Digital Pathology Podcast

237: Why Pathology Vendor's Don't Speak the Same Language?

• Aleksandra Zuraw, DVM, PhD • Episode 237

Share episode

Use Left/Right to seek, Home/End to jump to start or end. Hold shift to jump forward or backward.

0:00 | 33:08

Send us Fan Mail

Why are pathology vendors still speaking different image languages when radiology solved that problem decades ago?

In this episode of DigiPath Digest #46, I talk through four papers that all point to a bigger issue in digital pathology: we are not only dealing with better algorithms. We are dealing with interoperability, workflow design, explainability, and whether the field is actually ready to use these tools well.

I start with DICOM in digital pathology, because I think this is still one of the most important infrastructure questions in the field. Digital pathology has clear value for consultation, image analysis, archival, and workflow, but vendor-specific whole slide image formats still create silos. In the episode, I explain why DICOM matters, why adoption is still low, how the multi-resolution pyramid works, and why this is really about enterprise imaging and future-proofing, not just file conversion. 

Then I move into kidney transplant rejection, where the paper makes a strong case for multimodal precision diagnostics. Creatinine is late. Antibody testing can miss important biology. Biopsies can miss the area that matters. So the opportunity is not to replace pathology, but to combine biomarkers, biopsy, and machine learning in a way that is more useful than any one signal alone. I also talk about explainability here, because if a model gives a risk score, we need to know what contributed to it. 

The third paper focuses on perineural invasion in solid tumors, and I liked this one a lot because it shows how AI can help standardize something that is clinically important but still inconsistently detected and reported. Perineural invasion is not just a passive pathway of spread. The biology is more active than that, and the quantification can go far beyond a simple yes-or-no answer. This is a good example of where digital pathology can do something humans cannot realistically do by eye at scale. 

The last paper is on gastric cancer immunohistochemistry biomarkers and advanced quantification, including HER2, PD-L1, mismatch repair, and CLDN18.2. This section is really about complexity. We are now asking pathologists to visually score biology that is getting harder and harder to summarize consistently, especially when markers, spatial context, and multiplexing all start to matter at once. I make the case that computational pathology is becoming necessary here, not because pathologists are failing, but because the biology is outgrowing purely visual workflows. 

What ties these four papers together is simple: digital pathology is not only about remote reading anymore. It is about interoperability, quantification, explainable AI, and making pathology more precise in places where the old workflow is reaching its limit. If you are a pathologist, lab leader, or digital pathology trailblazer trying to figure out what actually matters right now, this episode will help you connect the dots.

Episode Highlights

 07:41 – Why DICOM still matters if we want digital pathology systems to work together.
14:39 – Current adoption of SVS, MRXS, and DICOM, and why DICOM is still lagging.
16:44 – How the DICOM whole slide image pyramid works and why it matters for workflow.
24:29 – Why kidney transplant rejection is still difficult to diagnose with any single marker.
29:18 – Why perineural invasion is clinically important and still inconsistently reported.
34:44 – How AI can quantify tumor-nerve relationships more consistently than visual review alone.
46:39 – Why gastric cancer biomarker scoring is getting too complex for purely visual workflows.
54:55 – Multiplexing, spatial biology, and why explainable AI matters in biomarker interpretation.
01:04:01 – What is really blocking digital pathology adoption: cost, workflow, regulation, or mindset? 

Resources mentioned

Support the show

Get the "Digital Pathology 101" FREE E-book and join us!

00:00:01
with my team. Um because uh I get this error message issue. Something went wrong. Refresh the page. I mean I can refresh the page. I don't want to refresh the page. Anyway, let's hope we are alive and that I can present with this voice of mine. I just woke up with a sore throat. I like had semi sore throat and I thought it's going to go away, but it didn't go away. But maybe it's going to go away as I speak. So, let's do it. Let's confirm with my team. Um. Ah, shoot. We're not live anywhere.

00:00:57
Um, not great. Not great. Why can't I stream? Okay, let's leave. How about now? Are we live now my friends? Are we live? And if you are listening to this, then let me know in the chat that you're here because it kicked me out twice already uh from my own live stream. Hello. Good morning. Dr. Alex, I'm a veterary pathologist and a digital pathology educator. And today we are doing our weekly Digipath Digest, our journal club style presentation of what's new in digital pathology and medical AI. I have

00:02:24
a sore throat. I just woke up with a sore throat and um I decided I'm going to do the live stream anyway. So, let me know in the chat if you're here. It looks like we are live. Just give me a signal signal. Um that you actually can hear me and can see me. Um 6:05 a.m. from Fairfield, Pennsylvania. And we have a few cool papers to discuss today. And by few cool papers, I mean four. Exactly. We're doing four. I decided So I always get these PMT alerts um like 10 more or less and then I pick the four best ones that

