Digital Pathology Podcast
Digital Pathology Podcast
97: DigiPath Digest #5 (AI in Modern Medicine: Diagnostics and Healthcare Outcomes)
DigiPath Digest #5 is ready as audio!
We explore how AI and image datasets can accelerate medical education for both radiology and pathology.
I review comparisons between the GPT-4 vision model and convolutional neural networks for neuropathological changes in the brain.
We explore how AI can potentially reduce healthcare costs, particularly in cancer risk discrimination.
Additionally, there's a focus on AI applications in digital urine cytology for bladder cancer diagnosis.
I also share personal updates, upcoming podcast guests, and my plans for utilizing YouTube content to create an educational course.
The episode wraps up with a lively discussion on integrating AI in clinical workflows and prioritizing patient care.
TIMESTAMPS:
00:00 Introduction and Podcast Updates
03:41 Guest Highlights and Personal Updates
06:33 Digital Self-Learning in Radiology
12:14 AI in Breast Cancer Risk Assessment
18:36 Comparing GPT-4 Vision and CNN in Neuropathology
21:58 Challenges in Lesion Identification
22:59 Few-Shot Learning in Neuropathology
24:42 AI in Bladder Cancer Diagnosis
29:48 Innovations in Digital Pathology
38:48 AI-Powered Clinical Workflows
44:42 Conclusion and Future Directions
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. I did not upload last digit puff, digest as podcasts, but today's DIGIPASS digest is ready for you right now. Other than me coughing. But I was able to meet myself almost on every platform, not on tick-tock or Instagram. But other than me coughing, we covered some cool topics. We covered. How AI and the image data set could accelerate medical education specifically, and the paper was about radiology, but that would apply totally to pathology as well. Then we talked, there was a comparison of a GPD four vision model with the convolutional neural net for neuropathological changes. In the brain and that was cool as well. We talked about How AI can help with. Your cytology evaluation. How much faster the pathologists and, Cytotechnologists where we're doing that. And we talked about, okay, will AI really reduce cost of healthcare? And that was actually, on an example of and prediction models. That was actually on an example of. Cancer risk district discrimination. From H a knee compared to standard of care based on genetic testing. So without further ado, let's dive into the content of today's DIGIPASS digest. 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. Amazing to see you here again. Good morning. 6am in Pennsylvania, in Fairfield, Pennsylvania. I am trying to go live across platforms and waiting for you to show up. So whenever you are there, I'm saying hi in the comments. Hi from. Fairfield PA. Let me know where you are dialing in from. I love to hear which like, and tell me what time it is. It's 6 AM here in Fairfield and it's going to be different times in different places. I'm trying to go live on different platforms, but Instagram and Tik Tok are giving me a hard time. Maybe I need to, lemme just give me one second to see if I can do Instagram from the phone. To click Go live. And let's see. Live. And go live. I don't know if it's gonna be working. Oh, we, Roberts so great to have you and you are dialing in from. Main, as always, and I don't know, so yeah, I'm trying to do on my phone for TikTok and Instagram if it's gonna work, nobody knows, but that's okay, because we're gonna figure it out for you. next week. My super coffee cup. So I know who was last time here from Poland, my friend Agnieszka. She works in Switzerland. She's a veterinary pathologist as well. And she recognized the cup. So as I'm waiting for you to join let me tell you what happened. Last week I had cool guests on the podcast. I had one radiologist, Dr. Nina Cutler, and I had also one veterinary pathologist who works for Aphoria. Aphoria is a sponsor of Digital Pathology Place. They have been sponsoring Digital Pathology Place for, I think, over three years already. So since the very beginning when I started this as a company. When you are joining, let me know where you're tuning in from. And last time I was telling you also about that. My running coach wanted to quit and I'm like, Oh, why should I run then? He didn't quit. So he has new motivation and all good. So I have new motivation for running and this is amazing. And then another adventure, it wasn't really an adventure, but I, since I read this book it's called by Dr. Lion. Her last name is Lion. Anyway, he, she's a physician and nutritionist, forever strong. And there I learned, oh you will, forever strong. You start losing muscle mass when you're 30 and I'm like, Oh, okay. I'm going to turn 40. How about I start working out? So Richard, so great to hear you. Fantastic. Okay, good. People are coming in. So let me finish this thought and we're going to go to the papers. So anyway, I'm like, okay, I need to do weight training because she recommends you need to actually train with weights, not just like cardio. And that's. Like nutrient and you