168: Smarter Slides: How AI Is Reshaping Kidney, Thyroid & GI Pathology

If artificial intelligence can match—or even surpass—our diagnostic accuracy, what happens to the role of the pathologist?

That’s the question I explore in this episode of DigiPath Digest #30, where I break down three fascinating papers showing how AI is changing the way we diagnose, classify, and predict outcomes in renal transplant biopsies, thyroid cytology, and gastrointestinal cancers.

These studies don’t just prove AI’s potential—they reveal what it means for us, the humans behind the microscope.

Study 1 — Renal Transplant Biopsies: Precision in Every Pixel

A Japanese team examined how deep neural networks and large language models improve diagnostic consistency in renal transplant pathology.

They highlighted how the Banff Digital Pathology Working Group is retraining AI models alongside updated Banff classifications—creating a dynamic feedback loop between human expertise and machine learning.

In the U.S., over ten digital pathology systems are now FDA-cleared for primary diagnosis, showing that AI can support both accuracy and accountability. It’s not replacing us—it’s working with us.

Study 2 — Thyroid Cytology: From Overdiagnosis to Optimization

As someone who’s personally experienced thyroid cancer, this study hit close to home.

Researchers in China developed AI-TFNA, a multimodal system that combines whole-slide images and BRAF mutation data from over 20,000 thyroid fine-needle aspirations across seven centers.

The model achieved 93% accuracy, reducing unnecessary surgeries and improving clinical decisions. What’s especially impressive is Image Appearance Migration (IAM)—a technique that helps AI adapt across scanners and labs, ensuring reliable performance worldwide.

Study 3 — GI Cancer: Prognosis Reimagined

An international collaboration of over 2,400 patients introduced a Deep Learning Pathomics Signature (DLPS) that merges nuclear features, tumor microenvironment, and spatial single-cell data.

This AI-driven model predicted patient survival and therapy response more accurately than traditional TNM staging—even identifying which patients are most likely to benefit from chemotherapy or immunotherapy.

It’s precision medicine powered by pathology.

Reflections:

Each of these studies made me think about the balance between trust and technology.  We’ve reached a point where AI can truly enhance diagnostic precision—but it also challenges us to stay actively engaged, curious, and informed.

Because the real risk isn’t that AI will outperform us—it’s that we’ll stop thinking critically once it does.

That’s why collaboration between pathologists, data scientists, and industry innovators matters more than ever.

AI isn’t replacing us—it’s redefining what excellence looks like in pathology.

#DigitalPathology #AIinHealthcare #ComputationalPathology #RenalPathology #ThyroidCytology #CancerDiagnostics #DigiPathDigest

Transcript

00:00:03 - 00:00:31

Good morning, Trail Blazers. So cool to be with you again at 6:00 a.m. in Pennsylvania. I have music on. Um, but when I see you joining, I'm going to switch it off cuz it's a little bit distracting. And I see you here. Welcome, welcome. Thank you so much for joining me. Uh while you are still joining, I have a few updates because there are conferences coming up.

00:00:33 - 00:01:03

Uh let me switch on the chat and say hi in the chat. And please say hi back and obviously tell me where you're tuning in from and what time it is um in your wherever you are. Uh so uh I was recently talking to uh a future podcast guest. Um um maybe I'll later tell you who that is, but basically she said that oh she's not waking up at 6.

00:01:03 - 00:01:33

 She doesn't have little kids. She doesn't have to do that, but she always uh listens during her lunch break when the recording comes out. So, um, if you are listening to the recording, let me know as well. Uh, and I'll look into the comments later and I'll just say hi and, uh, say everything. Um, just respond to your comments, right? So, there's going to be a couple of announcements, couple of things that I'm doing today. So, let's start.

00:01:33 - 00:01:59

 I have the links um, sorry, the QR codes for you on the screen. If you don't have the digital pathology 101 book which is this book to start everything you all no sorry all you need to know to start and continue your digital pathology journey. Uh that's the QR code below and uh later we're going to uh see what's new in the digital pathology store.

00:01:59 - 00:02:23

 So I'm going to take away the store. Richard, welcome. So great to have you. It's you're in Great Britain, right? Let me know what time it is. Um, okay, quick updates and let's do the papers and then there's going to be more updates. So, um, we start with the next conferences. The first conference is ACVP.