00:03:12
uh we should discuss in the stream. So let me share the first one with you. And we have our we have the papers, we have infographics. H we have everything today. And I love the infographics. Uh so the infographics, let me show you. We're going to start with DICOM. Uh I'm just waiting for a confirmation in the um in the chat that you guys hear me and see me. Um and if you are not able to comment then I will just continue because it looks like um it is okay. My team was asking, "Hey, what happened

00:04:01
to your voice?" They thought it was a different person speaking. No, it's me for real. Look, I have the earrings, the signature earrings. So, how can I make myself big full screen? I just click on myself. No, I stopped sharing. So, I just came back from the Indica Labs user group meeting. And so, we were recording a podcast. There's going to be a podcast about companion diagnostics coming out soon. And I think there is a podcast with a guest with Michelle Mitchell about um cancer patient advocacy and the role of digital

00:04:39
pathology coming out this week. Maybe it already came out um in the morning. Uh so that's an update and the user group meeting right because I showed you the earrings and I had uh they had a beautiful gift for me and a well I made it beautiful even more beautiful because they gave me stickers as well and they gave me a water bottle in the collabs water bottle and what do I do with the bottle at the airport? I leave it in the restroom. Yes. So there went my beautiful gift from Indabs. But um I also left the uh female members

00:05:19
of the team earrings and I put them on the table and some other people started looking at them and I said these are my gifts. They they are also available on the these are my gifts for the Indicabs team. They're also available uh on the uh on the website. Um, and people were like so curious and they look at me, they look at the earrings, look at me, and then when I say, "Oh, they're available on the website," they're like, "I follow you. I know who you are." So, that is always so sweet

00:05:51
when somebody recognizes me um in real life from online. So, big uh big heart and shout out to Indicabs for a great meeting, for the work together. you're going to see the results. And now let's go and discuss the papers. And if you have just joined, let me know in the comments where you're tuning in from. Uh me, Pennsylvania, Fairfield, Pennsylvania. Oh, sorry. Made myself small without sharing anything. Girl, I will master it next year. I promise you. I'm already so much better. we have

00:06:31
a lot less u technical hiccups although today we had one anyway uh so Fairfield Pennsylvania just one more thing from Indicab's uh also an anecdote so I always like um highlight Fairfield Pennsylvania but Fairfield Pennsylvania is a very small village it's like 500 inhabitants and I talk to Adam from Indicabs and he's like I know where Fairfield is and I'm like yeah right he knows where Fairfield is probably like Fairfield, California or whatever other Fairfield. And he says, "No, no, we went for track

00:07:07
meets to Fairfield." And I'm like, "That tracks actually because they have a great the high school has a great track. I sometimes run there." Um, I think I was committing to running more on this live stream, which I didn't. Anyway, so he knew where Fairfield was. You can Google. And now let's talk about DICOM. My friends, I have a personal relationship with DICOM. Um, love and hate relationship. So, the evolution role of DICOM in digital pathology. Um, here can be drawn

00:07:41
evolution the evolving role of DIY in digital pathology. Um this is a publication from medical college of Wisconsin and Leica bios systems which by the way uh partnered with IND labs for uh the software part of digital pathology and the analysis image analysis part of digital pathology. So like a bios systems right and we have published this in journal of pathology informatics which I saw that it only has an unofficial 2-year impact factor of 2.35. I don't know what an unofficial 2-year impact factor is. I thought all

00:08:18
journals had an impact factor like low or high but maybe I'm mistaken. So if you know let me know in the comments uh to make me a little bit wiser. Okay. So digital pathology has emerged as technology with the potential to transform anatomic pathology. I strongly believe in that that it does have the potential and it enables different things right it enables remote consultation. Can I make it even better bigger? We'll see. Um, okay. It enables remote consultation, computational analysis, streamlined

00:08:59
workflows, and more efficient archival of histopathology slides. Yes to everything. And interest in digital pathology has steadily grown. I believe in that as well. H and in the United States, 10 whole slide image scanners have been cleared or approved by Food and Drug Administration. And I have a little star here because it what it says it does not only say oh you have a bunch of medical devices that are cleared. It also says the same category of medical devices like 10 of them passed the rigorous uh

00:09:37
test of being approved or cleared by the FDA. What does that mean? That means digital pathology is good enough. FDA said 10 times yes you're fine. I mean you still need to keep uh keep clearing these devices that is the path of a medical device but they did it 10 times different companies different scanners it's good enough right so whenever I will hear the push back oh it's not good enough I'm not going to see this or that on digital pathology I'm going to say you know what go pull out the validation

