235: From Cytology to Omics: Where Pathology AI Gets Harder

DigiPath Digest #45 asks a practical question: can AI in pathology move from correlation to real clinical use? In this episode, I review four papers that push on that question from different angles: computational pathology moving toward morphology-driven molecular inference, the current state of digital cytopathology and AI, multi-omics and precision oncology in hepatocellular carcinoma, and AI literacy in veterinary education. What ties them together is not model performance alone. It is the harder question of validation, workflow fit, quantitative use, ethics, and human oversight.

In the first paper, I talk about computational pathology as more than pattern recognition. The focus is on morphology-driven molecular inference, digital biomarkers, and why spatial omics matters as biological ground truth. I also discuss why continuous quantitative scoring is more useful than forcing biology into rough scoring buckets. 

The second paper focuses on digital cytopathology. Cytology was early for FDA-cleared AI in cervical screening, but non-gynecologic cytology is still much harder to digitize because of specimen variability and workflow complexity. I also cover telecytology, rapid onsite evaluation, automation, and quality control. 

The third paper looks at hepatocellular carcinoma and AI-driven precision oncology. This part is about using AI and machine learning to integrate genomics, transcriptomics, proteomics, metabolomics, radiomics, and pathology to support biomarker discovery, tumor microenvironment analysis, and treatment stratification. 

The fourth paper may be the most broadly useful. It proposes an AI literacy curriculum for veterinary education that covers AI fundamentals, machine learning evaluation, LLMs, ethics, liability, and academic integrity. I think that matters far beyond veterinary medicine, because if clinicians are expected to use AI tools responsibly, AI literacy cannot stay optional. 

Highlights
00:01 Welcome and overview of the four papers
03:02 Computational pathology and morphology-driven molecular inference
11:01 Digital cytopathology, telecytology, and QC
20:47 AI/ML in hepatocellular carcinoma precision oncology
31:04 AI literacy in veterinary education
47:42 Final takeaways and Digital Pathology 101 update 

Resources

1. Computational Pathology as a Mechanistic Discipline: From Morphology to Molecular Data
   https://pubmed.ncbi.nlm.nih.gov/42052846/
   
   
2. Advances in Digital Cytopathology and Artificial Intelligence Applications
   https://pubmed.ncbi.nlm.nih.gov/42046894/
   
   
3. Navigating the Labyrinth of Hepatocellular Carcinoma: Leveraging AI/ML for Precision Oncology
   https://pubmed.ncbi.nlm.nih.gov/42065059/
   
   
4. Curriculum Framework for Artificial Intelligence Literacy in Veterinary Education
   Front Vet Sci. 2026;13:1801756 

Transcript

00:00:01 - 00:01:41

Trail Blazers, welcome digital pathology trailblazers, I believe we are live. Let me know in the chat if we are and where you are tuning in from and I will just confirm with my team that the streaming is actually happening because LinkedIn is not showing up. Let me confirm that. And okay, I see some people joining. This is amazing. Welcome. Good morning. 601 in Fairfield, Pennsylvania. Um, let me know where you're tuning in from. I'm getting some error messages. Okay, we have a confirmation from at Trailblazers that

00:00:54 - 00:02:23

we are live. Why does the text on screen look crazy? So, um, as I wait for a few more of you, uh, let me, uh, do we have people from city? Is what? I don't know what city, which state is city. Um, it's not Colorado. Colorado is Connecticut. Connecticut. Is that Connecticut? Um, okay. So here my struggles of today I was trying to find a nice um mouse pointer didn't find it. Uh my pen I don't know if you know this story already. I have a pen for my tablet. There is a tablet where I

00:01:38 - 00:02:53

display the um the papers is still not working. I have to have somebody remote in because I could figure it out but I just don't have the patient. H and yes um okay after like 3 minutes I figured out it's Connecticut um and Kenya we have people from Kenya joining this is amazing okay so thank you so much for joining I'm Dr. Alex, I'm a board-certified veterary pathologist and I run the digital pathology place. It's an online platform about digital pathology and medical AI. And today is our weekly digipath digest