00:02:23 - 00:02:51

 That's for my uh, trailblazers from the veterary pathology site. ACVP is the annual meeting of uh, annual sorry uh, American College of Veterary Pathologists. This is where I'm certified. So I'm going to be speaking there h about and the title of my presentation is a blackbox tool how to ensure a reliable and liable AI application and there's going to be a presentation on that but uh at ACVP I cannot record these things.

00:02:51 - 00:03:15

 So, what I'm going to do, I'm going to actually practice it. And here, let me tell you how fantastic you are for me. Because if it wasn't for you, I probably would just like go and speak to my slides and uh would deliver a suboptimal experience to my audience. But because I want to record it for you, I'm going to practice recorded and then I kill two birds with one stone.

00:03:15 - 00:04:19

 So, Blackbox tool, how to ensure reliable and liable AI application at ACVP in New Orleans. I'm flying on Sunday. So, if you're there, just say hi. And I'm going to just be going in and out for the presentation. So, it's not going to be so long. Then another update. Uh, SISY. Sitsy is the uh society of immune therapy in cancer and this is in national harbor Maryland Friday, Saturday the 7th and 8th of November and there I'm going to be with Hamamatsu which is a digital pathology place sponsor and they have this uh immunofllororesence imaging device Moxiplex. Moxiplex makes amazing multiplex images, right? Imunofllororesence images. They're going to be showcasing their uh partnership with Bioare. Bioare specializes in staining, iminohistochemistry, staining, basically tissue staining and they uh are helping companies be digitally ready

00:04:22 - 00:04:49

with all the pre-analytics, right? And do we have Scott? You guys are regular. Scott, welcome. Atlanta 6 am. He's my 6 a.m. companion. Thank you so much for joining. Uh and okay, so Hamamatsu booth 451. Let's see if I can write it on the screen. Let me do this. And here I'm supposed to be able to write 451.

00:04:51 - 00:05:38

If you are at Sity, sorry. 451. Stop because maybe it's 4:15. Let me check. What happened to my note? Yeah, 4:15, girl. 415. I mean, Dr. 4:15 415. 415. Not 415. Okay. So, Sity 415 with Hamamatu if you want to. And I'm I'm preparing something super cool for you, which is stickers. I know it sounds a little bit like what are you a teenager? But I love to put them on my like these hydration jars and um I have this I don't show you.

00:05:38 - 00:06:40

 See, I should have taken them later. I show you. I have these trailblazer stickers. So, we're going to make Hamamatsu specific trailblazer stickers. So if you are a trailblazer interested in IF immunoncology and you're going to be at SITY then we can see each other there because I'm going to be there um I think both days booth 415 and let's start with our papers and then there's going to be a few more updates just okay ACP later I'm going to tell you about how to make a presentation with AI and what I'm preparing for you about that. So stay tuned. Stay right here. And now let's do the paper situation. Okay. I need to be careful with this eraser because then it erase my erases my other markups. Sorry for that.

00:06:43 - 00:07:12

We're just going to erase and We're going to start with a public in nefron about I see more people than the people that are commenting. So, give me some comments. Show me some love in the comments. So, uh prospects for artificial intelligence-based pathological diagnosis of renal transplant biopsy.

00:07:12 - 00:07:41

 So, what are we talking here about? that AI can help with renal biopsies basically and the background is um AI initiated in the 50s well we didn't do digital pathology there that much at least maybe we're taking no no um that it had matured um after two setbacks into deep neural networks and large language models.

00:07:41 - 00:08:05

 So yesterday I was talking to Andrew Janoik who's going to that's another podcast that um is going to drop I don't know when within a month but he has this paper super cool I'm going to tell you what they did with this paper but uh he says that basically like the level of AI and large language models and all that is like discovering fire in the computational pathology space.

00:08:05 - 00:08:39

 So yeah, we had these large language models and it was like discovering fire and so now digital pathology can employ the deep neural networks and large language model and this is advancing globally and they say hm Japan lags behind and here a comment I was actually surprised when I talked to a doctor from Japan h at US cap last year Dr.