00:10:11
study from the FDA for all those 10 whole slide scanners here's the DOI of the paper that decides that Uh I didn't invent that. Go check. Right. So here my little rant on not believing in digital pathology. But yeah and in short it is reflecting a growing confidence of this techn technology. But we have a problem. Proliferation of proprietary vendor specific call light image formats. Yes. H commonly used hostlide image formats like SVS is syntax BIF. I don't know this one. M RXs and NDPI. Uh so this was

00:10:50
Leica, this was Hamamatsu. I think this is Phillips. And these you need to tell me in the chat which one these are and we can Google later. Um but they are adequate. They're like good enough, right? The validations uh was where with these proprietary formats, they're okay. But the key word here is this is a siloed deployment less well suited for integration into enterprise imaging solution. Another keyword enterprise imaging solution which also I learned that can be the key to unlocking the

00:11:26
potential of digital pathology for departments that maybe as an entity don't have the budget to do it but the organization department may have an especially if it's integrated in enterprise imaging this is like not a standalone justify everything but part of an ecosystem that is already justified by other components of the system. So, I need to explore that. H currently talking with um some potential enterprise imaging sponsors to come uh on the show on the digital pathology podcast and explain how that could

00:12:02
revolutionize revolutionize. I'm saying it. It's not AI saying it's me and but revolutionize is kind of AI speak a marketing speak. Anyway, so we have diccom right to solve this problem. We could do daiko h which radiology already adopted over three decades ago and offers a vendor neutral whole slide image form. So let me just give you an analogy. Am I big or not? Yes, now I'm big. So let's say you have a phone. I have an Android and I call my husband and it doesn't go through because he has an

00:12:41
iPhone. Can you imagine like not being able to call from one type of phone uh to call another type of phone because the like signal transmission is not compatible? It's ridiculous, right? In consumer tech, it is ridiculous. Of course, you have Google Play, you have the Mac store, whatever, but you can get the same apps, they work, you can send the same images. That is not the case for scanners. Um, so very bad. Very bad. And uh one time I was involved in a validation of a digital pathology solution. Um and this validation was

00:13:21
we're going to put on the infographic here because we're going to look at this. This validation was um I was so hoping to be able to adopt DICOM. So what happens? Software doesn't ingest DICOM. scanner doesn't produce native diccom and I'm like what do you mean I wanted to be so pure future proof I wanted to use my purchasing or validator boss or whatever I was project manager or something power to promote diccom and I couldn't do it that was three years ago maybe times has have changed um let

00:14:01
me know what is your experience with diccom um and just to let you know if I see you comment it is just so heartwarming to me. So if you are in able I know that somebody some people are just listening but if you can comment let me know that you're listening if you have any question or um any comments and now let's look at some information from our notebook LM infographics and if you pick up that it's AI let me know where because I have spotted a few uh spots but they don't um like they don't take away

00:14:39
V value of this infographic. When I look at them, they are so like full of text, but when you walk through them, you actually have a good structure. So, um what we want to do, we want to thaw the data silos. Ah the problem here guys the interoperability gap in DICOM or adoption or in general in these uh proprietary uh formats is huge because 59 have SVS 40 have MR MRXs and only 11% have DICOM and you know what even with Leica when they cleared the scanner uh GT450 they made it DICOM native they opened

00:15:22
the patents that they had on um file formats because like it's unheard of in medical device and they did it so that we could all adopt dyom and I hope it's going to happen soon and please keep demanding diccom when you uh buy scanners but here this is what the 2024 survey shows not too much diccom but radiology did it in 1994 shoot My infographic is moving too much. Um, radiology did it uh in 1993. Can you imagine? And we are in 2026 and pathology is still doing proprietary. Uh, but I hope we really are hitting the

00:16:09
same inflection point and we will get this universal standard. We already have it. Um and the technical architecture of the slide obviously radiology slide gray and black and white h not Google map style. So this is our Google map style of hostlide images. The multi-resolution pyramid um the Google map model is that dicon fetches only the necessary data tiles from the zoomed in um for the zoomed in detail like a map app. Right? So when you zoom in into the cellular structures, it's not going to pull all

00:16:44
these zoomed tiles and it's just going to put pull what you need like Google maps when you zoom in on Google Maps. And here uh so we have uh like kind of two components of diccom this pyramid and here um like the naming or or categorization or metadata right we have the staining protocol um section ID and here I think AI was a little prolific with block ID twice unless this is part of DIY but I don't think so because here it shows us study ID twice so still loving notebook LM for these infographics uh study ID, patient

