169: AI Across Organ Systems: Kidney, Liver, Colon, Bladder, and Beyond

Can one AI system learn from every organ — and teach us something new about all of them?

In this edition of DigiPath Digest #31, I explore how artificial intelligence is transforming pathology across multiple organ systems, revealing connections that help us diagnose faster, more consistently, and more accurately than ever before.

From glomerulonephritis to hepatocellular carcinoma, AI is no longer confined to a single specialty — it’s becoming the connective tissue between them.

What’s Inside:

1️⃣ AI for Bladder Cancer Classification
We begin with a multicenter study validating AI models for urothelial neoplasm classification using over 12,000 whole-slide images. Both CNNs and transformer models achieved high accuracy (AUC 0.983, F1 score 0.9). I discuss why the F1 score matters — and what it tells us about model balance between sensitivity and specificity.

2️⃣ AI in Colorectal Cancer Care
Next, we explore multimodal AI — integrating histopathology, radiology, genomics, and blood markers to modernize colorectal cancer workflows. AI now helps detect adenomas, infer microsatellite instability (MSI) from H&E slides, and predict treatment outcomes. I highlight the critical need for external validation, interpretability, and governance as AI enters clinical use.

3️⃣ AI for Glomerular Nephritis Diagnosis
A deep learning model trained on over 100,000 kidney biopsy images identified four nephritis types — FSGS, IgA, MN, and MCD — with over 85% accuracy. This technology could ease workloads and improve turnaround time in renal pathology. Still, I share why AI support may feel both empowering and unsettling for many pathologists.

4️⃣ AI in Liver Disease (MASLD & HCC)
AI is advancing noninvasive fibrosis staging and risk prediction in liver pathology. From large consortia like NIMBLE and LITMUS to predictive models for HCC therapy response, AI is moving us closer to precision hepatology. I also discuss the challenge of translating these tools from research to regulatory approval.

5️⃣ Lightweight AI for Domain Generalization
Finally, we look at one of pathology AI’s biggest challenges: domain shift — when a model trained on one scanner or staining style performs poorly elsewhere. The new Histolite framework shows how lightweight, self-supervised models can generalize across data sources — trading some accuracy for reliability in real-world use.

My Takeaway

Across every study, a single message stands out:
 AI isn’t replacing pathologists — it’s amplifying our vision.
By connecting kidney, colon, liver, and bladder insights, AI is teaching us that medicine works best when it learns across boundaries.

Episode Highlights

• Bladder cancer AI validation (06:41)
• Multimodal colorectal AI (12:38)
• Glomerular nephritis deep learning (19:29)
• AI in liver pathology (29:55)
• Domain shift & Histolite framework (38:17)
• Halloween wrap-up + SITC preview (46:18)

Join me next time for updates from the SITC 2025 Conference, where I’ll be live at Booth 415 with Hamamatsu and Biocare, discussing how AI and spatial biology are converging to drive clinical utility.

#DigitalPathology #AIinHealthcare #ComputationalPathology #CancerDiagnostics #LiverPathology #RenalPathology #FutureOfMedicine #DigiPathDigest

Transcript

00:00:00 - 00:01:37
Aleks: Good morning. Do you hear the music? Let me know if you hear the music. Oh, why am I recording on this? Okay, let me know if you hear the music. I'm gonna just send you something in the chat because today is Halloween and um now just let me know if you heard it cuz I cannot talk about this over this music. Welcome my digital pathology trailblazers. Once I don't play the music, I don't need the headphones either. and just give me a thumbs up or some chat reaction that you can hear me,


00:00:54 - 00:02:23
you can see me. And I was um getting ready and I look at the date and it's the 31st and I'm like I'm totally unprepared cuz today is Halloween and I like don't have anything. But then I uh open my live stream studio and there is stuff here. So, you will see something on the screen that looks a little bit Halloweeny. And I also brought a friend with me. This beautiful pumpkin. I just took it uh from outside cuz we bought a bunch at the pumpkin patch. And I have another friend. You can vote white or orange. I don't


00:01:38 - 00:03:12
know if they will fit both here or if I'm going to cause some disaster. Studio disaster. Let's see. Can they be like that? We'll see. Vote for your favorite pumpkin. Let's see if we can move them without. Oh, wait. One fell. We can do it, guys. I can do it. and they will just happily sit here with us for the Halloween edition of DigiPath Digest. Okay, you are here. So, let me know in the chat that you are here. One more thing that I wanted to do is to put like this pumpkin sticker on myself, which I