should eat protein. Anyway, but the point was I wanted to have something like we have for this life where you show up, somebody is life and is doing the exercises and showing the exercises, but it's not just pre recorded thing. If it's a pre recorded thing, I'm like, have I have a million other things to do, and I'm not going to watch it. Which by the way, if you're not live, please watch it, find time to watch it. But basically I was looking for something similar like this, like a, like an online gym and I couldn't find it. I just had a, the only thing I found was a personal trainer, which I took one class with, and that was too much for me. I. Didn't want a trainer. I just wanted like accountability and showing up and by showing up, reaping the longterm benefits. And I didn't find it. I found the pre recorded thing. So it's still on me to show up and do the exercises. And I only did it once this week, but that's okay. Better once than none. And. Welcome! Hello! 64 Codon Private Limited. That's from YouTube. Perfect. Great to have you. Okay! Let's do it then! We have new papers. And Sorry. Let me coordinate my tablet. The first one was not important, but I fell in love with the second paper from the PubMed Alert. And the second paper was titled, Improving the Diagnostic Performance of Inexperienced Reader for Thyroid Modules through Digital Self Learning and Artificial Intelligence Assistance. And I thought, wow, this is amazing. If you want to do a start up Let's do this. So this for education, like a digital pathology startup. So this one is in, I'm supposed to be red. Why is it black? Whatever. We're going to be black. Anyway, so it's front endocrinology, frontiers of endocrinology, something, an endocrinology journal. Let's see if it works. Okay. It works for blue. So I don't know why red is here. Okay. And this group from. Korea, this is a radiology group. So we have radiologists here working with us. Can I do a little video for you? Yeah, that would be cool. Okay, and I have a superstar here because, and a heart! Look guys, I even have a heart. I love that even though it was for radiology, what happened here, Their objective was to evaluate efficacy of digital self learning, which is fantastic for image based specialties when you learn. So image basis, pathology, radiology, I don't know what else. I get, I bet you have some imaging for cardiology, but mainly radiology and pathology. And you basically learn by looking at a lot of images. And mostly it's like in the olden days, it was, you sit with another pathologist and you look at the same slides through a multi headed scope. And now you can do it digitally however, but basically you just look at images. They figured out, let's do a tool that is a computer aided diagnostic tool, CAD, it's called CAD. And what they did last year, they had 26 readers with less than 20. A year experience less than one year of experience in thyroid ultrasound. So we're talking ultrasound, but it doesn't really matter. I think this concept is brilliant. That's why I'm thinking, Oh, this is a, this is an idea for a startup. From various departments, six hospitals participated in this study. So readers completed online learning session. So first they looked at like 3000 nodules. They looked at it online. So that's what that was totally virtual. What they did then they were asked to assess a test. So then they got the test of 120. So they learned on 3000 then assessed 120 thyroid nodules and they knew the surgical pathology for this. And after this learning session, they use this AI CAD, AI computer aided diagnostics. The, and so let me re restart. Okay so they did the test was before and after learning session. So before they got to 120, they learned for 3000, they got another 120. And then they used the CAD and made their final decision on their nodules. And then. Diagnostic performance before and after self learning and with AI CAD assistance were evaluated. So what happened here? What are our results? The area under the receiver operating characteristic curve improved after the self learning session, and it improved farther after the radiologist referred to AI CAD. It was 0 6 7 9 versus 0 7 13 for n versus 7, 5 8. So anyway, it was improving. And then the residents who are in very much in a learning mode, they improved the a UC and accuracy after self-training. But the other readers didn't, and the other readers did not improve after self training, but when they had CAD assistance, their sensitivity and accuracy improved in all readers. So regardless whether you were a resident who is tasked to work on it and learn it, or whether you were not a resident. Then you improved when you had the CAD assistance. I think it's brilliant for training. Imagine we had like for diagnostics or for anything, right? Usually we just have one image from a book or. And then you like do it by experience when you see some cool cases, you share them. Imagine if we had a whole training data set with, I don't know, one type of cancer whichever, or toxicopathologic lesions for me for toxpath or anything else, like a whole set