00:17:25
identity, right? Um so yes, this is Dao. Um but there is a problem. I want to find this problem here because the problem is that you can um like have files that are um like they can be dicom and non dicom like SVS and DICOM. Um and I'm wondering where we see it here on this infographic. Give me one second in the timeline. I'm looking for it. You can comment. Oh, here. is this dual personality crutch because we can use files that act as TIFF and DICOM because people were like okay if we already have these

00:18:22
proprietary then let's make it proprietary and DICOM and what is happening h there is a risk of mismatched metadata where these like identifiers this metadata information then doesn't match between one file or the other um and also so the first part of my personal story with DICOM was this validation story. The other one was um the annotations. So there is a supplement and working group 26 for DICOM um that that deals with annotations and I was part of well part of this group I don't think if I was

00:19:02
officially part of this group but I was taking part in these meetings very technical meetings and I don't think I contributed too much I was probably like asking more questions to people who were already involved for a long time and doing an amazing job in understanding also concepts that were super complex. But anyway, there is an annotation uh group which is now important very important in digital pathology and they do a great job on figuring out how to uh code in DICOM. So because there's of course different annotation

00:19:41
formats as well. Um yeah anyway Giojson and some other I think Giojson is the one that they are using or like DICOM version of that. Anyway uh they took care of that as well. So our DICOM I'm still a friend. Um I didn't manage to introduce it to the organization I was doing the validation for. Very sad about that but maybe one day it's going to happen and I hope it's going to happen soon for you. Okay, let's do our next paper. Let me know in the comments. Even if you're just watching the recording, if

00:20:22
you're not here live, let me know in the comments. Um, are you doing DICOM? Do you have a scanner that produces DICOM? Are you working with DICOM or not? DICOM, yes or not? Uh, give me your um your experience. And let's do the next one. We have a lot of clinical papers. Okay. My voice is holding up. clinical papers but which is good. This one rejection focused precision medicine in kidney transplantation and the full paper title is rejection focused precision medicine in kidney transplantation biology biomarkers and

00:21:23
artificial intelligence. So we have several of these uh the next three I think are more clinical um where the artificial intelligence and digital pathology are a component um of whatever new science is happening there. So this one is um live from Basil and this group is from Portugal and the impact factor of this one is 3.5 four and uh here we are talking about kidney disease and the kidney disease is the chronic kidney disease is rising worldwide and transplant ation remains the preferred modality of kidney

00:22:09
replacement therapy. I actually knew a person who had um kidney transplants but because of kidney cancer and not because of chronic disease chronic kidney disease. Anyway, long-term graft survival continues to be limited by chronic aloimmune injury particularly antibbody mediated rejection and its chronic active form. M and this is a narrative review that synthesizes contemporary evidence on the immunopathogenesis, epidemiology, diagnosis and management of kidney alligraph rejection. And we're going to

00:22:51
dive less into that. But they also examine how artificial intelligence and machine learning may support digital pathology, multimodal risk prediction and data integration while recognizing the current challenges of biological interpretability, external validation and clinical implementation. and they proposed a rejection focused precision medicine framework and outline key research gaps including multicenter validation trial ready endpoints and governance of AI enabled pathways. So let's look

00:23:29
at our infographic here. Okay, we have people who can comment. Thank you so much for joining Scott. This is amazing. Now I know that you're here. This is good. Okay, let's switch to sharing something else. Here is our next infographic. I look at it and I'm like overwhelmed. But we don't have to be overwhelmed. So we can just go through it and find the important things for us. Can I do this and still draw on it? Yes, I can. Amazing. Okay. So what is happening here for kidney diagnosis? They have this diagnostic

00:24:29
triad. I need a different color. Is this one better? Let's do red. Red is good. Diagnostic triad. Um so there is serum creatinine. Um then um antibbody testing and uh basically biopsies and they say okay there are problems with each of these. Serum creatinine is a lagging indicator. it only rises uh after the nephron scarring has occurred. Um then also that uh antibodies there is something called the sponge effect that standard DSA test can yield false negative uh or miss bound antibodies sponge effect and then also

00:25:22
that a biopsy can miss the area where we want to actually sample and and they um also talk about molecular Um but here what I want to talk about is the integration engine AI and machine learning because they have all this data coming from so uh different types of biomarker they call biomarker intelligence that's our biomarker from DNA from um other inflammatory markers cytoines then they have this funny marker well funny it's very innovative I have not seen of this kind of bio biomarker where they uh