00:02:27 - 00:03:51
have, but it belongs to my son, and he's still asleep because it's 6:00 a.m. 6:04 in Pennsylvania, and I would get in trouble if I would use up his pumpkin sticker. So, I brought a real one. Welcome, my digital pathology trailblazers. Um, couple of updates. I do need to hear from you because then I know I can move on. Um and I think my comments are coming going through. So let's start with a few updates and I see you guys are joining. Um update number one is the next conference. The next


00:03:08 - 00:04:40
conference is sits uh society of immune therapy and cancer and that is going to be in uh near DC national harbor 7th and 8th of November which is Friday and Saturday and the important thing is Saturday because I'm going there with sponsors with Hamamatsu and with Bioare our booth. booth is going to be 4:15 and I'm going to be drawing this on the board. But something I want to let you know about uh let's see if I can share it and let me know where you're tuning in from and what time it is. Where are my


00:03:55 - 00:05:30
regular stop screen sharing? Okay. And another screen share. Okay. Here. Um, you should be able to see my screen now and it's a LinkedIn event that you should be seeing. Let's see if I can make it bigger. Okay. And let me know where you're tuning in from. Okay, I'm going to make it full screen. Okay, so we have from spatial biology to clinical utility. Uh it's a talk by Carlo Bulko who is very active in the uh in this spatial biology space and he was working with markers with lung cancer


00:04:53 - 00:06:01
markers. So I am super interested about this one because um usually the the notion or the idea is okay I have imofllororesence and um this is more like for research and then if you want to take it into clinic you'll ra you'd rather translate it into um what brightfield right because it's less uh cumbersome but here they're going to be talking about from spatial biology to clinical utility I'm going to be there several people are already saying that they're going to be there. So, I hope


00:05:27 - 00:06:33
you're going to be there, too. If you're in Sity, leave me a comment uh that you're going to be there and definitely come and visit booth number 415. I'm going to be there with Hamatsu and Bioare. So, that is update number one. I'm super happy about this because I can just drive. I love the conferences where I can just drive and I don't have to fly because I had to fly to ACVP American College of Veterary Pathology annual meeting where I gave a talk on uh blackbox how we can


00:06:00 - 00:07:23
uh get from AI um like from the blackbox to AI reliability and utility and I'm going to record this talk for you and post it on LinkedIn. The only thing I wanted to tell you about that conference that so it's for veterary pathologists but we are so not lagging behind amazing talks about uh research applications might talk about more like tool deployment and um regulatory environment uh reliability and liability of AI in this space and also everything in between um a lot about domain shift and there is a paper


00:06:41 - 00:08:12
that we're going to be discussing. Let me share the right stuff with you now. There's a paper that we're going to be discussing that covers the main shift. Um, and if you are here live, you let me know that you're here live in the chat. And without further ado, let's go and cover the first paper of ours. Let's see. Everything working? Yes, everything working. Amazing. Okay. And I have people telling me everything cool. Amazing. Thank you so much. Um, I do have a little bit of a cold, so I


00:07:32 - 00:09:06
might be muting myself to cough. Hopefully not too much. Righty with coffee. the friends are accompanying us. So let's start with multi-institutional validation of a models models for classifying euro e urothelial neoplasms in digital pathology and what do we have here? Well, uh the deep learning approach is for classifying normal, non-invasive and invasive eurothelial neoplasms uh through digital estopathology images. H and they say well AI is almost everywhere but where is it in bladder cancer? Let's do it for bladder cancer.


00:08:21 - 00:09:47
So they developed convolutional neural networks and transform based models. So um important thing important um like I don't know if it's a trend I guess it's not a trend but something uh so now like the big new thing are the transformers right and when I started in this oh and we have people party so great to have you here I have colleagues joining and going back to my thought was that every time I would learn about a new modality, so first we started with and that was also something I covered in my ACVP talk


00:09:04 - 00:10:09
that we started with rule-based approaches, right? So we knew like what in intensity a cell had and we had to define it and then like depending where we would put the thresh threshold we would lose cells or have too many cells whatever but we knew what was happening but it wasn't fantastic and then convolutional neural networks came and I was like now we can give examples now we can throw this other thing away and then I look at the results and I'm like could we combine uh and now I see uh the same happening


00:09:36 - 00:10:39
Well, it's not rocket science probably. I just always get like overenthusiastic about the new thing and forget about the old thing. But here we have the CNN's and we have transformers and they were trained on a lot of images. 12,500 whole slide images. So, I don't know if you remember like the beginnings of this space. You would have 100 and you would be like, "Oh my goodness, that's huge data set." And now we have 12,500. And this is amazing. Although for training you don't always


