102: DigiPath Digest #10 (From HER2 Low to Gleason Grading: AI in Digital Pathology Research)

Welcome to the 10th edition of the DigiPath Digest. Today, we discuss essential updates including the free availability of my 'Digital Pathology 101' book and the podcast now accessible on YouTube and YouTube Music. We dive deep into the weekly abstract, focusing on advancements such as sex-specific histopathological models for gliomas, leukocyte identification tools, and automated Gleason grading for prostate cancer. We also explore the potential of SciSpace, an AI tool for interacting with scientific papers. Interspersed with live interaction, we discuss the importance of consistency in histopathological grading and the challenges faced by pathologists. J

00:00 Introduction and Announcements
00:55 Live Interaction and Updates
05:01 Abstract Review: High-Grade Gliomas
11:45 Abstract Review: Leukocyte Identification Tool
13:24 Abstract Review: Gleason Grading in Prostate Cancer
16:31 Abstract Review: HER2 Low Prediction in Breast Cancer
24:01 Event Announcements and Closing Remarks

THIS EPISODES RESOURCES:

📰 Sexually dimorphic computational histopathological signatures prognostic of overall survival in high-grade gliomas via deep learning
🔗https://pubmed.ncbi.nlm.nih.gov/39178259/

📰 A Digital Tool Supporting Pathology Practice and Identifying Leucocytes
🔗https://pubmed.ncbi.nlm.nih.gov/39176939/

📰 Assessing the Performance of Deep Learning for Automated Gleason Grading in Prostate Cancer
🔗https://pubmed.ncbi.nlm.nih.gov/39176576/

📰 Weakly-supervised deep learning models enable HER2-low prediction from H &E stained slides
🔗https://pubmed.ncbi.nlm.nih.gov/39160593/

▶️ YouTube Version of this Episode:
🔗 https://www.youtube.com/live/06QXmwojxDE?si=q59PjGkHbXCUFhwI

📕 Digital Pathology 101 E-book
🔗https://digitalpathology.club/digital-pathology-beginners-guide-notification

