Digital Pathology Podcast

241: Foundation Models in Pathology: Strong on Paper, Ready for Labs?

• Aleksandra Zuraw, DVM, PhD • Episode 241

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 42:07

Send us Fan Mail

Are pathology foundation models actually ready for labs, or are they still stronger on paper than in practice?

In this episode of DigiPath Digest #49, I unpack a timely review on pathology foundation models and ask the question that matters most to me: not just what these models can do, but what has to be true before they are genuinely useful in real pathology workflows.

I walk through how pathology AI moved from narrow, task-specific models into the era of transformer-based foundation models. That shift matters because pathology is no longer only about looking at H&E in isolation. Today, pathologists are expected to integrate morphology, immunohistochemistry, molecular assays, genomics, and clinical context. That growing complexity is one reason foundation models are getting so much attention.

In this discussion, I explain how transformers entered pathology, why image patches are treated like tokens, and how shared embeddings can support classification, regression, segmentation, and multimodal retrieval. I also go through the major pathology foundation models mentioned in the paper, including Virchow/Virchow2, Mayo Clinic Atlas, UNI, CONCH, H-Optimus, GigaPath, and TITAN, and why scale alone is not the full story.

A big part of this episode is about the gap between benchmark performance and clinical readiness. I talk about the persistent limitations in training data diversity, the overuse of TCGA, and why public benchmarks can still miss what real pathology practice looks like. I also cover where foundation models still struggle, especially in cytopathology, hematopathology, and underrepresented disease areas, along with the real-world problems of artifacts, domain shift, concept drift, infrastructure burden, regulatory complexity, and workflow disruption.

For me, one of the most important themes is this: AI in pathology should augment, not replace, pathologists. The future is not about handing diagnosis to a model. It is about building tools that support pathologists better, fit real workflows, and can be validated in ways that deserve trust.

I also spend time on what comes next: explainable AI, counterfactual explanations, conversational interfaces, retrieval-augmented systems, multimodal fusion, and the need for deployment-centric validation rather than paper-only excitement.

If you are trying to understand where pathology foundation models really stand today, this episode will help you separate the promise from the practical barriers.

Episode Highlights

00:01 – Why I chose this paper, what is changing at Digital Pathology Place, and why foundation models are worth paying attention to now.

02:15 – The core questions: what pathology foundation models are, where they are, and how difficult they are to apply in pathology.

04:50 – Why pathology is becoming more cognitively demanding, and how multimodal complexity is driving interest in scalable AI.

07:02 – From narrow AI to transformers: how pathology moved beyond single-task CNN models.

10:16 – How transformers work in pathology: image patches as tokens, self-attention, embeddings, and downstream tasks.

14:16 – Why multimodality matters, and what kinds of data foundation models may eventually integrate.

15:27 – Timeline of key model developments, from “Attention Is All You Need” to gigapixel-scale pathology foundation models.

17:13 – The leading models and what scale really looks like: Virchow, Mayo Clinic Atlas, UNI, CONCH, H-Optimus, and GigaPath.

19:51 – Why dataset diversity matters more than sheer volume, and why TCGA is not enough.

23:17 – Where foundation models still struggle: cytopathology, hematopathology, rare disease, artifacts, scanner shifts, and pen marks.

28:06 – Explainability, counterfactual explanations, and why trust in pathology AI needs more than attention maps.

30:17 – The real deployment hurdles: regulation, infrastructure, workflow fit, and economics.

36:32 – Why AI should augment pathologists, not replace them, and which tedious tasks pathologists would gladly hand over.

38:36 – Retrieval-augmented and conversational AI in pathology: where interactive systems may actually help.

40:51 – Vision-language models and multimodal fusion with histology, radiology, genomics, and clinical notes.

42:16 – The path forward: deployment-centric design, prospective multi-site validation, and human-AI collaboration.

44:08 – Closing thoughts on AI literacy, community learning, and what needs to happen next.