00:10:08 - 00:11:27
need so much especially if you do like cell based stuff where you have a lot of cells in one image you can still get away with uh you know just 20 or something right so but they had a lot over 12,000 five institutions and then um they included stain normalization and patch extraction techniques and did cross validation um against expert annotated annotated labels and um they had different models and the the one that was the best was efficient net P6 and when I looked at the metrics okay accuracy 0.913


00:10:48 - 00:11:58
I'm like not surprised when I see accuracy so high in these papers sensitivity 0 over 0.9 specificity over 0.9 I kind of like these are metrics that you have to like know what they are and know what it means because you can have like uh super high accuracy, but you claim everything is negative and the prevalence of the disease is super low and you're still going to be super accurate. But when I looked at this F1 score of 0.9, I'm like, is that really so good? So, it was funny because at the


00:11:22 - 00:12:42
ACVP um and if you're not familiar with these metrics, don't feel bad because I have to learn relearn them every time. But um these F1 is one of them that that comes um like returns over and over again. So I kind of like know my numbers and at ACVP we're talking about data sets and about measuring all these metrics and uh one speaker Kristoff Bram who was also the chair one of the chairs of the um of my session the AI session h gave this question at the beginning 0 F1 score of 0.7 is it good or bad and you know if


00:12:03 - 00:13:14
you don't do it on a regular basis you have no idea and I'm like 0.7 is Good. And now I see here 0.9. So I hope it's really that fantastic. But that's something that I'm like hm if I was studying this deeper I would go and check like how did they evaluate this? But basically uh let's believe the results and we say that they demonstrate effectiveness and generalizability of AI based bladder cancer classification. So that's good. Um, let's do another one and then I'm


00:12:38 - 00:13:50
going to show you uh a little surprise that we're going to get for uh get for everybody who who's going to join uh at City modernizing coloractyl cancer care with artificial intelligence real time detection radiomics and digital pathology. So we have multimodality is entering the scene complexity increases right. Um I'm just checking I didn't check exactly like the impact factors and I didn't do the affiliations. I'm going to do better next time so that you know exactly what


00:13:14 - 00:14:36
uh where these um publications come from and uh where the groups that are working on them come from as well. Okay. So AI is reshaping the colorctal cancer pathway. That's an interesting statement. Pathway pathway to me it's always like biological pathway. Uh but it is boosting lesion detection um expediting molecular triage and enabling quantitative multimodel decision support and evidence shows that computer aided detection increases adenoma detection and lowers miss rates. I think we already covered um or or there is a


00:13:54 - 00:15:05
podcast episode on the uh AI supported or AI aided detection of polyp in um in colonoscopy and how it can contribute to deskkilling of pathologists. That was like a discussion. Um we had a pretty cool discussion in the chat when we were discussing that paper. So uh whenever this live stream is published I'm going to link to it or I'm going to link to it in the show notes. Um but you know when you take it take the AI away people get worse when they detect this coloractyl the polyp but with AI they are better so


00:14:31 - 00:15:44
let's hope AI works and they can keep being better to help patients um also digital pathology can infer microatellite instability from routine H& slides prioritize confirmatory testing so this is the molecular testing well molecular prediction uh from H& a big thing and then you just do confirmatory testing and also we have uh CTM MRI so we have radiology segmentation radiomics based risk risk stratification nodal staging and response prediction and then we also have blood and genomics driven models


00:15:07 - 00:16:25
and that extend non-invasive screening and prognosis a bunch happening in colorctal cancer that is not just digital pathology um and you know we are using a lot of data. So um we need high quality data, we need external validation, we need interpretability, workflow integration and robust governance. So not really rocket science here as well that we need all these things to deploy AI. Um something that uh I'm thinking okay this external validation do you really need it if you're deploying it in just one


00:15:46 - 00:17:04
institution? Maybe not. If you want to have a tool that uh is a tool that you can commercialize definitely yes. Um and all the other things interpretability workflow integration of course. Um so but if you we would want to have a deployable tool we would need multic-enter perspective studies life cycle performance monitoring. I'm going to do a exclamation mark here. I mentioned this life cycle performance man mon monitoring last time when I was talking that um I had Andrew Janoik and um on a podcast this podcast is going to