Show less

Transcript

0:00

Welcome my digital pathology trailblazers. This is the 10th edition of the digit pap digest. Thank you so much for following the serious and regularly showing up. Today, we're going to start with a few announcements. Then we're going to discuss the abstract as every week. And there's one tool that they want to share with you. That's going to help your independent, exploration of papers as well. AI base for me, AI is leverage. So I wanna show you this tool as well. 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. Welcome, welcome digital pathology trailblazers. Today is the 10th DigiPath Digest. We are live, supposed to be live on through live stream. If somebody registered and I see you guys coming in, say hi and let me know where you're tuning in from. We should be going live on Instagram. And yeah, so great to have you. Hey, hello. Welcome India. I see hearts liking, everybody is liking. Fantastic. Whenever you join, let me know where you're tuning in from. And let's start with a few announcements. Okay, so the book, right? I have a book, it's called Digital Pathology one-on-one. And you know how many people, downloaded this book? It's a free BDF on my website. So if you wanna get it, you go to digital pathology place.com and then there's gonna be like a button, you click it and you get the book. 1,600. 93 people downloaded the book. So if you have not downloaded it yet, it's right there, available for you. Guys, it is the 10th DigiPath Digest. We said at the beginning that if I get to 10, if I stay consistent, I get to 10. I'm gonna get myself, a new tablet to reviewing the papers. Currently I'm working on a Lenovo tablet, which is fine, you know, you can draw on it and it's fantastic. I was thinking of a Wacom or something else. We'll see next week. I'm not gonna be here. This is the last time I'm, broadcasting, to you. From this office of my cousin in Poland, because next week on this day, on our broadcast date, I'm going to be traveling back to the U S so, and I see more people joining, let me know where you're tuning in from, say hi and let me know where you're tuning in from. So last time from Poland for this particular time, I'm going to be going back to Poland and this is the beauty of digital pathology that I can actually do it. I can download my digital slides and work from here. So. That's one update we did, the first part of every endeavor for me you showed up every single day for this DigiPath digest, already nine times today, the 10 times and I see people coming in. So we show up, right? I show up, I have our, abstracts and everything. We are past the hurdle of showing up. I'd like to believe the next thing is, to get better, to get better for me specifically. What does that mean? That, I would, I will be preparing more for this. I'm going to try to make this content better. I'm going to dive, deeper into the, papers that were into the abstracts that we're reviewing. And for you, it's probably going to stay in the same form, but I will be more informed about the paper because I will research everything that's not so clear from the abstract. Welcome, Monica. Hello to Edinburgh. Monica is a regular and a guest on my podcast. So if you want to listen, what she has to say about digital pathology in clinical trials, it's also on the digital pathology place website. For me, stepping up the game, past showing up is going to be, getting better in reviewing those papers for you. And, I'm gonna, because the. Those episodes are also on the podcast and a quick update on the podcast. Together with my team, we just submitted the podcast to YouTube as audio as well. Google podcasts, doesn't have the podcast option anymore. They integrated into YouTube music, which is a little confusing because those audio episodes show up. on normal YouTube as well, people sometimes just, listen to stuff. And those who were listening through Google podcasts now have this available for YouTube music. We have this podcast available for YouTube music as well. I'm going to keep Putting it out as an audio podcast only. And leverage AI to edit my ums clumsiness, and the ands that I overuse. There is a tool that I'm going to show you later, but let's dive into the abstract. Oh, we have guests from Pakistan. Hello. Welcome Pakistan. Anybody from the U S today? If somebody joins from the U. S., let me know that you're here at 6 a. m. EST. So let's dive into our first paper and then, based on this paper, I'm going to show you a cool tool. And thank you so much for the likes, for the hearts. Whenever you do that, it takes us to more people who might want to benefit from this. Welcome Instagram. Oh my goodness, Robert, you're here. I missed you last time. I didn't see you last time. Robert is joining from Boston this time, not from Maine. The first paper, and I have them all marked in green, the ones I want to, talk to you about. Sexually dimorphic computational histopathological signature prognostic of overall survival in high grade gliomas via deep learning. This, is from, science advances, and I checked the impact factor is 11, decent impact factor, Ruchika Verma et al. This is from a group from New York, Mount Sinai, and also Cleveland Clinic, Cleveland, Ohio. Here they say, High grade glioma is an aggressive brain tumor. Yes, it is. And sex differentially affects survival outcomes. They used an end to end deep learning approach of a H and E, H and E, scans, and they identify sex specific histopathological attribute of the tumor microenvironment, and created sex specific risk profiles to prognosticate overall survival, they had a two stage approach using ResNet 18. A deep neural net for image analysis. First, they segmented the viable tumor region. Second, they built a sex specific prognostic model for prediction of overall survival. And they, used the C index, concordance index, as their performance metric. It was, all around 0. 