where you get 3000 to learn from. I don't know if I have seen 3000 of a single lesion in my life. And that would be amazing. So anyway, then you get, first you have the set and then you have an AI tool that helps you for your new cases. Amazing. I guess in radiology, you just have this thyroid nodule, so everything is a nodule. And in pathology, it's then subdivided into different diagnosis. But still. What a cool application. I think it's fantastic. That's why I have a heart here. I decided we're not going to be talking about tooth segmentation today, even though we covered some dentistry applications. And what else didn't I want to do some other radiology, but I have other cool stuff we have here. A paper titled US payer budget impact of using an AI augmented Cancer risk discrimination, digital histopathology platform to identify high risk of recurrence in women with early stage invasive breast cancer. So this is from a different journal, journal of medical economics where they actually check, Hey, is it going to be cheaper? Is it going to cost less if we use those AI tools? Because everybody says yes, because then you don't need to do molecular, but will it? And by how much? This particular study was done by here this is a company Veranex, no, probably, Precise DX, that sounds like a company, I have to look them up. And then we have Brigham and Women's Hospital, Dana Farber Brigham and Women's Cancer Center, Boston, Massachusetts, and Mount Sinai Health System, so they did that. And the goal here was, our goal was to Use of so so basically check. Okay, the normal standard of care versus with AI so use of gene expression signatures to predict adjuvant chemotherapy benefit in women with early stage breast cancer is increasing so we use more gene expression signatures. And of course we have high cost of this and limited access and then eligibility. I don't know what the eligibility criteria are, but there are some for these tests. Make it less accessible for people. So The study evaluates the cost impact of Precise DX Breast, an AI augmented histopathology platform that assesses six year risk of recurrence in early stage invasive breast cancer patients. So this is like a risk stratification, it's not a classical, computer aided diagnostic where you show the pathologist where the cancer is, where everybody asks, like, how much faster can you be with that? They already trained and know how to how to diagnose everything. But anyway, I still think there is value, but here what's the risk, right? And there are material and methods. So they had a decision tree model that was checking the normal versus, the standard of care that we're using now versus this AI tool. So to compare cost of treatment guided by. Standard of care risk assessment. So what does that mean? Clinical diagnostic workup with or without Oncotype DX. And Oncotype DX is a genetic gene expression, signature gene expression test. That is based on what lab molecular stuff versus this PD. XBR. I love those names. PDXBR with standard of care in hypothetical cohort of us women. So this was modeling with early stage invasive breast cancer. And they compared cost of everything. Let me get rid of too many lines here. Cost of testing, adjuvant therapy, recurrence, adverse events, and surveillance and end of life care. And what I need to do now is cough. My apologies, because my kids go to the summer camp, and I don't know how, but it's in the middle of the summer and they brought some cold home. And of course I got it I think, oh no, it's for kids. And then one week or two weeks later, I got it. I get it of course as well, but question is how much cheaper was it? Was it cheaper or not? Let me tell you it was spoiler alert, it was cheaper. But interesting like how they were calculating it was the PRD, PRDX use in prognostic evaluation resulted in savings of four million in year one compared to current standard of care but In how many people in 1, 000, 000 female females members, so 1, 000, 000 people when you do that for 1, 000, 000 people, you're going to save 4, 000, 000 compared to the current and now the numbers where the savings increased to 12. 5 million again. My apologies for the people on TikTok and Instagram, where I cannot probably mute it. But 12 million in the pre treated patient costs in one year. Amounted to 19. 5 thousands. Oh my goodness, sorry, guys. My apologies for that, it always happens. And then I start crying and there went the livestream, but give me a chance, give me a chance, guys. I can recover. I'm not sure. I should make short content for the moments that I'm crying. Anyway, I'm actually very happy to be here. I'm not sad. Let's summarize this one, right? How much money did they save? They saved 12 million in year ma in year one. And this amounted, no, sorry. Savings increased to 12. 5 million. The pre treated patient cost in, this is the patient cost. So one patient in year one amounted to 19, 000, 19. 5, 000 for standard of care and 16. 