222: From Slides to Survival: Can AI Close the Gap?

How close is pathology AI to making decisions that matter in real workflows, real trials, and real patient care?

In this episode of DigiPath Digest, I review five recent papers that approach that question from very different angles. We look at multimodal survival prediction in cervical cancer, pathology-driven response assessment in neoadjuvant immunotherapy for head and neck squamous cell carcinoma, AI-assisted Ki-67 scoring in pulmonary neuroendocrine neoplasms, automation and AI in hematologic diagnostics, and AI-based qFibrosis readouts from the Phase 3 MAESTRO-NASH trial.

What I liked about this set of papers is that they do not all tell the same story. Some show clear progress. Some show where AI already works well as an adjunct. Others make it very clear that validation, governance, reproducibility, and workflow design still matter just as much as model performance.

Key topics and timestamps

• 00:00 Introduction, Easter edition, and community updates 
• 00:51 USCAP recap, signed book giveaway, and free Digital Pathology 101 PDF 
• 02:04 Partnerships, lab automation preview, and what’s coming in this episode 
• 03:25 Multimodal deep learning for cervical cancer survival prediction 
• 13:00 Why pathology may be a better response endpoint than radiology in neoadjuvant HNSCC immunotherapy 
• 23:09 Ki-67 scoring in pulmonary neuroendocrine neoplasms: pathologists vs two AI systems 
• 33:46 AI, digital morphology, and automation in hematologic diagnostics 
• 43:29 qFibrosis, digital biomarkers, and the MAESTRO-NASH Phase 3 trial 
• 51:57 Closing thoughts, community updates, and Easter promotion 

Resources

1.  Deep Learning Can Predict the Overall Survival of Cervical Cancer Based on Histopathological Image, Gene Mutation and Clinical Information
    https://pubmed.ncbi.nlm.nih.gov/41902378/
   
   
2.  Modern Pathology-Driven Strategies in Neoadjuvant Immunotherapy for Head and Neck Squamous Cell Carcinoma: From Residual Tumor Quantification to Spatial and AI-Based Biomarkers
    https://pubmed.ncbi.nlm.nih.gov/41899621/
   
   
3.  Ki-67 Proliferation Index in Pulmonary Neuroendocrine Neoplasms: Interobserver Agreement Among Pathologists and Comparison of Two Artificial Intelligence-Based Image Analysis Systems
    https://pubmed.ncbi.nlm.nih.gov/41898274/
   
   
4.  Molecular Pathology, Artificial Intelligence, and New Technologies in Hematologic Diagnostics: Translational Opportunities and Practical Considerations
    https://pubmed.ncbi.nlm.nih.gov/41897649/
   
   
5.  Quantitative regression of qFibrosis with resmetirom: Exploratory histologic endpoints from the MAESTRO-NASH phase III clinical trial
    https://pubmed.ncbi.nlm.nih.gov/41895606/

Transcript

00:00:00

Aleks: Oh my trail blazers, welcome to the Digipath Digest number 42. We've met already 42 times to discuss abstracts about digital pathology and medical AI. And um today is Good Friday edition. Uh it's Easter edition of the of the live stream. And let me say hi in the chat. who is joining us live today. Let me know where you're tuning in from H and what time it is. It is 6:02 in Pennsylvania. And uh as I wait for everybody to join, let me give you a few updates. So, happy Easter everyone who's celebrating

00:00:51

Easter. H I hope you have a fantastic Easter time. The last conference was US Cap and uh let me know in the chat if you've tuned into the streams because they were pretty popular. The conference live streams are pretty cool. People like them. People um who were there came to the Hamomamatsu booth. We had the book giveaway. There was uh I don't know how many books I've given away but definitely over 30 books. Uh I have signed them. And if you have not gotten a book, um, let me get you a QR code for the

00:01:27

free digital version, you should see a QR code on the screen when you scan it. If you have not joined the list of digital pathology trailblazers, you can join it h and get the free PDF of the digital pathology oneonone uh book in the process. It's going to come into your inbox and then you will get all the updates about what happening at digital pathology place including the announcements for conference streams uh different things like book giveaways uh or whatever else we are organizing together with digital pathology

00:02:04

trailblazers uh maybe meetings in person whatever we do. So uh the highlight well there are many highlights but one cool thing um at US Cup other than uh working together with Hamamatsu uh on showcasing their partnerships. So that was actually part of these partnerships uh was with an um lab automation company um Etherai Etherai Etherai I think you pronounce it. I'm going to show you uh the robot that they have um after we review one paper where they mention actually robots. Um so if you don't have the

00:02:45

book, get the book. Um and we will have a an Easter discount code to the store, but I'm going to show you the discount code after we review some papers because we have sorry this is a different thing. Um we have five abstracts to review today. So, let's start. If you're here, let me know in the chat that you're here. It always makes my day when I see a comments live from you. What time is it? Where you're tuning in from. I'm always super excited to hear that and to see that in the comments.