Resources Mentioned

Support the show

Get the "Digital Pathology 101" FREE E-book and join us!

00:00:01

 Good morning, my trailblazers. Good morning from Pennsylvania, 6:00 a.m. on Friday. What's Yeah. 6:00 a.m., my friends. Uh, if anybody is here, it means you are real digital pathology trailblazers. I'm going to drink my coffee in my Polish cup because this week we're going to Poland with the whole family. So, the next month, we're going to skip the live streams um cuz I'm going to be working on something exciting in Poland. Uh, what I want to do is to revive the membership where all the digital pathology


00:00:50

 resources that we created live and mm design a better learning experience. Uh, I Uh, so what happened is I stopped sending emails every live stream and >> [clears throat] >> uh because it was a lot of emails and we realized that when I posted on social media, also the same amount of people knows about the digital pathology live streams, about our journal clubs. So, we stopped bombarding the email list, but then I started getting emails, "Hey, put me back on the list. I don't know if you're doing the


00:01:31

 paper reviews or not." So, I'm going to be redesigning that a little bit and next month, while in Poland, I'm going to focus on finally finishing the book, Digital Pathology 101, All You Need to Know to Start and Continue Your Digital Pathology Journey, which you can get from the digitalpathologyplace.com book for free as PDF. So, if you don't have the book yet, go get the book. And today in today we're going to review a paper. Well, a paper and a topic about pathology foundation


00:02:15

 models. So, the original paper cover is this one. Let me share it with you. And yeah, what about these foundation models? What about the foundation models? Like so much hype, where are they? And I'm going to tell you where they are based on the paper and based for on some information that I have from other digital pathology trailblazers on the vendor side. And but we're going to be discussing pathology foundation models evolution, current landscape, challenges, and opportunities and clinical perspective. Opportunities from


00:03:07

 a technical and clinical perspective. That's what we're talking about today. If you're here at 6:00 a.m. 6:03 in Pennsylvania, let me know in the chat. I'm going to say hi. Let's see. Does it work? Yes, my messages work. And let's see. Yesterday I was trying to like do some um some sound effects. So, if there is going to be a comment, somebody's going to talk to me in the chat, then I'm going to do some sound effects. But, now let's see what about these foundational models,


00:03:59

 right? What about them? Where are they? What's happening? Like, how difficult is it to use them? Layout. Okay, here. This how difficult is it to use them? Where like I love them, or I could potentially love them, but where are they? How can I use them? Some are open source, some are not. Uh, so that's what we're going to be talking about about the evolution and current landscape. Obviously, there are challenges and opportunities. So, let's get into it. Why? Why? Why did they even


00:04:50

 start existing? Well, uh, first of all, I bet it was fun for computational pathologists, for computational pathology scientists, to start using the newest methodology. And we're going to talk, uh, about it. But, from the like, global perspective, uh, the rising volumes of anatomic pathology cases, right? Contemporary anatomic pathology faces increasing case volumes. And then, not only this, ancillary studies, routine specimens, now require require other types of testing, like molecular assays, and obviously,


00:05:31

 immunohistochemistry. So, it's not just H&E, we know that. Uh, cognitive load, because of this, of these ancillary studies, the pathologist is now not only looking at slide, recognizing a pattern, saying this or that, or um which is also different from uh, what it was when they they trained because the WHO changed the criteria. That's another topic, but basically the cognitive load is pretty high because multimodal data integration increases turnaround time and interpretative burden for


00:06:06

 pathologists. It's difficult uh including like all the other ancillary studies. So, there is is uh a need to shift from purely morphological assessment to uh systematic um high variable complex data. And the leverage can be pushed by scalable AI and pathology foundation models. Where is my ink so so thin? This is better. Yes. This is better. Ooh, okay. Okay. I got the chance to do the clapping. [applause] Yes, because we got the we got the comment. [applause] Hello Scott. Welcome back. Welcome back.