9 for PDXBR. Okay, we have, what do we have a 3, 000 difference. I was hoping for more. I'm sorry. I was like, Oh, maybe it was 20, 000 and now it's just 5, 000 or 1, 000 and not, but still 3, 000 per patient. So for 1 million patients, you save 4 million. However, they calculated that, but the limitations of the study was recurrence was not specified and they performed these analysis as best as they could with it, with the assumptions that they had. And obviously when the standard of care fails once or twice, then you go to a different type of treatment, which is like more expensive. So they calculated that into it. Conclusions pDXBR demonstrated robust overall savings, but still healthcare costs. In the U. S. is another, in the world, but in the U. S. is another topic that books were written on. And thank you so much for staying with me, and I see more people are joining. One more cough, and we're going back to the next paper. So when you have just joined, let me know where are you dialing in from. And also, if you have any questions, both to the papers and any other questions, let me know in the chat. I would love to answer them. And let's look at the next paper, Evaluating the efficacy of a few shot learning for GPT 4 vision in neurodegenerative disease histopathology, a comparative analysis with convolutional neural networks. Before the vision models, before chat GPT and the generative AI convolutional neural networks was the super, super amazing thing in computer vision. And it was used for pathology as well. So no cure here. They are comparing let's unpack because I'm like, okay, what's few shot learning. Is there a lot of shot and no shot? Yes, there is. So if you should learning in general, you have a pre trained model, pre trained AI, and then A zero shot is when it doesn't get any additional domain specific data. So you're just like, you would take this vision model and try to diagnose something from neuropathology. Then few shot, you give it a few examples and a lot of shot, or I don't know how they call it, but I think no, sorry, zero shot, one shot, few shot because if you have to give a lot of examples, then, it's, what's the point of having a pre trained model that is supposed to adapt quickly. But basically this was a few shot and we're going to learn exactly what the few shot means here. And Neuropathology Applied Neurobiology is the journal that we're talking about. And the group is from the us. We have Maya Clinic, Jacksonville, Florida, and University of Pennsylvania. Hi to the University of Pennsylvania group because I'm in Pennsylvania. So what happened here? We have this large, language model like chat GP for GPT for vision. So this is the this is an option of chat GPT that also does images, right? And this study evaluates the accuracy of CHAT GPT for vision in image classification tasks. So this is important classification, we're classifying images histopathological images, so we're back to our pathology core competence. And it compares its performance to a convolutional neural network. So what have we took? 1,520 images, including h and e staining and tau immunohistochemistry for patients with various neurodegenerative diseases. And these were Alzheimer's disease, progressive supra nuclear palsy, and corticobasal degeneration. The only one I'm familiar with is probably Alzheimer's, but that's not the point. The point is that they used multi step prompts to determine how textual context influenced image interpretation. So textual context would influence image interpretation, and they employed this Fusion learning. Can I do? Trying Colors. Fusion learning to enhance improvement in GT four vision, diagnostic performance. And they were aiming at three specific lesions. Astrocytic plaques, neurotic plaques, and. Es and they compared the outcomes to a CNN model, YOLO V eight. They had a CNN model that does that as well, classifies these different lesions. So what happened? The results accurately re the G PT four Vision accurately recognize. and tissue origin, but struggled with specific lesion identification. And I'm like, Oh, my first thought is like, Oh, it cannot detect lesions. Then like it's useless. But then I'm like, how much time does it take students to learn the. Correct and recognize the correct tissue and recognize the staining. You're always like, okay, I want to recognize the tissue I want to know the stain then I look for the lesions. So this one is already Three the two out of three steps in but that's okay. So then when it was and it was When it was presented with specific images And provided influence when it was influenced by the textual context this sometimes led to diagnostic inaccuracies. What do they mean here? So when they presented with when they were presented with images. of the motor cortex. The diagnosis shifted inappropriately from Alzheimer's disease to the other diseases, but they gave it a chance. They did this few shot learning. Oh, and we have we haven't, we have George or George from Norway. Great to have you here. And few shot learning helped. Did it help? Yes, it did help. And the few shot here in this case, it was 20 shots. 