not really the pathology data, like some screenshots, it has less, weight the data that went into the training is of lesser quality than if it was whole slide images. Oh, we have a question. We have a question from Thomas. What are the outcomes of such training? Foundation models. Are they useful for the diagnosticians? Other question in Talkspot. Where are we starting a foundation model with normal captions with big picture? So fantastic question. And that was my question as well. So that, The outputs, it depends like what the goal is. If you have like multi modal data. So this, the vehicle one, they even have a biomarker prediction. They can identify cells. it depends a little bit, what you ask it to do, but the other model was okay. Image captions. If you have a patch, it's going to tell you what it is or the other way around. If you write something, it's going to, tell you. Find you a patch that shows this. These are the outputs, segmentation, class, segmentation, classification. Maybe like description. It may be search, because those models can search as well, so it it's supposed to do everything that a pathologist does, but obviously it's not going to be everything. And ToxPath. ToxPath is where we probably, where we both work. So ToxPath, where are we starting a foundation model? So there is one. Oh! ToxPath. I forgot an important part, but let me answer the question, but we're not done yet, guys. Even though we're done with the abstracts, I have something to show you. A big picture at the moment is an image repository. It's not if you want to start a training, a model on this data, when it's available, then you can, this is basically this, effort to aggregate a lot of images so that we can do foundation model. there is a foundation model for life sciences. we can actually ask our tool that I'm going to show you about this. model. There is one for life sciences and the company that did it. I don't remember the name of the company because I remember the name of the company. This other company came from. So how can is the company where this spin off came from? I will find it. and if not, it's going to be in the comments, but basically big picture. So for those who don't know, it's a European consortium that is supposed to aggregate a large amount of images and they have a clinical, part of those. So clinical MD diagnostic pathology. They also have pharma part. So all the pharma companies that are taking part in this are contributing images from preclinical, from animals, from drug development. and when we have this repository, then you can train a foundational model or something similar like this. I wanna show you a tool. How about I show you my screen as well? So how do we stay informed? because there's going to be a new one and there's going to be one for life sciences and hopefully they're going to be published, right? we will be reading scientific literature or at least we will be reading abstracts, right? That's what we're doing. before I show you I get those PubMed alerts. So we set up a PubMed alert. So let's see if it actually shows. Yeah, I have PubMed. So we're going to check foundation models in pathology because PubMed we know is keyword based. Oh, and Google has path foundation model. I've heard about this one as well. I think I watched a little video or, a little presentation, but it was pretty high level presentation. so let's see, we have page model, Thomas Fuchs group. We have this other one, that was image pairing. We, now put the path foundation from Google. Jose is, Jose Carlos is, giving us this information. so we have seen a few papers. Let's see. if we put foundation models in pathology, do we want to foundation models foundation? I don't know how to call it differently. Like I know what I mean when I talk about foundation models. So let's see what PubMed says. We're Gonna go to check what PubMed says. And do we find any of the papers that we actually discussed, right? They're good papers. Definitely. And this is foundation models and pathology. You have a lot of information. results. we can do most recent. It is actually for most recent. we can restrict it to different stuff, but let's see. Deep learning based modeling for preclinical drug safety assessment. No. burnout among European veterinary students. So that's interesting for foundation models and pathology. I need to read that one, maybe to inspire some veterinary students. Okay. Integrated network pharmacology. I don't know why it's giving us this. in equalities for women diagnosed with distal arch and descending thoracic aortic aneurysm okay, of the top five, nothing is about foundation models. I think we need to be more specific for PubMed. Let's just check this other tool. It's called Undermined. Undermined, am I logged in? Yes. It's a different type of literature research tool. Let me know if you have seen it or if you have used it or if you have, Path AI has named it as Pluto. Path, okay, so Path AI has one that's called Pluto. We have Virco, we have Path, what's Google? I don't know, Path Foundation, Pluto. Of course, right? I should put it. behind me, like all the names of those models. I love them. and then you like take random letters from the words that you, it's supposed to be an acronym. Okay. But going back to Undermind, this new tool. Let me know if you have seen it, if you have heard of this Undermind. It's