00:26:08
gauge the activation of immune um immune system by h the how fast this torque tenno virus is replicating. So if it's replicating fast, the immune cells, the immune system is sleeping like not super active, which is good uh for a kidney transplant because there's not going to be a rejection of the transplant. But when the replication of the virus goes down, it means that the immune system goes up. So there is danger for the transplant. So that was a I thought that was a cool biomarker. I

00:26:48
never heard of a virus that is um harmless virus that that we have is a biomarker like that. So let me know what you think about this. um and then you know other biomarkers, blood expression for T- cell genetic genetics and and all that stuff right but we now can leverage AI and machine learning for fusing disperate data layers and uh also they focus on this so AI algorithms like random forest synthetizes nonlinear variables into a unified risk score um random forest. I have a experience with random forest for image analysis. I

00:27:35
guess this is not the best um tool for image analysis. Convolutional neural networks are networks are a lot better but for uh other variables maybe random forest is good. So let me know in the comments what you think. Um what is the official stand on random forest? I know that it used it is used for different more like text tabular um number data but for image analysis nope. Um and uh we don't want it to be in a black box. So um we want to reverse engineer AI decisions showing which variable voted

00:28:11
for a high risk prediction. So they have all these variables together and then uh they calculate the risk score and say okay this contributed so much this contributed so much because of this and that. So kidney transplant is relying or can rely on AI. There was another thing. No, that was in the next paper. H where we have a novel immunoistochemistry solutions. Okay. Our next paper is Come on. Let me make myself a circle. What do you prefer? Square or circle? Let me know. Depends how much text is

00:29:18
covered by me, right? Uh so the next one is perinural invasion in solid uh tumors with AI and machine learning applications. The full title is H perural invasion and solid tumors, biological foundations and the immer emerging integration of machine learning and artificial intelligence. So we are super happy about that. That's a hard by the way if you didn't know because this is published in frontiers in oncology impact factor 3.3. Uh but again it is oncology journal that is talking about digital pathology. So what

00:29:58
I'm seeing is this is now being seen as a method. I mean pathology is part of oncology of any type of diagnosis right but now digital pathology uh the the infrastructure the tools are tools that enable other um medical branches medical areas as well. So what are we talking about here? uh perinural invasion PNI. This represents a distant root of cancer spread in many solid tumors and its presence correlates with aggressive tumor behavior, local recurrence and neuropathic symptoms and uh reduce survival across selected

00:30:41
tumor types including pancreatic prostate, head and neck, coloractyl and gynecologic mancies. But that's not an exhaustive list. A and despite its prognostic value, PNI remains inconsistently detected and reported. And not only this h so this is like the standard right pathologists look at something and they differ in opinion h because it is our eyes that are limited in especially if you want to quantify something but also there's always interobserver variability and intraobserver like in the morning your

00:31:19
cognitive and visual acuity is going to be um maybe better than in the afternoon or the other way around depending on your circadian and circadian rhythm cycle whatever right anyway but the next thing is that college of American pathologists cancer protocols and clinical decisions are in and clinical decision- making are incompletely integrated uh but not only this uh CAP differs in opinion what to uh consider a perinural invasion from uh World Health Organization WH has different criteria. Anyway, um

00:32:04
so advanced advances in tumor nerve biology have reframed perinural invasion as an active birectional phenomenon driven by molecular cross talk between cancer cells. So before it was just like okay we see it here it is but now actually h it there is biology behind it. So, this information is even more important. And the granularity of this information is super important, too, as well. I like square more. I'm going to just keep myself square. Um, inside the box. I'm going to be inside the box

00:32:41
today. Not outside the box. Well, I can be a circle outside the box, but today we're going to be inside the box. Anyway, uh, so what happens? Schwan cells, neurons and the surrounding tumor micro environment um they do this molecular cross talk. So yeah there is if you have ever seen these um tumor micro environment graphics our notebook LM infographics are nothing compared with tumor micro environment infographics. They're like so full of different types of cells. And now we have another type of cell and

00:33:21
that the the Schwan cells, neurons and others, right? But but mainly Schwan cells and neurons. The others were already in the tumor micro environment. But parallel advances in digital pathology, machine learning and artificial intelligence have opened new opportunities to standardize PNI detection and quantify its extent. And we like that as well. We like standardize. We like quantification because we're digital pathology trailblazers. Um, and this rep this review provides a synopsis of current

00:33:56
knowledge on the biology and clinical relevance on in solitude with the emerging integration and application of machine learning and AI assisted approaches in histopathology and molecular profiling. So can I like Now I have to share something else. Just stop and share something else. This sharing experience is suboptimal. That's okay. Looks more dynamic on the screen when there's a lot of clicking. I hope you're not too annoyed. Um, okay. So, the center of this infographic is our nerve