00:07:02

 Amazing to have you here. So, if you're just joining, let me know in the comments. And if you have specific questions, let me know as well. So, what happened? Like what was the evolution? The evolution was um in the past um we had very task-specific models and now we have we are supposed to have task-agnostic, meaning it's going to do it all like a pathologist. It's supposed to be brilliant. So, before we had the early models when um convolutional networks convolutional neural networks entered the scene, I was


00:07:40

 ecstatic because I didn't have to like deliberate with my image analysis uh colleagues, what is the threshold and where do we put the threshold because if we put it too high for for example and DIB stating in IHC, then we lose uh faint cells. If you put it too low, then we get uh false positive cells. Where where do we put it? And you couldn't put it all the time in the same place and then CNNs appeared and that was a miracle. Uh but CNNs uh would give you a single task output. So, that's a little


00:08:15

 bit limiting because then for every single pathology task, the narrow AI is going to just do this one task. So, there needed to be something better. Well, there wasn't the technology to have something better, but once the technology, also known as transformers, appeared, um we And by we, I mean the broad digital pathology and computational pathology community uh and specifically the computational arm of this community, um they started working with uh transformers. So, uh and uh transformers are based on


00:08:57

 self-attention mechanism. So, they like learn from the image what's happening in the image. So, they can perform segmentation, classification, they can even tell you about prognosis, they can extract features, predict outcomes, and work on biomarker discovery, and every now and then you're going to have a paper that oh, there was this data set. We already had it analyzed but didn't tell the AI and the AI figured out what the results were and like discovered you know, we know the word validation,


00:09:33

 qualification, and validation too well uh in the digital pathology space, but the I'm always excited about this potential, right? So, the self-attention mechanisms, they capture a long range of contextual dependencies without handcrafted features. And now we can input the huge whole-slide image repositories. So, that led to a breakthrough. The pathology foundation models, they learn to generalize and then would learn these generalizable representations that are reusable from massive whole-slide imaging repositories.


00:10:16

 I don't know how these people get the amount of slides necessary to train them, but they do get them. The main repository that is accessible for everyone is still the Cancer Genome Atlas, the TCGA, which has its own challenges, but it has just like I don't know, tens of thousands. At some point I knew the exact number, but the numbers necessary for these models are bigger, a lot bigger. But, okay, so how do you encode the whole-slide image? Because the transformers, they came from text and you don't have text. And you encode


00:10:57

 the text in tokens, which are not words, but like little pieces of text that contain some meaning, right? So, what are our tokens? What are our tokens? Pathology tokens? They are going to be patches. And patches is nothing new in digital pathology because patches and patching and like dividing these huge images into smaller pieces was necessary for any image analysis task. So, actually like the the the patches tokens for text, that was a new concept. But the had patches. So, these are our tokens now.


00:11:35

 Um, and then we have the transformer encoder. Um, it's called multi-head. Well, it is described as multi-head self-attention models. Um, the the that can um, like they are multi-contextual. They they can um, long-range contextual relationships between patches. Um, so, the key information here is that they can do um, this between patches. So, um, like before, I think they still use the computer vision all the computer vision capabilities like edge detection and all that stuff. But now, um, they're even


00:12:19

 better, right? Uh, and then then we have something called shared embedding embedding. So, this produces a compact feature vector representing the entire image. So, um, basically, we're encoding the patches into a vector of numbers. And what can we do with it? A lot. And this is amazing that we can do so much. We can do classification. We can do tumor subtype. We can do grading. Um, so, the classical computer vision tasks that you do. Um, and they're described in the book as well. Um, so, grab the book if you don't have


00:12:58

 it. digitalpathologyplace.com. You're going to find it there right now. Um, then regression. So, the regression we're going to have the survival prediction, biomarker scores, segmentation, tumor regions. So, what are we doing here? Dividing tumor from stroma. Super important for Let me put myself like this. Uh, like this. I think this is going to be best. Um, segmentation tumor regions. Um, tumor from stroma, so the tumor epithelium from stroma, and then finding nuclei there that's going to be useful for


