205: What Makes AI Useful in Pathology Beyond the Demo?

What happens when AI looks strong in a paper, but the workflow still isn’t ready?

In DigiPath Digest #40, I reviewed five recent papers across kidney pathology, oral and maxillofacial pathology, glioma biomarker prediction, digital twins in neuro-oncology, and a major European colorectal cancer cohort. A common theme kept coming back: good performance is not the same thing as real-world readiness.

We started with kidney biopsies and the challenge of assessing interstitial fibrosis and tubular atrophy, where AI shows promise but still does not fully agree with humans. That led into a bigger point I keep seeing in digital pathology: our “ground truth” is often based on human interpretation, and human interpretation has variability too.

From there, I looked at AI in oral and maxillofacial pathology, where the field is still early and one major bottleneck is the lack of strong public datasets. Then I discussed a systematic review on adult-type gliomas showing that multimodal models performed better than unimodal ones, which makes sense when you think about how pathologists actually work: we do not diagnose from one input alone.

I also covered a systematic review on digital twins in neuro-oncology. The idea is exciting, but the paper makes it clear that reproducibility, public code, multimodal integration, and external validation are still limiting factors.

And finally, I talked about a paper I really liked: a large European colorectal cancer cohort built across 26 biobanks in 12 countries. That kind of harmonized, quality-checked dataset matters. A lot. Because better AI starts with better data.

In this episode, I discuss:

•  Why AI vs human comparisons are harder than they first look 
•  the “gold standard paradox” in pathology 
•  Why multimodal AI keeps outperforming unimodal models 
•  What is holding digital twins back from broader use 
•  Why curated multicenter datasets are so important for digital pathology research 

Resources mentioned:

•  Digital Pathology 101 pdf copy
•  Pathology AI Makeover Course 
•  DigiPath Digest AI-powered paper summaries 

Papers discussed:   

• https://pubmed.ncbi.nlm.nih.gov/41830415/
• https://pubmed.ncbi.nlm.nih.gov/41826004/
• https://pubmed.ncbi.nlm.nih.gov/41824546/
• https://pubmed.ncbi.nlm.nih.gov/41823607/
• https://pubmed.ncbi.nlm.nih.gov/41820399/

Transcript

00:00:00
Aleks: Trailblazers, good morning from Fairfield, Pennsylvania. 6 a.m. today. I am saying hi in the chat to everyone who's joining right now. And as always, let me know uh where you're tuning in from. What time is it for you? I love seeing people from all over the world. And maybe one day I will stop doing these live streams at 6:00 a.m. And then let's see who comes. Or maybe I'm going to keep them. Let me know if you want them in the morning or you want them h at some other time. And uh coffee ready.

00:00:41
Um as we wait for you to join. Okay. And let me see if we have problems with connecting. Let me know which platform are you on cuz last time we did have a problem with connecting all the platforms. So, uh we should be on LinkedIn, Facebook, um Instagram and YouTube. So, just drop me a comment where you are uh watching it. And I'm just going to give a quick question to my team to ask if we're actually streaming everywhere because last time I was like waiting waiting and not usually more of you show up um and that was due

00:01:27
to just not streaming to all the platforms. So one second for that. Um, and we might be troubleshooting if we're not streaming everywhere. But we kind of like troubleshoot. Uh, we did the hisystologology live stream. So, that's a new initiative. A couple of updates before we dive into the papers. Uh, new initiative that we're going to be doing. We're going to be doing hisystologology live streams because um there are several hisystologology videos on the YouTube channel uh on the on my

00:02:03
YouTube channel, Dr. Alex and Artificial Intelligence and Digital Pathology. and they uh are very popular. That means that there are trailblazers that or early trailblazers that benefit from histologology teachings. So that's what we're going to be doing. If you're interested in that as well, feel free to join. And if you're on my list, you're going to be getting this information as well that we stream histologology. Let me know which platform you're uh you're listening in on.