00:08:39 - 00:08:58

 Robert Osamura and he says that they have to do both like if they want to do digital they have to do digital and glass. So those people who are doing digital there are very motivated because they have to look at the same thing twice with different modalities. So they say here that unfortunately Japan is lagging behind when it comes to digital pathology.

00:09:00 - 00:09:23

um and they want to standardize diagnostic interpretation and reduce the workload workload of pathologists and it into renal transplant pathology RTP. And so they say that Chameleon 16 and 17 by the way demonstrated that AI can achieve diagnostic accuracy comparable uh or surpassing our uh pathologists.

00:09:23 - 00:10:27

 And they also say, and I have it in orange, guess why I'm having it in orange, that in the United States, over 10 DP systems have been approved, cleared. And uh yeah, the the FDA people are sticklers to this terminology. They have been cleared, but we know what they want to say. Basically, digital pathology, they were cleared as a class 2 medical device uh for primary diagnosis that it's happening, right? H and um US law assigns liability for AI related misdiagnosis jointly to pathologists and institutions uh promoting both accuracy and legal protection for pathologists which is good. And then in uh 2019 a working group was established. So this BF digital pathology working group what is BF? I did find it for you. B is oh sorry h it's like a consortium or how do I open

00:10:30 - 00:11:03

my I cannot open the comment okay yeah B foundation for transplant pathology transplant pathology so this is a foundation and they have this digital pathology working group and um they decided that uh basically AI can enhance uh diagnostic objectivity and then they have these band updates. So B they have like a scale or like a grading system of I don't think I need the headphones guys.

00:11:03 - 00:12:03

 Let me showcase my my my earrings of course. Um, BMP uh has these this um like a way of evaluating okay is the kidney um good for transplant or not and um in in general tissue for transplant but they're focusing on kidney kidney so they have these um B updates with AI retraining enables continuously updated globally standardized digital pathology for the transplant pathology and also um they say something that Andrew in our podcast yesterday said as well that advanced digital pathology requires close collaboration between trans transplant pathologists, AI engineers and cutting edge um GPU resources and you know if you want to have it in a regulated environment or as a digital pathology tool than even more people. Uh it was great what he um what he said yesterday that so they published this paper about validation of

00:12:06 - 00:12:32

um digital pathology tools and not just AI tools any digital pathology tool. So he he created a tool for H pirate lorry detection um in the Geneva hospital where he works and he says yeah um make this tool that would be like two or three weeks of work and the validation and implementation and putting this tool in the clinical workflow that was three years of work.

00:12:32 - 00:12:58

 So a lot of people were working on it and I could so relate to it. uh for the validation. Let me take away the store for the validation we did uh where I work at Charles River Laboratories and I am on the verge of submitting the paper also like two years later so I feel a little bit less bad um for submitting it since like so much time after uh we actually validate.

00:12:58 - 00:13:40

 So this podcast is going to be dropping within a month. The next paper that I have for you is sorry. Yes. Um, and let me know any comments, any questions, anything that you want to ask or in the comments because I see them all. We're back to our old streaming system. So, and the second one, I only have three today.

00:13:43 - 00:14:08

Psychological classification diagnosis for thyroid nodules via multimodel deep learning. Um, do you guys hear me well? I assume so. If you could confirm cuz I didn't put my uh fluffy microphone from the camera that I have and I think I'm using this one. Just give me a a sound okay in the comments and I'll I'll know that the sound is okay.

00:14:08 - 00:14:38

 And so here personal story about psychology sorry thyroid if you see my scar I did have a thyroid cancer and they took I don't know how many biopsies at least three sorry fine needle aspirates at least three and they were all like non-diagnostic and I had like a thing size like an of an egg. Give me a sound check in the comments please. that would be fantastic.

00:14:38 - 00:14:57

 Anybody who's listening live, you're still there. So, uh I think you hear me, but just to confirm that my microphone is okay. So, I had like a I don't even know left side, right side. It was like an egg and I would go and get these fine needles, a fine needle aspirates taken and then it would not be diagnostic.