7, both for females and for males, more or less 0. 7, C index, across training and three independent validation cohorts. So for all the cohorts, they had a separate C index. And, what does that mean? It means that, and to end the deep learning approaches using routine H and E slides trained separately on male and female patients with high grade glioma may allow for identifying sex specific histopathological attributes of the, tumor microenvironment associated with survival and ultimately build patient centric prognostic risk assessment models. So everybody wants to have this, Specific to patients, right? Whether it be specific to female versus male, different demographics, different race, different patient cohorts. They are not, all the same. Here they did for male and female, but what I wanted to show you is this new tool. So we have this science advances and, that tool that I'm talking about is called, size space. And I already installed it as a plug in for chrome. I just want to, and in the short it lets you chat. With this paper. We just reviewed the abstracts, and when I was reviewing it, I'm like, okay, what is this C index? It works very well. If you have the full paper available, because if you only have the abstract, it's just going to base the answers on the abstracts. And what is it? It's Like an AI chat bot to chat with the content of the paper. I'm just checking. We have, the full, paper available. And, let's, let's do this. Like, what does C index, because we have always different metrics in those papers. When I prepare it I'm like, I need to dive deeper into those papers so that you don't have to, you come out of this session with actionable information without having to go and Google stuff. Understanding C index. The C index or concordance index measures how well a model's predicted risk scores correlate with actual survival times. It ranges from zero to one. One indicates perfect prediction and 0. 5 suggests no prediction. Um, The C index in this paper, good. Is it good that it's 0. 7? Like, is it a good thing? So, the C index values reported in this paper are generally good, indicating effective model performance for the female cohort 0. 696 to 0. 736. A C index above 0. 7 is typically considered acceptable, suggesting that the models can reliably differentiate between low and high risk patients. I want to know, what features they, identified in this tumor microenvironment, that were specific for males and specific for females. So let's ask it. What histomorphological tumor microenvironment features were specific for males? Males versus females, it's just makes life so much easier whenever you, read papers. Here we have in males, high risk patches were associated with micro proliferation, pseudopalysating cells, and in females, high risk patches were linked to infiltrating tumors, infiltrating lymphocytes and MVP, which is the microvascular proliferation low risk patches in males were related to the peritumoral region, where in females they were associated with stromal regions. If this is something you're interested in, it's called SciSpace. I have a free account, so you can sign up for a free account. If SciSpace decides to work with Digital Pathology Placed, then I'm going to bring you a better, more, in depth tutorial it was a cool way to interact with a paper leveraging AI. For me, AI is leverage. Let's go back to our normal papers. If you have any questions, let me know in the chat. Let me know in the comments. I see comments mostly from LinkedIn here. If you are on YouTube or somewhere else, let me know that you're, doing it there. Our next paper is a digital tool supporting pathology practice and identifying leukocytes. So that was an interesting one. It's in a journal called Study of Health and Technological Information. I checked this one, but this one was impact factor less than one. So, and there, I think they might have had a serious of, different papers on digital pathology because I have a bunch, we're gonna look at this digital tool supporting, pathology for, leukocytes. Dimitru Christian Apostol et al. from Romania. Created, an all in one package that contains both the part where pathologists can manipulate the data as well as predefined models. This was supposed to, provide the numbers of leukocytes and the mass of the cell, which is interesting. Only from the image. With minimal intervention from the pathologists, they did this, as a prototype, as a proof of concept, and, the models correctly identified 89 percent of the actual positive instances, this tool provides visual scripting, reducing the learning curve for pathology analysis techniques, visual scripting is a kind of programming language, with pictures, not, coding. I would want to have it even more streamlined than visual scripting. I would like to have a user interface, and click on stuff, rather than the visual scripting. You can use this tool. It was a proof of concept, for, lymphocyte quantification. Our next paper is, assessing the performance of deep learning for automated Gleason grading in prostate cancer. We're talking not just like prostate cancer detection, which there is an FDA cleared, algorithm for that, but now we're talking about grading. This is a group from Augsburg, Germany, Dominic Muller et al. Prostate cancer is a dominant health concern calling for advanced diagnostic tools and utilizing digital pathology and artificial intelligence. This study explores, explores potential of 11 deep neural network architectures for automated Gleason grading in prostate carcinoma, focusing on comparing traditional and recent architectures. They used standardized image classification pipeline, based on the AUCMEDI framework, and this facilitated robust evaluation using an in house data set consisting of, uh, 34 1264 annotated tissue tiles, This AUCMEDI framework is an open source software that allows fast setup of medical image