188: AI in Pathology: Biomarkers, Multimodal Data & the Patient

Is AI in pathology actually improving diagnosis — or just adding complexity?

In DigiPath Digest #37, we reviewed four recent publications covering AI-based biomarker quantification in glioblastoma, real-world digital workflow integration in prostate cancer, multimodal AI combining histopathology and genomics, and patient perspectives on AI in cancer diagnostics.

This episode connects technical performance with something equally important: trust.

Episode Highlights

[00:02] Community & updates
Digital Pathology 101 free PDF, upcoming patient-focused book, and global attendance.

[04:07] AI-based image analysis in glioblastoma
AI showed strong consistency with pathologists when quantifying Ki-67, P53, and PHH3.
Significant biological correlations (Ki-67 ↔ PHH3, PHH3 ↔ P53) were detected by AI — not by manual assessment.
Takeaway: computational quantification improves precision.

[09:28] Real-world digital workflow + AI in prostate cancer (France)
AI-pathologist concordance:
• 93.2% (high probability cancer detection)
• 99.0% (low probability slides)
Gleason concordance: 76.6%
10% failure rate due to pre-analytical artifacts.
Takeaway: infrastructure and sample quality still matter.

[15:58] Multimodal AI (MARBIX framework)
Combines whole slide images + immunogenomic data in a shared latent space using binary “monograms.”
Performance in lung cancer: 85–89% vs 69–76% unimodal models.
Takeaway: integrated data improves case retrieval and similarity reasoning.

[22:13] AI-powered paper summary subscription introduced
Structured summaries for busy professionals who want more than abstracts.

[26:17] Patient roundtable on AI in pathology (Belgium)
Patients expect:
• Better accuracy
• Faster turnaround
• Stronger collaboration

Trust is high when:
 • Algorithms use diverse datasets
 • Pathologists retain final responsibility

Clinical validity mattered more than full algorithm transparency.
 Privacy concerns focused more on insurer misuse than cloud transfer.

Key Takeaways

• AI improves biomarker precision in glioblastoma.
• Digital pathology implementation works — but pre-analytics can limit AI performance.
• Multimodal AI represents the next meaningful step in precision diagnostics.
• Patients are not afraid of AI — they want validation, oversight, and governance.
• Human–AI collaboration remains central.

If you’re working in digital pathology, computational pathology, or precision oncology, this episode connects evidence, implementation, and patient perspective.

Transcript

00:00:02 - 00:01:17
Aleks: Good morning. Welcome, digital pathology trailblazers. Uh, good morning. 6 a.m. from Pennsylvania with coffee waiting for you with papers. Um, whenever you join, let me know where you're tuning in from in the chat. I'm going to say, "Hi, Trailblazers in the chat." and whenever you're here. Okay. And I see you amazing. Let me know where you're tuning in from, what time it is for you, and I assume you can hear me well. If not, let me know if anything, you know, fonts too small, not


00:00:42 - 00:02:06
hearing uh anything, let me know immediately in the chat. And before um we start couple of updates. So a book the book is a highlight on so like a hit on social media book is fantastic I mean what do I mean by book there is the digital pathology 101 and I recently posted a few more uh posts on social media that uh the PDF is free and this um paper version is available on Amazon and a lot of people starting their digital pathology journey are downloading the book. So, I'm placing the QR code as well if you're


00:01:27 - 00:02:44
interested. If you don't have the book yet, uh, grab the free PDF. Um, and I am working on an updated version. Um, slowly slowly progressing. Once it's updated, whoever has this current version is going to get the new version automatically. You don't have to do anything else. So, um, there is this book and there is another book I'm starting to work on. Well, I already started. Um, it's going to be for patients who just had a biopsy and would like to get access to their hisystopathology images, digital


00:02:04 - 00:03:21
hystopathology images. So, um, if this is of interest uh to you, let me know in the chat. Uh, put a comment. Let's say book two. Let's say book two. And book two is going to be um called B9 or malignant. How to get access to your hystopathology images to your biopsy images. Uh there is a this title is a work in progress. So you're going to see probably some polls on social media what we should call it. I'm going to uh bring you in on the process of writing this book. Um, and the announcement is