00:34:44
and and Then we have a nerf bundle. So the paradigm shift uh that they are referring to is before it was like oh the tumor like passively leaks like water if there is a crack it's going to leak and then go to the nerve. But um now the biology um and the the biology uh supports a different uh theory that actually the um nerves are not victims. They're complicit and schwan cells undergo henype reprogramming to guide cancer cells and dissolves dissolve physical barriers. So they um they secrete um

00:35:32
nerve growth factor. Yes. Yes. Nerve growth growth factor. And then cancer cells get like sucked into that space. And here I I just want to show you AI traces instead of phenotyping phenotypis. That's notebook LM guys. Cute. Can I make it bigger for you? Yes, I can. Here. I just think it's it's just cute because such a powerful tool and then you have these like little traces of what and there's a little bit uh more of that. So, I do look at these with my own eyes and try to catch everything that is

00:36:22
incoherent. I do uh we go through the abstract. This is like the original abstract. And then I also listen to the um AI generated podcast episodes which by the way are going to be on the podcast. We're just picking two out of the four. Um and I will provide intro and outro h but there's going to be an AI generated podcast. Um so a little story about these podcasts as well. And so I I find them extremely valuable. So, I'm going to keep them on the podcast because even if there's another person,

00:36:56
uh, they they definitely get like less um, listens than the normal episodes where I have guests or even the Digipat Digest, but the story about them is that they are so helpful for um, being on top of the literature and then like you listen to them. It's um learning is not just u obviously a concept process complex process and you know to really learn it you should do uh spatial repetition make flashcards are uh flashcards out of all these papers and who's going to do that nobody because everybody has a life and

00:37:38
a job and needs to perform at a decent level and also for their family right so but still we need to stay on top of that so this is why I um started doing the podcast with notebook alarm and I use them to prepare for these live streams. They are so valuable to me. They save so much time and they keep you on top of the stuff. So even if you uh listen to these two that we're going to be publishing, there's a lot before you can go like a month or two um before back in the podcast catalog. There's a bunch of

00:38:11
them. They're still valuable. They are still um current enough. Uh, right. But even if you just listen to the two every week, you're going to be so much ahead of the curve. So, here ending my shout out to the a generated episodes of mine. Let's go back to Is this our last one? Yes, I think it's our last one. No, no, no. We have one more. Um, so without further ado, let's go back to the infographic. And if you're still here, let me know in the comments what you're thinking. Uh

00:38:50
what what do you think about the infographics? What do you think h in general? Just leave me your thoughts. I want to have your thoughts shared. I'm sharing a lot. Okay. Um what did we focus on on AI making spelling mistakes and that is not the point of this paper. The point of this paper for me is h our AI revolution. Right. So here, let's go to this particular spot on the screen. They basically decided that it's relevant enough to make image analysis for finding these um cancers uh cancers

00:39:30
that are wrapping around the nerves. Um yeah, and we know the diagnostic dilemma. The dilemma is that a pathologist has to look at the 2D image on the slide to evaluate something that is 3D and it's challenging and you do your best and then you still differ uh a lot and the agreement the kappa values are ranging between 0.67 and 0.75 which is considered substantial and I am actually happy that we're substantially agreeing but we can do better uh when we use AI and image analysis. So we have

00:40:06
two stage semantic segmentation and semantic segmentation is basically uh like delineating whatever we want to delineate on the image. Uh stage one maps normal nerve structures and stage two zooms in to hunt for specific I'm obscuring our thing. Um specific tumor nerve interaction filtering out visual noise. Um, I'm just going to put the star here, which I shouldn't because now you're going to be looking here, but it's just for me. And uh, with this particular um, solution that I'm not even showing

00:40:44
myself, guys. With this particular solution that they uh, designed, they had 97.5 accuracy rate. Obviously, accuracy is just like a umbrella metric that can be high even if the algorithm is not fantastic. So let's uh look into the paper to know exactly what they did but um they uh finally like were able to segment that and then you don't have to hunt for these um and now the there we are going beyond the binary reporting before it was like oh is there PNI or not and now AI can extract micrometer level proxim proximity data

00:41:27
proximity that's AI doing fine iness, proximity data, nerve caliber and absolute nerve involvement counts that humans cannot feasibly measure. So we are actually measuring then what can happen and what they actually propose and suggest is that this can be a biomarker with you know or maybe even a well it is a biomarker already but maybe it can be a companion diagnostic. So they are mentioning specific uh drugs and let's go to our agentic future conclusion because they are mentioning agentic AI copilots for example path