00:13:36

 immuno-oncology applications. And we can also do multimodal retrieval. So, something I heard about uh in 2019 from a Google representative at PathVision. So, 2019, 7 years ago, right? Uh I heard about Smiley, similar image retrieval, whatever, right? And there is a multitude of these um abbreviations that are invented for every paper to sound funny. And Smiley was something I was hopeful about. I'm like, "Put it on the web." And I don't know where they put it, somewhere on GitHub uh or um


00:14:16

 somewhere where the computational arm of computational pathology has access to. Um but basically, this was image retrieval where you could like paste an image and then retrieve similar cases, which is like a mainstream thing for um for consumer AI where you can like crop an image. Go on Pinterest. I go on Pinterest like there's a cool piece of clothing, like a nice dress, and I I crop it and it shows me a bunch of other dresses like this. So, this multimodal retrieval um is uh for searching similar cases. And they can


00:14:56

 also be linked to genomics and clinical text. So, a bunch of different types of data. Um and just not to overwhelm you. There's not so many types of data. There's image, there's text, there's probably voice that gets translated to text, and video and tables. So, that's that. And video is image as well, a version of image. So. But it's multimodal. The more uh if it's more than one, it's already multimodal cuz you have two modalities. By the way, how do you like my


00:15:27

 Trailblazer mug? Okay. So, let's look at the timeline um 2 gigapixel scale, which is where we are right now. Um the transformer situation started with in 2017, there was this paper uh attention is all you need. I think it was by Google as well, and they described this self-attention mechanism. >> [clears throat] >> And then uh 2020 2021, vision transformers uh and data-efficient efficient image transformers, and they were enable uh they they enabled the scalability to whole-slide images. And 2022,


00:16:24

 we had something called masked autoencoder that allows for self-supervised pretraining, and here we are more or less we're already 1 year past 2025, and we had the emergence of gigapixel scale foundation models trained on millions of whole-slide images. I think I had some videos or information on this when this first came came out because I'm super excited by these new developments. So, um leading models by the numbers, we have vision-only models um VirCo VirCo 2. I think they came out from the Dr. Fox lab.


00:17:13

 Um don't quote me on that. I think so. And but And then we have Mayo Clinic Atlas. You can quote me that this one came out for Mayo Clinic by the name. And then we have Uni. And the important thing is how many images were they trained on. So, over a million for Virchow. Uh 1.2 million for Mayo Atlas. And then Uni was on tiles. So, this is whole slide images. The 1 million for Virchow and 1 million over 1 million 1.2 for my Mayo Clinic whole slide images not patches. So, because like the trick when you read


00:17:51

 these papers is like so and so many images were used. And by image they mean a patch. So, you're going to create a bunch of patches from one whole slide image. So, they always say, "Oh, so so many patches." And like 20 whole slide images. I'm like, "Really?" Uh but here actually 1 million of whole slide images. I don't know where they come from. Um from the hospitals. Then So, these were vision only. Then multimodal vision and language we have Conch. Um and this is over 1 million


00:18:27

 whole slide images. Um and it is This Conch is with paired text superior for complex biomarker like microsatellite instability. So, you have the report. You have genomics that says that this particular slide has microsatellite instability. So, we are entering the space of molecular predictions from H&E which is also something I'm super excited about and I want to see it being deployed. Um and I'm super excited about any like case report or mention of um this modality actually getting closer to clinic. And


00:19:07

 then we have other ones um that they're called so-called hierarchical where they work for their different tasks at different levels of the image. And here we have H Optimus and GigaPath. So these are the main players. There is more. And probably as we speak, they're developing even more. So have you used any of them? Do you have like access to any of them? Let me know in the chat. And so how do we evaluate clinical readiness of these? Well, the golden rule would be the pre-training data


