Digital Pathology Podcast
Digital Pathology Podcast
95: DigiPath Digest #3 (AI and Digital Pathology: Innovations in Disease Detection and Prognosis Abstract Review)
The third episode of DigiPath Digest just took place live, but I have an audio version for the listeners.
DigiPath Digest is a review of digital pathology and IA publications abstract review that I host weekly as a live stream (on YouTube, LinkedIn, Facebook etc.)
Here is the video version if you learn more visually
Today the abstracts we discussed centered around innovations in disease detection and prognosis powered by digital pathology and AI.
TIMESTAMPS:
00:00 Welcome and Introduction
00:35 DigiPath Digest Overview
01:14 Engaging with the Audience
06:09 Abstract Review: AI in Liver Fibrosis
11:21 Abstract Review: AI in Prostate Cancer
16:43 Abstract Review: AI in Glioblastoma
23:02 Abstract Review: AI in Red Blood Cell Analysis
28:38 Upcoming Events and Announcements
34:18 Closing Remarks and Future Episodes
TODAY'S ABSTRACTS & RESOURCES:
Become a Digital Pathology Trailblazer get the "Digital Pathology 101" FREE E-book and join us!
Welcome my digital pathology trailblazers. Today was the third digit path digest abstract review. And my friends, this is becoming a thing. I see people who are joining over and over again. And. It's at 6:00 AM in the us. So I know that people, well, no, I take the back because I had one person who, uh, is joining. I think he's he joined. I already like the third door, the second time. Anyway, so there are people in the us. Cool. The morning. So thank you so much. And when I was reviewing the abstracts today, there is a theme that these publications are not really from pathology journals. So I'm not going to be giving you more intro because there's intro already in the digital digest. So 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. Good morning, everyone. It's 6 a. m. in Pennsylvania, and I welcome you to the third edition of DigiPath Digest. When you log in here, let me know when you are tuning in from. Give me a hi and when Where you're where you're at right now. So, um, let me tell you a story. Of course. I have my beautiful, fantastic new camera here. That didn't work. So I'm just doing my normal, um, computer camera. But let me know where you're tuning in from. I see you're joining and let me just say hello in the comments and welcome you. Um, we have a few abstracts today. Three. Oh, Erica. So great to have you here. Hello. And Robert, you made it from Maine. Yeah. So there was a little bit of a glitch yesterday. I think I had it initially for yesterday, um, at 6. 00 PM. And, uh, it's always at 6 AM in the morning, but, um, Robert sent me a message that, Hey, is it yesterday or today it's today. And we have Jason from Baltimore. Baltimore is not that far from me. It's one and a half hours. So fantastic to have you here guys. Um, I know there's going to be more people joining, but for those who already were here on time, um, here agenda guys, I'm serious. Um, so. For those who are here, but don't know me, I'm Dr. Alex Rudaf and I'm a veterinary pathologist and the founder of Digital Pathology Place. And I'm on a mission to teach you all you need to know about digital pathology. And part of this mission is my book. If you didn't grab the book yet, there is one when you go to my website, digitalpathologyplace. com, you can download this one for free and Let's dive into one question before, um, because, uh, so now what I'm doing, I'm reviewing the abstracts that come from PubMed alerts, but I also, uh, discovered this, um, software for literature review, literature research called Undermined and I showed it to you last time, but I'm going to share it with you one more time because I want to ask you something. Step screen, unsure screen, uh, so undermined, you should see it now. So this is this, uh, research engine. So, um, what, um, last, this week I had a, uh, ergonomic training and. Welcome. Ergonomics training at work. And, uh, I was like, okay, I'm working remotely. So I do have some stuff. I have like a standing desk. I have a walking treadmill for things because when I started working from home and I have the watch and I have the ring to measure my activity, the activity goes, uh, a lot. Like it's so much less than when I was actually driving to work and I had to go from my car to somewhere anyway. And, uh, welcome Milena from the Netherlands. Great to have you here anyway. So, uh, long story short with this tool undermined, you can do research on specific topic. And I did a research on, um, pathology, digital pathology ergonomics. So if this is something you're interested in me covering, I mean, I'm going to cover this, um, At some point anyway, but if you're interested in any other topics, leave me a comment, what topics you're interested in, like specific topics, like, I don't know the scanners or, um, the AI models that are available, uh, open source or something, whatever, uh, you're interested in and hi to India. And welcome, Ahmed from Saudi Arabia. Great to have you here, guys. So any topic that you think, oh, I would need, like, I would want something specific to that. Not just what comes from PubMed every week, but just specific to that particular topic. Let me know in the comments, the topic. And without further ado, let's move to our abstracts. Need to share a different screen, which is my tablet where I can also draw, right? Yes. Okay. Perfect. So today the first topic is AI based digital pathology provides newer insights into lifestyle intervention induced fibrosis regression in MASLD. Thank you. So, um, this