00:03:25

I think audio and everything is okay. So let's just start with um deep learning can predict the overall survival of cervical cancer based on histopathological images image gene mutations and clinical information and let's just highlight we have couple of things right it's not just prediction from histopathology we have hisystopathology gene mutations and clinical information so Let's see how that went. Did they actually do it? Did they not do it? They did it to a certain extent otherwise they would not be publishing.

00:04:11

Uh so the fusion of data features uh from different modes like pathology images, sequence data uh has the potent and sequence data has the potential to predict overall survival of patients with cervical cancer and that is like we want to know how the disease is going to um behave in the future. So um that's what they like try to extract out of these multi multimodel data. Um and uh they developed a novel prediction model uh for overall survival that incorporates pathology images, clinical data and

00:04:52

molecular data. Um and the model underwent training using the cancer genome atlas tcga. So that's interesting because they used TCGA for training um and they included 119 patients um but then they independently validated the model and they used um a data set sets from Pecking Union Medical College Hospital. So they had real u [snorts] hypo well not hypothetical u that tcga patients uh tcga data for training but then they used a totally independent u validation set of 53 patients with cervical cancer um and they

00:05:38

identified with lasso cox regression they identified relevant features associated with overall survival and what were the relevant features were apparently a lot of genes, 484 genes uh including RGR, DBN1 and um CCR or C A LCR um but also a lot of image features uh and so um they found these like relevant features and then they use the relevant features to build build this multimodel deep learning model um to effectively classify the overall survival of patients with cervical cancer into two categories short-term and long-term

00:06:27

whether short-term less or equal to 3 years and long-term is over 3 years. uh and so here we have to pay attention because uh it was based of integration of pathology images and clinical features uh and they developed a model reasonably with a reasonably good prediction uh accuracy uh AU 0.783 and they say so I'm going to reveal this like mystery that's uh coming across in this abstract that I would have not picked up on if I didn't listen to the um to the AI uh podcast AI paper summary

00:07:11

um that I created. I'm going to show you what that is, but um here so they have this AU right and they have combination of pathology images with clinical and molecular data enables the creation of accurate and reliable prediction models for cervical cancers. So let me change the color maybe. Um here they say pathology images and clinical features and here they say pathology images clinical and molecular data. So when I listened to the summary erh it was important. So the actually pathology images and clinical features uh gave

00:08:03

this uh AU uh of 0.78 when h molecular features were included it went down to 0.6 six something something. And so they used molecular images and clinical for creating this model. But then when they were like testing which features and what was most important actually pathology and clinical data was the strongest uh with all these genes uh actually lowering um the the AU. So um and I want to show you where [laughter] how I know this. Well, um the best way would be to read the paper, but uh I do not have time to read full five papers

00:08:47

and but I do have time to listen to podcast about them. So I created something that is AI based uh paper summaries. Okay, let's see if Yeah, you can see it now. Um, AI based paper summaries are uh what the name says AI based paper summaries and when you look at let's see if I can make it better um bigger. Yeah. So when you look at uh the digital pathology podcast uh feed in uh my podcast host it's called Buzz Sprout. Um you will see these episodes that have they should have a padlock.

00:09:35

Yeah, they have this padlock here and among other episodes. So, let me show you other episodes. So, there's a bunch, right? And there's uh there is these that are like uh more downloaded and the ones that have a padlock. They are behind the subscriptions, a subscription. This subscription is pretty affordable. It's $7 a month. Um, and I'll show it to you in a second, but basically what I have is AI powered summaries of all these papers. So, which one was that? The survival. It's the 216 episode 216.

00:10:12

It's 22 minutes and I listened to them at uh higher speed because they they speak relatively slowly. Um, so in like 11 minutes you can have the content of this uh full paper for the price of maybe a Starbucks coffee. Uh, I don't know how much Starbucks coffee is cuz I don't have Starbucks in the vicinity, but like other fancy coffees definitely are going to be over five um dollar. So let me show you that first. Um, you can uh check it also with the QR code, but I want to show you the

00:10:52

I want to show you the the page where I have it. Uh, this one. Yes. Uh, so here. Okay. Yeah. So you can uh get these summaries for uh it's a subscription um based thing. Why did I put it behind the subscription? Because it's AI generated. I don't want to have the uh podcast feed flooded with AI generated paper summaries. Uh and this is for a very specific audience which is you my uh digipath digest listeners because you're interested in literature. So this is for you h if this is something you're interested to try

