239: Can AI Copilots Keep Up with Pathologists?

Can AI copilots really keep up with pathologists when the cases are new, the workflow is messy, and the benchmark is actually protected from leakage?

In this episode of DigiPath Digest #48, I focus on one paper: DALPHIN: Benchmarking Digital Pathology AI Copilots Against Pathologists on an Open Multicentric Dataset. I chose this paper because I think the field needs more of this kind of work. Less hype. More evaluation. Less “look what AI can do.” More “how do we test it in a way that actually means something?” 

In this session, I look at what makes DALPHIN important for pathologists, lab leaders, and digital pathology trailblazers trying to make sense of pathology AI right now. The paper benchmarks three models against human pathologists: two general-purpose models, Gemini 2.5 Pro and GPT-5, and one pathology-specific model, PathChat+. The dataset includes 1,236 images from 300 cases, covering 130 diagnoses, 14 pathology subspecialties, and cases from six countries. Human performance is benchmarked with 31 pathologists from 10 countries. 

What I like about this paper is that it does not stop at top-line performance. It deals with the benchmarking problem itself. The authors built a sequestered, indirectly accessible ground truth so the evaluation data could not simply be scraped into model training. That matters because without that protection, benchmarking can become an illusion of genius rather than a real test of generalization. 

The results are interesting and more nuanced than a simple win-or-lose story. PathChat+ reached expert-level performance in four of six tasks, Gemini in two of six, and GPT in one of six. That tells us something important already: pathology-specific training matters. But it also does not mean pathology is solved. In organ recognition, expert pathologists still outperformed all the models. In rare cancers, none of the models reached expert-level performance. And in ambiguous cases, the models still struggled with something human pathologists do all the time: expressing uncertainty. 

I also spend time on one of the most practical parts of the paper: model behavior. Gemini tended to overcall. GPT tended to undercall. PathChat was more balanced. That matters in practice. A pathologist using a copilot needs to know the tool’s calibration bias before they can safely interpret what it is telling them. I also talk about anchoring bias in conversational interfaces, where early hallucinations can propagate through later answers if memory is not reset between questions. That is not just a technical curiosity. That is a workflow and safety issue. 

Why should you listen? Because this episode is really about a bigger question: What kind of evidence should pathologists demand before AI copilots enter real workflows? If you want to understand validation, data leakage, rare-case performance, uncertainty, and why these tools should still be treated as co-pilots rather than autopilots, this is a useful paper to know. 

Episode Highlights

01:20 – Why I chose the DALPHIN preprint and why benchmarking matters right now. 

05:38 – What is in the DALPHIN dataset: 300 cases, 130 diagnoses, 14 subspecialties, 6 countries. 

07:57 – Top-line performance: PathChat+ reaches expert-level performance in 4 of 6 tasks. 

09:41 – The benchmarking trap of data leakage and why DALPHIN’s sequestered ground truth matters. 

12:19 – Why real pathology diagnosis is not text-only and why macro + micro context matters. 

15:26 – Tissue recognition, neoplasm detection, ambiguity, and conversational memory: how the testing was structured. 

21:29 – The diagnostic personalities of the models: overcalling, undercalling, and balanced behavior. 

24:36 – Rare cancers: where AI copilots still fall short of expert human performance. 

28:00 – Why binary outputs are not enough when pathology often lives in uncertainty. 

31:37 – Anchoring bias and conversational memory: how early hallucinations can keep propagating. 

37:11 – Why these tools should be treated as co-pilots, not autopilots. 

40:29 – Resources for beginners: Digital Pathology 101 and continued AI literacy. 

Resources mentioned

• DALPHIN preprint: arXiv:2605.03544v1 
• DALPHIN evaluation platform: dalphin.grand-challenge.org 
• PathChat+ pathology-specific AI model discussed in the benchmark. 
• Digital Pathology 101 free eBook by Dr. Aleksandra Zuraw. 
• Educational streams on tissue recognition and computer vision literacy mentioned in the session.