20. And it helped Versus zero shot learning. So zero shot was with without the previous examples and the accuracy was only 40, but when they gave it 20 examples, it was 90 percent accuracy matching, matching the performance of this YOLO model, which requires 100, which required 100 shot learning to achieve the same accuracy. Hey, 20 slides only versus 100. This is fantastic because as we know, challenges In pathology and specifically neuropathology is we don't have enough images to train these models on right when we think of the trainings that those natural models get, sorry, natural images and by natural images. The normal images that you can recognize in, like dogs, cats, whatever, right? These are natural images, and also natural language is basically the language we speak. So not like code or any other although code is also included in language. So yeah. GPT for vision works well for neuropathology if you add some some examples, which is fantastic because there's always this question. Oh, how domain specific are they? I'm probably not that domain specific to pathology, but if you add Some examples, then it can take all the knowledge that it already has on all the image properties. It did the tissue, it did the stain, like that's already huge for a non pathology trained model. And then you add examples, few labeled data, and it works. I think this one is going to be our last paper and it's evaluating artificial intelligence enhanced digital urine cytology for bladder cancer diagnosis. This is also done by a company, AIXmed, which I love when companies do that because they know that you need to publish. And this was published in Cancer Cytopathology. And that's like the first step when you wanted to take a tool on the regulatory pathway. One time I was looking, I was talking to a digital pathology company. I was talking to a startup and they wanted to be a startup in the digital pathology space. And I asked them, Oh, do you have any publications? Can I look at your publications? And they said We are not focusing on the scientific aspect, we, this is a commercial company, and I'm like, okay, then you don't know who you're talking to, because anybody in the digital pathology company is going to ask you where you publish your data, how you validated it, and they were like no, we just want to make software that does this and that, and I'm like, Okay, good luck. When you have a publication, you can talk. Anyway, so let's talk about this one. What happened here? I apologize for this cough. It's the worst. Anyway, AIXMED from California, and they also have here a group from Taiwan. There's several group, several groups, several departments that they work with, and maybe that's a part of their company, probably. So what happens here is we evaluated the diagnostic effectiveness of AIX Euro. Did we not just have AIX BR, but from a different company? Yes, we did. So anyway, AIX Euro platform, artificial intelligence based tool to support urine cytology for bladder cancer management. So what does it mean to support? I checked the images and it, the user interface is you have a whole slide scan or like a picture of this urine cytology with different cells and it shows you and classifies those cells into Whatever the classification system for your cytology is, and they mentioned a specific one. I'm going to tell you in a second. They had one cytopathologist and two cytotechnologists, and they reviewed 116 urine cytology slides and they, Also reviewed the whole slide corresponding images and they used three diagnostic modalities So they just used microscopy just normal microscope whole slide images and AI X URL support And here is the system that they used the Paris system Of, for reporting urinary cytology criteria, which I don't know what these criteria are, but our cytopathologists and technologists know, and they the authors of this particular software, AIX Euro, also knew and incorporated this into this particular tool, right? AI x Euro. The results were and they did of course, performance metrics and the results were long story short here from the results, it was better than without, but let's discuss it. This AI X Euro improved diagnostic accuracy by increasing sensitivity. This is crazy because sensitivity, I look from 25, 30 to 60. Three. And I'm like, how can you have this sensitivity? And then I remember, this is cytology. Cytology is not anatomic pathology, it's not tissue. So you like work with different sensitivities and specificities and different metrics for this particular test. So here, the sensitivity improved to 63%. Amazing. And. And they have a different things that they have to evaluate atypical ureteria urethelial cells and then suspicious high grade urethelial carcinoma And everything improved the sensitivity the positive predictive value the negative predictive value And they had these specific distinction between different types of atypical cells, and they also had the binary decision. Is it atypical or not? So the binary was easier for this tool. And they also demonstrated, this was funny, that the cytopathologist demonstrated higher intra observer agreement than the two cytotechnologists. I was like, Oh, pathologist