99: DigiPath Digest #7 (Exploring AI-driven advances in digital pathology from T-cell signatures to synthetic images)

In this episode of DigiPath Digest, broadcasting from Poland,  we delve into advances in digital pathology, including AI applications in bone marrow evaluation, classification of hematology cells, and the use of synthetic images for data augmentation. Additionally, we review a survey on pathologists' perceptions of ChatGPT and consider the feasibility of GANs for enhancing medical image analysis.

00:00 Welcome and Troubleshooting from Poland
00:21 Live Stream Challenges and Conference Details
02:21 Digital Pathology Podcast Introduction
02:51 Technical Difficulties and Audience Interaction
06:18 Exploring Digital Pathology Papers
06:43 Advances in Bone Marrow Evaluation
09:03 AI in Hematology and Pathology
12:28 Colorectal Cancer Prognostication
19:34 Pan-Cancer Xenograft Repository
25:16 ChatGPT and Pathology Survey
30:55 Synthetic Image Generation in Pathology
36:35 Upcoming Conferences and Courses
42:27 Closing Remarks and Future Plans

THE ABSTRACTS WE COVERED TODAY

📄 Advances in Bone Marrow Evaluation
https://pubmed.ncbi.nlm.nih.gov/39089749/

📄 Digital Imaging and AI Pre-classification in Hematology
https://pubmed.ncbi.nlm.nih.gov/39089746/

📄 Evaluation of CD3 and CD8 T-Cell Immunohistochemistry for Prognostication and Prediction of Benefit From Adjuvant Chemotherapy in Early-Stage Colorectal Cancer Within the QUASAR Trial
https://pubmed.ncbi.nlm.nih.gov/39083705/

📄 A Pan-Cancer Patient-Derived Xenograft Histology Image Repository with Genomic and Pathologic Annotations Enables Deep Learning Analysis
A survey analysis of the adoption of large language models among pathologists
https://pubmed.ncbi.nlm.nih.gov/39082680/

📄 Clinical-Grade Validation of an Autofluorescence Virtual Staining System with Human Experts and a Deep Learning System for Prostate Cancer

Transcript

0:00

Welcome my digital pathology trailblazers. I am broadcasting to you today from Poland. I already came here last week. And I keep troubleshooting. I'm always dreaming of live streams from nature where I discussed papers from the forest. But. Traveling. Is just disruptive to the routine. So what happened on Wednesday? As I had a live stream and my internet connection. Wasn't that great. So I got kicked out twice of the live stream. This is a livestream with Bianca call-ins. She's the organizer of a conference, actually over summit digital diagnostic summit that I'm going to. And I wanted her to join me and tell you what this summit is about. So the recording is on YouTube, but basically I was kicked out twice of this broadcast. But then I managed to come back and she had a discount called. The discount code is. Summit tan. On the Wednesday, she only had 10 spots left. So I don't know if there's any spot left, but if you happen to join this conference and by the way, you can get some more information about that at digital diagnostic, summit.com. Then we're going to see each other in person. And for everyone who uses this code, I'm going to bring one of my books, a hardcover version of digital pathology. One-on-one all you need to know. To start and continue your digital pathology journey. And I will sign it for you. So if we meet there, you're going to get a signed copy of my book. If you use the discount code. So after that live stream, I thought, okay, now I need to go to a place where there's better internet and my cousin has an office. And he has a fiber optic. So I went to my cousin's place to record this very digital digest, number seven and. The internet was great. But I forgot about the. Time change and they scheduled this livestream. So in Poland it was 12:00 PM. I scheduled this and usually in the U S it's 6:00 AM. I scheduled it for 12:00 PM us. So that would be. Afternoon here, which I cannot make a anyway. I confused the time and I sent everybody a recording. But here it goes the audio version as well for you. 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. Good morning. Good morning so much. Uh, I had trouble joining today and so I hope you didn't go home because it's already seven minutes past my, um, my normal time. I think some of the platforms were not synchronized because I'm in Poland now, so it's actually not 6 a. m. anymore. Uh, but It's twelve, twelve, oh seven. So I see people joining. Whenever you're here, please let me know where you are dialing in from. Today, I'm in Rudnica, Poland. Uh, if you were on the live stream on Wednesday, with Bianca Collins from Luma, you can see that I have a different background here. And, uh, I am now in Luma. And my cousin's place. I went to my cousin's place. He has better internet. And this is where I'm going to be taking my live stream from right now. So whenever you're here, please let me know where you are dialing in from. And, um, It may be that we're not going to be so many today because of the, um, because of what happened when I started. So I wasn't there. It's, uh, you had to wait for me. So I don't know if people actually waited. Um, we're going to do it anyway. Uh, if you're here live or viewing the live stream, let me know, uh, where