00:11:36

out you know try for a month and see if you like it. If you don't like it, great. If you keep liking it, fantastic. And so, you're going to have deeper insights into the into our papers. And now, let's go back to the papers. Let me know your thoughts on that. If I know um some people have already subscribed, which is amazing. So, if you are here, if you subscribe to these AI summaries, let me know what you think because to me, this is like such a timesaver. It's like I don't know if there's any other tool

00:12:13

or AI enablement that saved me so much time as the paper summaries uh to stay on top of what's happening in the digital pathology and medical AI literature. And now let's go back to our papers because we have a lot to discuss too. I'm going to put this code on the screen uh a little later as well. So just at the end you're going to get access to it again. Um and now let's move to the next paper. Okay, here. Oh, this one was super cool as well. this one and also benefit I benefited

00:13:00

very much from listening to the AI summary um because there's a lot more depth that what we have in the abstract um and this one is titled modern pathology modern pathology driven strategies in neoaduventant immunotherapy for head and neck squamous cell carcinoma from residual tumor quantification to spatial AI based biomarkers and the full paper has a lot of details so If you are into immunoncology, spatial biology, uh tumor micro environment, this one is for you and I would definitely read it or listen

00:13:38

to the AI summary as well. But what's happening? Neoadriven strategies in head and neck claim cell carcinoma are reshaping therapeutic partings shifting emphasis from anatomical staging towards biologydriven response stratification. So what's have what's the neoaduvent strategy? Neaduvent therapy is something being given before the surgery to shrink the tumor to make it smaller, make the surgery easier. Um and here the transition from induction chemotherapy to immune checkpointbased and

00:14:15

combination regime has transformed the peroperative setting into a translational platform that enables interrogation of tumor immune interactions and clone selection under therapeutic pressure prior to surgery. That's like a mouthful, but uh basically it just used to be indu induction chemotherapy that was supposed to shrink it, make it smaller and now we are using um immune therapy. And what's happening there, right? Uh so we'll see what's happening. So in this context, pathological response

00:14:52

assessment has emerged as a robust surrogate endpoint surrogate endpoint overcoming the limitations of radiologic evaluation which often fails to capture immune mediated pseudo progression and specially heterogeneous regression. So what does that mean? uh you take an you you do radiology whatever imaging it is X-ray or whatever uh and you check if the mass is smaller because that's what you want to have before the surgery and then you see that it's not smaller h so that's what they called um

00:15:34

pseudo progression immune mediated pseudo progression because what's happening there if you're using uh immune modulating agents imunotherapy there's going a lot of uh stuff happening there. We want the immune cells to go into the tumor and basically fight the tumor. Uh so it's going to fight uh while creating inflammation. Inflammation is going to come with uh edema. There are like four four different um features of inflammation. Color which is going to be um hot uh red. We're not going to see that. um

00:16:09

tumor rubles hot tumor. It's going to be it's going to grow. So that's the pseudo progression. It's going to grow because there's going to be edema. So it's not going to grow because there is more cells. Well, there is more cells, but these cells are immune cells and you're not going to see that in radiology. So they says say pathological response assessment has emerged as a robust surrogate endpoint. Obviously that's uh is associated with taking a biopsy, right? But h it's better to take a

00:16:40

biopsy than to take out something that pseudo progressed and then look under the microscope and say hey it didn't really sudopress pseudo progress. You didn't have to take it out. So um complex concepts super interesting uh paper. It's a review paper. And thank you so much for saying hi. If you're here listening live, say hi in the comments. It makes my day. So [clears throat] um they um developed this so-called residual viable tumor and this provides a reproducible metric of therapeutic

00:17:20

efficacy while characterization of immunated regression beds tertiary lymphoid structures macrofase polarization stages and compartment specific nodal responses. I'm going to explain that. Um again because I listened to the AI summary of this paper [laughter] um so um what do we have um characterization of immune related regression beds so where we have a lot of immune cells but not uh so many tumor cells tertially lymphoid structure. So this is uh what's happening when um immune cells come in they they organize.