00:02:39
Okay. So, I think everything is working. But if you're joining on any of these, let me know and let me know what time it is for you right now. Update, conference update, US Cap. US Cup, we're going to be going. I'm going to be going. I'm going to be going on Saturday already. And I'm going to be joining a sponsor, Hamomamatsu. And the booth that you can find me at is 312. Let's see if I can like show it on the screen so that you guys know where to find me. Three. What? Okay, that didn't work out. Um,

00:03:24
doesn't matter. 312. Let me put it on the screen. Um here. Okay. US Cup this booth 312. And uh what are we going to be doing at this booth other than hanging out with the Hamamatsu team uh that has amazing scanners that are produce high quality images uh for different purposes. There is going to be a book giveaway. So, if you are interested in getting the physical book copy, if you don't have it yet, um, signed by me, then this booth uh, is going to be where we're going to have the books. I think they have been

00:04:18
delivered. I just got an email that the books have been delivered or are in the process of being delivered to San Antonio. Uh, so booth 312, I'm going to be there Monday, Tuesday. uh come and grab your book. We're gonna have several of them. So, it's not just, you know, three books and they're gone. It's going to be tens of books. Uh so, if this is interesting, okay, and I see YouTube is working. Fantastic. Thank you so much for confirming. If there is anybody on LinkedIn, let me

00:04:53
know. I don't know if Instagram comments work, but okay. So, uh, at this booth books, there's going to be cool stickers that you can put on your water bottle or on your mug. Uh, but the main thing that we're going to be talking about, and that's also the theme of the conference is connections. Uh, connections between the uh, healthcare providers and the vendors, between different vendors. So, that's going to be the main theme because digital pathology doesn't happen in isolation. It's not just the scanner.

00:05:26
It's a lot more to that and as we're going to see in the papers that we are um going to be reviewing today that it starts a lot earlier than the scanning. It starts with the staining with a basically digital ready specimen. So if you're there and you see me, wave at me. Uh don't be intimidated by the camera. Just say hi. And uh when in doubt, come to the booth 312 and where I'm going to be working with the Hamamatsu team. We're going to be recording some podcasts, uh recording some social media

00:06:02
clips. So give us some love on LinkedIn specifically and on other social media. Um and I'm going to give you more updates, but let's dive into our papers. For those who are already here, thank you so much for joining. I am so happy that you're here at 6:00 a.m. Let me know if 6:00 a.m. is actually a good time for you. Maybe we could double the number of people if it wasn't at 6:00 a.m., but currently and this is the time I can I can do it. But in the future, maybe there's going to be a different time.

00:06:34
So, give me your preferred time in the comments. And let's look at this. We have a couple of abstracts today. And I started doing these uh podcast, these AI podcasts. They are um a subscription uh version of the podcast where the AI hosts dive deeper into those papers. And now I'm realizing how much more there is to the paper than the abstract. I mean I knew that but um just highlighting that and I'm going to give you more details in a second. So assessing interstitial fibrosis and tubular

00:07:09
atrophy in kidney biopsies artificial intelligence intelligence versus humans who's going to win can you anticipate the answer uh I think by joining me a couple by the way it's the 40th digipath digest if you have joined me let's say at least three times you will see the pattern so the AI is undergoing intense study and applications of this are also true for interstatial fibrosis. And why is this important? Because this is a surrogate measure of chronic kidney disease progression. And obviously as

00:07:48
with any visual assessment, we will encounter interobserver variability among human pathologists. Um this has been demonstrated um as with every visual assessment of quantity. I call it gueststimation but this is also standard of care. Um and recent findings are that computerized assessment of interstitial fibrosis including with AI has been um they assessed it alongside with pathologists and it um was measuring the fibrosis. Well, so it has shown interstitial fibrosis measurements and indirect

00:08:29
assessment through kidney compartment segmentation. But the studies has have shown lack of complete concordance among computerized methods and humans. Um does that surprise us? No. Uh because first of all humans are not concordant with each other. uh and we are comparing kind of um what do you say oranges and apples like different types of fruits uh but because the pathologist assessment is still our ground truth we are in this trap um and I have mentioned this publication several times there by a colleague of