00:02:16 - 00:03:43

journal club where we discuss um the newest papers about digital pathology and medical AI mostly digital pathology. So, um, when I share my screen, I discovered something cool that I can actually have a different layout like this. These little things give me joy. If you have little things like this that give you joy, let me know in the chat. H, but let's start with our first papers. We have four today. And um is stuff working more or less? It's more or less working. Okay. So we have computational pathology as a mechanistic

00:03:02 - 00:04:09

discipline from morphology to molecular data and we also have the infographics. the infographics were hit hit uh last time because the abstracts as we said last time they don't give justice to the publication and then I listened to them on um podcast from notebook LM and there's so much more to the uh to the paper than the abstract but in this particular one um computational pathology as a mechanistic discipline from morphology to molecular data the authors say that pathology is undergoing

00:03:35 - 00:04:51

a shift from or form molecular interpretation toward computational integration of molecular mechanisms encoded in tissue architecture. Um this is really high level explanation but we're going to dive into it. So such as morphologydriven molecular inference may enable biomarker prediction and potentially generate therapeutic insights direct directly from routine hisystologology. So what are we talking about here? Here we're talking about molecular predictions from H& so um but there is a

00:04:12 - 00:05:31

caveat uh that hey why do we even believe these predictions it's just like a correlation um so they explore that actually the paradigm has important clinical implication for quantitative biomarker testing patient stratification design of digital biomarkerbased clinical ical trials and they are mentioning a digital biioarker. We're going to mention it as well. Um maybe you already know which one this is. If you know, let me know in the chat. So the most current artificial intelligence models remain correlative.

00:04:52 - 00:06:05

We just mentioned that. H and they the clinical impact depends on rigorous validation integration into workflows and ethical governance. And we have such a good paper at the end about AI literacy. So definitely stay till the end and if you can't watch the recording and uh fellow veterary pathologist of mine wrote this one. Um but when I saw this ethical governance that's why I remembered this paper because there is a part on that as well. uh and obviously addressing addressing these challenges will be essential for

00:05:29 - 00:07:21

computational pathology to turn uh to mature clinically meaningful discipline and not just a bunch of like single um applications. Let's have a look. Can I share something else? How can I do that? Uh, I can do that. But almost. Bear with me. What did I do with myself? Need to move myself to this different screen. Yeah, these little things show you that it's actually live. I hope you enjoy it. Oh, okay. So, uh, screen No window. This one. How about now? Now we're talking. Um, and that's the wrong

00:06:30 - 00:08:14

infographic. But the first one, um, can I draw here? I can, but it have Let's do this. Okay. Let's just do this. Can I move myself? No, I can't. I can only do this if I want to show you the full thing. Um, okay. We'll do our best here. So, um, what I like very much here is, uh, they have these different levels, right? And we are starting with level one. Level one is our morphology layer, the visible uh what we see, right? Then we have level two is the AI computational layer. So this is like the uh molecular

00:07:35 - 00:09:04

prediction from H& is happening with this computational layer. But uh what they say we also need the ground truth and I am not talking about annotations by pathologists. we are talking about spatial omics um grounding the AI. So we have spatial uh transcrytoics protoomics that you put on top of the uh like on the uh tissue right you can see what transcripts what um proteins are there and then when you uh query the AI our level two uh level two where did you find this information and let's say uh

00:08:18 - 00:10:01

the information was found Here it's a stroma let's say in the stroma and we have a ground truth from spatial onyx in the same region stroma uh then we have an indication that actually took the information necessary for the decision from the right region is it like 100% no but that's as close to 100% as we can get. So um now what we can do we can get from these subjective categories um also um that's kind of a separate thing um because now uh we used to do we used to score IHC with these buckets 0 1 plus 2