00:17:59

So in the um lymphoid uh lymphoid system uh reticle lymph reticular system which is going to be your thymus your lymph nodes um where the immune cells are supposed to be like when nothing is happening they're organized in specific formations like um round centers uh follicles follicles right and then um when they come in to do something in the body where they were originally not present. They form these tertiary lymphoid structures which also are like follicles. Um there's a definition like how big they're supposed

00:18:36

to be. Um also macrofase polarization states and this is this these are things that we're not going to see on H& this is something where we need to do IHT or we need to do uh multiplex imoflloresence. So this is where spatial biology enters meets oncology and uh digitization and digital pathology and image analysis is not an option any not a um elective anymore. It is basically mandatory to understand all these um dependencies. But so macrofase polarization states because uh depending like what the

00:19:13

polarization state is they're going to either uh promote the tumor growth or suppress the tumor growth and um that's happening in uh nodal responses uh lymph nodes uh where you can have micrometastasis that kind of are protected by these macrofasages polarized in a way that does not destroy cancer cells. So a lot of depth, a lot of spatial biology, a lot of tumor immune micro environment uh in this paper they explain it. Um but um the the [clears throat] short story long story short and these give

00:19:52

them mechanistic insights into tumor clearance and resistance evolution. And there is evidence from phase 2 trials, single cell sequencing, spatial transcrytoics, and as we said, multiplex immune profiling that supports the prognostic relevance of pathologydriven endpoints, which I'm going to put a heart here because pathology provides more information than radiology uh in this case. And being a pathologist, I'm happy about that because usually you would consider well radiology is less invasive but like the

00:20:29

outcome of the radiological um exam here when you see pseudo progression would be that there is a very invasive surgery and um instead you do a biopsy and check the pathological end points here. Um so integration of digital pathology and AI assisted image analysis further enhances reproducibility and enables high resolution mapping of residual disease and immune architecture and uh this within this modern oncologic framework the new adivant treated specimen functions as a dynamic biomarker platform guiding response

00:21:08

adopted surgical strategies and biomarkerdriven clinical trial designs. Um, and this was designed as a narrative review. So, they did review different um, papers, but it is pretty thorough. I highly recommend you reading it or listening to the AI powered summary. And thank you so much for saying hi. Hello. Hello to Facebook Trailblazers. And we have people on Facebook, on YouTube, anybody from LinkedIn or Instagram, let me know. [clears throat] Okay, we need to speed up my friends, my trailblazers.

00:21:51

We're on paper two out of five. Let's go to the next one. I think that was the the most the longest. And now we have a a classic in a new way. a classic in a new way. Uh because every now and then we're going to have papers about Kai 67 and how it was used for um like validating some image analysis and I'm like have I not read this Kai 67 paper uh yet uh because I think a week ago or a couple of weeks ago we had uh something about K67 but no this one is new and has cool insights. So a K67

00:22:30

proliferation index in pulmonary neuroendocrine neoplasms interobserver agreement among pathologists and comparison of two artificial intelligence-based image analysis systems. So it's very valuable because um often questions come up oh which image analysis system is better and uh finding these papers that actually compare some commercially available algorithms in peer-reviewed literature um is a nice find to guide your decisions. And what it's going to show us is also pretty interesting.

00:23:09

Um this is I think a publication out of Turkey. So for these um so Kai 67 is not formally incorporated into the grading system of pulmonary neuroendocrine neoplasms. So this is not like breast cancer where you have to do it. Uh you can do it. Uh and that's what they do. It is widely used as an adjunct marker to reflect proliferative activity and support diagnostic stratification. So Kai Kai 67 is a marker of proliferative activity and uh as we know manual Kai 67 as we know anything [clears throat] manual in

00:23:53

pathology assessment is subject to inter observer variability and method methodological limitations. uh and this study aimed to evaluate the reliability and performance of two AI int AI based image analysis systems in K67 index assessment and to compare their results with expert pathologist evaluation right so like a classical setup we have experts that uh used to do it they they still do it this is like our center of care and then we have AI we compare and we see are they at least taste similar or good enough can we use

00:24:34

AI and how right so we had 63 pulmonary neuroendocrine neoplasma cases um and there were typical carcinoid atypical carcinoid and large cell neuroendocrine carcinoma and they were analyzed and k7 proliferation indices were independently assessed by four pathologists so I'm already impressed you've got four pathologists to do it for Um, amazing. Usually like getting one to do some work for paper is difficult. They had four. So, this is good and um important within predefined hotspot regions. Why do I

00:25:14

know that it's important? Because I listen to the AI powered summary. Uh, so they had this predefined hotspot regions counting approximately 2,000 tumor cells per case. >> [clears throat] >> Four pathologists counted 2,000 tumor cells per case and congratulations to the research team to make them do it for this publication. I am impressed. So the same regions were analyzed using two AI based image analysis systems rash upf 67 and Verasoft Viright Kai 67. So the these systems basically count uh the K67

00:25:55

positive cells but they have like different ways of counting them and in the paper uh it's explained like how different they are in uh what they show um but when it comes to the results and and it like the way they work the AI uh powered systems um is going to be important to you if like you want something that works like you. So one of them and I don't remember which one was that it was just doing a binary call positive negative. The other one was like grading oh high intensity, medium