Transcript

00:00:03
Welcome 6 a.m. from Pennsylvania, Fairfield, Pennsylvania. Let me know when you are joining. I'm just going to say hi in the chat. And now, am I even live again? I think we can do it. I think we can do it. But I need to chat with my assistant. It should work. Uh let me refresh the page. Every week something. Yesterday we had a webinar. Let's see. And for most of the the webinar, everything went well and then we moved to the Q&A session and it was less well cuz I was cutting out. Um, and but we made it. Um,

00:01:20
okay. Let's do it again. Okay, we are live now for real. Are we live? Let me know in the in the chat. I had to leave. Now I am back and I see you have joined. so amazing uh for sticking with me and let's welcome you again. Welcome my trailblazers. Today we are going to be talking about a paper. One paper. So I switched the format a little bit uh to focusing on one paper and going a little deeper. We're going to be working with the paper with the abstract and then we have our uh AI created presentation.

00:02:42
uh and I'm going to be perfecting it. So I hope this new format is going to be good for you. Then uh if you prefer to listen to it, you can then later listen to it on the podcast. I'm Dr. Alex Zurf. I'm a veterary pathologist working in the digital pathology space for over a decade and teaching you about digital pathology because I believe patients have the right to fast diagnosis and the gateway to fast diagnosis is digital pathology. So let's dive into our paper today. Um I am having some technical trouble

00:03:23
again. Just stay with me. I'm going to be here till we finish the paper and just give me some reinforcement in the chat that you actually uh see me and hear me well. Um and I don't know if all the messages from all the platforms are going through. We are streaming on LinkedIn, YouTube and Facebook. And today's paper is dolphin dlin dolphin let's call it delphin benchmarking digital pathology AI co-pilots against pathologists on an open multicentric data set. This is our paper today. It is a

00:04:09
preprint from archive. And can I highlight this? And why did they take a pre-print? So, a friend of mine uh and a former colleague T Bujoui posted this funny thing on LinkedIn where he was sitting at a table saying the real science is not published in uh high impact factor journals anymore. It's published on the preprints and prove me wrong. and he was sitting at this table and I'm like hm let me think about it because I was like I don't want to read any preprints because it's not

00:04:53
peer-reviewed but then of course the peer-reviewed takes time and there's a lot of other things involved in that so I thought let's take a pre-print and also I know a few of the authors I know uh Danna Monteuma I think we have Franchesco Chih who I'm waiting for him to join me on the podcast and uh checking here. Oh, there is a Polish sounding name. Mateosh Manfki and Adam Kovalefki H. Anyway, so I know a few of them and it's an interesting paper. Let's see. Are you here? Let me know in

00:05:38
the chat. So abstract tells us that the foundation models with visual question answering capabilities for digital pathology are emerging. Yes. So a foundation model with visual capabilities would be a chat GPD and this is one of the ones that they were testing here. H such unprecedented technology requires independent bench benchmarking to assess its potential in assisting pathologists in routine diagnostics and pathologist and checking okay am I going to be replaced or not and that was another post on LinkedIn by uh the OG of

00:06:27
digital pathology communication Dr. Keith Kaplan where he was playing with Chad GPD for diagnostics and um he stated that no we're not going to be replaced yet. So Delin is the first multicentric open benchmark for pathology AI co-pilots uh comprising 1,236 images from 300 cases and they are spanning 130 rare to common diagnosis. six countries and 14 sub specialties. So um I want to highlight this like in a different uh color anyway but like six countries and 14 subsp specialties. Congratulations.