00:19:51

 diversity. So what do we mean by that? We want to have I'm just checking if I'm not frozen. Yesterday I was streaming about histology and I was frozen. So the golden rule, pre-training data diversity. The problem is one of the problems is that the data used previously for training AI was not that diverse. So in the paper they cite that the TCGA, 80% of this data set is from patients of European descent. So no data on Asian, black, indigenous, or any other population. It was mostly European. So


00:20:35

 here we the golden rule is to have pre-training data diversity, different anatomical sites, cancer types. This outweighs the sheer data volume for downstream success. And here the top performers work out two, Conch, H Optimus one consistently lead across 60 plus clinical tasks on standardized benchmarks. And we have two benchmarks mentioned in the paper. Um, there are more path bench and path of bench. The path of bench is specific for foundational models and um what is cool about these models because


00:21:19

 they're like huge, big billions of parameters which I don't even know like, you know, I come from the classical uh handcrafted feature computer vision tasks like how dark should a nucleus be to classify it as positive. So, I I'm like uh I was trained on sliding one parameter at a time. These have billions of parameters. So, it's beyond my comprehension, but so like adapting every one of these is a no-go. This is not happening, but there is an option to efficiently adapt them. Um,


00:21:57

 they call it parameter efficient fine-tuning uh and one of these frameworks is called Laura. And this adapts like one layer of these specific these models, massive models to specific tasks with minimal computational cost. So, they're like pre-trained on what is an image, pixels, like all this general computer vision knowledge, edge detection feature extraction, whatever, right? Um for the list, you go back to the paper. But, basically, they already like have the knowledge how to interpret image and


00:22:34

 then you train the top layer. That's that's how they refer to it to um train it for a specific task and make it a lot more efficient. But, where do they struggle? We have the They are not perfect, nor are the pathologists, but you know, I I'm a pathologist, so I like to believe we're more perfect than the foundation models, but if we can join forces, we're going to be even better. Mhm. So, uh where do they struggle? The two usual suspects, cytopathology and hematopathology. Why cytopathology?


00:23:17

 The disease stacking. Cytopathology, where for some areas of cytopathology, like Pap smears, or maybe even um like simple tasks for um blood smears, but that would be hematopathology. But everything that you can make flat and nice and unified, uh also um urinary bladder cytology, where you can do the thin prep, cytospin, and thin prep, they look nice, and they have been models designed for that. But any other thing from cytopathology, where you like squirt a fine needle aspirate on the glass slide, stain it, and then you have


00:24:00

 different levels of tissue, that's going to be challenging because um the models are trained on flat anatomic pathology images. And then hematopathology has a different problem because it relies heavily on multimodal ancillary studies, and uh these are flow cytometry and cytogenetics. So, if a model was not trained with this data, you cannot deploy it on a hematopathology. Basically, well, the the main thing is these specialties, or these slides, look different than the training data than the uh


00:24:41

 million whole slide image models were trained on. So, you just didn't train on this. If we had a model trained on millions of cytopathology images, maybe we could be better there, and the same for hematopathology. So, um yeah, the gap is that AI development has largely focused on common tumors and leaving rare disease and complex preparation under represented. Why? Because there's less data. You go where the data is to develop these methods and you check is your method good enough for the data


00:25:15

 that you have and then you try to deploy it or like train it or get more data for the places where it's under performing. Um there may be enough data, there may not be enough data, but then like you will have underperformance on the missing data. And what else is problematic? Something artifacts and concept drift. Um so domain shifts and shortcut learning. Do you see this pen here? I heard about this pen in 2019. Uh the first time and I was like, "Really?" Yes, really. So models fails fail due to fixation