00:09:08
mine fam at al the gold standard paradox where there is this paradox the pathologist evaluation or annotation is the gold standard But actually in every single abstracts of the ones that we are reviewing there is this statement there's this sentence there is inter observer variability right so our gold standard is has bias h anyway so uh they say there's lack of complete concordance and studies studies have still shown the persistent value of human assessment in many circumstances and here is our conclusion that you

00:09:48
would have predicted after at least three digipath digests. Ultimately, humans working with AI may provide enhanced analysis for more effective patient care. And that's what I'm doing to update the book. Let me give you the QR code for the book and I'm going to tell you like if I was starting today digital learning digital pathology. So, if you don't have the um where is my book code here? You should see a a QR code for the book on the screen and digital pathology 101. The QR code is for the digital

00:10:28
version. So, um because I get questions from trailblazers who are just starting in digital pathology, they're asking okay which course which uh how to do it and there is a general framework pick something and stick to it because this is a super dynamic field. So that was uh the reason why we started digipath digest right because every week we get like 10 papers 11 papers or more uh that have the keywords digital pathology and AI so there's a lot happening h and by following it over time uh you can see

00:11:00
the trends you can see which aspects are more advanced which are less advanced so if I was starting today I would start with the book uh which you have the QR code uh down there then what I would do I would take the like and I am mentioning my resources right if you find different resources all power to you right uh but the fastest thing that I would do to take the book first then I would take the um pathology AI makeover course which talks in general and not only in general um specifically about uh

00:11:35
but gives you intro and a little bit more depth to uh AI in pathology and then what I would do um I would do the AI powered paper summaries cuz uh we started doing this with my team three digipath digests ago or four and I'm reviewing them and oh my goodness this is a closest to reading several papers a week um as it can get without actually reading these papers. It doesn't replace the reading but it gives you a lot of competitive advantages. So that would be my framework. I would start with the book obviously

00:12:17
and the book is free. Uh so putting this on the screen one more time. If you don't have the book get it and if you are at US Cup you can get a physical copy. Now let's move to our next paper. If you have any questions, any comments, let me know in the chat. Also, what time is it? Where are you tuning in from? Give me some comments. Show me that you're here, that you're actively listening and not just falling asleep because it's very early. Okay, let's do this one. Codes can go away for now.

00:12:56
I'll put them back at the end of the stream as well. Um and the next one is artificial intelligence and its applications in oral and maxulacial pathology. I like these um other applications uh of of AI because usually it's surgical pathology uh and uh sometimes we get other publications in this oral and max facial pathology. uh and they state very uh clearly that AI in oral and max facial pathology is relatively n in nassent stages and but can be used to assist pathologists in detecting certain

00:13:38
lesions for example squamous cell carcinoma classifying tissues and prognosticating the behavior of tumors. So uh unfortunately the comparing uh comparison of performance is uh difficult because we don't have a robust publicly available data set. So we don't have it even for surgical pathology although there is a data set that was created in Europe. That's our last abstract. So stay till the end. Um just opening chat to see your messages. Um so maxel facial pathology not so many data sets. Um but uh this is important

00:14:17
due to the large degree of pre-analytical variance in the creation of digital pathology images. So that's a theme that's um coming up over and over again. Um so that's it I guess. Well yes for this one we learned that uh it's relatively in nent stages in maxular facial pathology. There is obviously more in the deep dive um AI based summary but let's move to the next one. This one is cool. I love this uh type of publications because they are kind of transformational. They're not um what do

00:15:01
I mean? So the title is the performance of AI in classifying molecular markers in adult type cliomomas using histopathology images. Uh this is a systematic review. I like these as well because the team uh went through the effort of actually sifting through all the other publications. Uh and we have people joining from different places in the world. Let me highlight you. I am so happy when you comment on these. Uh so we have 614 from Atlanta. James, thank you so much. We have 11:14 in Sweden. Any other countries? um give me some

00:15:40
some more information where you're tuning in from and what time is it? And I see uh LinkedIn is working, YouTube is working. So all good. If you're on Facebook, let me know. Um and Instagram, let me know in the comments as well and um we'll see if they come through. But so um going back to the publication, right? Uh what is transformational? this prediction of um molecular markers from histologology. Uh wow we have people from Qatar. Amazing. Fantastic. So um why is it transformational? Because you can