was better. I don't know if it's better or not. But also in addition to everything was better with the AIX euro, it significantly reduced screening time by 52 to 83%. That is nice. That's really good. So the next thing I would want to do here is oh, can we do this? economical analysis for this tool and see if it's cheaper because maybe we can screen more and take into consideration also the How much you it costs to scan and all these different things But the conclusion here in this one is AIXuro demonstrates the potential to improve both sensitivity and efficiency in bladder cancer Diagnostic. Let me check what else. No, we had this latent diffusion last time so Here are the papers for today and there are other papers there. There are other things happening in the world. I don't I need to catch up on reading them because there are a lot of foundational models coming out. And one recently came out by page. So I need to read that paper. I already started reading that paper. And there is a tool. I think I already mentioned this tool to you. Oh, and we have more people. Hello, Nigeria. Amazing to have you here. So I think I mentioned this tool to you, this literature research tool called Undermined. Undermined. Yes. ai guess. It's Like in contrast to PubMed, PubMed is a keyword based. So the keywords that I have now in PubMed are digital pathology and AI. And every week I get those alerts and we're looking at the email that I get from from PubMed. And I check, okay, do I like these publications or not? And for this undermine. ai is a different type of tool that searches literature. It's more for when you need to gather relevant information for a project you're going to be working on for a PhD, for grants, for, I don't know, maybe market research, because you can do market research with scientific literature as well. You, you give, you basically talk to it, to chat GPT and tell this tool what is the problem. And it finds you relevant information, a relevant publication, not just based on the keywords, but also on the relevance with, they have some kind of model that assesses this relevant and gives specific points and the authors of this tool. These are two MIT graduates who are now starting to develop a company around this tool. So they basically sat and made this tool and decided, Hey, it's going to be useful. And it is extremely useful. I already used it for many things, but they weren't, they want to join us. They want to show you the tool next time. So next time they're going to be at the live stream. I still have this live stream. So I'm going to be in Poland. The live stream is going to be in Poland. Run at the same time at the moment, and they are in California, so they will have to wake up very early and it's gonna be significant. It's gonna be like, what, 3 a. m. I don't know, early to show us this tool, but they were very excited that they will get the chance to show this tool to the digital pathology trailblazers. So that's one thing. The other thing I was talking about is my YouTube course. Yeah, I have the platform to host it. I figured out how to, excuse me, scrape all the information from YouTube. And now I'm going to be working on developing the curriculum. If anybody is interested in binge watching YouTube, Content in the specific order based on all the almost 400 videos that I have on my YouTube channel, this course is going to be coming out soon. Let me know if you're interested, right? Course in the comments, if this is something it's going to be it's not going to be a high price point. That's going to be like something that's affordable for everyone who is interested one time one time investment and every year I'm going to be updating it because. As the year goes, I create new videos and new things are happening. And now this hot topic is foundation models, vision models for pathology, and now maybe I'm thinking, okay, I'm like still every other paper is Oh, annotations are a bottleneck, but now you start hearing, Oh, they are, there are these zero shot models, one shot model. 20 shot model. Oh, maybe you just need like 20 annotations because it already has this vision knowledge to recognize different things. And I'm like, Oh, that would be so cool. Some cutting edge things already Being obsolete. So this is crazy. I love it. I love how fast it's developing and the podcast episodes that they are going to be coming out in, I don't know, a few months because I do have a backlog of podcasts, but Dr. Nina Kotler, she is The AI main person at radiology partners and radiology partners is the largest radiology practice in the U S with 3000 radiologists with where the average one is 15 or I don't know Few. And I got introduced to her by my friend Joe from Apridia and Apridia is a digital pathology place sponsor as well. And he said, you need to talk to this to this person because she's amazing in terms of AI and obviously radiology is ahead of the curve. A head of pathology in going digital. And obviously she has like enormous knowledge publications has this position is very active in different as a chair of different radiology