00:07:14
I hope it shows up in the super high impact factor journal soon because coordinating I'm always impressed when authors like look how many authors are here. a lot. Coordinating this effort is like a medal on its own. Uh when I have to write a paper with like three or five co-authors, I'm like overwhelmed with how much uh logistics there is uh which there is none like you know they re they help you they write their their papers their part but here respect anyway. So uh dolphin designed and uh the dolph

00:07:57
delphin design and data sets are introduced alongside a human performance benchmark of 31 pathologists from 10 countries like another thing 31 pathologists like 31 gather and coordinate 31 amazing congratulations uh 10 countries and they report the results of two general purpose charge GPT Gemini 2.5 pro and one pathology ology specific compiler path chat plus which is fantastic that they included um a pathology specific one for sequential and independent answer generation. So uh they observed that no statistically

00:08:36
significant difference from expert level performance in four of six tasks for path chat two out of six for Gemini and I'm sorry GPT but just one out of six. So GPT didn't do so great. Um, so let's move to my I have a presentation that is cocreated with AI and of course changing what they want to share is a challenge. But I can do it. I can do it. My trailblazers this one. So before you try to use chat GPT or Gemini or even Path Chat for your pathology diagnostic tasks, let's have a

00:09:41
look what happened and what was the result of this ultimate. Can I write ultimate diagnostic stress test benchmarking AI copilots against human pathologists in clinical reality uh and as we said 300 cases amazing so they address a lot of interesting data science points in this paper and I highly recommend you actually read it or listen to it. Um but the there is one core problem when you benchmark them for benchmarking these models that uh it's so-called data leakage and the illusion of genius. What do they mean

00:10:33
by that? that models like Gemini and Chad GPD they ingest a lot of data right the entire public internet uh and standard benchmarks the answer so if it was something uh online where there was questions and answers it already learned it so a delin solves this via a sequestered and unscrapable testing vault what do they mean so um the open pipeline Everything h is here including is in the training data including the test questions and the answers. H and the chat is memorizing it. So basically like cheating because it memorized it.

00:11:20
Uh whereas when we look at the delin way of training it what they did they sequestered the training data. So AI was training on sequestered data and then ground truth data was um made so so-called unscrapable. What does that mean? That the AI model could not scrape it and train with it. So this is why they consider this a true evaluation. Um and also what they realized that uh they wanted to simulate the human diagnostic workflow. What does the human diagnostic workflow look like? Uh when a pathologist

00:12:19
you look diagnosis decided that they are relevant. Right? So here for example this particular place uh it looks from far away that there is like some inflammatory stuff and sure enough these are all inflammatory cells. I need a different color. Inflammatory cells and they their insight was that AI cannot rely on text alone and they already use well the capabilities of the models are there because these are vision and text models but it's not just text or just vision it's det. So it must

00:13:02
synthetize macro architecture with micro details and uh like a doctor does at a microscope. Um so in addition to text right and then it can produce text. So here is who was tested Gemini 2.5 Pro. Do you use Gemini a lot? It comes with a Google subscription. I have access to it, but I have not used it so much. And maybe I should because it definitely outperformed Chad GPT. Well, in so it outperformed uh it matched the non-expert non human non-expert in two out of six task whereas GPT just one. So

00:13:52
um and the pathet is pathology specific. So uh here was who was being tested human pathologist and they were resident was the non-expert and uh there was an expert right and what they found in this paper I don't see anything in the chat I don't know if chat is uh working today I'm going to send you a Hello. Okay, my message is showing. Uh if you're if you can comment, just let me know that you're in there live and that it's working. Okay, so um and they checked how in general these models behaved. So

00:14:43
they um they stated that okay Gemini was a generalist and the diagnostic baseline was open web and uh there were couple of rounds of testing. They tested uh how it handled ambiguity in round three and anchor bias vulner vulnerability and we're going to talk about what that is in round four. There were just four rounds of testing and chat GPT also generalist trained on open web path chat is domain special it is multi-model pathology this is the diagnostic baseline and everything was um tested

00:15:26
and then we had our human pathologist which training parding is biological experiential years of tacid knowledge and was also tested for um ambiguity handling and anchor bias vulnerability in round three and four respectively. H. So, round one, I love it because in round one, they tested basic orientation, which is how it um recognizes tissue, which to me it was like that's easy like you need to know it. Pathologist needs to know it. But I was also surprised how um this that this is a perishable skill