00:16:20
skip a step so to say. You have the image and then for molecular analysis you need to uh use more tissue do additional tests the molecular tests. H and here the computational models can actually predict the mutational status um from the tissue. So obviously uh there is these are predictions right that would need to be confirmed but there's plenty of places in the world where this molecular analysis from the tissue is not possible. So in those if I was in those places h I would be happy with a prediction versus nothing. uh

00:16:59
obviously uh assuming that the prediction is better than tossing a coin. So let's see if our prediction here is actually better than tossing a coin and uh when is it good and when is it not that good? So uh these adult type glomomas are among the most prevalent and lethal primary central nervous system tumors and the molecular classification particularly the detection of specific mutation isocitrate dehydrogenase mutation and 1p19q codilion has become crucial for accurate diagnosis and prognosis. So uh when I

00:17:41
was listening to the AI summary uh I learned that this uh was an update to the WH guidelines where you cannot do morphology only for diagnosis anymore. You need to run these tests to be able to diagnose. Right? So that's like a mandate. Um and existing reviews mostly focus on radiology rather than hystopathology. uh but here the objective was to evaluate the performance of AI models in detecting and classifying IDH mutation status and one P19Q gene codelation in adult type clomomas using histopathology

00:18:19
images and uh this systematic review searched seven databases Medline um psycho info mbase and all the other uh different publication databases and the publications were from 2015 to 2025 and they found many many reports. They found 2,453 reports and you know how many they identified if you're watching on YouTube and not just listening you know that it was just 22 studies out of 2453 22 studies were meeting the inclusion criteria and uh out of those 22 studies that pulled average accuracy sensitivity

00:19:07
specificity and area under the curve across studies were 85.46% 46% 84 uh.55 and 8603 uh and uh 86.53 respectively. So we say uh in the medical field like it has to be above 90 for us to trust it enough. Um when it was just a unimodel model model uh it wasn't over 90 but hybrid models demonstrated the highest diagnostic performance accuracy uh over 92 almost 93% sensitivity 89.62 and AI models that used multimodel data outperformed those that use unimodel data in terms of sensitivity and AU. Uh

00:20:03
are we surprised by that? Are you surprised that multimodel achieved better accuracy than unimodel? I'm not surprised uh because it basically has more information, right? It's mimics the clinical workflow where the pathologist uh is not only looking at the image, they have the clinical history, they have other type of uh information, they have the radiology and basically everything uh that happened to this patient. So they also use multimodel data types. Um and here multimodel um AI was better than unimodel even though I

00:20:40
guess we still like hope that uh there's going to be patterns found in the images that people cannot discern visually which can be the case. Um but what I have seen so far in the last couple of digat digests abstracts is multimodel performs better than unimodel but it's more challenging logistically to run these models. So um models had a better overall performance in identifying IDH mutation than one p9q codilution. uh and all models designed for binary classification h exhibited lower performance than those

00:21:21
for multiclass classification in terms of accuracy and sensitivity. Um and here when I was listening to the uh AI summary uh I learned that um it is because when you have multiple classes uh the model has to focus on u more more features of the image. So it kind of like learns the image better uh and then has to distinguish between these classes and um has better performance than just binary classification like um mutation yes or no or I don't know what the classes were but it's in the AI summary.

00:22:00
Um and the conclusion is that models show strong potential um for the molecular classification of adult type glyomomas using histopathology images specifically or particularly for the IDH mutation detection. H but we are constrained by the limited number of studies uh there's only focus on adult type glomomas lack of meta analysis and restriction to English language publications. So um although most of the um most of the publications are in English that I get from these databases right but obviously that doesn't mean

00:22:37
that these are the only publications um sometimes we get other like French publications and other languages that because we do it in English we don't uh take into consideration so that's a valid concern um and AI offers v valuable diagnostic support must be integrated ated with expert clinical judgment and I basically see this integrated with expert clinical judgment in every single um abstract here. Uh also I'm now I am working on the uh next version of this book of add on edition two h and

00:23:18
I work heavily with AI and even for writing even just large language model based on transcripts based on the previous version um you would be surprised how much uh how many like shortcuts the AI does how many redundancies these even like the good paid models. So, I'm using uh mostly Claude for this and I have to argue with my Claude uh a lot. So, if you are interested in the book, I'm going to put the code on screen one more time. Let's have a look what else we have prepared for today.