you're dialing in from. Let me know if you hear me. I have my mobile set up right now. So, um, I'm going to say hi. Uh, in the comments. Okay, we have people joining. Thank you so much, Sabino from Portugal. No, from Peru. The last name is Portugal. Peru. Amazing. Amazing. Okay. Thank you so much for joining. So, uh, we were supposed to have guests and we were supposed to have guests from undermined, but Um, they are in California and we decided to do a webinar instead of a live stream because last week when we were, when I was trying to compare PubMed and Undermine, this new tool for literature research, I don't know, I don't think I did it justice. Nor did I do justice to PubMed because, uh, what you see here on the screen, this is, um, this is from PubMed, right? I get these PubMed alerts, which I want to keep getting. Um, ah, by the way, let me know if the mic is okay. Because it's my travel mic, so PubMed alerts, right? And there's this other tool for literature research called Undermined. It is AI powered, so it basically like scores the papers according to relevance and not just for keywords. So it's like a different way of searching. And I was supposed to have the founders on the live stream today, but they have pretty cool plans. They're, um, they're meeting with, um, With top level AI companies, uh, to talk about this tool. And so they're going to join us for a webinar. And when I think on the 21st of August, and by the way, if somebody enters and then exits, just ignore, uh, because I'm renting slash borrowing this, uh, space from my cousin. Okay. So without further ado, let's jump into our. into our papers, into our papers. And, um, they're actually, um, cytology or like not necessarily tissue heavy today, at least the first couple of them. So advances in bone marrow evaluation. It's a, um, it's a review. So the abstract is just telling us that digital pathology has the potential to revolutionize bone marrow. Assessment for implementation of artificial intelligence, which I think is totally the case because if you know anything about bone marrow evaluation, you have like a certain number of cells and my clinical pathologists who are on the line, feel free to to put in the comments how many of which type of cells you need to evaluate. But basically you have to count a lot. You have to go really high magnification and you have to count all the different types of cells that you can see in the bone marrow and then you give a diagnosis for whatever is happening there. So, um, But, and the advantage for, uh, image analysis or for AI on bone marrow is that if you have a good smear, you should have those cells separated. Uh oh, and thank you so much for the likes. Uh, I see a, like from Facebook, if you can, uh, give this live stream if you're here live, uh, or if you're watching the replay. And if you could give it a, like, um, that helps us bring this thing. Two more people. So going back to the paper, um, to the review or to the topic, let's, let's say this review is gonna set us for the topic. So there is the digital pathology has the potential to revolutionize bone marrow assessment through implementation of AI. Um, and this review goes into barriers towards Implementation, uh, which there are enough, um, and, um, I don't know if we always have this, like, conflict of interest, I mean, in every paper, but I did not notice it in our abstracts before, but basically, um, we have here, okay, uh, the authors, which, who are, by the way, we have to say, I'm Um, it's Joshua Le Lewis and Olga Kova, uh, Brigham and Women's Hospital, university of Pennsylvania. And, um, they are consults for, for Opio and Sysmex who are active in the space of evaluation, um, of, uh, of smears, of, of blood, uh, cells of. The non tissue part of digital pathologies, and that takes us to the next, uh, next paper, digital imaging, uh, and AI pre classification in hematology. So hematology is going to be blood, it's going to be the same cells that are in the bone marrow, um, maybe not the same cells because the in bone marrow, you're going to have like the precursors, the early cells that later, get released, and I see some more people joining. Thank you so much. Thank you so much for joining and thank you so much for liking this, um, this live stream. So what happens here, and they used AI and, uh, they, by they, I mean, uh, Bowers et al, Bowers and Nakashima. And this is Cleveland Clinic. And look aside differential of peripheral blood. So again, a differential is like you, um, talk about all the different types of cells that are there and you count them. And then you maybe do rash ratios and, um, things like that. I'm an anatomic pathologist. So, uh, apologies for. Not knowing the exact details, but definitely there is an application. So leukocyte differential of peripheral blood can be performed using digital imaging coupled with cellular pre classification by artificial neural networks. You can even do platelets and erythrocyte morphology. You can assess it and count it. And, um, there was a system used. Um, and we have another conflict of interest. Uh, somebody is a paid speaker for Sysmex. So I anticipate this is the company