00:16:14
once you specialize and uh I was surprised because um I'm a veterary pathologist. So this is and I work in tox in toxicologic pathology where you always have to evaluate all the tissues of the animals that you're evaluating. So I need to recognize these tissues. I mean I have it labeled but uh you need to recognize the tissue. So it was a given for me that everybody every pathologist in the world recognizes every tissue. H that is not the case and especially uh when you have highly highly specialized pathologists in human

00:16:53
medicine um they did not recognize. So, oh my goodness, I forgot to put my earrings on. And my test was the earrings. I would show the earrings and ask, "What tissue is it?" And not everybody knew. So, guess which tissue is it? And I'm going to tell you at the end of the live stream uh what it is on my earrings that I'm putting on right now. So anyway, they they tested uh ah and also I'm super happy about this um tissue recognition challenge or the necessity to train this skill because we

00:17:28
are also by we uh I mean digital pathology place team. So myself and my team supporting me in the background. Uh we are hosting pathology sorry hisystologology live streams where we teach how to recognize tissues and um we were supposed to do one in the tomorrow but maybe it's going to be next week we'll see but it's on a regular basis and one thing that I want to do is just basically have a bunch of tissues and rec have a session recognize the tissue with me. So if you're interested in

00:18:00
that, let me know in the chat and I can send you the link to that um live stream. But here human experts universally outscored all AIS. Let me do orange. All AIS on basic organ recognition even path chat. I'm sorry Path chat your domain specific but the human people human people humans were better by 24.4%. So uh what was difficult difficult was something that has an amorphous logic uh so-called soft tissue and I'm like what is soft tissue? Everything that is soft is soft tissue kind of like but uh in

00:18:44
pathology soft tissue is being referred to when you talk about muscle connective tissue h tendons and like the bone supporting structures that's why you have the bone and soft tissue pathology when I learned this uh term I was like why bo bone is hard and soft tissue is uh soft why do you put them all together because it's the muscular skeletal also including fat that tissue. So um yeah soft tissue was not the strength of AI skin was okay because it has high structural logic when you look you have

00:19:18
the keratin on top the epidermis connective tissue here in the dermis and then subcutaneous tissue but soft tissue is like a bunch of soft tissue so the chats were giving an error path chat completely omit the tissue prediction for 41.5% of soft tissue cases saying no residual normal tissue is recognizable. So soft tissue and um also the performance varied widely based on the structural predictability of the organ and representation in uh the training. and and there was different representation

00:20:07
of different tissues. Skin was good. Uh breast was relatively good. So the red ones were were good. The gray ones were um not that fantastic. Bone and soft tissue gray. Heattopathology not that great either. Okay. >> Want to tell me more about pathology issues you're having? >> My glasses are talking to me. It happened to me once at a conference where I was like touching I touched my um my glasses. So these are the meta glasses. Let me show you. Maybe I should take a picture of myself.

00:20:46
Yes, it's asking me if I want to know more about pathology uh questions that I have. No, I'm giving a live stream meta glasses. Um okay. So then so so now we are at round two, right? Neoplasm detection and AI personalities. So at the beginning we had this like oh chupity is generalist Gemini is generalist and um my AI uh co-analyzer of papers notebook LM uh classified them as uh one is an anxious student and the other one is the relaxed student. Anxious student overalls everything. So Gemini was

00:21:29
highly sensitive but low had low specificity jumps at shadows catches real tumor but frequently mclassifies B9 inflammation as neoplastic. So that's what Gemini is doing. Whereas Chad GPD is more relaxed. It's it's um it's not calling everything. It's highly conservative, ignores shadows but actively misses real existing neoplans. So this is high specificity but low recall and the balanced resident is our expert gold path chart and human residence we're here in the middle uh