00:24:01
Let me know if you have any questions um related to today's live stream, unrelated to the live stream, anything digital pathology. And let me know, are you coming to US Cup? Let me know in the uh in the comments who is coming to US Cup. Then after the live stream, I'm going to let you know. Um, well, I already let you know that I'm going to be there, but I'm just going to engage with you so that you don't feel intimidated if we have not met in person yet. I love meeting my trailblazers at

00:24:33
conferences because it's like as if we already met before. This is so cool. Okay, so digital twins. Digital twins in neurooncology as systematic. We have a bunch of systematic reviews of current implementations, technical strategies, and clinical applications. Um, so as for the other ones, I listened to the AI overview as well, which was so good. So, okay, I the more I listen to these AIs, the more I'm like, they're so much better host. AI is such a so much better host of a podcast than me. Uh, but I'm

00:25:10
not giving up, don't worry. But it's like the new way of learning like you can put it in and notebook LM is optimized. So I'm doing this with a notebook LM and it's optimized for for research tasks. So it's not going to be like other general large language models or things that are not optimized to this. It's actually optimized for academic notes. So I give it a prompt uh and it produces a really high quality podcast overview. Um after this one I'm going to give you the code to the to the

00:25:44
AI summaries as well. Okay. So digital twins uh this uh here the purpose was let me make it big. Let's do the highlights. So this was a systematic review and evaluating current digital twin twin implementations and h I disappeared from the screen. Uh and let me put myself back on the screen. Okay. Um digital twin implementation. So um the highlighting the clinical relevance and technical strategies also uh identifying opportunities to advance personalized personalized predictive care in neurooncology. So we are in the

00:26:38
neuroononcology space today. Um and there were also u publications from pubmed scopus web of science databases. They were screened also for English language original research articles. So this is important here. um original research articles. Um and what they did um they u they do the u what what was the time frame uh from inspect inception of this publication through June 2025 focused on digital twin development validation or patient specific computational models in neuroononcology and um they they also evaluated the risk

00:27:31
of bias and applicability uh were assessed using the prediction of risk of bias assessment tool. So uh they like take it ser seriously and that bias is uh a part of these publications and we need to evaluate. So um they had 73 articles that they found and 21 metability criteria. So percentage- wise better than in the previous one. H and digital twins simulated tumor growth, radiation response, immune interactions, and drug transport. So let's let's take a step back and talk about what these

00:28:07
digital twins are. Um it's basically like all your medical data um outside of you in a digital form like lab results, imaging, everything. And um the what's happening is there is modeling happening on this digital data and um also updating the digital data with kind of real life data. So let's say uh you have your digital copy of your medical history there in the cloud let's call it that way. Uh then um you get uh you get labs done then like as soon as your labs are um ready they're

00:28:53
being uploaded to your digital twin uh and you can keep modeling and uh giving feedback to the model in real time because you have real-time data. So it's kind of like a combination. It is challenging. uh so uh we would think that now with AI it's going to be less challenging but actually most models relied on mechanistics or biohysical frameworks um increasing adoption or artificial intelligence driven and hybrid approaches uh 12 studies focused on glyobblastomas or high-grade glyomomas 17 relied primar

00:29:31
primarily on uh MRI data tumor growth and treatment response simulation were most common uh digital twin applications. So uh how is the tumor going to grow and what's going to be the response to uh some treatment? Um actually only six studies provided publicly available code. Um so this closed loop calibration was reported in eight study. Closed loop calibration is the thing I just told you about that um you adapt or adjust the digital twin as soon as you get real life data from the patient and it was only in eight

00:30:10
studies. Um the predictive accuracy and correlation with clinical data were generally high. uh but real-time integration uh multimodel data fusion and also external validation were limited. So um in conclusion digital twins showed promise for advancing personalized neurooncology with demonstrated potential in modeling tumor behavior and optimizing therapies. Um but it was not mostly not AIdriven. It was mainly mechanistic uh AI intelligence artificial intelligence methods. So not like the deep learning