00:26:31

intensity, low intensity and you can kind of set your threshold. There's always like this those so K67 is a nuclear marker. There are always these nucleides. They're like grayish bluish. You don't know if they're positive but they're not like super negative. So you have to decide on your threshold right and not only am I impressed that and that but I'll tell you in a second that they have four pathologists they actually agreed very much um so they were um they were evaluating the

00:27:04

interobserver agreement h and um they also took into account domain shift so slides were digitized using two different halls scanners Ventana DP600 00 and Leica Apiria 82 and they used cor color normalization and quality control proc procedures were applied prior to AI based analysis. Nice nice work. They like took care of potential sources of bias and the Ki 67 indices were stratified into categorical groups based on tumor subtypes specific thresholds. So we have thresholds that are specific to these uh

00:27:43

different tumor subtypes. So there was low um less than 10% 10 to 25% was intermediate and 25 over 25 was high and then agreement between manual and a based categorical cost scoring was evaluated using coins kappa coefficients. Uh and the results were that the Kai 67 proliferation indices varied across tumor subtypes. Um which is to be expected. U typical carcinoids showing low atypical intermediate and large cell neuroendocrine carcinomomas high proliferative activity. So then we have like tumor biology. Okay. Some

00:28:28

subtypes have more proliferative um activity, some less. But listen to this. The interobserver agreement among four pathologists was excellent. It was 0.98. Like can you imagine four pathologists counting 2,000 cells and they had this agreement? Um, I'm impressed because usually we're going to like comb some across something like 0.7 0.8 maybe and here like almost perfect. Uh, but how how were they so perfect? Well, we did have the predefined hotspot regions, right? You're not going to have

00:29:11

that in your normal workflow. And they did count the cells for real for the publication. Usually, I don't know how many you count. uh you should but probably you visually estimate that's how it usually happens. Uh so strong correlations were observed between manual K67 assessment and AI derived indices. Uh and there was 0.961 for ROS AI and 0.904 for Veroft AI and 0.926 between the two AI systems. So they're like both over 90% concordance with the pathologist. They're good enough. I would say they're

00:29:54

good enough. Um so that indicates high concordance without systematic bias and categorical agreement analysis using subtype specific K67 threshold showed excellent concordance between manual and AI AI based scoring. So, um, then also in this AI powered summary, uh, they like they're cool because there's there's two hosts and one is always like throwing wrenches into the, uh, into the conclusions or like statements in the paper and the other one is defending and explaining it. It's

00:30:30

fun. It's fun. Like, give it a try. Let me put it the Let me put the QR code on the screen for you. uh just you know try it for a month and see what you think about it. I I like it saves so much time and it gives you so much deeper insight into these papers um than the abstracts without actually reading them. It doesn't replace reading but it's as close as it gets to reading without reading as possible as I have encountered so far. Um okay but let's finish the conclusion of this one.

00:31:08

Erh so this confirms that CL the clinical interpretability and reproducibility of AI based K67 assessment and K67 I also want to say it's something that like if you want to introduce AI into your lab if you want to um like gain confidence in these algorithms Kai 67 is something to start with because it's going to be used for every probably type of tumor if not as official part of diagnostic guidelines then like as an adjunct marker as they called it here and um it's pretty easy for AI to be good enough and then easy

00:31:48

for you to validate and but also gives you insight on how these u sorry how these tools work how to integrate them in your workflow. I remember when I was working in a a digital pathology company at the beginning of my digital pathology career, uh we wanted to create a Kai 67 tool um to have a 510k clearance on that to like enter the diagnostic test um world so to say. So conclusions. A based Ki67 index assessment shows strong concordance with expert pathologist evaluation and reflects biologically

00:32:26

relevant differences among pulmonary neuroendocrine neoplasm subtypes and um AI assisted K67 may serve as a reproducible and objective adjunct to routine diagnostic practice in pulmonary neuroendocrine tumors. Basically they say it's good enough. um they validate it and I like these proof of concepts uh of like classical things because it indicates um that people are adopting people are looking into it and Kai 67 is a very good uh candidate to start your uh AI powered AI [clears throat] image analysis journey also for

00:33:06

diagnostics. Um, and if you want to learn more, then you can subscribe to the AI powered summaries. But we are not done yet. We have three two more. And if you're still listening live, let me know in the comments that you're here. Uh, it definitely makes me smile and I like featuring you on the screen. Okay. So, oh, this one is I'm like everyone every every of these one is cool [gasps] and they really sound a lot cooler once you listen to the AA summaries. Um, and now I stop like repeating that. Let's focus

00:33:46

on the literature. So molecular pathology, artificial intelligence, new technologies in hematlogic diagnostics, uh translational opportunities and practical considerations. So hematologic diagnostics uh is I would say under represented in these digital pathology papers although there is a lot of algorithms for um for hematology. But let's look at this abstract. Diagnostics for hematlogic disease rely on integrated assessments of clinical manifestation, morphology, flowcytometry and molecular testing.