00:16:25 - 00:17:34
be dropping hopefully within a few weeks h about tool deployment digital pathology tool deployment um and that's something that needs to be incorporated into every AI deployment um and not because the AI tool is going to be changing the AI tool once we lock it is going to be locked until we make a new version and then we will have to lock the new version as well. But what's going to be happening is the underlying biology is going to be changing. So me coming from the toxicologic pathology space, we have


00:16:59 - 00:18:17
this um database of different findings for animals. H it's called historical control data. So we have a bunch of control animals where we um record all the findings and we can access this data. But uh generally the cutoff for usability of this data is 5 years because the animals biology changes. So uh we always want to have the most recent five years of data of these um of these um animals, right? So the same happens with biology uh of people of everybody, right? some sometimes I mean bacterial


00:17:39 - 00:18:42
biology changes a lot faster than human biology but it happens right cancer biology resistance to treatment whatever this can happen and you can start seeing a drop in performance of your tool even though tool stayed the same so this thing this performance monitoring is an important thing that I see coming up over and over again and we have a comment from Patty sometimes the technical scientific part of implementation AI in these workflows is the easy part and indeed H that is the case. So when I was


00:18:10 - 00:19:25
talking to Andrew uh last time he said that the like the creation of the tool itself uh took them maybe three weeks. So the the tool was for detecting helicoacttor pylori. So um bacteria in the stomach. Uh pretty easy from the computer vision standpoint from the biology standpoint also like easy visually for the pathologist. So it took them like 3 weeks to make the tool and three years to deploy the tool across the institution. So totally uh like you know we think the difficult part is creating these algorithms. But now the


00:18:47 - 00:20:10
difficult part well the whole thing is kind of not that straightforward. So moving on to our next I know I was supposed to okay after next one I'm going to tell you about the surprise at sity. So I'm crossing my fingers to not forget. Um now artificial intelligence and we're doing across organs today. What did we have? We had the colorctal cancer, we had urothelial cancer. So every kind of organ specific AI model uh we can think of and now we have glomeular nephritis. So we are in the kidney using


00:19:29 - 00:20:43
pathological image analysis approach and multic-enter model development and validation study. Um I want to check something. Did I have a dot somewhere to tell you about? I know I just wanted to highlight the life cycle monitoring. All good. Um multicenter right so complexity increases the better we are the more complex we more complicated we make the things as well. Um so the diagnosis of glome glomeular nephritis uh is based on kidney biopsy analysis by nephropathologists and and obviously as any human


00:20:06 - 00:21:19
assessment it's going to be subjective labor intensive and will include interobserver variability. We will not get rid of it. Doesn't matter how well you train the people to be aligned. If you have uh you know three pathologists uh if the three agree it's like very very um high agreement. Usually we like don't agree more than 0.7% when you look at these uh papers but so guys you are reacting to my Halloween look. Yeah. I look I literally look at the um at the date and I'm like I need


00:20:44 - 00:22:09
to have a pumpkin at least. Okay. But going back to the kidney kidney um so what are we talking about? We are talking about AI assisted model for diagnosing glomeular nefritis using hisytological images from kidney biopsies. Um and this was a model development and validation study and included patients who underwent kidney biopsy. Right. Uh and there were four different types of discomar nephritis. focal segmental glomeallo sclerosis IGA nephropathy membranous nephropathy and minimal change disease I love this name minimal


00:21:27 - 00:22:40
change disease there's minimal change but it's disease um and obviously I'm just making um fun of it but these are like official entities that have diagnostic features that the nephropathologists are very familiar with and can train the model on right so they um this was a study from China there There are three institutions h and again we have a bunch of slides there um nanfang jinu and um Hawaiian and I'm just on dualingo I am learning Chinese but I'm only score Chinese 10 which if you are familiar


00:22:03 - 00:23:25
with dualingo scores then you know and if you are not then it's very beginner um so apologies for my mispronunciation of Chinese city names But um the point here is that um they had a lot of slides and they're going to tell us how many but so they had a training cohort from one of the hospitals internal validation cohort 20% of these slides and then they had two external validation cohort external um validation cards sorry need to do a little better and patients with diagnosis of one of the four


00:22:44 - 00:24:00
specified glomeal nephritis types were included. So whenever one of these types was there, they were in the study and and the exclusion criterion was low quality biopsy images. So um if we were doing like the full paperwork, I would go there. Were like how low quality biopsy images were there? uh were they not enough quality for um for the algorithm or were they also not enough quality for human evaluation? That would be very sad because then you have to take an biopsy and it's not fun. But