00:02:43 - 00:03:55
probably going to go out to the list and on social media soon to pre-order. And okay, we have our guests from Silver Springs, Maryland. Hi, Sean. And also from New Zealand. This is amazing. What time is it even in New Zealand? It's like tomorrow or something. Saturday. Is it already Saturday? H. Oh, don't go to bed. Stay till the end. We're not gonna We only have four abstracts today. So, let's start before people uh in New Zealand go to bed. I think you're going to be proud of me


00:03:21 - 00:04:49
today because my pen is working. It should touch the correct screen. Yes, it is. I think you're should be very proud of me. Um so, let's do it, my friend. Oh, no. Why? that my mouse is not touching the right thing. Okay, focus girl. Let's share the screen big and okay and we have some takers for book two. Fantastic. Thank you so much. Uh hello everyone who's joining. Um if you're joining now, let me know in the chat where you're tuning in from. And we are going to start with paper number one


00:04:07 - 00:05:51
which is artificial intelligence-based digital image analysis for assessment of K67 P-53 PH3 expression in glyobas blastoma multiformat and um I do have some problem with this and I did prepare for you what these markers are for. So um the Kai 67 is proliferation index, P53 tumor suppressor status and PH3 mitoic marker. K67 is for proliferation uh during the whole cell cycle and uh PH3 is specific for like whichever phase of mitosis. I think it was G2 but um let's look at this publication from Turkey.


00:04:59 - 00:06:23
And the objective here was I don't like this blue color. Maybe orange. Let's do orange. Um to compare uh AI based image analysis for K67, PH3 and P-53 uh IHC biomarkers in glyobblastoma multiformat with conventional evaluation performed by experienced pathologist. like you know not too complicated study design and we've seen that I don't know last year a lot of these like image analysis algorithms being evaluated uh you know unimodel classical image analysis but people going digital and taking um this


00:05:42 - 00:07:03
first step of of using some uh of these markers right so that's what they do here um specifically what they did was a cross-sectional analytical study it's uh it's a study designed that um that um evaluates this variable and the outcomes at the same time and they had 20 cases of glyobblastoma and uh they they used our staining sky 67 PH3 and uh p-53 and they digitized the slides and two expert pathologist and AI based image analysis was used and u they use different areas one square millimeter


00:06:21 - 00:07:54
and 7 square mm of tumors. Why can't I like can I move it with this? Yes, I can. Okay. Um and the results that the the interesting thing was um Okay. Um inter intra method comparison showed strong consistency both AI and pathologist assessment at 1 millm millimeters and seven. So um strong consistency for both AI and pathologists. But uh in AI analysis there are significant correlations uh between Kai 67 and PH H3 which we would expect right Kai 67 is proliferation for the whole cycle and PHH3 is uh mitoic figures. So there


00:07:13 - 00:08:32
should be a correlation right the more cycle the more cells are proliferating the more mitoic figures we will have and also where's the other one uh we had between uh PHH3 and P uh 53. So p-53 is the tumor suppressor. Um and if there's no tumor suppressor there's going to be more proliferation. So like less tumor suppressor, more proliferation, more mitoic figures, but no correlation was observed between these markers at either field in pathologist assessment, which tells me let's do AI because there


00:07:52 - 00:09:18
should be a correlation and there is some kind of loss of precision maybe uh when we are doing it manually and you already know my opinion of on counting cells manually. Uh it's not a favorable one. I I don't support that. I think computers should count uh dots in images. Um so that's what they did. Let me check the chat. Okay. From UK. Okay, we have guests from UK. Fantastic. Um, okay. If you're joining right now, let me know in the chat where you're tuning in from. I'm going to get some


00:08:43 - 00:10:09
coffee. And did you see the the new trailblazer mug? It's not openly available. It's not in the store. There are other things in the store like my earrings if somebody is interested to getting them for their next conference. um or getting them for somebody. I'm going to give you a QR code to the store for a second. Um and we are going to move to our next. So I was um about to do this one assessing the influence of two deep learning assistance modes on pathologist and cancer identification. But um