00:34:30

And their current classification systems include the WH hem 5 and the international consensus classification. I think they call it IC later in the abstract and highlight the central role of genomics in defining disease entities and risk. So genomics is important and then at the same time laboratories face growing case complexity and staffing challenges automation collaborative robots and it's green on purpose. I'll show you something in a second. uh also [clears throat] known as cobots. Did you

00:35:09

know that word? I didn't know. I thought it was like invented at US Cap when uh when the um when the executive from Ether AI interviewed him, Joe, I'm going to show you the video in a second. And he says, "No, no, cobot is a collaborative robot. It's a an official industry term." So I'm going to show you a cobot after we after we discuss this paper and then also digital morphology platforms and artificial intelligence have begun to address these issues with issues the um

00:35:46

complexity of cases and staffing challenges. Here we examine the application of technologies in hematopathology and molecular diagnostics and consider their translational potential to improve diagnostic accuracy and ultimately patient care. It's so interesting because in every other paper you're going to have this improved diagnostic accuracy and ultimate patient care. And when you go to the conference uh like all the manufacturers have to be super uh cautious about making these claims.

00:36:19

So like often you cannot use these words when you are like in a real life setting with a device that does not that is researchers only does not have the necessary clearance whereas like every other paper is like improve diagnostic accuracy and ultimately p patient care uh but it's not a medical device paper is not a medical device reviewed by the FDA it's peer-reviewed by peers right so a review of peer-reviewed literature and technical reports published through December 2025 was performed and they

00:36:52

were focusing on digital morphology platforms, AI for peripheral peripheral blood and marrow interpretation, AI enabled flowcytometry, automated and robotic deployments in clinical laboratories and machine learning applications in molecular hematopathology. So the important thing here that does not come across in the abstract is that they reviewed and they will tell you specific uh company names in this paper. So if you're looking into solutions for hematopathology definitely look into this because they

00:37:27

mention specific automation platforms specific scanning platforms specific AI platforms like you can actually go and look up these companies and call the manufacturers if you're at the point of your hetopathology journey that you want to introduce any of these technologies and if you are let me know and the results Here were that digital morphology analyzers show strong concordance with manual microscopy and now serve as efficient platforms for a assisted differentials cell classification and fibrosis

00:38:04

quantification. I don't know why they quantify fibrosis like I don't know how you quantify fibrosis on the smear. Um but I bet they explained that uh deep learning applied to multip parameter flow sitetometry achieves performance comparable to expert review in distinguishing mature bell neoplans and acute leukemia and here automated solutions cobot systems and robotic armbbased slide scanning clusters have demonstrated substantial gain in throughput and pre-analytical consistency. We're going to get back to

00:38:45

this one. H and the AI models in molecular hematopathology increasingly assist with uh variant interpretation, genetic risk stratification and linking morphologic and genomic findings. So the conclusion is AI is being beginning to change how hematopathology and molecular diagnostics are practiced and successful translation will depend on disease specific validation. always always disease specific validation the development of multimodel models aligned with ICC and WHO frameworks and laboratory governance that maintains

00:39:23

expert oversight and let's look let's look at the cobot let's look at the cobot and I'm going to tell what it means well it's collaborative robot what does that mean collaborative it you can like interact with it. I think it's this one. Let's see. Okay. Yes. Yes. So, uh that was a use cap. H and um this is Joe. Yeah. From Ether AI H. And here in the background, you can see the robot, the cobot that is feeding slides into a Hamamatsu [clears throat] nano zoomer scanner. So it takes it's so

00:40:08

precise like you can see let me put it in let me put the link to this one in the can I copy link yeah I can put it in the chat no no no watch >> the topic of US cap is making >> I don't want this one okay I'm going to post the link to it if you want to like have a look at it h and look everything but basically cobot means like uh in you could touch it and it would stop working It's not supposed to take any more space than a person would h or it should take less. And when I [clears throat] was

00:40:43

talking to Joe, he was describing basically um like how that um departments don't want to go digital and this is like kind of tragic tragic and a big opportunity at the same time. So departments don't want to go digital because when they tell their uh technicians that hey now for eight hours a day you need to like put the slide into the scanner and then do visual QC look if it's blurry they're like I'm not doing it. I'm like not going to be working this job. So basically they fear