00:30:52
vision transformers but um the old school stuff maybe uh machine learning maybe um just basically advanced math and calculations. Um so despite the strong predictive performance and reproducibility um multimodal integration and external validation remained limited. Uh I am wondering if we have the validation paper here. I'm going to check in a second because there was a paper. Let's see. Let's see what we have next. And if it's not it, do we have it? Just No, we don't have it. But um because it was last week and

00:31:35
last week um we skipped the live stream cuz I was traveling so we didn't do the live stream. But um before we dive into this um comprehensive European coloractyl cancer cohort data set, let me mention you that publication and it's in the um let me show you the code for um for the AI pirate summaries because it's in the summaries validation. This word is like so confusing. Uh and actually the meaning evolved over time in the medical space and different scientists understand different things under the word

00:32:15
validation. And now this publication suggests that we don't use this word because too this is too much of an umbrella word. We have to be uh more descriptive and more specific. Um so now whenever I see the word validation in these abstracts I'm like what do they mean by validation? And when I was writing a publication on image analysis qualification and basically vetting if image analysis algorithm for a tox are good enough. We also had the same discussion like what does it mean to validate? Um and there are different

00:32:51
meanings. So let's just be okay with that. And if you're interested in the a powered paper summaries, get the code and subscribe to this. It's super affordable. I kept it as affordable as the platform let me and so that as many of you can get access to this as possible. And now let's review this last abstract of ours. Uh, sorry for the confusion here. Collecting too much. Okay, this goes away. I'm making myself small. And you let me if you have let me know if you have any questions. And oh, so

00:33:40
here I have a question. Um, I have a problem understanding pathology. Can you guide me? That's what we're going to be doing um in the hisytologology live streams because the basic like to understand pathology um in terms of uh visually understanding it you need to understand hisystologology need to understand what is normal to know what is abnormal. So we're going to have a um live stream series either every other week or once a month uh where we're going to be discussing hisytologology.

00:34:10
So, if you're interested in that, let me know in the comments hisystologology. And uh me and my team are going to be working on setting that up for you. And in the meantime, on the YouTube channel, there are already several hisystologology videos. So, you can go and check uh whichever organ you're interested in. And now, let's talk about the comprehensive European Hello Pam. European colortor cancer cohort data set. I am always so these publications basically describe how they assembled this data set and what is it

00:34:44
for and basically like inform you hey there is a data set you can use but I am always so happy when I see them because it makes they like finalize this tremendous effort of putting together a data set that people can use and actually can compare different AI algorithms on this data set and we have seen in several abstracts that this is always a limit limitation. There are no data sets to run these things on and by these things I mean AI models and actually compare one to the other compared to ground truth compared to

00:35:18
different things. So coloractyl cancer highlighting uh is leading cause of cancer related deaths worldwide. Hey come on work with me. Are you working with me? Okay let's see. Yes. So, the biiobanking and biomolecular ah my highlighter is giving me trouble. Why are you doing this to me? Okay. If not, I'm just going to use the underlining. Um, so the bioanking and no, it's giving me trouble. Apologies for my program. Biobanking and biomolelecular resources European research infrastructure established a

00:36:08
CRC cohort with European coverage uh contributed by 26 bio banks. Can you imagine 26 bio banks, 26 different institutions working together and putting this together and it doesn't end there. Uh they had a lot of hurdles but they made it. So I'm always so excited about this kind of stuff and I see people are interested in histologology. So whoever is interested in hisystologology um give it to me in the comments whether you're uh viewing this um as as live right now. So give me the comment right now so that I can see how

00:36:43
much interest there is for hisystologology. And if you're watching the replay also leave me a comment hisystologology and there's going to be more hisystologology videos if you want them. So show me that you want them. But now back to this cohort right 26 bio banks from 12 countries. Like really this is amazing. I'm like melting when I hear s hear and read publications like that. I'm like they they had the patience, they had the resources. Okay, but let's look what this cohort is and