is liver fibrosis. I'm going to explain everything to you, but, um, something that's interesting, I'm looking where these publications come from, not only geographically. So, um, this one is from China, but what kind of journals, uh, they are in. So this is like a liver internal medicine journal. Um, there are a few that we don't have, uh, uh, In pathology journals today, and they are from other disciplines where people use digital pathology and, and published there, which I think is cool because just like spreads digital pathology from this niche of pathologists, because pathology is the gateway to diagnosis. And, um, that's why I'm happy about that. So. What did they say here today? Life stay intervention is the main stay of therapy for metabolic dysfunction associated, uh, stato hepatitis. So MASH, this is what we're talking about. And fibrosis is a key consequence that predicts adverse clinical outcomes. So when it goes very bad, you have liver fibrosis. I want to make it even bigger so that you can see this. Um, anyway, so what they did here, they did imaging of biopsies of pre treatment and post treatment biopsies. And I'm going to tell you what the treatment or like, um, what the intervention was, um, but they use a different method than just our standard H and E. They used harmonic generation to photon excitation fluorescence microscopy with, um, AI image analysis. So, uh, it's, uh. non staining method. You excite the surface of the sample and you image it, you image the fluorescence and the method, um, to calculate the level of fibrosis. They called it Q fibrosis. And this is, this provides qualitative assessment. And so I thought that was also interesting because the non staining methods are getting more popular. prevalent. And if you're just, if you've just joined because I see the number of people joining and unjoining fluctuate, uh, be sure to say hi and be sure to let me know if there is a specific topic you would like to cover in one of these, uh, journal clubs. So, um, they examined those unstained sections. So here's the keyword unstained sections from paired liver biopsies, baseline and end of treatment intervention. So pre and post, and they had something that is called routine lifestyle intervention and strengthened lifestyle intervention. Um, And they actually saw fibrosis regression, which I think is a big deal because fibrosis is like, um, one of those processes in pathology that doesn't go away that easily, but they saw it with this lifestyle intervention. Uh, and. They basically say that using digital pathology, they could detect more pronounced fibrosis regression. Um, congratulations to them. Let's see. Where is it published? And we already checked in the liver internal medicine. Okay. So that is our first abstract without staining. I think this is cool because, um, so I told you the other day, or I keep telling people, uh, that I was super excited about the molecular predictions. And for a long time, like 34 years, uh, it was like only academia, only research, nothing into the clinic until there was a publication in 2023, that something actually made it in the clinic. The other thing that I'm super excited about is the stainless steel. Uh, imaging direct to digital imaging. Sometimes you can do direct to digital from, um, from tissue, like immediately without even fixating the tissue or just like cut it or freeze, cut it or different methods. And sometimes you need to fix it, but you don't have to stain it. So anyway, in each of these. Uh, options, you skip some of the analog stacks, um, analog steps of pathology of digital pathology. You go direct to digital and, and that's something that is super, um, like a huge advantage of digital pathology because radiology and all the other digital imaging, they have it. They don't have to go to analog and then to digital like pathology. And I'm like, okay, until that goes away, uh, we're not going to be as good as the other ones. So Uh, we are as good as the other ones, uh, but just have class slides. Okay. Now moving on to our next abstract, uh, which is artificial intelligence for, oh, sorry. Artificial intelligence for detection of prostate cancer in biopsies during active surveillance. So, um, I assumed that Male people know what that is, but I checked like, what's the difference between active surveys and surveillance and treatment and basically active surveillance is when you are not treating, you're just monitoring and you're checking. You still do. Um, biopsies pre and post at some point, um, sorry, no, you do the 1st biopsy and then you have the results that are telling you that there is no cancer or inconclusive for like, low. There is a specific Gleason score for that. And, um. That's active surveillance. You don't do prostatectomy, you don't treat, you just monitor. And what happened here, they use an AI algorithm, which to me, it was like, okay, um, well, we even have an AI algorithm that is, uh, cleared by the FDA for prostate. So that's kind of like not that novel. Let me check the journal again. Because this one was also, Oh, so I didn't know this journal BJU. This is a British journal of urology. So this is again, a non pathology journal that, um, is actually publishing about digital pathology and has the keywords digital pathology. So Um, they evaluated cancer detect, uh, sorry to, uh, uh, the objectives were to evaluate a cancer detecting artificial intelligence algorithm on active surveillance. So they had, uh, and this is a group from Norway, they had 180 patients in the, um, prostate cancer research, international active surveillance priors. Did I tell you that the pathology likes those, uh, abbreviations? Um, I think last time I told you. So, uh, they had diagnostic. So that the initial diagnosis, initial biopsy and let me zoom in and their re biopsy slide. So they had 2 biopsies and the dates were from 2011 to 2000. 