00:23:21 - 00:24:46
they had uh over 100,000 glomei light microscopy images to develop this AI assisted assisted diagnostic model. uh and AI um had these three components. So they first glumber localization module for segmenting glomeili. So they go to a model goes recognizes glomeili in the kidney in the kidney and then we have feature extraction and fusion model for glomeular lesion lesion and then we have patient level classification module to diagnose four types. So um all of these um different types of glomeular


00:24:05 - 00:25:38
nephritis that we just mentioned um and then they evaluated the F1 skull precision recall and accuracy and again like our F1 score was 0 8 wait F1 score of 85%. So I'm it's the same as 0.85 85, right? But to me, this is like freaking high F1 score. Like really precision 83 and recall 88. Um, and what did they say about this? They say that by automating pathological diagnosing, the model may potentially be used to reduce reduce clinician workflow and improve efficiency, ultimately supporting faster diagnosis. And


00:24:51 - 00:26:01
obviously if you're a pathologist and you hear this supporting faster diagnosis, you always have mixed feelings about this because I'm like, "Oh, am I now supposed to like read uh twice as many slides in the same time? If I have this um support um and there are mixed opinions on that, but the fact is if you do have support, you probably are going to be faster." And I'm like curious automating pathological diagnosis that's also like something that triggers are you going to now like


00:25:26 - 00:27:00
not ne need pathologists and things of that sort are kind of happening. So at ACVP and the American College of Veterary Pathologist annual meeting there was a mention of it's for animal diagnostics for cytology uh one just one application of cytologology there was a mention that there is an option to get just AI generated results from um from this cytology right so the veterinarian can submit a sample and get an AI generated out for in like within minutes and only with additional um payment and like additional effort


00:26:13 - 00:27:16
will they get um pathologist interpretation. So uh oh what's going on now in the pathology world. We will see. I'm going to be following this. Uh it was just a poster at the conference. So I'm going to be looking out for the publication. Once they publish, I'm going to bring it in here and we can dismantle this publication and see and you can react to what do you think about this approach. And obviously it triggers me a little bit as a pathologist especially veterary pathologist.


00:26:45 - 00:28:10
But let's figure out how we can stay in the loop so that it doesn't go wild and crazy. Um, and before we move on to liver publication, let me show you the surprise that we're going to have. The surprise are going to be stickers, but not these stickers. So, I brought these guys. These are the Trailblazer stickers that I have um for my Trailblazers. And you know what? I didn't even know, but I want to put that on my mug. But they are dishwasher proof. I already put it on some of my


00:27:28 - 00:28:44
of my cups. And actually, so if you uh have gotten one of these from me at the conference and didn't stick it anywhere yet, that's like so crooked. That's okay. They are dishwasher proof. So, we're going to make special ones for those who love immunofllororesence. And let me show you how they will look cuz I have them prepared for you. My team already prepared mockups. I don't know which one. I'm not sharing. Let me share. [Music] Okay, here we're going to have the brain


00:28:06 - 00:29:15
trailblazer. I don't know. So, I'm um not yet sure which one it's going to be. If we're going to have all of them, you can vote which one you like. Uh, let's do one for brain, two for this trailblazer, the the one with the background and three for this one. Give me a one, two, three in the but so so these are the Hamamatsu now has this uh new device. It's called Moxiplex that does it's a fluorescent scanner uh for a lot of markers. I think in this one we have nine markers. So uh the Hamamatsu


00:28:42 - 00:29:57
team wanted to highlight this one where we have all the nine colors. H I also love this brain section because it's just so cool and I like this design. So, um, let me know in the comments which one you prefer and maybe you can influence the decision, uh, which one we're going to print. And then when you come to our booth, the booth number here 415. H, you can you can get one. And I'm going to make sure that they're going to be dishwasher proof and that you can put them on your bottle or on your mug or


00:29:21 - 00:30:29
wherever you want to put them on. Um, and you that's going to be our gift for you with these beautiful immunofllororesence marker. And definitely if you come to the to this event that they're organizing with Dr. Bookco, I'm going to have them there for you. So, everybody who goes there is going to get a sticker. I'm like, "Yeah, teenagers put stickers on." But not anymore. 41 year olds, almost 41 year olds put stickers on their stuff as well. And I'm referring to myself. Look


00:29:55 - 00:31:19
at this. Okay. So, I'm Okay. I see I see some preferences. Oh, some brain brain is winning. give me a give me a number one, two, or three for which sticker you're interested in. And I'm going to uh get back with this feedback to the Hamamatsu team and say, "Well, the Trailblazers said they wanted number one or two or three." You let me know or maybe we're going to have more. Okay. Now, going back to our publications. Uh, okay. No, not this one. We have the liver first. liver.