classification pipelines. It's for image analysis scientists that need to set up a pipeline, to analyze images, they investigated 11, 11, deep neural network architectures, and they had varying sensitivity across architectures, and they had conv next, conv next, some convolutional, neural net that they called ConvNext, demonstrating the strongest performance. And they say newer architectures, achieved superior performance, even with challenges in differentiating closely related glories and greats. So, what do I, understand here? I'm glad that newer architectures, achieved superior performance than the older. So that means we're advancing and second, challenge in differentiating closely related Gleason grades. Am I surprised? Not at all, because they used, human ground truth to train these things. And whenever it's something difficult to visually distinguish by a pathologist, then the ground truth is going to be inconsistent. And when the ground truth is inconsistent or when it's visually very, very similar, then the algorithms have problems with differentiating. And so, yeah, I'm expecting this challenges as long as we will have ground truth from people, it's going to be only as consistent as people. The concordance between pathologists, like across studies, if we reach 0. 7, we're like, wow, we're so concordant. Usually it's less. The next thing they chose for us is super cool, on Instagram, whoever is joining from Instagram, look at the stories. There was a question, in a poll where, I asked, which paper do you want me to review? The options were, this paper, which you can see that one, weekly supervised deep learning models enable HER2- low prediction from H& E stain slides. Or the other question was, which I scrolled through the other paper was, machine learn based analysis identifies and validates serum exosomal proteomics signature for the diagnosis of colorectal cancer. We had 30 votes and 23 of those votes went for the HER2 low prediction. So, let's look at the HER2 low prediction. And this is a paper by Renan Valiris, from Sao Paulo, Brazil, et al, and this is in breast cancer research. What is the background here? Human Epidermal Growth Factor Receptor, HER2 low, breast cancer has emerged as a new subtype of tumor, for which novel antibody drug conjugates have shown beneficial effects. Assessment of HER2 requires several IHC immunohistochemistry tests with an additional in sito hybridization test, ish. And. In, if a case is classified as HER2 2+ So when you look at the IHC, you can have a different HER2 status. You can have HER2 negative which is IHC. her two low, which is IHC one plus, or IHC two-plus with a negative fish. And you can HER2 positive, which is IHC two plus with a positive fish. And IHC. Three plus. So we have this HER2 low. That is a little bit problematic, which is IHC one plus or IHC two plus with a negative fish. So, there is this 2 that is, problematic. When we have, several IHC tests, which I don't know how many, like, one, right? And then you have IHC and when it's low, you need to do ish to confirm, then you already have two molecular tests, one of them is, fluorescence, so that's another level of complexity, if there was something that could streamline this, what would it be? Prediction from H&E, right? Everybody wants that, I want that, because then, You don't have to do other tests, or you can do them in a targeted way, right? So that's what they wanted. And, what did they use? They used a self supervised, attention based, weakly supervised method. That's interesting. Self supervised, attention based, weakly supervised, right? Method to predict HER2 low directly from 1, 437 histopathological images from 1, 351 breast cancers they, built six distinct models to explore the ability of classifiers to distinguish between HER2 negative, low, and high. in different scenarios, the attention based model was used to comprehend the decision making process aimed at relevant tissue regions. Our results indicate that the effectiveness of classification models hinges on the consistency of the model. Have we heard it? Consistency and dependability of essay based tests for HER2 as the outcomes from these tests are utilized as the baseline truth for training our models. Here we have, Something that you could consider. It's not really human generated truth because you have IHC and then you have the result from the IHC. Who creates the result from the IHC? The pathologist. When there is a borderline case, is it low or negative they say negative, low and high, right? So probably between negative and low. It's tough to distinguish and. So, I actually once was at the workshop for. Pdl one scoring, even though I'm a veterinary pathologist, when I was working for a company that would help pharmaceutical companies analyze human tissues, I was doing these scorings and they wanted to train us to be more consistent on PD L1 scoring. How does that work for PD L1? Depending on which antibody you're using, there is a threshold, like a cutoff below this cutoff, it's negative and above it's positive, let's say for one of those antibodies, it's 10%. They wanted us to be, more consistent around this, 10%. And if you have access to comment, please comment. How concordant were we? Like, How many, people in the room, and it was like, I don't know, 100 to 150, would say negative and how many would say positive after, an hour of training that we were supposed to be more consistent. After this training, how concordant were we? Like, however you want to put it in the chat, let me know and I'm going to move on and answer this question. So. Where was I? Okay, so everybody wants to skip tests or have less tests or have cheaper tests. So, they want to do a computational test on HNE instead of, doing some mutations or IHC or something like that, right? So the Effectiveness hinges on the consistency and dependability of the ground truth. Make an educated