00:09:10 - 00:10:21

plus a bunch of markers IHC markers is still being done like that h but actually it's a continuum um and you have to like arbit arbitrarily make the cut of where is where uh one plus ends and two plus starts maybe it's one and a half but you didn't have you don't have the visual capacity even capability physiological capability to do that uh but the computers have that and we can have quantitative continue they are referencing um the now I hope famous and if not yet famous there's going to be a

00:09:45 - 00:11:03

webinar about this quantitative continuous scoring uh for the TROP 2 marker uh that stratified patients and this scoring is um calculating a ratio from the membraneous staining of a cell to its cytoplasmic staining and it also normalizes the um the membrane staining and there's going to be a webinar about this with wash team. So, if you're interested in the webinar, uh, leave me a comment. Webinar, we're not promoting actively yet. We're still figuring out the details, but if you're interested,

00:10:24 - 00:11:39

I'm going to just send you the invite um when you comment, right? Uh, on whichever platform you are, just leave me a comment webinar and you're going to be invited to this one. And in the meantime, let us go to the next paper. Let me know if you have any questions about this. U let me know if you want the infographic as well. Uh what I started doing is posting it as posts on social media and we know that this is trust but verify. This is an AI generated infographic and on one of them I have like funky AI mistakes that I'm

00:11:01 - 00:12:19

going to point out. Um, but webinar if you want the webinar, infographic if you want the infographic. We're going to figure out a way to send it to you. And let's let's do this one. The next paper is advances in digital cytopathology and artificial intelligence applications. This psychopathology is something that I even though I'm an anatomic pathologist so I look at FFP for formal and fixed paraffin embedded and not really cytoathology I'm thinking the prep preparation of cytologology

00:11:44 - 00:13:06

specimen is so much easier but the digitization is so much more difficult but also the first applications of digital pathology where in cytologology so what happened in between and this is what this paper is addressing So let's look. It's a review and um the authors are Sattorwir Parwani and Lee from Ohio State. By the way um maybe you are there and dialing in from uh Ohio because they have now the a conference global engage digital pathology conference. So probably the authors are there. So if you want to

00:12:30 - 00:13:46

meet them then you would already have to be there. Let's look at the abstract. Any questions any comments let me know in the chat. H cytoathology is the first field of pathology in which AI models were successfully developed like FDA cleared models for um and there was a commercial they were commercialized for routine clinical screening of cervical cytologology h and this has been in place for the past two to three decades I guess depending on the applications I was wondering is it two decades or three

00:13:08 - 00:14:38

decades but probably uh depending depending on the application and um the AI applications for non gynecologic cytologology. So this was cervical cytologology. This is uh pap screening papsmears um cervical cancer screening. uh but non gynecologic cytology this has just begun and uh the problem is the variety of cytologology specimen types and preparations with associated unique characteristics presents technical challenge for the complete digitization of the cytology workflow and this is a very diplomatic way of saying they look

00:13:52 - 00:15:22

so bad sometimes compared to an FF specimen men like they don't look great because you have cells smears and uh the level of interpretative effort uh in when doing cytologology is so much higher than when doing uh anatomic pathology. Um so that is the problem because they are not thin enough, not even enough. The uh cervical cytologology they standardize the way they are um creating these um slides, right? They have the thin prep which is centrifuging everything away other than the cells um the cervical cells that

00:14:38 - 00:15:53

we're interested in and then it's it's being like put on the slide in a little circle. So it looks really really nice compared to the other specimen that can be like all over the place. Um but despite these challenges, few institutions have a complete digital cytology workflow um and technical investment have replaced conventional rapid on-site evaluation um by different telescytology systems. So we're still doing um live microscopy but now it's telescytologology. Some can be scanned but most are being

00:15:15 - 00:16:38

done with remote controlled microscope. And um there is one system hologic genius and that is cleared not approved. I checked uh but this is like you know FDA terminology regulatory terminology. This is the only one that is cleared by the FDA uh for routine clinical screening of cervical cytologology in United States. So, we're still in cervical cytologology, right? The beautiful um looking thin prep. Um and I see people are interested in webinar. This is fantastic. Thank you so much for letting me know. Uh if you're