00:26:05

 variations. So the domain shift is like that the slides look different from different places from batch to batch. We also have scanner artifacts and the models are exploiting non-biological features, pen marks. My friends, pen marks, my trailblazers. Where do these come from? Where a pathologist marks something under the microscope because how else are you going to find it or show it to somebody else if you don't have a pen mark on the slide and I used to do this too when I was not working on digital slides. Now I have


00:26:40

 digital pens, which I love. Really like thank goodness for digital pathology. It's um game changer for my life. I love it. You know that I love it because I talk about it all the time. Um So, yeah. But, it's not perfect, nor are we. There's also something called concept drift. I briefly mentioned it, uh evolving diagnostic criteria, evolving clinical criteria. Uh so, the guideline updates, the uh WHO guideline updates, they cause performance drops over time. Um not because the performance changes, but because uh


00:27:28

 now the labels or the reports or the information used to train the models is no longer the same information that we're applying to. So, like if there is a specific uh diagnostic term that changed, now the label changes. So, the performance drops because the model doesn't deliver the uh new label anymore. Anyway, so there is a solution, they propose a solution, a multi-site validation, and explainable AI. And there are different concepts to have uh diff- different [clears throat] like ways to have AI explainable. Um


00:28:06

 what uh they talk about is something called counterfactual explanations are required to diagnose blast black box fail- failures. Um and I liked this way of explaining, maybe because I didn't know it before. You know, I knew the um attention maps where the model shows where they took the information from, but this one uh is even better because it's shows you this the places where if the evaluation would change, and whatever the evaluation is, like scoring or some kind of like parameter calculation would change by a certain


00:28:48

 percentage that diagnosis the label or the task would be performed differently. So, let's just take a nucleus detection something so to explain okay, why am I detecting nuclei because if something would change in this area of the nucleus by so and so many percent I would not classify it as a nucleus and therefore this is where I take this information from. Um you know, one of the ways to check it and there's a whole branch of science called explainable AI which they abbreviate XAI like X archives


00:29:30

 Z or psi explainable AI big thing in clinical space in medicine because the black box problem is bothering medical community very much. So, we're trying to explain the black box always when we can. But we also have I call them so whole so called life hurdles. So, now you have this amazing model. It lives somewhere in our imagination or in the internet cloud or wherever it lives. I cannot take it and deploy it on my whole slide images that I have in my collection. Uh and if I'm a if I'm operating in a


00:30:17

 regulatory environment you know, don't even touch it. Um so, there are life hurdles. Um and one of these hurdles is regulatory. Although I think in this particular case they throw the FDA frameworks under the bus a little bit too much. So, the authors of this paper say that FDA frameworks struggle to classify continuously learning or task agnostic systems. So, continuously learning, this I have a problem with this because when you submit something to the FDA, this version of the thing is not continuously


00:30:56

 learning. You can then train it and have a different version and FDA has frameworks for that now and that are um what do they call it like predetermined change control? Um If you know the exact name, let me know in the chat, but they they are like Let me tell you where all these frameworks and regulatory inputs are easy to find. There was this seven-part AI in pathology series um by Rashidi et al. and we had a full series of Digipath Digest on every single paper and there was this paper about regulatory aspect of all these AI


00:31:45

 evolutions in pathology. And the amount of frameworks from different regulatory bodies is huge. So, once I saw the table like, "Okay, in Europe, in this country, framework for this, for that, for risk, for fast approval, for this, that, whatever." I'm like, "FDA has some smart people, fast-thinking and fast-adapting people on their team." And obviously FDA is huge and other regulatory agencies as well because that's their goal to provide to to maintain the safety of these


00:32:20

 tools. So, I would say it is a hurdle, regulatory always is a hurdle because you have to put a lot of effort into regulate these things, which is good, and FDA is evolving. >> [clears throat] >> Although probably slower than the science, but that is how it's supposed to be. Like you're not going to regulate every scientific discovery. It has to be good enough. It has to be qualified, and there has to be a path forward for this. So, you know, it's a chain. Anyway, infrastructure.