00:41:25

losing qualified, highly sought after uh professionals that are already in shortage by going digital because uh instead of like giving them a different job, they give them the job of like manually uh moving stuff for several hours a day. It's so funny the comment under this post. You can um you can read the comments uh when you go to the post directly. But a person one person was like that gives me scary flashbacks to a previous job I had. I worked at as a pathology assistant at the hospital. It

00:42:01

was I was essentially tasked with creating an organization system for the hospital's decades worth of tissue slides. Spent my days in the basement filing glass slides in file cabinets. Um, what they didn't tell me was that the basement was also a more wild job. Not positive experience, but he's happy that the robots are stepping in. Um, so yeah, that was an interesting thing. Uh, if you're on Facebook, this post is currently trending on Facebook. I sent you a LinkedIn um link. Um, so yeah,

00:42:33

that's cos are coming in, but we have one or two more papers. Stay with me. Stay with me, trailblazers. Let's go back to our papers. We can do it and then we can celebrate Easter. I think it's going to be the last one. Oh, this one is cool as well. I know I said it before every This one is cool. This one is cool. Yes, I chose them because they were cool. Hey, I did it on purpose. I chose the cool ones. Okay, so quantitative regression of Q fibrosis with resmet again quantitative regression of Q

00:43:29

fibrosis with reset exploratory hisystologic end points end points uh from the Mastra phase three clinical trial. So important things is we have a cl a phase three clinical trial. We have hisystologic endpoints and we have um a drug right obviously it's a clinical trial. So you are triing a drug. So we have like very close to reality research experience right? We are still triing. We are still trying to prove that it's actually working and we have hisystologic end points. I think yeah they were exploratory. So they were

00:44:14

checking um if they're granular enough good enough and we already had an example of these hisystological end points with the head and neck cancer um where we wanted to avoid pseudo progression and needed to look at pathology. And here this one is kind of complicated what they uh evaluated on histologology but um the a part summary explains it very nicely. So let's have a look at the abstract though. Resmetum a thyroid hormone betaagonist has been shown to improve metabolic dysfunction associated

00:44:48

statpatitis MASH and fibrosis in patients with noncerotic mash in a phase three serial liver biopsy study. So serial biopsy study is you take a biopsy at the beginning of treatment and then at certain points I don't know what their points were but basically you compare what's happening in the tissue um throughout the treatment and the impact of resetum compared with placebo on histologic fibrosis features was evaluated using AI based digital pathology Q fibrosis in the phase three master nash trial which I love it

00:45:26

because they are evaluating an AI based tool in the clinical trials they are doing it. That's an exploratory end point. But hey, let's check the results, right? So the methods uh first let's check the methods. Uh Q fibrosis was a secondary analysis of liver biopsies from 966 patients. A lot of patients with biopsy confirmed MASH and fibrosis stages F1B, F2 and F3 enrolled. So there is probably a definition. Well, probably of course there's a definition what these stages are. Um um and we would

00:46:00

have to go to that back to that definition. I don't have it right now. H it was a multic-enter double-bladed placebo control phase 3 master nash trial and biopsies from baseline in week 52. So here this is the like serial biopsy we're assessed using second harmonic generation and two photon excitation fluorescent microscopy. um and the prespecified assessment according to treatment was conducted of Q fibrosis continuous values. So QFC are the continuous values, categorical stages, QFS the stages and Q statosis and post

00:46:39

hawk assessment were conducted of regional Q fibrosis features and 30 clinical outcome associated Q fibrosis features including correlations of the 30 features with pathologist assessment fibrosis improvement and non-invasive tests. Right? So we correlate this hisytologology endpoints with other different endpoints including pathologist evaluation and we had uh different doses 80 mg and 100 mg and they led to the uh Q fibrosis score improvement. It decreased by one or more in 24 uh.4 and 22.3 uh more

00:47:18

patients than in placebo respectively for these two doses. H and then it uh reduced QFS worsening. So that was an interesting con concept as well. Okay, it improved something that was bad, but it also decreased the worsening, right? So it didn't didn't progress so much or didn't get as bad as it could have gotten with placebo. And then mean placebo corrected reductions in Q fibrosis continuous scoring were minus0.95 and 108 uh respectively for the doses. Conclusion is that Resmmet treatment led to

00:48:02

significant improvements in QFCE QFS and individual collagen features associated with fibrosis progression. These digital pathology findings support the anti- fibrotic efficacy and of resetum and demonstrate the potential of AI based quantification to help define the fibrogenic response in mASH and um the impact and implications. This is the first evidence that AI based digital pathology including Q fibrosis and region specific collagen features. So region specific collagen features it's something that is important here because