00:09:28 - 00:10:46
there is there was no abstract available here. So I need to skip this one. So we're going to move to the next one. The digital lab practical digital workflow and integration of AI for routine and pathology through the example of prostate cancer. So I like these um papers because they show real life struggles of people who who did it who went digital and now I think that digital pathology is not that niche anymore even though like in the medical like in general in medicine it's probably niche but in pathology like


00:10:12 - 00:11:20
everybody has heard uh about digital pathology people know what scanners are right so now um the shift that I'm seeing at conferences and in these papers it the people actually talk about their failures which is fantastic because then you can read this paper and avoid the failure like is that not fantastic I think it's beautiful because there's no uh sorry fiddling with these QR codes there's no uh such thing as journal of negative results uh and I struggled with this during my PhD when I


00:10:46 - 00:12:00
had like negative results and I was required to publish it nobody wanted to take my publication. Um in the end I did publish it somewhere. So there is a record of my PhD thesis. But uh here we are incorporating our failures into the story of our success right because there is no success without failures. Um you just have to get up and keep going until you succeed. And that's what they did uh here in France. This is a French publication. Um and let's let's talk about their successes and about their struggles. Um


00:11:23 - 00:12:40
so this is Randy University Hospital. Digital top authority has been re routinely deployed since 2020. Like kudos for that here a heart for them since 2020. That was co you are amazing. H and then AI solution for prostate adenocarcinoma um by ibex. Also shout out to ibeck um that they provided it. That was since July 2023. So now we are in 2026. They've been using it for almost around not yet three years, right? In July it's going to be three years. Uh and they talk about their experience. Um so


00:12:01 - 00:13:21
uh they assess the impact of digitization on the various professions within the department and prospective use of AI for routine diagnosis. So obviously they did validation and calculated their concordances and everything. H so concordance between AI and pathologist was 93.2 uh for detection of high probability cancer and 99 uh for low probability slides and then there is some intermediate probability and this their cancer was confirmed in 4.7 cases and Gleon grading concordance rate was 76.6 six and uh the integration of AI has not


00:12:41 - 00:14:07
changed the use of imunostochemistry. So that's interesting. Um they still use the same amount of imunoistic chemistry and then there was a 10% failure rate related to pre-analytical artifacts. Um so I think now in the era of fancy algorithms and foundation models and as like pushing the frontier of computational and digital pathology uh sometimes we forget about pre-analytical artifacts unless you are a researcher that works on the TCGA data set and you have a data set full of artifacts and


00:13:24 - 00:14:39
also frozen section images. uh so there you probably remember uh but uh here like 10% failure rate due to pre-analytical artifacts that can be avoided right so um they say that effectiveness of digital pathology and use of AI models specifically I would say that's my words are closely dependent on pre-analytical quality and its organizational integration and actually recently I gave a webinar uh about the importance of pre-analytics like so I will link it in the recording here uh so you can check out how


00:14:02 - 00:15:27
important they are. Let me know if you have any questions, any comments. We have guests from Nigeria and some compliments for my earrings. Thank you so much. You can have them too. I have different ones as well by the way. Let me make myself a little bigger. Uh I have these. This is Alian blue on column and I also have cartilage the other way around cartilage. My favorite are the multi-ucleated giant cells. I don't know if they're still in the store. You can check. Uh I'm putting the code. And these are the same uh that


00:14:44 - 00:15:54
they have on the mug. I like the multiated giant. So I don't know why I like them so much. I think I don't know the these are macrofasages that come together and they're usually um there is like something big happening like a foreign body or specific type of cancer or or mostly I know these from foreign body and they come together and try to engulf this foreign body and they're so inefficient um and just stay there in the tissue but I like the shape of them. Let me show you how they look.