00:15:57 - 00:17:09

interested, I'm going to send you the link. Just leave me a comment in the chat. And um here the recommendations for a validation and best practice guidelines for digital psychopathology are currently being developed. There is a task force in uh digital pathology association. Um and there are technical advancements in automation for sample preparation. Uh rose this rapid onsite evaluation and using telescytology. There is automation for screening of uh gynecologic and non-gynecologic cytology

00:16:33 - 00:18:00

specimens. Automated quantification of biomarkers, quality control and this is what this review is talking about in detail. Um now let's move to our infographic. This is we also uh Okay. I like to have it all on the screen. How about that? That looks decent. I always want to make it full screen, but then I cannot draw on it. Oh, I can draw on it. Amazing. discovered a new thing. Um so what is happening here? We also have level one uh which was the established foundation the gynecologic psychology. Uh then we

00:17:34 - 00:18:58

have level two what is currently happening uh digital infrastructure telescytologology for rose. I don't know why they like did this MRI picture for telescytologology. ah maybe because they're doing something with the patient and the patient needs MRI and then they do the uh evaluation. So um maybe that's why so in this level two we have evolution of virtual talcytology digital diagnostic solutions establishing validation standards. This is where we are right now. This is what we can do safely. uh and then we can go

00:18:14 - 00:19:33

to level three and expand our horizon to non-gynecologic uh samples and advanced AI. Um so the the transitioning to non uh gynecologic that's going to be um it has begun but it has challenges regarding diverse specimen preparation and unique characteristics. So like check out how these uh cytology specimen look and they are challenging for human eye. So if they are challenging for human eye um then they're going to be challenging for computer vision. But what is happening right now in this level three there is

00:18:55 - 00:20:17

three there is automated quantification and quality control. So um there for specifically for uh cervical screening there is already a workflow that um has like a triage built in and there is a cytochnologist looking at it and then uh the pathologist for the uh the ones that are suspicious and now um you can put AI on top of that and do a 100% quality control uh of the negative samples and catch anything that actually was not negative, was false negative, right? Um and uh if we could do if we could

00:19:36 - 00:20:44

achieve total workflow automation, that would be amazing as well. Let me know if you have any questions. Let me know if you're actually doing cytology, if you're looking for solutions for cytoathology. um you will need to be more creative than in anatomic pathology. In anatomic pathology you have a lot of cleared devices. Uh here we just have one and just for a specific use case. So um the important thing is that it's for a specific use case, right? That it's cleared. Uh and then if you want to use

00:20:13 - 00:21:23

it for other use cases, if you see that it's working for other use cases, you can do it. But then um you're kind of uh doing an off label use of the device and you can still validate it. You can still figure it out um in a through the clea uh framework or whatever regulatory frame framework you're working through. Uh you can do it. So like don't think that oh if there is no clear device or approved device we cannot do it. Yes, we can but then we switch the regulatory framework and we have to be compliant um

00:20:47 - 00:22:30

whenever patient health is uh considered right. Okay, that being said, let us move to the next paper. That was our second out of four. Please, I can scroll off screen. Okay, this paper, navigating the labyrinth of hepatitis Cellular carcinoma, leveraging AI and machine learning for precision oncology. Um this is more how AI is being used in oncology. Um so the authors Manan and Ilas are telling us that the hypatic cellular carcinoma is a significant global health challenge and we're going to see it in

00:21:41 - 00:23:04

the uh oh do you see it? Am I sharing it? Come on, now we are seeing it. Okay. Yeah, you need to let me know if there's no audio and no screen immediately into the chat. Uh because I don't want to waste your time. Okay. I'm like here speaking from my uh from my tablet and nobody's seeing it. Okay. Um, I'll wait for the comment that you see the screen. I see it on my end right now. Um, when you see it, let me know. Okay. No. Yes or no screen? Okay. Let me stop sharing and start