00:30:43 - 00:31:47
And I have tiny ones as well. I put them on my headphones, on my uh iPhone headphones, uh which I don't have an iPhone, but I have the iPhone headphones because I never know which box is mine and which is my husband's. And I put a a normal one like not a vin vinyl one, not not this like special dishwasher proof on it. And after a few days of having it in my pocket, it got destroyed. So this one doesn't get destroyed. And it's a special size for uh ear earphones, headphones. Okay,


00:31:15 - 00:32:41
I'm waiting for the numbers 1 to three. And in the meantime, let's talk about uh AI for predictive diagnostics, prognosis, and decision support in MASLD, hepatellar carcinoma, and digital pathology. So again, we're we're doing multimodality because it's and digital pathology. And what is our MASLD? This is metabolic dysfunction associated static liver disease. Do you probably like pronounce it muscle D or something? Anyway, this metabolic dysfunction static liver disease and what's happening there AI has been


00:32:02 - 00:33:23
integrated into largecale consorcia such as nimble litmus target Nash statite um so a bunch of different efforts uh happening to incorporate AI to improve diagnostic accuracy and patient management and that the consorcia utilize AI to derive and validate non-invasive biioarker in fibrosis staging. So why do we want non-invasive biomarkers? Because uh to do the fibrosis staging um oh and we have a vote for two but variety is good. So maybe we should do more. Depends like on the printing arrangements. If there


00:32:43 - 00:34:02
is a majority for one number it's going to be probably just one but maybe we can do more. But going back to the liver, so a non-invasive biomarker, right? How do you stage fibrosis? You need to take a biopsy and taking a biopsy is invasive. We want to avoid invasives invasive. Um so is there anything we can do that is noninvasive for establishing this? Um so um also u there is AI based models to uh enhance the detection of hypothesize ballooning metabolic dysfunction associated statitis minimizing interobserver variability. So


00:33:22 - 00:34:40
you know what whenever there is an observer aka pathologist or you know radiologist whoever um there is variability and uh we want this to improve clinical trial enrollment criteria h and also there are a applications to different differentiate masld from alcohol associated liver disease and they use gut microbiome and metabolic profiling. So this is so cool because hisystologically they look super similar uh but you can have this type of change in people who don't even drink and it's very much metabolic associated and then


00:34:01 - 00:35:29
you profile the gut microbiome. Interesting. I need to learn more about that. Uh in hpicellular carcinoma the AI improves risk stratification, diagnosis and prognostication. Um and there's also a based models uh on liver stiffness and clinical parameters um to stratify patients for um for habit cellular carcinoma development and enhanced imaging techniques, radiomics and histopathology powered by AI improve the accuracy of detecting indeterminate liver nodules, predicting microvascular invasion and


00:34:45 - 00:35:59
Um AI also improves treatment response prediction for therapies such as this was this is called trans arterial chemolization. Um I was not aware of this technique. Um and I guess if you don't have liver cancer you pro or are not a liver researcher or something you don't know about it but it's something like you can go through an artery and give chemo directly to the tumor h and then embleize it and um meaning cut off the blood supply to the tumor and this is used for tumors that are not uh you


00:35:22 - 00:36:49
cannot do surgery on them. So that was pretty interesting that you can use AI for that. Uh and also for immune checkpoint inhibitors. So immune checkpoints, immune oncology, that's going to be covered at CISY for sure. So um what do we have in digital pathology for this liver disease? AI has redefined fibrosis sta staging, donor liver statosis assessment. This is super important. uh and disease diagnosis. So donor statosis assessment. This is something happening when people donate livers and often you that happens


00:36:05 - 00:37:46
when people are in a coma about to die and you take a biopsy or and you assess it. If it's good then you can transplant. If it's not good uh you cannot right. Um and we have uh some exceptional platforms um fiberstess and why am I like talking about it with a kind of um different voice cuz I have personal experience with this type of stuff um that's how I know anyway uh platforms platforms fibonst and q fibrosis are two exceptional aa platforms terms and we are happy because the field of MASLD HCC hypotoellular


00:36:58 - 00:38:13
carcinoma and digital pathology is advancing toward precision medicine and I want to tell you auh about the conversation I had um at some point it was maybe two years ago or three years ago I was talking to someone who was doing this type of liver research and they were asking asking me like how can we do more non-invasive? How can we get rid of this requirement of needing to do a biopsy to uh assign somebody to a clinical trial? I actually don't know if this has already happened, but at that po at that