for literature research need an account, but for PubMed you need an account as well. See, Sveta says a different thing. A first result for my PubMed is a visual language foundation model for computational pathology. So that is not my first. With the same query, Sveta, let me know if it's the same as foundation models in PubMed. pathology. Why don't I have it? I don't know. But other people have it. So I'm not totally, I'm not dismissing palmitas at all, but I just want to show you the other one. So if heard of this, and thank you so much for checking. This is, this is super cool. And because I, of course, cannot check on a different computer than my computer. So we have one. Not so great results from PubMed. One good result. feel free to go to PubMed and check what, what for this query and let me know. Okay. so let me know if the other ones are relevant and we're going to check and I'm going to show you what undermines does. It's a compliment. I wouldn't say at the moment it's to replace anything, but we're going to search now and we're going to take the same query. Oh, it disappeared. So it's going to take some time to search, because it's like a chatbot. I'm going to take the same query and, it greets me with welcome. I'm your research assistant. So basically this is a little bit of a research assistant. it asks you to talk to it, to chat GPT. I want, publications about foundation models in pathology. Let's see what it, okay. Yeah. So Thomas, what's wrong with my PubMed guys? See, Thomas, because Thomas says he also has, the same result. Do I refresh it with the same exact query that I have foundation models in pathology the same? Maybe I should check Maybe I should log out Let's see. Maybe if I log out then it's gonna be better. Let's try No, I get the same stuff. I don't know. Okay guys, that's okay But thank you for checking. And I don't know maybe my, PubMed is biased because I already maybe looked for stuff that was drug development, related, but okay. So our undermined assistant research assistant, let me know. Have of undermined? Give me an undermined in the comments. if you already heard about this thing. What does it say? I was like, give me the papers. The first time I was working with it. And then it's not designed that way. So it's, I understand you're looking for, I want publications about foundational models in pathology. Can you be more specific? Imagine explaining yourself fully to a colleague. So it always says that. And I'm like, my colleague would know. What would you tell them? I would tell them exactly that having this detailed information is crucial to help us find exactly what you're looking for. Let's see if I put the same thing again, what it's going to do to me, unlike my colleagues, no, my colleagues are on the live stream with me and they exactly know what foundation models are okay, but. It's going to ask us clarifying questions. You're interested in foundation models suggest you're keen on exploring pre trained models like BERT or GPT applied to pathology, perhaps for diagnostics or predicting disease outcomes. okay. Anybody have heard of undermined? give me an undermined. If you have, Or no undermined, if you haven't, then, you'll see how it fits. But basically, let's see what undermined says. Could you clarify whether you're focusing on a specific pathology? Cancer, infectious disease. See, undermine knows that not all pathology is cancer. So histopathology images, genomic data. Okay, so now. We're going to put our query. I want publications in pathology that used histopathology images for training. Are you looking for more papers to discuss the theoretical framework and development of these models? Or are you interested in empirical studies that demonstrate their application and effectiveness? Okay, great. I've revised the search criteria and copied them below. Take a look and edit them if necessary. Then when you're ready to start the deep search, click submit. And this is going to take us some time. I want to find publications about foundation. Okay. It basically gives me the same query that I come up with. So here, like being totally objective, regardless whether my PubMed is biased or not, and this took us a little bit more time than just the keywords, right? And it's going to be searching. So guys, I'm live on multiple platforms, but it's doing pretty fast job. in the meantime, we're just going to go to PubMed and see, maybe histopathology. What is going on with my PubMed guys? My PubMed is like not giving me the right stuff. There is also another thing we can do. It's Google Scholar. Let's go to Google Scholar because Obviously, that's what I would do after I would see those results in PubMed. I'm like, what is this? Nothing that I want to read about here. We're going to do this and we're going to just do the same query from PubMed. And what do we have? this one I already read. So can I make it bigger for you? I should be able to, huh? Ain't that beautiful. We have go to Google scholar and do this as well because, maybe you guys have different results. Everybody has different results. Now in the generative AI age, everybody has different results from their search, but that was not yeah. Like it's what it is. It's like learns from your preferences, right? That there's algorithms running in the background that give us the searches that they think we might be interested. This