00:22:35 - 00:23:55

again. Okay. I should be full screen now. Now I'm trying to share and I should be visible with my paper. Let me know in the comments if you see it. And it may be that I like need to wait a second or two for your comment to go through. Okay, perfect. Thank you so much. Thank you for prompt reactions. We can move on. So let me start again. What are we talking about? We are navigating the labyrinth of hpatic cellular carcinoma leveraging AI and machine learning for precision oncology and habitatellar carcinoma is a

00:23:21 - 00:24:39

significant global health challenge. We're going to see it on that infographic like how much they are predicting it's going to rise in the next 20 years. uh I was very surprised by how much and then the problem here therapeutic efficacy and advanced stages often is often limited by underlying liver dysfunction and adaptive resistance. So this is a review reviewing the landscape of molecular targets and combinate com combinatorial strategies um and they critically examine it. they focus on the transition from preclinical

00:23:59 - 00:25:28

discovery to clinical application. Um molecular they they uh take into consideration the molecular heterogenity. Um and what they say here is that computational paradigms are redefining target discovery and therapeutic stratification in hpostoellular carcinoma and uh they see the role of AI and machine learning as integrative tool for translating highdimensional multiomics data into insights for uh HCC management. Um and also they described AI driven frameworks to analyze complex data set derived from genomics,

00:24:44 - 00:26:08

transcrytoics, proteomics, metabolomics and epigenomics. Uh so we are doing multimodel here very much. Uh this is now relevant for um cancer therapies different cancer therapies in this case habit cellular carcinoma. Um and they are also um discussing can AI enable identification of novel predictive biomarkers, patient subgroups and rational uh drug combinations. Um they acknowledge the convergence of AI with spatial transcripttoics, digital pathology, single cell technologies. H so pretty comprehensive review. um if

00:25:26 - 00:26:43

it's going to be very clinically focused and oncology focused but u we're going to focus in the infographic on the AI part and this infrastructure um that we have the AI infrastructure it can be used for decoding tumor micro environment interactions and spatial heterogenity and tumor micro environment interactions this is this knowledge about them is non-negotiable for modern immune oncology um and uh what they claim is AI enabled multiomics driven approaches are instrumental in advancing uh HCC

00:26:04 - 00:27:46

treatment. So let's dive into the infographic here. Why is it not making me big when I change these? Maybe in another 50 digipath digests I'll figure out the tech, but then I will change the computer and it's going to be different again. But here what I was surprised that they have this projection of 55% increase by 2040. Like that's a lot. Why? Uh so they say that they they are diverse ethologic drivers um viral alcohol metabolic and non-alcoholic um hypotic stattosis and also aphletoxins

00:27:04 - 00:28:40

right so um different things different uh molecular drivers of these problems of of of the cancer itself. Um and they uh have this timeline about uh management of this disease. Um where we are now present 2020 um what we have is immunoncology era combination therapies and then in the future they are envisioning gene therapy era with different gene therapy modalities. Um but what I want to focus on let's see if I can make it even better for us. Can I do this? Uh, it's not showing the way I want it

00:27:58 - 00:29:42

to show. That's okay. I can still do it. And what I want to focus on is this lower part. Let's do this and then we're going to move on to our next paper. But uh yeah, so here um there is this AIdriven precision pipeline. What do they mean by that? Uh there is the data acquisition part. I'm wondering if this is even anatomically correct. I have to Well, there is a gallbladder. Um, that's a digression about the quality of the infographic. Uh, I'm going to show you where AI did some funny things. Um,

00:28:58 - 00:30:44

so we have data acquisition, right? Uh, clinical data, multiomics, radiomics. Um, and then we have this AI uh architectures and training. we have convolutional neural networks that um are good for imaging. Then for multimodel we have some other deep neural networks. Um and we can then what we have um unified molecular profile and we can do a lesion detection with this and also um predictions. so-called virtual biopsies. Um, and the decisions that we are trying to get from this is as we said lesion detection for