00:22:04
where so so they were what we want to strive for. We don't overall we don't underall we are just right the expert gold. So uh and what the authors of the paper say that without knowing your model's inherent calibration bias a doctor cannot safely contextualize its advice. So the next time you try Gemini or Chip PT uh try to remember that okay Gemini was an overall whereas Chaj was an underller. Next time uh when I present I need to bring all these LinkedIn posts uh that inspire me like Keith Kaplan's and Too's

00:22:52
posts um about how I run the live stream. So behavioral matrix precision versus recall. So what we want we want high precision and high recall where so PHA GPT uh has kind of high precision uh but the recall is very low whereas Gemini has high recall but the precision is pretty low. Pathjet was the closest to our human resident. uh it approaches expert recall but precision and specificity remain comparable only to non expert human residents. Um so interesting path chat is already at the level of a non-human expert. Obviously

00:23:42
we want higher level but still that's um interesting for this particular uh chart. We're going to move to a different screen cuz my AI didn't do a great job. Well, it it did a great job showing like what we want to show, but the graph does not represent reality. So, let me share the paper again. And if you have any questions, any uh comments, anything, is it worth reading? What do you think? Why is it not showing my screen? Share screen. No. Share window. Share. Why are you doing this to me? Okay, now

00:24:36
it's not sharing. Let's see. Share something else. Entire screen. Let's see if we can do it. Okay, now it's doing it. We need to go to figure three. This is our figure three. Okay. And we have here common cancers and rare cancers. Uh, and Gemini is blue, chat GPT is this orange, and past chat is red. Experts are in dark gray. Non-experts are in uh less dark gray, clear gray. Uh, so this is expert reference and non-expert reference and common cancers. So expert past chat was like really nice for

00:25:40
common cancers but for rare cancers it missed the mark right so these are the experts but for uh non-experts it actually was on par with the non-expert so what do we mean by rare cancers it is less than six per per 100,000 cases um And what? So the promise I'm going to go back to to the presentation. I'm just going to ignore our charts here. Mhm. So the because the promise that uh everybody is talking about oh AI is marked marketed as this ultimate tool for saving solving obscure cases right

00:26:38
so uh if you don't have a specialist you just feed it into AI and it's better than not having anything look at it well not the case not even for path chat uh none of the models sorry none of the models not even the dom main specialist path chart reach expert human levels on rare cancers. H and the insight here is that neuronet networks are pattern recognition engines. So if a rare entity lacks a massive statistical cluster of training data hello there these are rare cancers so by definition they're rare

00:27:18
enough not to have a massive statistical cluster of training data. AI cannot reliably generalize its reasoning. So the the yeah there are limits to the knowledge and it's not an oracle. It's not an oracle. And also this was super cool. H there was this concept of the illusion of certainty. H the models manufactured certainty. Uh what does that mean? So when you look at cases and if you have ever if you have written a pathology report you know for sure if you have read a pathology report you

00:28:00
will see that as well a lot of we call it hedging language and the hedging language is not for legal purposes to hedge yourself from liability. It's expression of uncertainty because you're looking at um biological structures that may not have boundaries may not be like super certain when you look at it because uh it is the property of its biology. So especially in cancer diagnostics what's happening with the tissue the tissue becomes abnormal from normal. So the closer it is to normal, the easier it is to

00:28:45
recognize. The farther it is from normal, the more funky it looks because the cells change their criteria of malignancy of that the nuclei uh change. Then you have the differentiation of tissue. it basically like changes form and if it changes form a lot then even a trained human observer may not be able to 100% say oh this is what it is um and that's what we do we express the uncertainty uh with hedging language with with with definitions right whereas AI is trained to um do binary calls so path chart predicted

00:29:31
exactly zero cases as incit or uncertain. So there are different um like codes for the level of uncertainty. Uh and these two categories institute and uncertain are are the ones that uh expressing that express uncertainty. It aggressively forced binary nuance of human biology into um sorry the every nuance of human biology into binary B9 or malignant bucket. This is this is how they are trained. I um love that the paper is addressing this because um that's also what you evaluate the performance on. Oh, did it say B9 or