00:33:02

 Infrastructure, I probably would not be able to deploy this um um model uh or train a model that is based on gigapixel images, and this demands high-performance GPUs, and then LIS integration, and you know, you can argue this is difficult, but you can do it, but cloud, whatever. It is a life obstacle that needs to be overcome in whichever way you want to overcome it. Infrastructure is infrastructure. Labs uh were not built with the main Sorry, the um newest infrastructure in place right now from the get-go, and they


00:33:44

 don't change things every week. So, interoperability, infrastructure is a hurdle. Then, workflows. Disrupting time-sensitive pathology workflows. We want to have nothing of that, and I experienced it firsthand every time uh I work with like a software upgrade, and then it has some glitch. So, that's tough, because as much as I love digital pathology, I don't love to be slowed down. And then, in many instances, and me, the digital pathology enthusiast that wrote the book digital pathology 101 is saying this. In


00:34:36

 many instances, if like it takes longer to open the image, if I had the glass slides, I would be putting it under the microscope and just reading it. But, you do either or the other. Some you know, in the transition periods you have both. Um anyway, the workflow, if you slow down the pathologist, you're slowing down care. You don't want to do that. So, um important, right? And economics, economics means you have to buy these things. So, every time you have to buy something, you have to figure out where


00:35:08

 the money comes from. And there are ways to get it, but they these are not the classical ways. So, obviously, lack of classical ways and the authors site reimbursement pathways uncertain. [clears throat] I think they're being generous. Lack of reimbursement pathways not for everything, right? So, you need to be creative for the economics to work or and wrap it up into a larger maybe enterprise deployment framework. Like, you need to think how the economics are going to work. And there is no guarantee that they're


00:35:49

 always going to work, right? Um and the last concept is we are augmenting, not replacing. And that's actually for everyone, I think, not just for pathologist. So, we're augmenting, not replacing. Not just for pathologist, for lab personnel as well. As with any change, as with any disruptive technology, and I consider the foundation models a pretty disruptive technology, there's fear. Hey, will my job is going to still be there in 5 years, 10 years? Uh and some jobs will probably this Well, no. Not jobs.


00:36:42

 Some tasks within a job will disappear. And I'm happy for them to disappear. Like, for example, as a pathologist, like if you make me count cells, I'm going to flat out refuse. If you make me guestimate a biomarker, ooh, that's going to cost me a lot of energy. So, if this task disappears for me, I'm all for it. Uh for for example, lab personnel, um I think if the task of cleaning slides and loading them into the scanner disappears because a robot is going to do that, I think they're going to be


00:37:21

 happy because they can um do then like more cognitively rewarding task than putting slides into the scan Well, I think I don't know if people still do that, but at some point in the digital pathology evolution, that was the case and um not everything is automated. But anyway, the point I want to convey is the jobs will change and you need to adapt. We need to adapt to change and that's life. Um so, but we don't want to replace anybody and the AI output uh serves as powerful decision support tool and not


00:37:59

 autonomous replacement. And when we say decision support tool, there is also caveats to that because there is something called um confirmation bias and different types of biases. So, everything is multifaceted, but um it's it's a positive change. Decision support tool, not autonomous replacement. Then uh pathologist refinement. where we'll have interactive frameworks, retrieval augmented and or conversational. This is fantastic option, I would say. Um so, how I would use it? Obviously, I


00:38:36

 would just chat with it and talk to it. I have voice-to-text software on my computer for every possible app. Uh so, for my pathology workflow, I would do the same and and retrieval of um images. That's so cool because um I remember, oh, in one study, I saw this type of lesion. What did I call it? What did I and and um toxicologic pathology, you have like a specific glossary. It's similar to clinical, but like there is a special list of terminologies, uh the international nomenclature of these


00:39:13

 preclinical terms. So, you want to stay consistent and but there is always more terms for one lesion. So, I would retrieve the image and check uh what I called it before. Um so, that would allow pathologists to query, refine, and validate outputs and continuous learning. Um interpretability maps uh model outputs to recognizable histologic features, for example, nuclear atypia uh they to build trust over time. So, interpretability is super important if that would be something that a pathologist can verify.