20. And they had a lot of these biopsies because they had 4, 000, um, 4, 744. And they were scanned and analyzed by an in house AI cancer, uh, detection team. Algorithm. So they, they developed their own, they didn't use any commercial tool. They just developed their own. And what happened, uh, also they were, uh, then, um, the goal was, uh, that To analyze this, um, this algorithm for a sensitivity specificity and accuracy to predict the need for active treatment. So their surveillance and if there is a need for active treatment and retreat, what's the prediction and, uh, other prognostic properties like cancer size, prostate specific antigen PSA, um, level and a PSA density and diagnosis were evaluated and the results were. And this particular algorithm had the specificity of a, um, specificity, sorry, sensitivity of 0. 96 and specificity 0. 73 respectively. Um, And their reference method to check how good it was, was the original pathology report diagnosis. So that was what they were comparing against. Um, and the conclusion is that this cancer detection algorithm could be used to reduce pathologists workflow. So, Could be used, I would have to go into the, um, original paper and see, okay, like, are they planning to use it? Or did they just have the good specificity and sensitivity and sensitivity and specificity to use it? I guess, um, but basically it could be used to reduce pathologist workflow in the active surveillance cohort. Um, And the detected cancer amount correlated with the cancer length measured by the pathologist and the algorithm performed well in finding even small areas of cancer. Are we like suggesting here that we run this algorithm and then, uh, the pathologist just checks and we have like a screening tool. I think that would be cool. Um, So, and what they say is that to their knowledge, this is the first report on an AI based algorithm in the world. An active surveillance cohort. So normally it's just like diagnostics and but here they were doing this active surveillance and it worked. It worked. So the 3rd paper today is from, uh, scientific reports and scientific reports, especially. Part of the nature portfolio. So these, when I look at these papers, they're usually like very impactful and pretty complicated. Uh, but that's okay. It means, um, and this group is from Wuhan, China, Wuhan, China published in. Scientific report. So the title is matrix metalloproteinase nine expression and glioblastoma survival prediction using machine learning on digital pathology images. So let's start with what are these matrix metalloproteinases and why is this expression important? So these are enzymes that, um, they To change the tumor micro environment, for example, they, uh, can digest proteins, digest collagen and restructure, rebuild the tumor micro environment. For example, the tumor stroma and, um, what's happening in the micro environment is often important for a prognosis and, but this is an interesting thing. Set up as well, because they applied, so the study aimed to apply pathomics, I'm like, pathomics, what do you mean with pathomics? What is this? Um, and pathomics is like, Using pathology expression data and like multimodal data, including pathology. So they'll use this pathoma to predict ma matrix metalloproteinase expression in glioblastoma and um, to investigate the underlying molecular mechanisms, mechanisms associated with paths. So I like heard this word, but I didn't know it was like a. Like an official word. I had to look it up guys. Um, so they had 120 glioblastoma patients and 78 of them were allocated to the training, uh, randomly allocated to the training and test scores for pathomics modeling. And, um, so then they calculate the prognostic significance of this metal, uh, matrix metalloproteinase. So just to give you a little bit of context, the normal way of measuring this is, um, RNA, uh, expression, the RNA levels. So you have to destroy the tissue and, um. Find what, what the levels are. Right? So here, um, they used pyrodeomics. They used pyrodeomics. I'm going to tell you what pyrodeomics is because I had no idea. I had to look it up for you. Pyrodeomics was used to extract the measures of H& E stained hall slide images. And I'm like, pyrodeomics? What is this? Apparently, it is some open source model. Um, I want to read you the real definition. Pyradiomics. It's an open source model, uh, open source Python package for the extraction of radiomics data from medical images. I'm like, okay, that's Python. Pathomics, but they use Radiomics, I guess it worked well as well. So anyway, they did feature selection and, um, used different parameters and they created a prediction models, uh, using support vector machines and logistic regression. Um, and, uh, they assess the performance and what Uh, they state. So the performance was assessed using ROC align analysis, calibration curve assessment, and decision curve analysis. Uh, and the MMP nine. The metal metal, um, metal proteinase nine expression was elevated in patients with glioblastoma. So, uh, and this was an independent prognostic factor for Glioblastoma. So this is like independent thing. And they had other different features, but, um, I went into the introduction of this and basically what they did, they used the TCGA, uh, Tumor Cancer Genome Atlas