00:30:08
malignant? Because maybe if it says B9 then at some point pathologist doesn't have to look at it or it the case becomes dep prioritized. Uh, and there was no uncertain category. The human gold standard, what I already alluded to, human experts comfortably navigate ambiguity utilizing the ICD3 uncertain codes when biology defies strict categorization. AI is mathematically rewarded for definitive answers, lacking the self-awareness to express doubt. Sometimes I think that Claude is expressing doubt to me. They didn't do

00:30:51
Claude. I don't know if Claude does that well with images. Maybe that was the reason. Sometimes when I ask Claude, it's like I don't know. I was not trained. I cannot retrieve. Go and retrieve yourself. And I'm like, that's why I'm using AI so that I don't have to crawl the internet, but sometimes I have to still. But I'm digressing. Let's do round four. This was a very I don't know if it was innovative. It was new to me. H to even ask yourself this question. Does conversational

00:31:37
memory build a richer clinical picture or trap the AI in a cascading ecochamber of its own mistakes? And um what um what they found in the paper is that we do have memory trap mistakes. What does that mean? Uh if you does it say here yes um that if you have a query in a chat and then like it's it's this uh interface and they call it the interface trap and anchoring bias. So anchoring is the key word here anchoring bias. So if a um chat invents something uh at the beginning of the conversation then it refers to itself as

00:32:32
if this hallucination was true. So what they um stated is that uh sequ sequential chat user interface wasn't that great. independent clean user interface where the chat wipes out its memory uh and doesn't refer to itself from the beginning of the chat uh was more beneficial because then independent evaluation prevents failure. So treating each query independently, wiping memory between questions, prevented late stage failures, highlighting the hidden danger of intuitive chatbot interface in

00:33:16
medicine, right? Intuitive is like you just chat with it and it remembers uh what you what what you told the chat and what the chat told you. And that's the problem. If the chat remembers its hall house then it's going to be referring to it and um it's going to be cascading the error um and there was an arctinic kerattosis case where um a slide image was presented this is this is time then there was a hallucinated misdiagnosis ai hallucinated something and then uh a question was asked about the lesion

00:33:55
factor ual question regarding the lesion regarding the real lesion of actctinic keratossis and the I love this word the chat rejected reality chat rejected reality I love it how often do you reject reality I sometimes reject reality when I I don't know when uh you mostly when uh mostly with my kids I like want them to be compliant and different and the reality is that they are not always compliant and they are their own people. So like me wishing them to be different um is basically rejection of reality.

00:34:37
It's what AI does when it invents. So that's like me. I invented this ideal picture of my kids and now I reject reality that they don't match this ideal picture. Um yeah. So let's not reject reality. it makes things more difficult. So, uh when later presented with multiplechoice option containing the actual ground truth, both models uh refuse to answer deeming the factual question inapplicable, right? My kid defined it's inapplicable. How can my kid be defiant at school? Well, that's rejection of reality. Okay.

00:35:21
So, uh yeah, both Gemini and GPT hallucinated a wildly incorrect diagnosis early in the chat thread. Uh when later presented with mult multiplechoice options containing the actual ground truth, both models refused to answer, deeming the factual question inapplicable. So, uh the takeaway here is that AI anchored so hard to its initial host nation that it mathematically rejected contradictory reality. that is you know you know what let's put it in red here because here it was a test right in this case it

00:36:01
was a test and we tricked it or or like tricked it we tested it we wanted to check this feature if it anchors and it does and we knew how we were testing and we knew what we were testing this is just a testimony of because we know if if you have ever used any of them, the chat GPT being the most famous because it was the first one, you know that it can present you information with high confidence. It can be if it's like grossly and accurate in your area of expertise, you're going to recognize