college of radiologists positions and all that things. All these, so I knew I'm going to be talking to a really high level expert, but the thing that I loved about this podcast most is like how clearly she was explaining these concepts like to non experts without dumbing it down. Like I was. Mesmerized by the way she explained the concepts and we talk about things like agents, which is also like another new hype where you can basically have like a program that programs other programs within this generative AI. But she had a better explanation anyway. This is my way of recognizing somebody who's an, amazing expert. You have different levels, right? And she is beyond the technical knowledge and jargon when she explains things. Often you'll hear experts or people who are very knowledgeable. It's a natural process, right? The more familiar you become with a terminology, with a topic, the more you Start using this language because you're often talking to people who also understand this language, but then taking it a step further and being able with the same words to address experts who know the jargon and non experts without the experts feeling like they are being patronized or talked down to, or or any any sort of this Oh, do you know, I have this expertise and that's. And amazing skill. I thought this was fantastic. And we ran out of time because obviously she's super busy. And I also had meetings scheduled. Maybe I'm going to meet her live in October. She's going to be at a conference in Boston. So maybe I'm going to go there. It's going to be more imaging, not just pathology specific. And she's also active in the Society of Imaging Informatics, SEIM, where radiology plays a huge role, but also pathology is another imaging specialty. That was an amazing guest. And yesterday I talked to another guest. He's a veterinary pathologist. He works and he works for euphoria, but he does not only work for a Foria Air Foria as an image analysis company. I mentioned at the beginning of the livestream that they've been sponsoring digital pathology place for a long time. But he also does diagnostics. And he got in touch with euphoria for an AI diagnostic Pro project that they were, they were doing in his diagnostic lab and he fell in love with AI. He decided this is something I need to learn. This is something I need to get more involved in. But he still has two jobs. He's a diagnostician and then he does the AI and helps other get the AI for diagnostic purposes up and running. And the A4EI is working with both human hospitals, veterinary hospitals all different types of diagnostics and diagnostic applications. So he does that. He was interesting. I could relate to him because he's balancing two jobs and he says it's tricky, but he didn't want to give up the diagnostics. He said, this is, he's a diagnostician. In the heart, but AI is going to be the tool that's going to leverage diagnostics. So that's that. And next week I'm going to be in Poland. So you're going to see a different office behind. And also I wanted to tell you before we finish this live stream that this is like my favorite thing favorite thing during the week. I was watching another a person I follow on YouTube for a different topic, more the digital communications topic. And he read this book that I need to read as well. Something about what's worth doing and what like lights you. Gives you energy. What is the heck yeah thing and this is my heck yeah thing I love showing up and I'm aware that this is a very Not great time for most of the people that are my readers which are in the US and this is 6 a. m But I consistently have Robert on the stream and actually he was yesterday. He was like, are we doing this because there was no invite and I think Something went wrong with the emails but Thank you so much. Oh, we have a question. We have a question. Thank you so much for the question is hi, doctor, what are your thoughts on clinical workflows powered by AI? So I have a lot of thoughts. Give me give me a specific thing that you're interested in. And I'm going to start talking with this thing. I think clinical workflows should be powered by AI and they can be powered at different stages. Sorry, I dunno why I thought something was falling. Maybe my cup was falling. They're, they can be powered. You can apply ai. So ai, I see ai as leverage in general. So to answer this in general and everybody who needs to sign off, thank you so much for joining. But lemme answer this question. AI is leverage. How can you use this leverage? In your workflow. So this my guest, Richard Fox from aphoria, he had identified one problem and that was Kai 67 counting and like evaluating and he decided I could do that. And he applied AI there. And then and that's his advice, basically identify the problem where you want to leverage AI, like your lowest hanging fruit in terms of everything, in terms of a workflow integration, in terms of price, in terms of regulatory concerns. Identify this one thing and start there. You're going to have to go through hurdles to deploy it. You're going to learn from your own mistakes