guess, how many were, below and how many were above the threshold, like what the concordance was. If you have consistent ground truth, then your predictions are going to be good. It was 50%. Half of the room said, Oh no, it's negative. Half of the room said it's positive. And I'm laughing at this because it's kind of a ridiculous ask all the pathologists in the world are constantly being asked and I know I have a few pathologists on the line here. Let me know what you think about it. I know it may be like it must be, Not mainstream thought because if that was mainstream thought then we wouldn't be made to score in that way But basically all the pathologists in the world are made to guesstimate Thresholds and then you know when it's Very positive. It's easy to say that. But there's always this cutoff where you visually as a person as a human being are not capable of distinguishing whether it's 11 percent or 9 percent So around that cutoff, it's basically guesstimation and that's what they said here in this paper, even though there is, an, like, possibility to, to predict something from H& E in the correct direction, and IHC also is a method that has a lot of variables to keep it consistent, even though it's a mainstream method used widely for diagnostics, okay, so, that was it today. Let me know if you have any questions regarding what we talked about today, regarding your thoughts on, pathologist scoring, relying on percentage estimates. If you're interested in exploring, Scispace, you can do that too. I'm going to be exploring it and, telling you about it. One update. I started reaching out to people, because I want to organize an event about the immigrant doctors changing US health care. And that's going to be obviously pathology edition because we're digital pathology, trailblazers. And before I move on, I'm going to read Monica's comment. Borderline cases are always challenging. Second opinion is important and the benefit should go to the patient. Totally. My problem with this is like, okay, second opinion. If it's quantification, to me, the only way of a reliable quantification of something in the image is with computers. Most of those tests are not relying on computers. The majority, and that's also a trend that you're going to see in papers. Yeah. We should have a lot of pathologists agree. Then obviously the evidence is stronger. And like Monica says, benefit to the patient always. So the event is going to be around Thanksgiving. And it's going to be about what are immigrants doing in the American healthcare focused on pathology. I already reached out to a few speakers. If you are an immigrant doctor, regardless whether you're an immigrant in the U S or in a different country, let me know, if this would be something you would like to join, let me know because I'm gonna be putting, up a registration page together. And I wanna invite, many prominent pathology movers and shakers that are from abroad. I very much relate to this because I'm from abroad. So that's what I'm going to be doing. And, at the beginning, I already said that the podcast, this is going to be available as the podcast, it's audio only, and the audio only is also available on YouTube. So, the channel, the digital pathology place channel had, I don't know, 370 videos. Now that I added all the audio podcasts, it's like double that. And the podcast comes like from my podcast host. So it's only audio. I'm going to be sharing videos. on the email list. Now that we are 10 DigiPath Digests in, I'm gonna stop bombarding everybody who is subscribed to Digital Pathology Place, which is currently over 8, 500 Pathologists and digital pathology professionals, I'm not going to be sending you emails every time I go live. I'm going to send a survey who is interested in getting those emails that you can join live directly from your inbox. Which is something I personally use very much when I register for other webinars as convenience, but I know that people have People are busy, inboxes that are super full, if you get, three emails, about, Hey, I'm going live in 15 minutes, if you are not planning to go to this live, then it's irrelevant for you. So I'm gonna send a survey about who wants to get those emails, you're going to be hearing from me more or less once a week, about the content that I'm releasing. The educational video, the podcast that I'm doing, every now and then there's going to be something additional. There are going to be, promotions when I release new products or organize events like immigrants changing, American healthcare. Other than that. I want to be respectful of your inboxes, give you a piece of valuable content once a week. And maybe I'll send a few surveys if somebody wants to hear more often than that. I'm going to be including some information in the PS section of the emails, if there is more announcement than just once. Than just one. Thank you so much for joining me today. And I talk to you in the next episode, which is not going to be next week. It's going to be in two weeks because next week I'm traveling back to the U. S. Thank you so much for staying till the end. Thank you so much for constantly listening and supporting the show. If you have not grabbed that digital pathology one-on-one book yet it's available at digitalpathologyplace.com the moment you entered the website, you're going to see a button to download it. It's absolutely free in the digital form. There is a. An option to buy it on Amazon, the paper form as well. If you're interested. And if you are coming to the digital diagnostic summit, In September, I'm going to be there and I'm going to have a few. Books. To give away. So feel free to reach out. When you see me, I'm going to have books and I'm going to also be giving away the books from two live streams that I'm going to be. hosting from there. So I'm looking forward to seeing you there. And I talk to you in the next episode.