that provided it. Um, but basically this system and you can train your own, uh, model, whichever software you're using. Uh, you can buy a system that already has that, this, um, but basically here. Um, the message is the system performs comparably to traditional manual optical microscopy, which is kind of our gold standard. Uh, a pathologist looks at something, recognizes, and this is, uh, what our ground truth gold standard is. And then you check, okay, does the automated tool also do that? Is it comparable? Um, and they say, yes, it was comparable, uh, but. It is designed and intended to be operated by a trained morphologist. So, um, it's like you need to know when it's wrong and then basically visually discard what it's doing. Um, but that's in general for all type of image analysis. Um, unless maybe those, those, those are going to be so. Um, mature that you can just let them run and flag you some stuff for purification. Um, but basically they are mostly designed for people who know what they are looking for. Um, but need more speed, need to be accelerated, need to have more quality. So, Um, the next one is actually in tissue because this is for, um, for early stage colorectal cancer. I'll take red here. And we're evaluating CD3 and CD8 T cells, uh, immunohistochemistry, so they're stained with IHC, for prognostication and prediction. of benefit from adjuvant chemotherapy and when I looked at this title like, why are they doing this again? Why do I, well, I said again because I didn't Um, read this part prediction, uh, sorry, prognostication and prediction of benefit from adjuvant therapy. Um, because I knew the work and this is, um, the group of Jerome Gallon, um, invented or discovered or, um, put together something called immunoscore and where basically you look, uh, into the numbers of CD3 and CD8 cells and you did. determine, okay, is the prognosis, uh, going to be better or worse? Um, so what this group did, they, they took those numbers and they were checking if there's a benefit from adjuvant chemotherapy. So this, this is Williams et al. And this group is from Leeds and also Rush Diagnostics in different places, Arizona, California. Um, they did it together and So the first part, high densities of tumor infiltrating CD3 and CD8 T cells are associated with superior prognosis in colorectal cancer. So if you have a lot of these. You, you're a better off and they are superior prognosis in colorectal cancer. But do they mean something when you would evaluate if you would benefit more from adjuvant chemotherapy? They didn't know. So that's what they decided to analyze. They, um, used tissues from 868 patients in the quasar trial. Um, this was adjuvant. Fluora, flu, war, war, you Russell for linic acid, um, versus observation in stage two and three colorectal cancer. Um, and it was analyzed by CD three and CD eight immunistic chemistry and pathologists assisted by artificial intelligence. And here we have this theme. Pathologists were assisted by I, um, they calculated CD three and CD eight cell densities. So how are you going to like, Calculate densities, and you probably like visually or manually, you have, you would know the area of your field of view, and then you count the cells and you get the density number of cells here. They have, uh, Uh, over a millimeter square. So, um, if you have an AI assistant, then, Oh, we have more people. Let me know whoever is joining, whoever is joining midstream. Let me know where you're tuning in from. That would be fantastic if you would comment and let me know where you're tuning in from. And it can be that most of the people are going to be watching the live stream. So if you're watching the live stream, like, and comment as well. And that helps to bring this live stream to more people. Who might be interested in this topic and don't really have a similar resource, uh, talking about those abstracts. So going back to the paper and the densities, uh, number of cells over the, um, uh, unit area, uh, and in several places, uh, core tumor CT and invasive margin. What's the invasive margin, invasive margin. Let's say this is my tissue. Um, And, uh, wait, let me, let me do it differently. I need a, I need, my mouse, my mouse is going to be my tumor. So, a mouse is the tumor, and my hand is the tissue. And here is where the tumor ends. And this is the tumor. Exactly. This place is invasive margin. So it's, um, it said and shown that if there are a lot of, uh, cells there that they're actively fighting the cells, immune cells. So in this case, uh, CD eight CD three, they are actively fighting, um, the tumor or, you know, there are different theories. If there are too many, then they cannot get in and, um, all these things. Yeah. But it's important to check. Um, so let's take my tumor, uh, the numbers both in the invasive margin and in the tumor itself, because if they went into the tumor, then maybe they're more active. So, um, and, um, the outcome that they were, uh, using for the study was primary outcome was recurrence free interval. And the results were, um, Well, in the recurrence rate in the high risk group was twice that in the low risk group for all