00:37:13
not just like why I love it. Okay, so this was a retrospective multic-enter study and contained structured and curated clinical data uh supporting research on biomarkers for early detection, prognosis and treatment. So they had this uh phenotypical clinical data model h it was defined and individual level data from drum roll please 10,780 colurectal cancer patients have been collected uh at the BBMRI ERIC in the central data deposition service hosted as part of its services uh so BBMR or I Eric is this um

00:38:01
this infrastructure and the participating bio banks host additional data which can be accessed on request and used to derive additional data. This mechanism has been used to extend the collected data with scans of histopathological slides to support research in AI in digital pathology and with whole genome sequencing data to pilot a use case uh of the upcoming European health data space. This is so cool. And they present the methodology. Can I highlight? I want to highlight. They present the methodology, the

00:38:40
quality assurance mechanisms and the implementation of fair and fair health principles applied to build the cohort. So fair it means findable, accessible um and I will need to Google that or you can put it in the comments as well. Fair principles. You're always amazing when I forget some acronym. Okay. Findable, accessible, interoperable, and reusable. findable, accessible, uh, interoperable and also reusable. So now let me like be in awe of this effort again because um we have 12 countries right let's take

00:39:54
one country US and different hospitals like they have different systems different scanners have different file formats like and here we have this at scale 12 different countries h and and what did they have 26 institutions and They made it and made these data uh findable. What am I here? Findable. Findable, accessible, interpretable and uh reusable. And when I was listening to the AI summary, which I'm putting um putting on the here a summary. So when I was listening to the AI summary of this

00:40:37
one, h they explained like how much how many errors there was in the original data. So like for example things like uh the dates of treatment for these patients that were in this cohort like that the treatment ended before it started or that there was a surgery on the left side but the record was for the right side in one place and they made all these institutions go back find the original records and um and basically uh correct it and Some of the people got um like not annoyed but like didn't have the resources but they

00:41:21
stayed till the end and they actually did it uh to review and review and review and review. Uh but what it shows is basically that the curated data uh where these AI studies run on uh is very curated whereas real life clinical medical data is messy is not always interoperable. So that's why I'm super super super happy when I see publications like these. You know, it's just a description of the effort. But I am so happy about the effort. So now and that was it for today. What I would love

00:41:58
to know also before you get off is are you going to US cup? Because I will be there. I will be there. Uh what's happening? I'm not sharing. Um, I will be there at the booth 312 with Hamamatsu. Uh, oh, we have people in Jordan joining. Uh, and people who want hisystologology. So, uh, if anybody is at uh is going to be at USCAP, leave me a comment under this live stream or send me a um send me a message u on LinkedIn. LinkedIn is the best one because I'm going to be checking LinkedIn for sure. Um, and I'm

00:42:43
going to be there at this booth. Hamatsu, if you want the book, the book is going to be there for at least 50 people, I think. So, Monday, Tuesday. I'm not going to give them all away on Monday. I'm going to keep a certain amount uh for Monday and certain amount for Tuesday. Um and so to remember it's the the booth 312 just write it down in your phone and let's meet there. H I would love to see you. Um and again if I was starting today if I was starting my digital pathology journey

00:43:25
today I would start with digital pathology 101 the book. uh then what I would do I would take the pathology AI makeover course to get like a framework what's happening in AI and then once I have a grasp on that I would go and start listening um start joining the digipath digest abstract reviews and also start listening to the AI powered summaries. This is the closest to reading all the papers that we're discussing just the abstracts uh as it can get without actually reading them. Uh so if you value your time, this is

00:44:09
definitely a resource for you. I am benefiting tremendously from it uh when I review it for you. So highly recommend. Uh gives a fantastic overview what's happening in the field, what the trends are, who's done what, where are the gaps. Um, also if you're like working in these collaborative projects, um, and so definitely super affordable. Uh, you can just, you know, try it and if you don't like it, you just stop using it. But, uh, I highly recommend. Thank you so much for joining me. Thank you so

00:44:43
much for joining me every week. Um, get the book as well. You can get different things in the store and I talk to you in the next episode and I see you at #USCAP2026.