00:37:36 - 00:38:53
point 3 years ago, I'm like, "Oh, I don't know how you're going to get this through the FDA and change these guidelines because it's such a herculean effort to do this." But people do it. people like publish update guidelines they are not set in stone. So if you are part of this type of research definitely um effort worth publishing and publishing from publishing to deployment there's a lot of um ground that needs to be covered but uh publishing is the first step right there's so much more


00:38:17 - 00:39:18
happening before publishing so if you already publish it's out there can be cited and Um, I remember that conversation and I was like, "Oh my goodness, how are you going to do this? It's probably doable." It was just like the beginning of these uh vision models, transformers, and multimodality in pathology. And I'm like, "Yeah, you could probably label your your ultrasound or whatever with exact corresponding hystopathology images and you have like direct labels." Um, we


00:38:48 - 00:40:09
were basically discussing how how we would approach this type of study. Um, I don't know what happened. Maybe I'll hear from this researcher after this episode. One more thing to cover the lightweight supervised learning framework for domain generalization in histopathology. Yeah. So I want to I told you at the beginning that at ACVP there was the main generalization discussion and there's always somebody so the main shift uh is um when you have a model and and let's do the uh hisystopathology or um pathology example


00:39:33 - 00:40:41
right uh we need digitiz uh digital images for this we need digital slides right so if you develop a model on digital slides from one scanner and then you change the scanner you kind of shift the domain and your model does not perform that well on this new slide and there's always one person who says like what do you mean uh it's not going to perform I see a glomeilus here and I'm going to I see a glarilus here inflammation here it looks the same as inflammation here and then he like went


00:40:07 - 00:41:22
on to saying oh inflammation in human is the same as inflammation in mouse or in dog and whatever yes it is But uh for you as a human observer and you're not a computer that reacts to every single um every single variation in domain including you know staining uh thickness of section and these um slide scanners right we are pretty good even though we have such huge interobserver variability we are going to recognize these glami uh in very different many different um shapes and forms. terms.


00:40:46 - 00:41:56
So anyway, there's always this discussion like how come this domain shift influences so much and how come you have to test this always if it looks the same? Well, it doesn't look the same. That's why we need to test. And the question is here, okay, is lightweight self-supervised learning framework uh going to be good for domain generalization? There is a surprise in this in this abstract. I was smiling when I realized this surprise. So large foundation models in histopathology. Foundation


00:41:20 - 00:42:32
models are basically huge models, big models that can do a lot instead of just like very specific one that detects glami. Right? So we have these foundation models for histopathology and they have demonstrated significant potential in advancing computational pathology very much. Now you know basically you can potentially remove the bottleneck of annotations or reduce it so much. Um you have like companies that are uh hiring pathologists to uh make annotations for these type of projects. I get like inquiries on LinkedIn. Oh 50


00:41:56 - 00:43:06
$50 an hour for annotations and you need to have an MD and the best or or DVM or PhD. I'm like yeah no my job is okay. I don't need to do additional uh annotating jobs but that was what was required and what still is required to have um good performing models right so the those foundation models have the potential to overcome the domain gap between training and test testing data sets so domain gap can be like anything different scanners different labs different stains different whatevers right we don't even


00:42:31 - 00:43:41
probably know like it's a little bit like magic and any of my computer scientist this please comment on domain on the domain shift because when I hear like what kind of domain variability sources can there be I'm like this is a little bit like IHC magic domain shift magic is another magic domain in this digital pathology space anyway however but they are huge right so they rely on vast computational resources and that often limits accessibility and widespread adoption of foundation models


00:43:06 - 00:44:08
And now we have this Histolite a lightweight self-supervised learning framework designed to enable domain invariant representation learning in histology. This domain discussion is like okay um if you know that you're going to have slides uh from different places and images from different places. You want to make a model generalizable. So you want to make it kind of like immune or resistant to domain shift. So you want to include as many domains in development as possible. Uh and that's what happened for these foundation


00:43:36 - 00:44:55
models right um so but we are now having this light model histlite histlite utilizes customizable autoencconders auto encoders within self-supervised learning parading that learns generalized and transferable features in an efficient manner. Um, we evaluated the proposed framework uh using breast hallled images and benchmarked performance with state-of-the-art uh foundation models for domain generalization and we're going to learn which these foundation models are in pathology. Um so they had