was the one that we already, talked about medical Twitter, that's a different one. Why is it, why is the, why is my, why is our paper from. Page not here. Let me know if you guys, what do you have the page paper in Google Scholar? I do have the page paper when I just do normal Google search. I was working in the background. I don't want to use the term veerho, or give it the keyword that directs it immediately to where it should go because, I'm doing this little test. Of course, I know what I'm looking for, so I would just use the right keywords. So it's an artificial thing. where is this one, right? Yes. That's the one, nature. com. computational pathology as a search word, Monica says, yeah, let's do that in computational pathology. So here, just to compare and contrast the two tools, here, regardless, what you guys had, you had the better result, but me, I didn't have a great result and this is life. I didn't pre search it. It didn't find anything I wanted for my query. So now I'm like looking through different websites and changing the query with better keywords, right? In Undermined, it took a different route. it prompted me to have better search, term. in the first place. I was a little stubborn and left the one that they had in the first place. Computational pathology, digital remote access. So it still gives the same first one. let's do just five years. let's do one year. it still gives me the same stuff. Okay, feel free. Sweeta says Google Scholar gives the same results. So for this Google Scholar search and feel free to give me better keyword searches for my PubMed. We need to have this PubMed, get this PubMed to work. okay, Svita says Google Scholar, is the same. Robert says check the article the articles. I know where this article is, so this is just an exercise to compare two tools. and, yeah, I have the link to this, article already. I downloaded it, but, Undermined, this, is done, right? What do we have? we'll see. Okay vision transformers, Rudolf V. It doesn't have Virchow, no yes, but no, because this is, the, RRX Viv version and we have it. So here's, let me give you an, overview, what it does, topic match. It gives you a relevance, relevance of the topic. So basically. And it tells you, publications about foundation models trained on histopathology images for pathology applications were found specifically referencing 1, 2, 3, 7, 8. and here it gives you, okay, so far I've closely analyzed 30 of the most promising paper and I found 17 to 18 that are relevant, which is probably 42 percent of all that exists. Interesting. Okay. So that means that, we trust the tool, that means that, okay, it already found a 17 to 18. And, if we want more, it's not going to be like thousands more that we have to go through. this is 42%. So we're going to have another 20, right? Or whatever, 25. I think it's a useful info specifically for, those who do literature research, not just like paper search, literature research, and let's see, domain specific optimization of diverse evaluation of self supervised model for histopathology Rxvid, Rxvif, oh sorry, okay, where is nature? Where is our nature? Okay. I don't know. Nothing of nature. Guys, what I'm gonna do, I think it's a tie. Let's see. Nature on PubMed. We add nature. I don't know, guys. I'm lost. anyway, so what I am encouraging you to do, so these are, which ones are histopathology foundation models enable accurate ovarian cancer, subtype classification. So here, foundation models, foundation model by pathologist Rudolph V. So that's like a Virgo version. And often the papers that we have, in other journals, they are published in this are. Viv, this is like a pre-print publishing, place. And what I wanted you to try is to try this undermined. This is, so they have a, for whatever it's gonna be worth for you. for example, if I was preparing a, presentation, like a webinar on those foundation models and I wanted to cite a lot of papers, I would probably do a combination, at the moment. I like that these Papers have this topic match from undermined have a topic match and it actually like matches 100 percent and then it goes down right and then I know okay, whatever goes down I'm not gonna be even touching there's enough that has 100 percent that I can check and explore and which PubMed doesn't give me I will have to judge it by the title which I can there is Google Scholar, there are different tools. Just wanted to show you this new tool what I'm going to do one day, maybe the week after I'm going to have the authors, the founders, the creators of this tool, with us on the live stream. what we're going to do, I'm going to say, Hey, this other day, watch this live stream. Look, PubMed was crazy. Your tool was, okay. but I was looking for a specific paper and I didn't find it. I only found it on like page, eight, in Google Scholar. And if I didn't know about this paper, what would I do? How can they're going to join us. They're going to talk about this. in the meantime. When you need an account, and there is a free one, this is the one I'm using. You can actually get an account if you feel like it, to test if, especially if you do a lot of literature research and for next time when an account, then you can do the exercises. If they're going to show me, they have a comparison between PubMed and Undermined. And of course it came up better for whatever they were