measures. Um, CD3 CT, uh, and CD8 CT in the middle of the tumor, CD8 invasive margin. Um, don't they have CD3 invasive margin? I know, basically they say, uh, if you are high risk, then you're going to have twice. uh, the recurrence than low risk. That's like logical, but I had to read this abstract twice. Um, I had to read it twice because this was kind of like, uh, we already know that, right? If there's, uh, if their high risk group was twice that the, uh, the recurrence rate was twice than in the low risk, but what about the Uh, chemotherapy. Well, recurrence rate in the high risk groups are twice those in the low risk groups. We already said that. This is the conclusion. And proportional reductions with chemotherapy are similar. So they didn't really come up with like a number of those cells that would guide chemotherapy, but, um, there are some numbers necessary to treat. Um, so, so you might think, okay, those based on the numbers of the C3 and C8, those who are high risk, so have less of those cells and might benefit from chemotherapy more because they need, uh, additional support. Their immune system is not that great because there are not so many of these cells, but in general, like, so I guess, um, If we say the high risk group with less cells, they would be candidate, but I had to actually like read it twice. Okay. The next one is, uh, pan cancer patient derived xenograft histology in the repository with genomic and pathologic annotations enables deep learning analysis. So, um, this is a cool thing because, so pan cancer patient derived xenograft histology. Okay. image repository. So patient derived. And this is white at all. And this is a group from Jackson's laboratory. Um, the main one I think is in Maine. I once, well, I didn't visit the lab, but I was there in front of their lab. And I think I was taking selfies because, um, they work, uh, they do a lot of mice work and I'm a veterinary pathologist. So of course that was like something cool for me. And also Singapore, Frederick, Maryland. Leidos biomedical research. I live close to Frederick Maryland in the US when I'm not in Poland, Utah. Salt Lake City, Bethesda, the National Cancer Institute and California. So those patient derived xenografts, they call them PD x's. Uh, I like to just spell everything out. Um, model human intra and intratumoral heterogeneity. So, uh, in the context of the intact. Tissue of immunocompromised mice. So you inject, uh, some tumor cells from people to immunocompromised mice, and they grow a piece of this tumor on the mouse and, um, what they have here, they developed, and then you can do all kinds of research on this without having to do this research. Um, well, you are not able to do this research probably otherwise, unless you do cell cultures. Um, anyway, so you can, you can inject it to the mice and they developed an extensive pan cancer repository over 1000 of those, uh, xenografts. Paired with parental tumor. So the original tumor was there. Um, the H and E image, the, uh, the H and E image of the xenograft, and not only this, they had also associated genomic and transcriptomics data, clinical metadata, pathologic assessment of patients. cell composition and in several cases they have detailed pathologic annotations of neoplastic, stromal, and necrotic regions. So, um, so they have this super huge fantastic repository and what do they want to do with it? Uh, I mean, It needs to be useful for something, right? So they had a couple of, um, like applications. Applications number one was the development of a classifier. Develop of a classifier for neoplastic, stromal, and necrotic regions. So they had annotations, uh, for some of them. And then they could develop a classifier. So, uh, why is this classifier useful? important because you want to analyze things in certain regions. I've seen image analysis projects where there was a bunch of stuff quantified and like classified, quantified, and like extensively described. in the necrotic region of, uh, the cancer, in the necrotic meaning dead. So that was irrelevant. So if you can classify, uh, the neoplastics of the epithelial neoplastic cells, stromal and necrotic regions, then you can run your analysis in the corresponding region. Then they wanted Um, predict like development develop a predictor of xenograft transplant lymphoproliferative disorder. I guess that's in the disorder that can, um, that can happen. Um, so they wanted to predict it and, um, then application of published predictor of microsatellite instability. So this This is a genetic, uh, alteration genetic mutation, and they want to check if they can predict this particular mutation on this data set. Um, yeah. That's what they did. So, uh, significance, a pan cancer repository of over a hundred, sorry, thousand patient derived xenografts, H& E stained images will facilitate cancer biology investigations through histopathologic analysis. Yes, I agree. And because we have a lot of people from the government, look how long this conflict of interest, um, So we have government and we have Roche. So we have a, um, looks like government slash company. I don't know if it's a consortium collaboration, but look, the