00:29:51 - 00:31:59

maybe radiology or pathology as well. Um, what's going to be the PR transplant recurrence, immunotherapy response and five-year survival prediction. So, um, this is what would be amazing to achieve with this Aadriven precision pipeline. And let's look at our last paper. Last paper I left the very actionable thing h as our last thing to discuss because that is basically going to affect all of us me including all the time. Um, let's see. Yes. Um, this one. Okay. Uh what we have here is

00:31:04 - 00:32:43

the curriculum framework for artificial intelligence literacy in veterary education and it's in come on highlight yourself literacy and venerary education you can basically substitute this uh with anything else with uh medical educ education with well there are specific things that are unique for veterinarians here and the authors mention it. Um so the authors eating hang and candis chew and I want to show you um I want to highlight that last author. Let me share something else with you. Who is Candice Chu?

00:31:57 - 00:33:39

Kis Chu is a fellow veterinarian, fellow veterary pathologist. She's a clinical uh pathologist and she's an assistant professor at Texas A&M University. um we briefly met um we were like colleagues at some point and then um when so she entered at the place uh where I'm working but then she uh did so much amazing work in veterary medicine on um AI AI literacy AI applications um and I also uh had her on the podcast let's see if she's going to join me again um about this particular paper but

00:32:46 - 00:34:13

she is so involved in uh teaching AI to veterary professionals. So I just wanted to highlight her because I think her work is really remarkable and also not only this she's uh very active in promoting her work through social media platforms um on Instagram you can find her on Instagram you can find her uh on LinkedIn and other platforms let's see here she's LinkedIn Facebook uh Instagram obviously Google Scholar you going to have a look at uh what she's publishing. But let's go into the paper.

00:33:39 - 00:35:08

I need to be with this sharing. And then we have an infographic uh about the paper as well. uh but what um the authors are saying in this um particular paper is that AI is rapidly uh evolving very fast um and um it is increasingly incorporated into veterary medicine. So veterary medicine let me tell you how it works. It's a lot less regulated than uh human medicine. So application of different things are going to happen faster there. So digital pathology the first system deployed uh in like a corporate

00:34:24 - 00:35:54

environment was at the veterary pathology uh laboratory at scale. They tested everything and then only two years later they actually um set up all the trials and filed for FDA um clearance of the system or approval or whatever that was. But uh so so the work the cuttingedge technology application in veterary medicine is happening a lot faster because there's less regulations but uh the literacy of veterary professionals in this space need to be really high because they are still liable for whatever the outputs of AI are and they

00:35:09 - 00:36:18

are highlighting this uh very much in the paper and in their curriculum. So, uh given the demand for a integration and accountability placed on licensed veterinarians, uh there is an urgent need to integrate AI literacy into the doctors of veterary medicine curriculum. Um so they're talking about US uh but it is applicable everywhere. Five core modules for a comprehensive AI literacy curricul curriculum for DVM students can serve as a foundation for institution specific course development. So you know you can

00:35:44 - 00:36:55

take her framework and basically develop a course at your university and that's what they would like you to do. They urge institutions to provide the resources outlined in this article to facilitate the development of a literacy courses in veterary schools. And let me tell you, it's not like a superficial thing. It's deep. I believe that if I would um I would take part in this course, I would learn a lot even though I'm here describing uh discussing papers with you. So, um, let's move to the

00:36:21 - 00:37:48

infographic because it's super cool how they design it and really it is deep and it prepares you. Okay. So, here we have it. Can I draw on it? Yes. Five model framework for venary education. This is from April last month, right? Um and I saw it on LinkedIn on Candice profile and I'm like I need to discuss this one. Um high quality work, great design curriculum. Uh so again we have layer one. Layer one is they call it the physiology or fundamentals of AI. Um what are we learning here? Here we are