and another database, to train this, uh, This model against the RNA levels, and then they used it, uh, RNA levels for metal protein is so pretty fancy. Another, um, thing that gives me like hope. I don't know if hope is the right word, but basically another thing where you can predict something from the image and the image is always there. The image is part of your pathological diagnosis. So one, one part is okay. If you can do some imaging without, um, the, the slides, that's fantastic. That's like the next level. It would be something that you will be using for intraoperative procedures. Uh, I don't know, maybe something on the skin where, I don't know, something where you would not need to take out the sample, but for this type of diagnosis, the pathology on glass is often going to be the. Um, the thing that's going to be, uh, still going to be used and this image is always there. Right? So if you can avoid some downstream tests, or if you can guide or narrow down the tests that you're going to be doing with, um, after the pathology is already done, after the biopsy is already taken, that's, uh, all power, power to you, right? Fantastic. So this is another thing where they, where they did that and published in a journal. Nature, and these are the three ones that, uh, no, there is one, one more that I want to share with you. Um, it was interesting because I was looking, so, so the title of, uh, the next one is 1 million segmented red blood cells with. 240k classified in nine shapes and 47, 000 patches of 25 manual blood smears. So it's an interesting title, right? They had 25 manual blood smears and they managed to have all those thousands of pieces of data. Um, and, uh, this is also, Scientific data. I'm checking here. This is also nature portfolio. I'm like laughing at this title because it's more like a marketing title, uh, for people to click on. Um, but you know, that's a good enough title for scientific data. And this group is from Egypt. And this is actually a group that has a commercial entity pathologics. I was looking up this company. If anybody from pathologics, um, Is listening to this, feel free to reach out to me on linked in, um, in great publication, but they couldn't find the website of this company, but my computer is acting out. So what happened here? I want a different color. I want a green. Okay. So, um, 25, 20 percent of complete blood count samples necessitate visual review using light microscopes, um, or digital pathology scanners, right? We can scan as well. Uh, so there is no currently no technological alternative to the visual evaluation of Red blood cells, which I thought was interesting because I thought that, um, the blood analysis, but it's, I think it's, um, white blood cells, like all these morphologies that, uh, uh, AI algorithms actually started in cytology to differentiate white blood cells and different blood cells from each other. So I was surprised to read this, that there is no technological alternative for red blood cells, morphology shapes. So, Um, the erythrocytes. Um, so, um, and, and what are the problems with erythrocytes? Um, t True, sorry. Wrong pen. There is, um, true non artifact teardrop shaped red blood cells and schistocytes or fragmented red blood cells are commonly associated with serious medical conditions and can be fatal and then increased ovalocytes are associated with almost all type of anemias. Interesting. So. What they did, they took 25 blood smears, um, and each of those blood smears was from a different patient. Then they were manually prepared, stained, and sorted into four groups, and each group underwent treatment. Imaging using different cameras integrated into light microscope with 40 X microscopic lenses. So they did it with microscopic camera. They didn't do this with scanning. And what happened then was that they, uh, had a lot of patches. So because, uh, with the microscopic camera, you'll probably have to go field by field of view, by field of view, by field of view. And, uh, had, they had a lot of patches, 47, 000 plus field images or patches. And then two. hematologists processed cell by cell to provide 1, 000, 000 plus segmented red blood cells with coordinates and classified 240, 000 of red blood cells into nine shapes. So we have classification of nine shapes and this is, um, um, Manual data labeling, and they created a data set of safety RBCs for AI and safety is one of the authors of this paper. And then you can, you have this data set that enables the development of testing and deep learning automation for red blood cells, morphologies and shape examination. Um, and you can also include specific normalization for blood smear stains. So the blood smears and the cytology, these are slightly different stains. Um, then the H and E that we are all used to, all who look at tissue are used to, um, pathologists a lot. Um, so they prepared this data set and they provided also codes, uh, for one for codes, meaning the code. I don't know which coding language they used, but one was for semi automated image processing, and another for testing of deep learning based image classifiers. So that is interesting. Guys, which one do you think is most interesting? Liver, blood, or What else we had? Prostate active surveillance and, and see, I don't even remember what we talked about and I just finished talking about it. And glioblastoma, the metalloproteins. If you have a favorite, give me a 1, 2, 3, 4. Um, in the meantime, Thank you so much for joining, but I want to tell you a few more things today. So, uh, what's up in general in life? I'm going to Poland in