00:48:44

regions like microanatomic hisystological regions where things are happening not like the overall quantification not like the size like we had in head and neck or anything but region specific stuff in this case collagen features uh before it was immune cell infiltrates and the profiles of different cells right sensitivity detects antifirotic treatment effects of resaterum in mash beyond the conventional ordinal staging based on the phase three pivotal trial. And these continuous metrics highlight

00:49:16

nuance pathophysiological mechanisms of fibrosis progression and reversal in mass LD including early matrix remodeling allowing for more granular reproducible and biologically plausible analysis. I'm going to put the dot here as well. Uh near biologically plausible and the ability to quantify distinct collagen remodeling patterns especially in portal tract and zone 2 regions could help refine the understanding of the anti fibrotic effects of treatment for mash including the impact to reduce

00:49:52

fibrosis progression to cerosis and clinical outcomes. So this is a lot a mouthful again but what it's basically telling us we have visual biomarkers that can be quantified by AI that give us a lot more nuanced information about what's happening with the patient and how the patient is reacting to the drug than the categorical scoring which was the best we could have before. So, uh now unlike Kai 67 where you can like directly reproduce it uh visually and if you're like good and diligent you actually

00:50:38

count these 2,000 cells here you cannot do it. H and it is taking us into the it has already taken us into the era of um digital biomarkers. digital biomarkers. What do I mean by that? Um is biomarkers, visual biomarkers on images that can be these still can be confirmed by a pathologist. Like pathologists can tell you, oh yeah, it actually is in uh zone 2 regions. It's actually is in the portal tract and I see this red modeling pattern, but they cannot visually quantify it. So you need a tool to

00:51:18

quantify and put all these features together um to be able to uh derive insights from this whereas pathologists can describe it, can see it. Um and that makes it a super cool tandem uh super cool collaboration and a good collaboration of the pathologist and the machine where um you have the expert in the loop confirming verifying that the quantification is of the right things. So that what happened in this clinical trial and it happened in a clinical trial which like gives us even more real life validation.

00:51:57

Um if you want to dive deeper into these abstracts the next step is going to be the AI powered summaries. The QR code is on the screen right now for you. Um then of course if you're here for the first time or if you have not grabbed digital pathology 101 yet that we gave away at US Cup. Uh so if you have a signed version let me know in the comments that you were there and uh it was so nice to meet you all in person. I'm going to be posting some videos of our book signing event at the Hamamatsu

00:52:36

booth at US Cup. It was so nice. It's like so heartwarming to see you. There's always like this moment I guess in every author's creators or somebody who puts stuff out there's always like what are you gonna say? are they even like interested in this? And so many people came and said, "Thank you so much. I want to take it home for this and that person." One person told me that uh oh, they already know this book, but of course they want a signed copy. But uh when it first came out, they printed the

00:53:08

PDF at work and highlighted everything. And I'm like, really? It made my day. So if you have the book, let me know in the comments. It's going to make my day. If you don't have the book yet, I'm putting a QR code on the screen right now. Changing the QR codes and the PDF is free. Um, and you can just see what's happening in digital pathology. And I am working on the second version, second edition. Uh, and I'm actually committing to having it ready in two weeks. Why in two weeks? Because I'm going to be

00:53:42

traveling for two weeks. Uh I'm uh starting a sabbatical next week to write publications and to basically write uh accelerate my scientific writing. So the new version of this book is going to be out uh after this trip and I'm going to be streaming for you from the travel from the travel. I'm not going to tell you yet what kind of travel it's going to be. You're going to see in the next live stream. Uh, I just am going to tell you that I needed Starlink to be able to do this. So,

00:54:20

uh, I hope you join me next Friday because it's going to be from somewhere in the US. It's going to be in the US. And let me just check if we have I forgot. We have a promotion Easter promotion for the store. So, here code to the store is in the left corner. I'm putting away the book code and there is a code promotion code. So if you go to the store and you at checkout use the code Easter 25 sorry I have a special tool for writing Easter 25 h on anything that is in the store you're going to get a 25%. So, if you

00:55:08

have been like on the on the fence whether to take the AI course that's there now, you're going to get a 25% discount. Uh, or if you wanted earrings here, the ones I'm wearing for every conference or whatever you see, they're in the store. Uh, everything is discounted by 25% until uh, April 8th. Um, so take advantage of that if that was something you always wanted to do and now you have some time. Um, it's going to definitely save you time >> [clears throat] >> uh versus doing it yourself or like

00:55:48

digging up other resources. Um, I think anything else? No. Thank you so much for joining me. Thank you so much for staying till the end.