00:39:55

 Um that would help tremendously, I think, with uh regulators and with the um trust of the community in the validity uh of the tool and that uh feeds directly, of course, into that this would be decision support. So, it's a cycle and it has to be deployed responsibly. And what about converging clinical data? Um it would be fantastic to have this comprehensive disease modeling to include histopathology, radiology. These two are image-based, right? So, they're perfect for this. Then we have genomics and and


00:40:51

 clinical notes. So, here we have this multi-modality. And to align with real-world diagnostic reasoning models must must integrate beyond just image. So, the hope is in vision language models that offer a promising pathway towards fusing clinical data. Vision language. Would we cover everything with vision language? Yes, this is language. Sorry. Language, vision, vision, genomics. I'd say it's letters, so let's call it language. I don't know, it's a different, but letters are used for the


00:41:30

 basis. So, the authors authors propose a path forward. We had this shift the models have moved computational pathology from narrow algorithms to scalable reusable frameworks. So, here a point to consider. The narrow algorithms didn't disappear and they're still useful and they all these new technologies exist in parallel, right? The challenge We have a challenge because bridging the gap between controlled experimental benchmarks and highly variable real-world clinical workflows is challenging.


00:42:16

 Um So, it's like taking an experiment from a very clean sterile lab into the real world. Your sterile control conditions disappear, and that's the same on the computational pathology side. Um and they also propose a solution that the field must prioritize deployment-centric design, prospective multi-institutional trials, and human AI interactions. Deployment-centric design. This is good. Um prospective multi-institutional trials and human AI interactions. For this to actually reach the clinic, I


00:42:58

 like to call that I'd like to say that life is a funnel, right? You have to do a bunch of science at the basic level to see what sticks, what works, and then the stuff that works, okay, can you then think of deployment? It's like drug development, right? You have so many candidates and only like one or five make it to the bottom of the funnel and then are going to end up in patient treatment. You're not going to go and validate everything that's at the top of the funnel. It has to pass all


00:43:31

 the other steps to be at the bottom to actually do clinical trial on this particular drug candidate. So, here it's the same. Instead of drug candidates, we have like model candidates, algorithm candidates, or whatever the um computational equivalent is to this chemical compound that goes into a drug development pipeline. And then once it's far enough down the pipeline, let's do it. Let's prioritize deployment-centric design, multi- multi-site trials, and the human AI interaction. So, for this human AI


00:44:08

 interaction to be possible and to be useful. We didn't know we need to know enough about AI. Uh so, that's why you guys are here. That's why we're doing the Digital Pathology Digest. Uh and at the beginning, I said I'm going to be re- invi- reinventing the membership that we used to have like a year ago, maybe, 2 years ago. We have uh I had a lot of resources there for you to study. I am reinventing this membership so the learning experience is better. So, um that you have a


00:44:53

 very clear path forward and to learn about AI, to increase your AI literacy in pathology. There's also going to be a separate path for histology and tissue recognition. So, if you're interested in that, it's going to be in there. Um nothing like there is no launch date or anything, but I'm letting you know. So, if you're interested in learning more or letting me know your preferences, let me know in the comments. Send me an email. Just get hold of me and say, "Hey, you said membership is coming back. I'd like


00:45:25

 to be part of the new version." So, let's see if we can How can we finish this? Do we have any like nice sound effect? No, I'm just going to clap to everyone [applause] who was here today. You are the real Digital Pathology trailblazers. You are the reason I keep going. And any every time you listen to this information and carry it forward, you are advancing digital pathology. So, thank you so much for being here and I'll talk to you in the next episode.