abstract is, this is the abstract and this is conflict of interest, like a long, long, so they have to disclose everything. They have to disclose everything and I picked one for you. Let me find it. It's actually at the end because it's very, uh, it's an interesting one about synthetic images. I'm like, oh, synthetic images. Are we going to be using synthetic images? What are we going to be using it for? I know, before, before synthetic images, actually one more. Uh, this is a survey about Chad GPT, our friend, our best friend Chad GPT. Um, what happened in this survey? So, I took part in this survey. This is a survey by, um, actually by, et. Um, but the author I know is Andrea Boff, so I don't dunno if you've seen, uh, look at some other live streams or my previous videos from this year, like around March. Uh, he published this on a poster at US Cap and he gave this poster to me and I had it as the background. Uh, so, uh, check that poster in other videos. But, um, I took part in the survey. I don't know if I was actually. counted because he mainly interviewed MD pathologists, but I think there was a specialty. I think he went for all pathologists. Um, anyway, I answered the questions that he had. And, and, uh, that was pretty early after their release of, uh, chat GPT, I think within one year. So, uh, he was checking, they were checking this group. Uh, they sent a Google, uh, Google doc survey, ask people questions and they had over. Uh, they had 215 responders, so 215 people took their survey, fantastic. I saw it on LinkedIn and I immediately took it. Uh, and they wanted to check the perception of large language models applications among pathologists. So half of them, almost, 46. 5 percent reported using them, especially chat GPT, for rational purposes. Predominantly for information retrieval, which at that point was maybe not the best use case. Um, But because it was now there are applications of information retrieval. Um, when you look at the web pages on Google or wherever when the first paragraph now is generated paragraph that answers your question and then below you have the websites where the, um, generative large language model took the information from so it actually retrieved information. This is information retrieval. Then they used it for proofreading, academic writing, and drafting pathology reports. Cool. And they highlighted significant time saving benefit. This is a huge time saving benefit for me when I use those tools. And academia, academic pathologists, demonstrated a better level of understanding of LLMs than their peers. Um, and they were considered moderately proficient concerning pathology specific, sorry, the LLMs. Um, they were, sometimes they can provide incorrect general domain information, but they were considered moderately proficient concerning pathology specific knowledge. Cool. And it was mainly used for drafting educational materials and programming tasks. Concerns were like, not unlike the rest of the society, information, accuracy, privacy, and need for regulatory approval. So I don't know why I need for regulatory approval. I don't know. I guess it's a powerful tool. So you don't, I mean, I know, but would I want to vote for this to be regulated at the moment? I don't know. So anyway, that was the concern from the survey. And, um, but the conclusion was that the large language model applications are gaining notable acceptance among pathologists, specifically that this was less than a year after the tool was introduced to the market. So that was a super cool survey that I took part. In and like I have the poster in Pennsylvania. So now our last one, uh, and the question about sharing those articles. So there is a question from the audience. Uh, if I can share those articles, yes, the articles, uh, are going to be shared. Okay. I don't, yeah, I can show it. Okay. So the articles are always shared when, um, the live stream has, uh, then a replay specifically on YouTube. I go in and I comment, uh, in the first comment, you're going to find the links to all the articles and they are also shared. Uh, And after this live stream, Oh, I see some more people joining after the live stream, I share them in a, an email blast. So if you are on my email list, then you're going to get it every week. You're going to get all the papers, um, meaning with all the links to wherever, um, I think I do PubMed or whatever is like most accessible. And if you are not on my list. Then, if you go to my webpage, digitalpathologyplace. com, digitalpathologyplace. com, there is going to be a button, uh, I could share it maybe? I'll share it later after we Let me try it. There's going to be a button to download an ebook, and once you get that ebook, I get your email address, and then you get all the, you know, videos, information, digital pathology information, uh, papers, uh, that we are discussing. And digitalpathologyplace. com. That's where it's going to be. So let's talk about this synthetic image generation. That's like, hmm, we're going to be generating synthetic pathology data. How accurate is it? Why would you want to do it? So, uh, I'm familiar or, well, I'm familiar with this as well