00:42:07
chat path chat was designed by Modella Modella is now part of Astroenica so the pharmaceutical companies that are creating the drugs can have a companion diagnostic for the drugs and that's like the theme of the month there's going to be a webinar on companion diagnostic with uh rash there is going to be the um the the podcast that I told you with Indica Labs with Doug Bowman Bowman sorry Doug Bowman uh from Indica Labs and uh all about companion diagnostics right because you want a stratified patient so this could be a

00:42:43
companion diagnostics and what you can what does have AI copilot chat uh have to do with companion diagnostic well uh what you can do now is you can interact with the results of image analysis So for example, if you run the image analysis on uh the slide, what you can ask now is show me all the instances where this uh PNI is for example over 40% or over 35%. Whatever whatever the image analysis uh delivers you as a result or like show me all the instances where you found it. So now you don't have to hunt. You have it's kind

00:43:19
of similar to um the way um cytologology is visualized like you you have boxes around these cells and then you have a panel um and you can see each of these cells. You don't have to hunt for them and then when you click on it you can go back to the slide and she see the context. So fantastic. and uh a quantified PNI will serve as a companion diagnostic for targeted drugs like tanzumab anti-NGF and we talked about NGF the nerve growth factor which is what's happening the nerves are giving it to the tumor and tumor then

00:43:54
comes towards the nerve because of this so we can have an drug that is anti this and also cell per catinib inhibitor another pathway Yeah. So, our power tools, image analysis, and then agents that can interact with multim model data. I am super happy about that. And let me know if you're happy about that as well in the comments. You're quiet today. You're quiet. I don't want to like harass you about the comments uh because I know that people are working or doing other stuff and just listening

00:44:35
in. But if you have a chance to comment um now or later, leave a comment. It always makes me feel super nice. And let's do the last one. I need to drink something and cough. Sorry, guys. I'm kind of proud of myself how well I'm doing with this voice. I was reading this book um the big leap and um the author Kay Hendris uh tells that oh the big leap is actually like a mental shift not working harder not working more and you have to exit your zone of excellence and go into your zone of genius and then you have

00:45:30
this upper limit problem and the problem is when you start sabotaging yourself with different things like, "Oh, everything is going fantastically." And then you start ruminating on something. Everything is going fantastically in your finances. And then you have a fight with your spouse. Everything is going fantastic with something else. And then you like sabotage yourself in another area. And one of the examples he says, "Oh, you get a cold." And I'm like, "Oh, did I have an upper limit problem? Am I

00:45:58
like exiting my zone of excellence and entering the zone of genius?" I hope so, but I also hope uh this cold goes away. Um, oh, thank you so much for commenting. This is so heartwarming. Just I know that you're there. I see that, you know, a couple of people are there, but then if you interact, it means a lot to me. So, let's do it. Let's do the last one, my friends, my trailblazers. The last one is gastric cancer IHC biomarkers and I quantification and uh this is this has been published in Medicina. It's a

00:46:39
Lithuanian journal and the full the full title is comprehensive overview of gastric cancer iminoistochemistry key biomarkers advanced detection methods and perspectives. And we as trailblazers are interested in advanced detection methods and perspectives. They talk a lot about IHC here, but they also talk about super cool ways of evaluating this IHC. So this is a publication from Romania and they say that IHCM in histoe is a keystone in gastric cancer management. um and the critical role of IHC markers

00:47:26
analyzing their uh in this review um they uh evaluate the critical role of IHC markers analyzing their efficiency in molecular subclassification and prediction of response to gastric cancer targeted therapies while also describing state-of-the-art IHC techniques and perspectives. So this is also important even though like dive you want to hear my story about IHC I did it once uh in my uh do during my PhD I had this project that was how can I make myself bigger without I don't know I just have to stop sharing and make

00:48:08
myself big. Um, so anyway, my my story with IHC, I did I for my PhD project, I didn't have to do IHC. I did PCR and I had negative results and I still got the title. So, all good. But, um, then after I finished that project, I was doing an IHC project for detecting scinessence in dermal samples. And I don't remember the marker anyway, some histone marker, whatever. I did IHC. Oh my goodness, my IHC technician, like I'm relatively organized, but the level of organization to do manual IHC,

00:48:48
she was furious with me because I like make stupid mistakes. Then in the end, it turned out okay. And I don't know if I had to like redo it or then I learned the process. I kind of pushed it out of my mind as a semi uh pleasant experience or more unpleasant experience. But um lesson learned, you need to be a lot more organized than me to do manual IHC one. Lesson two, respect to all the technicians that are doing this manually and are having consistent results like they have a superpower and IHC is magic.