00:15:19 - 00:16:34
This is a multi-ucleated giant cell. Okay, back to the core of this live stream. Trailblazers, let's move on. If you have any questions, let me know in the chat after this abstract. I'm going to review the chat again. Oh, now we have a now we have a heavy one. And I I was listen I was reading to this one. It's multimodel learning for scalable representation of highdimensional medical data. So, I'm like, what does this even mean? I'm going to explain you what it means. And then I like go through this abstract and


00:15:58 - 00:17:17
then I see who is one of the authors. Uh, Hamit Tizouch. And the first, he's the last author. And the first author is Ari Alsafin. I hope I pronounced the name correctly. And this is Kim Labio Clinic. And also, yeah, Mayio Clinic. Um so Ahmed's research is very computationally heavy and very uh clinically useful. That is like a focus of his is it has to be clinically useful and it sure is computationally heavy and well not computationally heavy let's say computer deep computer science let let me call it


00:16:38 - 00:17:57
that way deep computer science for pathology to be useful in pathology. So let's look at what they did. M so integrating AI with healthcare data is rapidly transforming medical diagnostics and driving progress toward precision medicine. We agree with that. We talk about it every week. And then um but multimodality effectively leveraging multimodel data is not that easy particularly digital pathology images because they're big images. So obviously from the uh computation standpoint they're


00:17:17 - 00:18:33
difficult and genomic sequencing remains a significant challenge also because these are two different categories of data image data and um it could be I don't know just like text data or or tabular data um genetic sequencing so so these are like two different types of data so they have this intrinsic heterogenity the these modalities have the intrinsic heterogeneity And um there is a need for scalable and interpretable frameworks. Um and existing models are unimodel data, right? Uh it's usually what we


00:17:55 - 00:19:35
just we just saw uh something on pathology images only, something on uh genomics only. There are combinations but it is not the norm yet. Um so they introduced Marb Marbix Marbix Marbix also um known as multimodel association and retrieval with a binary latent indexed matrix. Uh yes very creative uh and sounds like computational pathology very much. uh so but this is a self-supervised frame framework that learns to embed whole slide images and immunogenomic profiles into compact scalable binary codes termed monogram.


00:18:44 - 00:19:57
So they um have this model do embeddings on whole slide images and then do embeddings on genomic data and they combine them together uh to basically group together similar cases but similar cases not only in terms of okay what was the histopathological morphology or what was the genomics but the combination we have the multimodality h and uh here they optimize the triplet contrastive objective across modalities. So uh what this is uh you have like a um base case then you have a similar case and you


00:19:22 - 00:20:47
have a dissimilar case. So you have three three cases. That's the US way of showing three. This is the Polish way of showing three. Um so you have three cases right? The the case that is that you look for similar cases to a similar case and a dissimilar case. And this is this triplet contrastive objective um for both modalities and uh they captured high resolution patient similarity in a unified latent space. Uh latent space is uh like what a model learns not what we would see. So it learns something in both whole or slide


00:20:04 - 00:21:18
image genomics and then they create a space this unified space which is a space together. Uh and then we have this code telling us okay this is a similar case to another case and you can search by similarity and uh this enables efficient retrieval of clinically uh relevant cases and facilitating casebased reasoning. So you basically can uh you don't have to do oh show me similar host images and then go again and show me whole slide show me similar genomic you can have this one vector grouping them together h and then


00:20:41 - 00:22:16
in l cancer uh this marbix achieves 85 to 89% across all evaluation metrics outperforming hystopathology and imunogenomics. So we have the um a better thing a better thing than just each one of these right. So um the hisystopathology is 69 to 71 and imunogenomics 73 to 76 and the marbix is 85 to 89 and then uh they also did uh in in kidney there is real value monograms and um the binary monograms. So real value is like continuous values not just 01's and the binary is a simplified the version of these monograms these codes


00:21:28 - 00:22:57
that um is just 01 and the F1 for the real value is 8083 accuracy 87 to 90 and for binary is slightly lower 78 to 82 F1 um binary is easier to retrieve less less data heavy uh and the real value is a little more um slow to retrieve. Um I don't know the like magnitude of change but basically um they're good. They're over 80% and we have multimodality that uh could be clinically useful. So what I want to do and I want your opinion on that I kind of started but you know it depends on you if it's going


00:22:13 - 00:23:31
to be used uh or not because uh we only have so much time uh in the week everybody and if we wanted to dive deep into each of these papers uh or even like a little deeper uh we don't have enough time so what I created is um a a specific podcast like a special podcast feed with AI powered summaries and you may have seen uh my seven part of pathology audio bundle last week that I was um I created as well and was offering this as a Valentine's offer. Um but what I have now let's see if I can