00:15:00 - 00:15:27

Right? So let's talk about the situation. And then I was talking to Andrew Bitchkov who is a thyroid um thyroid pathologist and also to to Dr. Nikki Forov who is the person who who created World Tumor Registry. And I told them about my tumor. They're like h we don't even really classify it as tumor anymore but you know I didn't want to have the egg on my neck.

00:15:27 - 00:16:01

 So, I wanted to get it out and it was uh crossing the the capsule already. So, then I don't have a thyroid. Long story short, but going back to the paper is okay. And I thank you so much, Richard, for confirming that you can hear me. Um so the the the thing is and this is a publication from China and the thing here is the rising prevalence of thyroid nodules is straining limited cytoathology resources.

00:16:01 - 00:16:31

 Right? This is a recurring theme as basically that um there are not enough pathologists and cyto and the more specialized you go the less specialists you have uh resulting in excessive overdiagnosing or over and over treatment with significant patient and healthcare consequences. Um the to address this there AI TFNA developed AI uh thyroid fine needle aspirate.

00:16:31 - 00:17:10

 Um and this was an AI platform leveraging extensive clinical data to enhance diagnostic accuracy and clinical efficiency. And I am always so impressed like how many samples they get from China. 20,83 thyroid samples were collected from seven medical centers across different regions in China. And of these 4,421 thyroid fine needle aspirations um from three hospitals were used to train this AI ensuring strong generalizability and then they validated.

00:17:10 - 00:17:47

 What did they do? I should have a comment here. Ah didn't save doesn't matter but this is uh so how they scored this cytology they have this the Bethesda scoring system so they did validation of internal validation I guess there was an external as well and it was an exceptional performance the overall accuracy of TBS1 is 9327 what is TBS1 basically non-diagnostic or um yeah non-diagnostic or or something that is not diagnostic either like super confusing or non-diagnostic.

00:17:47 - 00:18:20

 Um and the sensitivity of TBS5 so the score goes from one to six six and six is malignant. This is suspicious of malignancy and like you go backwards towards non-malignant but one is non-diagnostic. So um the TBS 5 and six reaches 8537 and we're talking about sensitivity here and 8378 while the specificity uh of TBS2 is 9713.

00:18:23 - 00:18:55

Let me just check what is TBS2. I think this is like B9 and TBS2 thyroid is B9 in the Bethesda reporting thyroid system. So this is B9 and the specificity is 97.3. So, cuz this and I need to tell you something else because that's not the first time I'm hearing about um these massive cytology screening happening in China.

00:18:55 - 00:19:57

 They do it for uh papsmear as well and they have like AI systems that screen it. So, I'm going back. Oh no, I disappeared. um the they also incorporated uh graph mutation data into this AI and um the development of a multimodel model for their improved precision by significantly improving the differentiation between B9 and malignant thyroid nodules and that's what you need to know out of the fine needle aspirates um and they also used a technique called image appearance migration um a technique that uh substantially improves cross institutional model generalizability. So what is this uh AM image appearance migration is basically um data augmentation and making these images for training manipulating them in a way

00:20:00 - 00:20:31

so that the model is more generalizable. It includes data augmentation. Uh it also includes can include um the guns uh the generative adversarial networks and like different AI techniques to make the algorithm more general. We know that these AI algorithms are susceptible to domain shift and by domain shift I mean uh like coming from different scanners, different stainers, different wherever, right? There's a lot of uh going on in the pre-analytical space.

00:20:31 - 00:21:02

 So we try to generalize as best as we can by standardizing pre-analytics but also implementing some AI and help with that. So uh okay so sensitivity was improved by 190 and specificity by 812 and this AI TFNA offers rapid reliable decision support advancing thyroid nodule diagnostics. So, another discussion had with Andre and I'm like spilling all the beans before the podcast drops.

00:21:02 - 00:21:26

 But I hope that then you want to listen to the full thing. We always have a pretty good flow with Andrea. And let me show you something because I have a um I have a course. It's a free resource um that I did with him and it's in the store. So, before we go to our last paper, I'm going to show it to you.

00:21:26 - 00:22:08

 Uh, stop sharing and share again. Let me do it. Okay, you should be seeing this. And um, he is this histools histoe tools. Let's see if I can show you. Yeah, that's him. and um he is obviously author of this free tool and this is a free resource. So if you're interested in how to check your pre-analytics, this is definitely um the course or this the series.