00:37:17 - 00:38:35

learning about the uh architecture of different frameworks right uh they they are uh discussing how uh convolutional neural networks c pathology um and they because the clinicians must understand like they compare it to cellular physiology uh they must understand the normal what's happening in the cells to see the abnormal and basically this is the uh the premise like if you're going to be using these tools and if you're going to be reliable um liable for uh what the output is, you need to be able to detect when it's not

00:37:58 - 00:39:14

working well, how are you going to detect it? You need to know how it works. So this is module one. Then layer two, module two, the mechanics, machine learning and evaluation. And in this course they actually make at Texas&M they make students like design an image analysis application. So what they are doing here is handson with YOLO. YOLO is an image analysis framework. You only look once architecture for real-time image recognition and classification. They use the canine lymphoma uh application

00:38:34 - 00:39:56

canine lymphoma as the um entity that they're working with. I promised you that I'm going to show you where the infographic didn't go so well here. Like what is this FGTL? I assume it was supposed to be fail. Let me show it to you. This is just the fun part. What did Notebook LM do? Not so fantastic on this infographic here. This one. And there is something else here. Um, I don't know what this image is. Uh, but also the the the text says research byes and here's like research oes.

00:39:21 - 00:40:39

So trust but verify and that's very uh very fitting to this particular a literacy paper. Okay, that was just my me digressing on my notebook LM output. Let's go back to the content. So we were layer two, right? Um then what they say because you're going to have a company coming to you and saying, "Oh, we can do this that." So you actually have to be able to integrate the shiny brochure. Um and then uh you have to check the performance metric versus real world bias. And it's very evident in veterary

00:40:01 - 00:41:04

pathology. You can see like this dog is thin and doesn't have so much uh fur and the other dog that the h image and or whatever the algorithm was deployed on samples from another dog or images from another dog looked totally different. So like you have these breeds of dogs or different animals and if it was developed on one there is no guarantee that it's going to um be working on the other one. It's it's similar to different demographics in human medicine. And then what are we doing in

00:40:32 - 00:41:41

layer three module three and there is the large language models and prompting there's interacting with AI there's going to be a lot of text AI because there's going to be a lot of AI that um is going to be workflow focused speaking focused like um ambient scribe is one of them or there there are veterary um versions of ambient scribe where you talk to the owner of the animal h and the clinical notes are being created from this conversation. Um so um they emphasize that large language models are

00:41:07 - 00:42:20

pro probabilistic autocomplete. They teach you what it is. They it's not a person that you're talking to. Uh right and also they dive into so-called vibe coding and agents. I saw that and I'm like they are cutting edge because they're doing what's like really happening right now. this these CNN's were happening like what three four five years ago maybe that was the beginning and right and uh then now we're talking about vibe coding so vibe coding is um just like speech coding you can create

00:41:44 - 00:42:56

apps talking to an app creation platform you don't need to know how to code you probably need to speak English uh but I bet there are already vibe coding u platforms in different languages um so you can like code an app for your practice. Uh but then you're still somebody liable for the outcome of this app. So they're learning that and a agents workflow orchestration or orchestration of different steps of AI. So now we've learned all these things even the newest uh hypes vibe coding we

00:42:19 - 00:43:41

can make an app. But what about the guard rails? Ethics and liability. This is covered in module 4. Um and here I want to highlight something. There is um there are no FDA pre-market requirements for this uh AI software for a veterary medicine. for veterary uh drugs, veterary um medications, there's normal like FDA uh approval and this has to go through a specific drug development process. Uh in human medicine, there's this framework of software as a medical device. It's a medical device and then uh FDA regulates

00:43:01 - 00:44:15

these medical devices when they become products. In veterary medicine, nope. That means for the companies that they're providing this fantastic they can like that's their playground for the uh veterinarian very very high liability of something that um maybe is not that thoroughly tested yet right so they emphasize the human in the loop requirement um because the um veterary medical association in uh American veterary medical association so they are talking about the US um says AI is strictly assistive.