00:22:52 - 00:24:14
share it without glitches. Uh is yes on the podcast. Let me make myself smaller on the podcast. Um, we have these different episodes, right? And there is one that I that has this um padlock icon. It's subscriber episode. So that means that there's going to be a subscription, but uh it's an AI powered um summary of the papers. And I currently have one and I'm going to put um the ones that you want from this live stream. So, leave me in the comments which uh which one you want. You can


00:23:36 - 00:25:07
like go by number or by title. Um or if you want all of them, just put all of them. Um and you can try it out. And let me see what I have this at right now in terms of how much the subscription is. Um where is this one? Okay. Yeah. Um so it is $7 a month. um two and obviously if you want to pay more you can always uh but then there's going to be what I'm planning to do every time we do a live stream after the live stream uh what we're going to do is uh put all the papers that we're discussing in the live


00:24:23 - 00:25:43
stream so like pre-selected and already pre uh discussed in abstract in this subscription feed. So whoever thinks this is valuable for them um you can subscribe as well. Let me give you a QR code and uh they will have a different icon on the podcast. Uh they they will have I I'll make them distinct uh and they will only be available to subscribers. Uh part of this is I don't want to uh populate my uh podcast feed with AI generated content that is relevant to very specific audience basically to you


00:25:05 - 00:26:05
to the trailblazers who are joining the Digipat Digest. Um so if this is of interest uh you can check it out. There's already one uh one publication, one uh podcast there and I'm going to include everything that we have in today's live stream. Um and you let me know and I I will see if you like it, you like it. If you don't like it, you don't like it. Um I thought it's going to be useful because I want to listen to it. So basically all the things that I create, I create for myself first and


00:25:35 - 00:26:42
then I think, hey, why uh why not for you as well? So that's available. I'm going to put it back on the screen in the end. And now let's move to our papers. We still have a few a few or one. One or two. Let's see. So I definitely want to dive deeper into this one. So, I'm going to be creating an abstract also like to explain in a non-computer science language what they did and why it's useful because I already know it's useful. And oh, we have a cool one. I should have said at


00:26:17 - 00:27:46
the beginning that we have a super cool one at the end which is about patience patients patient um patient impression here. patient impression of uh the digitization in pathology and AI which is so important but so under discussed and it's going to be discussed more in the podcast I'm going to have patient advocates and uh we already had a few patient centric pathologists we had Dr. Lijah Joseph we have we had Dr. Marilyn Buouie. Um, so if you want to go to these podcast episodes, um, I will link


00:27:04 - 00:28:29
them in the show notes as relevant episodes and, um, for the recording, we can, um, also put them in the cards. Um, but the connection between patient and pathologist, uh, I don't think it's an overstatement, is almost non-existent. uh pathologist is kind of invisible to the patient and it it's not about the pathologist as as the person that needs to be visible as a profession which also is an important thing but the insights and the information uh that the pathologist brings to the treatment continuum. This


00:27:46 - 00:29:02
is basically the diagnosis, the image uh the report uh is what the treatment is based on. Right? So now we're using digital tools, computational uh pathology, AI, what do patients think about it? And this is a publication um this was a round table held in Belgium with patient advocates. Um it was a digital pathology and AI um let me use my so digital pathology and AI promise faster more accurate cancer diagnostic yet patient views remained and remain undocumented. And so they documented and I'm super


00:28:24 - 00:29:47
happy because this is something I want to like dive deep into as well. So they had a 2-hour moderated roundt with six Flemish cancer patient advocates. It was recorded, transcribed and analyzed using reflexive uh thematic analysis and patients anticipated improved accuracy, shorter turnaround times and stronger interlaboratory collaboration. A good thing is trust in AI was high when that's important. Algorithms were trained on diverse data sets and pathologists retained final responsibility. So here uh two aspects


00:29:05 - 00:30:24
here diverse data sets not just you know one population, one hospital, one uh group of patients and pathologists in the loop. So they trust the pathologist and then by proxy if the pathologist trusts AI and that should be the case uh because it should be validated uh correctly right and then the pathologist should have trust and the final say then patients are okay with this and there are a couple of other uh very interesting insights. So a clinical validity outweighed full algorithmic transparency.