00:22:08 - 00:22:43

 It was a webinar series that we recorded and now is in the course in the store. So let me give you the QR code to the store now and let's go back to the papers. But why did I even say that? Ah, because we were talking about because here this AI um helps you like make it's a decision support, right? And there is this semi philosophical or I would say like not even semi- phililosophical but philosophical discussion. Okay.

00:22:43 - 00:23:14

 Is going to make us be stupid. Did I already record that podcast? I might have. Um there was this paper from Lancet about deskkilling of pathologists um the sorry not pathologists endoscopies um that were doing colonoscopy AI they uh were they were using AI and then they took the AI um away they were using it for uh spotting polyp in colonoscopy they took it away and their performance dropped significantly.

00:23:14 - 00:23:40

 So obviously there is this discussion has been going on forever already that okay are we going to be like worse as doctors uh scientists or whatever because we have these tools and it's an ongoing discussion that I also had with Andrea uh yesterday in this um paper that we just covered made me think of it. Um okay we have the store. Yeah.

00:23:40 - 00:24:21

 So if you want to uh check the histoc in the left upper corner there's this QR code to the store and let's go to our third paper before I give you some more updates. Computational pathology approachment of prognosis and imunotherapy response in pan gastrointestinal cancer. This is also a group from China. And what did they do here? Um the background is that current staging method cannot accurately predict survival outcomes and therapeutic benefits in cancer patients.

00:24:21 - 00:25:21

 So um did implemented so-called digital pathics. Pathomics is basically like molecular combined with pathology. Um and you can do it directly from H& you can do you you can do fluoresence and all different things but pathomics is basically combination of pathology and genomics pathology and proteomics all the omics with pathology image right um and this is a rapidly uh evolving field and actually I had a podcast with a person uh with Trevor who has a company called Pathomics Um so what did they do here with the methods? There was an international multic-enter study that included 2463 patients with pang gastrointestinal cancer from 12 cohorts and seven cancer centers. Again maybe these were the same cancer centers that the thyroid study

00:25:24 - 00:26:25

maybe we can check in a second. And then uh 1,653 patients were diagnosed with gastric cancer. So um then they proposed a deep learning pathomics signature DLPS. Sorry for my handwriting on this tablet. deep learning pathomics signature DLPS uh by integration integrating information on three scales from the whole slide images of H&I tissue including pathomics features pathomic features related to the nucleus micro environment single cell spatial distribution so this is you know single cell spatial distribution is going to be probably a lot about the sitzy So, if you're there, let me know. The booth was 4:15 with Hamatsu. Um, so join me if you're there. That would be so cool. And I love that it's so close this conference to me.

00:26:27 - 00:27:28

Obviously, I can drive there. So um but yeah we had this single cell spatial distribution and they assessed the predictive accuracy of the LPS for prognosis chemotherapy response and immunotherapy response and what were their results and they found that the LPS um digital uh what what is this what did we digital pathomics deep learning pathomics signature yeah deep learning pathomics signature deep learning uh was significantly associated with overall survival in gastric cancer and pang gastrointestinal cancer exhibiting good accuracy uh area under the curve ranging values ranging from 0.723 to 0.840 840 H and they that indicated the the Cox multivariable COX regression analysis indicated that DLPS uh was an

00:27:30 - 00:27:55

independent prognostic factor and nomogram and I had to check what nomogram is. It's like a diagram but for more scales than just two. So diagram would be for two and nomogram that integrated the uh the LPS and tumor no stage showed improved performance in predicting cancer prognosis compared to that with TNM stage alone.

00:27:55 - 00:28:26

 So um the gastric cancer patients with low uh deep learning pathomic score but not those with high were exhibit sub substantial benefits from adjuvent chemotherapy and then um the objective response rate to anti PD1 immunotherapy was also significantly higher in low DLPS group compared to uh so that was 29.6% 6% and the high the LPS group was just 8.3.

00:28:26 - 00:28:53

 So that's like pretty big difference I would say in the immune therapy, immune oncology world. These so I was so naive um at the beginning. I thought that these um stratification tests will tell you like that those who are not going to respond like are not going to respond at all and those who are going to respond are going to respond like 100% of them is are going to respond.