00:43:37 - 00:45:08

The licensed veterinarian is 100% legally accountable for every diagnosis and medical decision. Right? So if something goes wrong, your beautiful vibe coded up. Um if it didn't work well, it's your fault. Um so you need to learn about ethics and liability. And then uh layer five, module five is the um research and academic integrity for uh scientific writing for synthesizing and and I have a question here. How is vibe coding relevant to pathologist or modern pathologist coding? So, VIP coding is like creating an app um with

00:44:25 - 00:45:49

your words and what what can this app be? Maybe an app to output um pathology like um clinic notes. I don't know if there's um something like this in medical pathology. Definitely in veterary pathology if you're uh or venary practice actually broader uh you can design an app you can design an app or like I would say maybe you can even u if it's workflow if it's not like uh patient decisions you can design uh the app for be for like family medicine maybe this is me guessing right u

00:45:07 - 00:46:20

because the question is how Is this uh relevant in our modern pathologist coding? I'm not coding, but I did uh do try uh to vibe code an application. It was just like a web app to uh offer my book by the way, digital pathology 101. Uh so yeah, I don't know if that answers the question. Let me know if you want to dive deeper. Uh let me know in the chat. But I thought this was um a a really interesting addition to this paper and that's also uh because this is for veterary medicine not just for veterary

00:45:44 - 00:47:08

pathology. she's teaching at a veterary college, right? So that would be like something for medical students, uh any kind of healthcare science students, um life science students. Um okay. And what they say about this layer five synthesis uh research academic integrity um the context window danger. uh we probably already heard about hallucinations multiple times fake citations there were papers pulled from nature or high I don't know if it's nature but very high impact factor journals for fake citations h but also

00:46:26 - 00:47:41

another thing is blend conflicting research papers into a single force false conclusion in the document length um if the document length exceeds their context window. So context window is how much information can uh a large language model like see at a time. How much um can the transformer kind of like oversee at the time and if the length is longer then it only sees parts of it and is um drawing conclusion on the part and um they are also teaching different tools AI enabled literature synthesis. uh what

00:47:04 - 00:48:21

they say there is that uh staying on top of literature manually now is very challenging if not impossible. A and they do use tools like notebook LM uh which I am using as well uh research bites to efficiently manage their rapidly expanding volume uh of veterary literature while maintaining authorship integrity. Um, Candice also has this um, writing framework, AI uh, powered writing framework that she gives workshops on and if you go on her social media, you're going to find her framework. You can download it. So, um,

00:47:42 - 00:49:06

I love this one. I'm wondering if anything like that is happening uh in medical colleges um or medical associations. Yes, probably. Yes. If you know like a specific course, you can send me a link in the chat um here even if you're watching the recording because I would like to just learn more. Okay. what is the the program official program by official associations and I'm going to do my research as well uh but if you know something if you have taken some courses that uh were good um I

00:48:25 - 00:49:37

obviously have my digital pathology 101 book that is for those who are starting the um digital pathology journey and there is a um a chapter on AI in the new version that is almost finished. By the way, if you want the new version, h get the book. The book you can find it on, uh digital pathology place.com. Let me put it in the chat. Um you go on digital pathologyplace.com and you're going to uh there's going to be like a popup. Uh you're going to find this book. you can download the PDF for

00:49:06 - 00:50:19

free and then I'm going to let you know every time I go live. Uh so uh but also you can see this this edition is from 2023 and I'm currently working on 2026 edition. It's almost ready. The PDF is going to be the first one to go out. So if you're already on the list, if you have the book, then you're going to automatically get the new version. Uh, I will be bragging about it on social media as much as I can once it's out. Um, but you can already be on the list to get the new version. Thank you so much for

00:49:43 - 00:49:54

joining me today. I appreciate you so much and I talk to you in the next episode.