00:36:36
it, right? Uh so it's always uh to be treated as a helper. But if it's outside of your domain and it's like semi- true, you may not recognize it. Um, so that's why I like the ones that show you the references and you can see, okay, where is this reference? Is it from somebody's private Facebook page or is it from PubMed? H I will uh trust the Pubmet reference more. So anyway, what what what what did they want to tell about this? just to be cautious because it's going to invent something and then

00:37:09
defend its answer. So, so the verdict for co there are these are co-pilots and Microsoft called calls its uh guy it's model co-pilot not autopilots. I love this. Not autopilots, my friends. Uh, and the next great leap in medical AI is not higher baseline accuracy because path chat did pretty well, right? It did as well as a non-expert. The next leap is the algorithmic capability to recognize ambiguity and explicitly hand control back to human expert. So let's say we have path chart that is at the level of non-expert

00:38:04
and the gap we're not going to like force train it to be better as an expert but we figure out a way for it to say hey this is in this uncertainty this gap domain between expert knowledge and me how about an expert looks at it uh so That's that's our message to take home. Uh co-pilots, not autopilot. And uh important thing is to know your model's inherent calibration over color versus underller. Uh and I bet there's more nuances to that as we start testing them as they become more prevalent. There's

00:38:47
only one available path chat uh that is domain specific. Maybe there's going to be more. Maybe maybe as I speak there are already more that are not publicly available. H then the other thing is beware of conversational interfaces that propagate anchoring bias and oh my goodness my trailblazers there are so many sources of bias that when I read about them I want to like switch off my computer but I was talking to a colleague of mine Dr. Candace True. She's a veterary pathologist highly involved in AI uh literacy uh

00:39:23
propagation and basically promoting AI as a tool. She says a tool is a tool is tool. It is a tool, right? So like don't throw away the tool. Just learn how to use the tool. H and if you're not comfortable using the tool for some tasks, just use something else for these tasks, but figure out what you're comfortable um with using the tool for. H. And the third takeaway, true medical intelligence isn't just knowing the answer. It's knowing the limits of your own knowledge. Oh my goodness, that's

00:39:56
also a tricky one. You don't always know the limits of your own knowledge. Um, but I don't remember the four stages of um learning stuff. First is like, you know, you don't know. Then you know a little bit, then you think you know everything, and then you realize I know something, but there's a lot I don't know. So that's like the fourth level. You have to reach the fourth level yourself. And what we are trying is to teach AI to basically recognize the gap that it has. And I bet there are computational

00:40:29
methods to do that. I hope I can ask about these methods. Um my future podcast guest Francesco. Um if you have any questions, let me know in the chat. If you're just starting your digital pathology journey, I have something for you. Uh this book, Digital Pathology 101. All you need to know to start and continue your digital pathology journey. Um this edition is the first edition, but you can get it for free. And uh let me show you where on my website. When you get this, you're going to be on

00:41:06
my mailing list, which means you will get the second edition uh as soon as it's out. And the only thing that uh blocks me from giving it to you is images. I need to create new images. Um so there are like illustrations. Uh I need to use show you some better illustration. My favorite illustration are the giraffes. Um oh here the giraffes explaining computer vision concepts to non-computer scientists which is something super important uh coming forward the webinar that uh we held with rash diagnostics

00:41:51
USA yesterday was talking about the changing role of a pathologist where they will need to have some computer vision literacy. Let me just show you where to find the book. And when you go there, it's I'm going to put it in the chat as well. I hope you're getting my messages. I'm not getting your messages today, unfortunately. Going to cry. Nobody commented. Maybe sometimes um I'm going to stop myself from crying uh just yet because what happens sometimes is the streaming software doesn't bring in the comments

00:42:40
and then I look at the recording and there was like this that from this country and a lot of questions. So even if you're watching the recording uh go ahead uh comment say hi it's always I always it's very heartwarming uh when I see live comments from you. Thank you so much for joining me. go grab the book uh so that you can get the updated second edition automatically.