00:28:53 - 00:29:54

 And then when I started learning that, oh, like the differences are like if you have a 30% difference, you're like golden. I'm like, we're talking about like different scale than I thought. Uh, so yeah, now I see like 29.6 versus 0.3 and I'm happy. And before I'm like, what are we talking about? Um so um they concluded that the LPS has the potential to enhance prognosis assessment and I'm going to comment on has the potential phrase in these papers and identify patients who are likely to benefit from adgivant chemotherapy and imunotherapy and pan to intestinal cancer. So has the potential is a common phrase in these um in papers right because has the potential and it goes back to the uh to the tool development digital pathology tool development that I talked about in the podcast with Andrew and oh my goodness there's so

00:29:57 - 00:30:30

much more going into developing in a tool versus into and I'm going to put my book here as well if you're if you don't have the digital pathology oneonone yet. You can get the free PDF version here through the QR code. Um, but so much more going into a tool than going into a publication and so much more going into a publication in a peer-reviewed journal than you know just having a concept and idea.

00:30:30 - 00:31:02

 So life is a funnel like drug development is a funnel and you know their sales funnel life is a funnel from idea to product to actually tool um because Andrew's group developed this uh tool that they did um for H pylori detection as an LDT lab developed test um and you know if you want to productize that's like the next level of complexity and or or like a different pathway that uh you have to um go through.

00:31:02 - 00:31:22

 So every time I go to a pathology conference, I'm basically impressed by the number of vendors that are there and those who have clearances from the FDA. It's like such an enormous effort and I always call the vendors um unsung heroes of digital pathology. Um and I'm always super excited about my sponsors.

00:31:22 - 00:31:47

 But what I wanted to tell you is another thing that I'm working on because um obviously ACVP I'm speaking right. So again, I need to prepare a presentation. H and I'm like, "Oh my goodness." Over and over I prepare these presentations and I have a I already use AI tools because without it I would like spend a lot of time on it and I don't want to spend a lot of time and I kind of most of the time know what I want to say, what my message is.

00:31:47 - 00:32:47

 Um so uh let me put this on the side. Can I do it? Yeah. So uh I'm like okay, how can I have a process? uh is it something I can make into a process and I did so I have a process how I uh do a conference presentation um it's it's more for like conference concept presentation than a scientific results presentation so if you are interested in my workflow my process with like the tools that I'm using and I'm going to make a separate podcast and probably a video on this but if you're interested in um my process for preparing a conference presentation, leave me a comment. H I don't have it like as a QR code yet. Um I don't have it in a funnel or like in an automated way um to share with you, but I can share it with you as a PDF and then hopefully by the next live stream I will prepare something that I can share as a code and I'm also thinking of uh doing a

00:32:50 - 00:33:05

workshop. The workshop is going to be a paid workshop. Um, so when I plan it, I'm going to let you know, but the PDF is going to be free. It's going to be something you can download and basically uh go and use the tools. Um, I'm going to tell you in a separate podcast how I use the tools and uh maybe make a video.

00:33:09 - 00:33:31

If you are uh listening to this on YouTube, watching it on YouTube, please uh give me give me this uh thumbs up and subscribe button. Uh I'm going to be responding to all your comments. If you have any questions even when you are listening to the recording, if you're interested in the my presentation process, uh book courses, you know where to find them. You know where to find me.

00:33:34 - 00:33:55

Never hesitate to reach out. Uh you are important to me, my trailblazers. Um you don't even know how much you motivate me to do all these things. And uh yeah, the last piece of motivation that I got was from uh Ingred from the company, Caris. she's going to be a guest on the podcast and she was like, "I listen to you during my lunch and I want to go with you for a real lunch.

00:33:55 - 00:34:20

" And that was so heartwarming. This is always so heartwarming when I hear these things from you. Um, also earrings, right? Maybe it's time to buy Christmas presents and maybe somebody wants these earrings. So, you know where to find them in the store as well. And with that, I'm going to let's put some music on to say goodbye.

00:34:29 - 00:34:29

not for joining