because I read it. But, um, something that kind of was less. Sounded less dangerous for me was that synthetic generation of tabular data and like, um, I don't know, cholesterol values, blood pressure values, everything. That's a tabular. That's a number. You can generate these based on, you know, your population data. You can generate fake ones that are going to be as valuable as the other ones because these are just numbers. But what about images? So here, Okay. What happened is there was a synthetic genital urinary image synthesis via generative adversarial networks. This is a type of, um, AI architecture that can generate images. And that, um, it's actually a system of two networks where one is generating, the other one is checking if this one, if the first one, uh, made those images. In a way that they resemble reality so the second one the second one that the checker knows how reality should look and Generator doesn't know but gets feedback from the checker all the time And if it get when it gets enough feedback, then it's generating images that I Actually are indistinguishable from real images, and that was, um, checked for natural images for artwork and different things like that. Um, and now we're starting to use it. This is also the principle for virtual stains, and now we're using it. For let's see what you were using it for so, um, it says artificial intelligence diagnostic precision But I didn't really find anything about diagnostics here. So this is fun bovin at all Where are we in Miami and oh in Hungary? Groups from Miami and Hungary, Florida, and, okay, so what happened here is, uh, this particular one, which journal is it, Journal of Personalized Medicine, so their problem is, it's the problem in the digital pathologies, in the medical imaging space, um, um, that there is a scarcity and restricted diversity of genital urinary tissue data sets and it poses significant challenges for training robust diagnostic models. And, uh, so what did they take? They take, they took those guns and they started generating high quality synthetic images of rare, underrepresented animals. Tissues so not really like entities or different cancers or different diseases, but just the tissues and and they have put hypothesized that augmenting training data with these gun generated images. And that would significantly enhance model performance for tissue classification, segmentation. Disease detection, so OK, they wanted for disease detection, but they do the. GU tissues, and they let the gun produce synthetic images of eight different, uh, genitourinary tissues. And what did they use to check if those tissues that, uh, if the images of those tissues generated by the gun, uh, look good or good, um, they use the score that I am not familiar with. It's called relative inception score. I assume it's a computer science way to checking. And if those images resemble the real images for significantly, and there is fresh at inception distance as well. But what I do understand is that additionally, the synthetic inputs received an 80 percent approval rating from board certified pathologists. This is like, this is familiar for me. A pathologist looks, says it's okay. And they're 80 percent of the time, I assume, or 80 percent of those images were approved by a. um, by a pathologist. And, um, so this study not only confirms the feasibility of using GANs for the augmentation in medical image analysis, Um, but also highlights the critical role of synthetic data and this addresses the challenge of data set scarcity and imbalance. So that the fake data I called fake data. It's synthetic. Synthetic is a good word. Synthetic data generation is entering the histopathology space. I mean, it was already there with, um, I knew this from, uh, virtual stainings. Where you, where you generate, um, one type of image from another type of image. Um, but here you're generating images from scratch, new, uh, images that then you're gonna have a model trained on. This is interesting. And, and these are our papers for today. I have a couple of updates for you today. So, um, I don't know if on Wednesday you were on the live stream and my live stream got, well, it didn't get interrupted, but I got kicked out of the live stream twice because my internet in my Polish home is suboptimal. And I was trying to stream like I do in the U. S. on all different platforms and on TikTok as well. This didn't work today. I didn't go live on TikTok. So any of the TikTokers watches it on YouTube or somewhere else, then, um, I apologize. But feel free to go watch the recording because I was talking about the next conference I'm going to and this conference is Digital Diagnostic Summit. And the website of this conference is also digitaldiagnosticsummit. com. At the time of the live stream, which was two days ago, Bianca, my guest, who is also the organizer, she's the marketing head person at Lumea, she has 10, 000, Spots still for the conference and when is this conference good for very much for networking and moving the needle in the digital pathology space and also, uh, kind of living an adventure because the lectures are only till noon and then the networking slash