and other mistakes as you do literature research. And I'm going to show you this tool for literature research next time undermine. ai. But you're going to learn through the process. And then in the process you're going to identify the next things that you want AI to address. But this is for an already established workflow where you're plugging in AI from the outside and there is an option to Start in a different direction like to have a digital first AI first workflow And what you're gonna need to do you're gonna need to think of image quality and there are technologies that are emerging That are actually direct digital pathology imaging. I'm talking to a company muse Microscopy or Smart Health dx. They are having a product like that in in, they are evaluating it in like official studies. And that is amazing because the pushback that. Is legitimate and is heard often is like, why should I even bother if this technology is not direct to digital radiology is direct to digital. Logically, they're going to be ahead of us. They're going to be like leading and they can do it because there is no like analog anymore. We are still doing our glass slides and then scanning them and then only applying stuff. And you can show numbers, you can argue, but with. The analog, taking outta the equation. That's digital first workflow and you can design everything around it. And then apply ai. Use AI for different things. It can be diagnostics it can be quality control like workflow centric ai workflow for case prioritization, workflow for sending cases. What is the thing that you wanna, use leverage for, like for for the chat GPT and the large language models, right? What is the thing I read one paper by a friend, Candice Chu. She wrote a paper about chat GPT for digital pathology. Sorry, no, not for digital pathology, for veterinary medicine in general. And she was analyzing both human and veterinary medicine and it and she gave percentages. How much time does it take? And I think it was above 35 percent of physicians or veterinarian or any doctors, dentists, whoever, time is spent on admin tasks on feeling feeling the documentation. Okay, can you use AI to help with extracting like from recordings? How much money will we save? Let's do analysis and let's implement there. So yeah that's a very. Multifast, like it wasn't a straight answer. I hope it was a good answer and let me check if we have any other questions here. I have even some more people joining later. So if you are watching the replay give me a high from the replay as well. And let me know where you were tuning in from where you were listening in from. And also for me to see. Who was there, but but every comment, every interaction on this live stream brings it to more people who would benefit from this. My goal is to get this information to everybody who may benefit from it. You would not. Imagine because we already emerge in this reading papers and everything how many people like have no clue what digital pathology is even In the pathology world. They're like, I would like to get involved in digital pathology. Where do I even start? So yes, you will start with my youtube videos. How about that? But also there's gonna be a course so anyway Oh, and then we have a follow up. This is interesting for prioritizing a prioritization in imaging. I can see use case in prioritizing patients in ed for junior doctors. Yes. Junior doctors, you can we just reviewed the paper. I don't know if you were online then they had an computer aided diagnostic tool that was helping with education of pathologists. So I'm showing them a lot of cases and then making them diagnose and then the tool would help. So basically it is okay, where do I want the leverage? What is the Bottleneck that is easiest to solve that would free up resources to, to do more care, to, to be able to provide more care. So I hope that answers the question. And next time, please come with questions. I love answering your questions. Maybe we're going to do half an hour of the papers and then half an hour of discussion. And then I can engage with you a little bit more in the chat because when I'm reviewing the papers I can monitor the chat less. Thank you so much Dr. Archford. And thank you so much for joining. And I will talk to you in the next episode. And now I'm gonna log off on all the platforms. Thank you so much for joining. As always, also thank you for joining on audio. And I see that people are listening to those a digit path digest on audio as well. So I am committing to not missing any other future uploads. Uh, because everybody consumes information differently. If you enjoyed this episode. There. At least two other digital digest episodes already ready for you, but also have a look at the last podcast episode, episode, number 96, achieving work-life balance in medicine. As a pathologist with digital pathology. And that was with Dr. Todd Randolph. Uh, it was a good one as well. He basically created himself. Digital pathology remote position when it was not that trend yet. So I hope to hear you in that episode. And I talk to you in the next one.