00:44:15 - 00:45:43
a novel data set um that was same tissue slides scanned by two different scanning platforms. So here we're focusing very much on the scanner um domain shift um and they did analysis of co-variate shifts due to scanner bias um and they evaluated for example differences in embeddings across scanners. Um and the the top models that they used was Uni Verco 2 Pro Gigapath. These are like kind of famous if you look at the foundation model uh literature and we did we cover I think we did cover uni or or vero one maybe um


00:45:00 - 00:46:16
in one of our digipath digests. So these are like the ones that are going to pop up and in general uh most foundation models were found to be susceptible to scanner bias and as shown by differences in embeddings and drop in performance on the held out scanner. So the one that this stuff was not scanned on. Histolite offer low representation shift in embeddings. Yay. Lowest performance drop on out of domain data with modest classification accuracy. So that's that's the surprise. I'm like okay it


00:45:38 - 00:46:59
was fantastic for domain shift but it was not accurate anymore. So, I don't think they solved the problem, but they indicated that the smaller model may exhibit a trade-off between accuracy and generalization. H So, you know, okay, that's what happens. They made a smaller one, but it was less accurate. Um, and again, these are all the nuances. it then like enters into the kind of like philosophical discussions on different things. But if you have not voted on the sticker that's going to be happening at


00:46:18 - 00:47:31
Sity uh next weekend already, next weekend, Sity near DC National Harbor. Um let me know. Do you want number one? Do you want number two? Do you want number three? And if you're not going to Sity, that's okay. you can vote for the sticker as well because maybe at some other conference I will still have them or Hamamatsu will have them and also maybe Bioare will have them because Bioare is another partner that uh is going to be uh co-sponsoring this coverage of Sity for them. Um also what we would like to do if you are at


00:46:57 - 00:48:08
sity then join for the from spatial biology to clinical utility. Um we are not going to be streaming this one. There's there are going to be live streams from the hamamatsu booth during the um during sity and I think one from the bioare booth. So, you can tune in live or uh or just come to the booth. You can you can visit me and see me live streaming if you want. Although, maybe we're going to do it before the exhibit hall open. So, we'll see. We'll see. But if there is a chance for you to just see me live


00:47:32 - 00:48:51
stream, you can do that. And am I sharing this? Huh? I think I am sharing it. Yes, I am. Okay, good. So if you can join, join and that would be fantastic to see you there in person. H and I wish you happy Halloween. Have fun. If you like dressing up, dress up. I'm like, this is the most I like to dress up and how you see me at conferences. Um I'm not I didn't really grow up doing Halloween cuz I grew up in Poland and it was not a thing. Now it's a bigger thing. Um but still I think there are mixed feelings


00:48:16 - 00:49:33
about this. But here US obviously my kids are already getting ready to put costumes on. Um I prepared this year. I was not sitting till midnight yesterday making costumes here. Pumpkin. Let me know which one you prefer. And also Trailblazers if you didn't didn't get the book. So what happened is I started sending messages on LinkedIn and asking people if they know that there is a book and you know how many people say they don't know there is a book. So if you don't know there is a


00:48:55 - 00:50:02
book called digital pathology 101 especially if you're at the beginning of your digital pathology journey or if you have like expertise in one domain. So when I was starting I came from the image analysis side of digital pathology. So that was like my kind of uh kernel of expertise and then I had to learn everything around and then I wrote this book uh so that you can have this not too long it's going to be updated uh the QR code is for you to download the PDF for free and if you have the PDF for


00:49:28 - 00:50:40
free uh you will be on my email list. That means whenever the new version comes out, which did not come out yet, although I wanted to do it last month, but it will come out before the end of the year, you will automatically get it. So, you know, let's uh eliminate decision fatigue. Get the code, get the book, and you're going to get the um the new version as well. And obviously if we're interested in getting these for yourself, meaning the earrings that I wear from for the conferences, my kind of signature


00:50:04 - 00:51:16
earrings, or if you want to um give them to somebody for Thanksgiving or maybe for Christmas already, if you are so like well prepared that you're already thinking about Christmas, you can do that in the store. There is a QR code in the in the other corner. Um, and anything else? If you're up for a course, um, there is an AI in pathology course, pathology AI makeover now in the corner. But actually, you can get to this course through the store. So, just go to the store, check it out, and if you don't


00:50:40 - 00:51:36
want to go to the store, the book is something that's going to make your digital pathology life easier. Thank you so much for staying till the end. It means you are a real trailblazer and I hope to see you at Sity at the Hamatsu booth 415. And uh let's see, can I do music to say goodbye Halloween music? Have a fantastic Halloween.