doing with this tool. I like it. I use it. I want to learn to use it better. And next time we're going to just challenge the founders of Undermined with how do we use it properly so that it outperforms other stuff and we don't need anything else. And whether we will use anything else or not, it's our decision, right? Because we can do whatever we want to do. There was one book I read recently. the words of wisdom were people will do whatever they want to do and they will not do what they don't want to do. I experience this every day with my husband. He's a doctor, he's an MD pathologist, clinical pathologist. And he is brilliant, he can do whatever he puts his mind to, or hands, or whatever. what he will not do It's the things he doesn't want to do. And it's with my kids as well. They are now five and a half and three and a half. And, I used to be naive and think, Hey, I will teach them all these skills that they should know in adult life. You can only teach them whatever they want to learn. So I'm trying with artificial intelligence. Let's see how that's going to be. Let me know if this session was useful. Let me know if it was fun. Thank you so much for engaging with this and checking, verifying what I'm doing, if this is real. Couple of things, that I wanted to tell you before we cut off this live stream. In our book, I'm reading is, atomic habits. it has been published already in 2018 atomic habits. And it says, don't set goals, build systems because with goals, you're going to fail. And with systems, the outcome or the goal or whatever is going to happen is to fail. A derivative of the system that you put in place. So this is our system. to explore different tools, to explore the by the way, Robert, I'm going to show, where to find those articles. we want to go directly to the source. a lot of articles I also take from, LinkedIn because people when published, they, are super proud of it. So I marked them on LinkedIn. You can mark, save stuff and then go back to it. but in those atomic habits, it's like set systems. So you know that this is already the sixth time. so I promised myself that in, if we do four more times, so on the 10th, I'm going to be getting myself a tablet for annotating. I have to start looking for the tablet because it looks like we're going to actually do it. thank you so much, Thomas, for joining. This is beautiful. So if anybody is watching the replay, feel free to because even if it's not live and you're watching it, it's gonna take it to more people. If you think that somebody would benefit from it, but you're like, you don't want to share with them. then, going live on Twitter right now, that is not good. Not smart, whatever. I found where to click. I don't think we went live on Twitter. Next time guys. And TikTok, I didn't figure out the TikTok thing. but yeah, this is our system guys. And this is our DigiPath digest system. Is it it's, is it perfect? We're just using, we're just reviewing abstracts. We're discussing things, on a high level, in a high level way. Is it useful to me? It's super useful. So I'm hoping it's super useful to you as well, or useful enough for you to show up every week, which I appreciate. I see the several people that are showing up every week. So I super, super appreciate you. if, you listen to this, if you want to review it, in the audio version, I'm going to publish this is a podcast I've been publishing several, not all of them, not all six are as a podcast. If you go to the audio version of the digital pathology podcast, this one is going to be there either today or tomorrow. So then you can listen to it in the car on your commute or, doing chores and just remember and absorb more. Different people learn differently, so I have it as an audio version, And I have a goal, to get 50 podcast reviews and I need your help with that. So if you listen on Apple podcast, Spotify, YouTube music or whichever podcast app, just to the audio version, if you could go there and give me 15 minute review. Of course, if you feel it's a five star, if it's not five star, don't give me a review, then I bet if you don't like it, you're just not listening to it. But if you like it, if you could give us a five star, review, I would love to get to 50 podcast reviews by the end of the year. I we can easily do it. we have under 10 right now. It would be amazing and if you can write something there as well Like why do you think it's good? That would be amazing There's another goal that we're having as the digital pathology place team on YouTube. And obviously if you've been here for an hour, you already contributed to this goal. The goal is to get 4, 000 watch hours. We already have 3, 600. And so we need 400 more. The thing with YouTube is, It has to be within a year. So tomorrow, all the hours that I gained, on the same date last year, they disappear. So people have to keep watching keep watching more than they already watched last year until I get to those four hours. 4, 000 hours. what happens when we get the 4, 000 hours, then the channel is monetized. So basically then, we may get some revenue to support these activities, for digital pathology place from YouTube. but I need those watch hours. Thank you so much for joining me. I appreciate you. I really appreciate the interactions and I talk to you in the next episode.

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.