adventure part begins like a horse riding quads or it's a different type of meeting. So definitely, if you're trying to. Influence something in that space, uh, be it, uh, with a product, with your knowledge, with your institution. Um, it's gonna have the movers and the shakers of the industry, and I'm gonna be there as well, and I'm gonna be streaming from there as well. Uh, if you're on the email list, uh, then you're gonna get all the updates. If you're not, go to digital pathology place.com to get the book and be on the list. Other things that I wanted to tell you was, okay, I was talking about, um, my YouTube course. Uh, so a course that has a curriculum, has a structured curriculum, uh, but consists of my YouTube videos. So there's going to be an option for this, well, for this course, all those videos are going to be free on YouTube for everyone who, uh, is not, uh, Doesn't want to pay for this. It's going to be for free and but the course and I already have the list of all the videos I am anticipating this may be happening in September when I'm back in the US and I have all the videos there is over Definitely over 300. I don't have 400 yet, uh, on YouTube, but basically it's gonna, I'm going to structure it into a logical curriculum, so you're not going to be bombarded with ads and you're not going to be, uh, having to, um, like navigate through playlists or things like that. It's going to be structured for you with a login. In the course, uh, like course environment where you're going to be able to attract your progress and there's going to be a certificate at the end. Um, so that's going to be one thing that I'm working on and I decided I'm going to, uh, improve the curriculum in the. Current membership, the digital pathology club with, um, some videos from YouTube, but a highly, uh, relevant vetted experts and also in form of a structured curriculum that takes you from A to Z of digital pathology and from, from the basics of, uh, you know, tissue processing to advanced things, including maybe some literature research. So my own content is going to be enriched with YouTube content. And again, this YouTube content is available on YouTube as well for everybody who doesn't want to pay for it, but it's going to be structured as the curriculum, um, curated and vetted for you in the paid option. So, you know, if. If money is tight, then you don't do it, if time is tight, then you do it, right? We pay for things either with time or money, and there's nothing free in this world, which is okay. That's how it works. But what inspired me to do it was the lecture I found about large language models. Wait, no, large language models and retrieval augmented generation. No, no, no, not large language models. Sorry, um, foundation models by Hamid Tissouche. Uh, so he basically in his lab, Kenya lab, put this lecture that he gave at the, um, Association of Pathology Informatics Summit. And I'm like, this lecture only has 300 views. This like this needs to be brought to other people. So of course I shared on social media, but then I thought, how many of these things that are like at conferences and somehow they got the permissions to publish it on YouTube. are just buried, uh, in the tons of YouTube videos that, uh, that are being published every day. So what I'm gonna do, I'm gonna go in and hand pick them, uh, check if they're really, like, given by experts, uh, check all the references and find them for you and include them in the membership. Full disclosure, they're from YouTube. You don't want to pay for this curation. You can find them on YouTube on your own. So these are the plans. I'm planning to do this, as I said, September, um, and plans for next week. We will have the live stream with papers on Friday, 6 a. m. Eastern, which is 12 noon in Poland. Uh, and I will make sure that all the platforms, uh, are synchronized and that we don't have a problem, uh, that they are scheduled for a different time. And, um, mid-August, we're gonna have a webinar with the founders of undermined this other literature research tool that is AI powered, that can be. An amazing complimentary tool for PubMed, or maybe it can replace PubMed. I'm gonna try to ask some challenging questions to the founders to say, Hey, is it better than PubMed? Is it different than PubMed? Why do we even need it? Why should we care? And I already have a free account. It's undermine. ai. And you can, you can have a free account there. And thank you so much for joining me. Thank you so much for, for saying hi, letting me know where you're tuning in from. If you are watching the, we're watching the, uh, replay, feel free to give it a like, uh, and let me know where you're tuning in from, because even when it's not live anymore, it's going to be shown to people. And I would very, very much appreciate if you could help me show it to more people. And I talk to you in the next episode. Just a very quick ask before we leave. If you're listening on apple podcasts and you love this podcast. Would you be so kind to give us a five star reviews? I am aiming to get 55 star reviews until the end of the year. And your help would mean a lot to me. Thank you so much.