00:29:44 - 00:31:03
So meaning they care a lot less about knowing what's inside the algorithm if it's clinically uh valid and then uh but they do encourage explanability research u then also explicit mention of AI in report what in reports was considered unnecessary if quality assurance was demon demonstraable demonstrable um so that's interesting as well I think there is a requirement If there's anybody from Europe, let me switch to chat. Let me know um what does GDPR um say about it or are European AI


00:30:24 - 00:31:37
regulations. I think whenever AI is used, it needs to be mentioned. H but patients said you don't have to explicitly mention it if there is enough quality assurance, right? They this is good. Uh this is like very positive sentiment uh to me because it it signals there's no fear. It is rather considered as the next another diagnostic tool and if you validate it if you have quality assurance they were fine doctors using it. We we still trust the doctor. If this is the tool they want to use let


00:31:00 - 00:32:25
them use it. Right? That's patient advocates in Ber in Belgium. Privacy worries focus on potential insurer misuse rather than pseudonymized cloud transfer. uh so so what I'm seeing is here like okay what are the things relevant for the patient uh like for the particular patient as a person more worrying is uh insure misuse of uh the data then pseudonmize clouds cloud cloud transfer right so so very like tangible very personal very okay what does this mean for me as a patient when AI is used


00:31:43 - 00:32:43
or not used and the dangers of AI which which is kind of like a a little bit different voice from what you hear in the industry and the representative requested future tools that translate technical reports into lay language and we have some of these tools already there are websites like that I'll try to find them for you and um or if you know which one uh these are put it in the comment even if you're watching the recording because that's going to be useful but I'm going to find


00:32:13 - 00:33:21
um that as well. But there are some of these maybe European people don't know about them yet. So I need to talk about them more. Uh and then suggested questions to support shirt decision making. So I want to see what questions they uh supported. Um but like another heart here because patients representatives were generally supportive of the introduction of AI and pathology provided the algorithms are clinically validated trained on representative data sets and deployed under clear professional


00:32:47 - 00:34:10
oversight. We need to like make a quote out of this. Um because it signals a positive sentiment towards uh AI where you still have this like fear. you still have uh there are still cases where like when AI is mentioned um people like are so afraid that they withdraw from clinical trials which is so sad uh because there are so difficult to get into one right uh but here they express positive sentiments and there is uh expectations regarding human AI collaboration data governance auditability and communication about AI


00:33:28 - 00:34:45
use I love the keywords AI digital pathology patients. So that's what patients say about AI and I'm so happy that this is published because how how are you going to like get uh to this type of this type of insights and I don't think I have anything else for today to get the recording my trailblazers um get the book because then you're in the newsletter uh getting the book for free The PDF is going to get you on the newsletter and then you're going to be have information. You're going to get a


00:34:09 - 00:35:24
lot of emails from me so prepare for that and you are free to choose which you read and which you don't read. Uh so but the book is um free the PDF the the um this paper version is available on Amazon. So um you can just Google it or I can send you the link as well. I should do that. Um, so get the book, you're going to know everything. If you're interested in checking out these AI generated paper summaries to go a little deeper, it does not replace reading the paper. Okay, let's be


00:34:46 - 00:35:50
transparent here. It's not that, oh, AI is going to tell you about all these things, but it definitely goes h a little deeper than just the abstracts that we're doing during DigiPath Digest. Uh this is something that's useful for me. So uh I assume it's going to be useful for you. You can check it out. There's already one uh in there. All the ones that I'm doing today are going to be in there as well. And uh with that, I wish you a fantastic day. H if you have any questions that you have not


00:35:18 - 00:36:06
asked yet or if you're watching the recording, listening to the recording, uh let me know in the comments. Uh if you're interested in my second book about how to get uh access to digital pathology images, um I'm going to be uh opening it for pre-sale soon. So if that's of interest to you, let me know in the comments. Um, and I talked to