August, so I don't have the live stream scheduled for August yet. I probably will skip one week. Um, and. Have it. No, I think I'm going to keep it at the same time because in Poland, it's 6 hours difference. So it's going to be, uh, 12 o'clock my Polish lunchtime. It's still going to be 6 o'clock for everybody who is in the same time zone as I am right now. So that's going to be happening. You will have see a different background behind me. And then another thing I'm working on. Let me know if you're interested in this, uh, it's gonna be, I have a bunch of YouTube videos, as you know, um, it's like. Almost 370 or over 400. And so anytime there was a topic I want to explain, I would make a YouTube video and one of my membership digital pathology club members suggested, Hey, could you like embed those videos in a kind of curriculum and, um, I am looking for a way to choose, uh, my YouTube videos. So, um, this is, you know, everything is already available for free, but I want to have a little small, um, Option like a paid version where they are systematized in the curriculum, um, that's going to be something that is less, uh, investment than my digital pathology club membership, which is at 97 a month. This is going to be like a one time course fee. So if this is something interesting for you. Give me a comment, say YouTube course, and I'm going to gauge interest. How many of you are interested in that in this? It's like, I call it mini course, but it's not going to be mini course. It's going to be like comprehensive curriculum, but from videos that are already available for free, that are going to be systematized, maybe some additional resources. And I will need to host it somewhere so that it's not, uh, I mean, the videos are going to be the same, but I'm going to host it somewhere where you don't have to scroll through YouTube, It's going to be a dedicated page. So if this is of any interest, drop a comment, YouTube course. And another cool thing, uh, that I witnessed or, or was part of, I was just listening to it. There was an AWS, Amazon, um, works services. What, what does AWS stand for? Even I know it's Amazon for cloud, right? AWS. Um, What is it? Amazon web services. Yeah. Amazon for cloud web services. Um, they had a summit in New York and, uh, there was a part of the summit where they were announcing different collaborations with different people, different businesses that you could register for and listen to. So. There's a lot going on with generative AI, right? So, by the way, if generative AI and chat GPT and stuff like that is of interest, and drop me a comment, uh, below as well, that this is a topic of interest, and then I'm going to go Um, into, undermine the AI and do a specific search for that. And then we're gonna go through papers on that particular topic. Um, I remember giving a webinar last year where, uh, there were no, uh, I was explaining chat, GPT and you can find it on YouTube as well. Maybe I'm gonna link, uh, in the comment. But basically there were like, I was doing literature research and there was nothing out. And now every week. Um, there is something new, uh, pretty high publication. Uh, it was tested how pathologists would like to use it. Um, so, um, if that's a topic of interest, give me a comment. Let me know chat GPT or generative AI, uh, whatever you want to put in there. Uh, so, uh, what's happening, they are working, AWS is working with Anthropic. Anthropic is like, uh, I don't know if they're, I assume they're competitor to OpenAI. They have this other model, Claude, in contrast to ChatGPT, and I use both. By the way, uh, so they are working with Claude, uh, they're working on solutions for enterprises that are, um, secure that are, uh, being able to be fine tuned, uh, and, uh, just use your secure company data without leaking everywhere. So a lot of, um. Business solutions that will help different type of businesses, um, get leverage. And I, of course, was on the lookout for anything in the biotech space or pharma or pathology. Um, and I've heard that Pfizer is already using using Anthropic for something. I was looking for some press releases and it only says that they are collaborating and doesn't really say, uh, what. Um. What they're doing with it. I couldn't figure that out. And okay, Erica, perfect. You want the chat GPT version. If anybody else, uh, let me know because then I can prepare something super specific. And if you have any other topics that you would be interested in, rather than just the abstract reviews, let me know in the comments and in the meantime, you have a wonderful rest of your day and I talk to you in the next episode..
Just very quickly because you are listening to it. I just wanted to let you know that everything that I'm like asking in the comments or everything, you can just send it to me by email. If you're subscribed to my newsletter, you're gonna get, you're getting. All the content that is going out there. And you can always respond to those emails and I am super, super happy to hear from you anytime. And if you like this show, there is also an option to support the show. Either on the podcast website. Or if you're watching this on YouTube, there is an option. To give super thanks. And if you are listening to this podcast, because you want to hear interviews with people, don't worry. There are going to be interviews next week. We're releasing another episode. With a digital pathologist, Dr. Todd Randolph. So stay tuned. We're going to have both an amazing guest and we're going to be reviewing abstracts to stand top off that digital pathology, AI science.