00:49:22
Uh because like I don't know, humidity whatever antigen retrieval all these things they influence everything and then you still like want to have something diagnostic. So whoever is doing IHC, respect to you. I hope I don't have to do IHC ever again. Uh this uh experience taught me about IHC right about all the steps and um everything that needs to be regarded and how complex this method is. And when people tell me like oh digital pathology image analysis is complex I'm like have you

00:49:53
done manual IHC? Anyway, so here for gastric uh cancer IHC and in general it's part of the molecular diagnostic uh workup IHC, right? So um what they say here and and they cover a lot of uh new IHC detection methods, new antigen retrieval methods and all the advances in IHC. There is one that I'm a spec especially fan of because it yields itself amazingly to image analysis. Uh so considering the role of IHC and DAB the following topics were successfully addressed in seven sections. um gastric cancer key

00:50:35
biomarkers such as her two I had to like and her two is human epidermal growth factor receptor uh PDL1 program death liand PDL1 and DNA replication mismatch per system and they allow uh direct correlation between tissue morphology and protein expression this is why we do IC right to see it in the tissue uh that's the um advant advantage of IRC over molecular that you can see where the signal comes from. Does it come from the tumor? Is it does it come from necrosis? I had an adventure like that

00:51:13
as well where um a a data team was like mining a signature of very strong IHC signal and then the pathologist uh looked at these images and I'm like and and they said well it all comes from a necrotic area. Do we really want to quantify necrotic areas for this particular signature? And the answer was no. So here IHC shows you where the signal comes from. The main directions were focused on the integration of AI algorithms for digital quantification of IHC signal and also on expansion of panels to new targets such as claudin

00:51:49
18.2. So there is some new target and we can now address it with IHC which is great. Uh which redefineses treatment approaches in advanced stages. So other than it's cool that we found new biology uh and cancer biology is pretty complex um I also experienced um a couple of presentations at the in the collabs um user group meetings and the users are so advanced and especially when they're like staining all these uh cancer pathways right um but my message from this okay I'm not going to be an expert

00:52:26
in cancer pathway um unless I like enter this area of research which I'm which I'm not planning um at the moment but the biology we are discovering is now so complex that our visual assessment is not an adequate tool for it anymore even though you want to simplify the tests as much pos as possible because uh the simpler it is the easier to deploy it is in the lab and clinical setting but we reach a point where it's no longer possible Um and then you have computer uh computational pathology, biomarker

00:53:05
assessment, computational biology, companion diagnostics. So it's happening, it's happening. I'm seeing this happening in all these publications. Um so imistochemistry remains indispensable in modern gastrooncology. The evolution towards digital pathology and the refinement of scoring criteria will transform will transform IHC from a complimentary test into a visual tool that is essential for personalizing oncological treatment and I actually believe this is going to happen and it's actually already

00:53:39
happening. It's going to scale up. So, moving on. Before I move on, I'm going to show you a comment that is so heartwarming. It's just a thank you. Big thank you for all your work. Whenever I have the chance, I recommend your channel. Greetings from Brazil. Thank you so much. Thank you. This is This is so nice. This is so sweet. Thank you so much for these comments. one is enough. You don't have to like flood me with fantastic comments, but I ask for them just to, you know, have a confirmation that this is useful.

00:54:19
I'm obviously entertaining myself here in front of the computer because I'm laughing. I hope you're laughing with me. Um, anyway, last infographics, my trailblazers. Let's do it. Let's see if we have any AI mistakes on this one. Um, maybe yes, maybe not. We'll see. I might put myself inside the box. No, actually we're going to do I don't know. Maybe we're going to do this one today. Then everything is visible. Okay. The digital pathology revolution mapping the

00:54:55
gastric tumor micro environment. See again micro environment. um this tumor micro environment is so relevant and so complex that we need digital pathology for this. Um so let's look at the middle. I'm just going to move my key a little bit. Um so in the middle we have the cellular micro environment. What happens in gastric cancer? We have uh first inflammation uh to atrophy. Then uh we have metaplasia to dysplasia and then we have invasive adenocarcinoma and this biology um also the this entity used to

00:55:36
be in older patients now it's in younger patients so there is more of it h and the treatment is going to save more years of life so we covered the uh the three important biomarkers her two pdl1 oh my goodness PDL1 and this score CPS. I need to Google what CPS stands for. But I know what the score is. Give me one second to just like uh see because it's score PD1 combined positive score. Uh the other one is two more proportions score. And this one is combined um combined positive score. It's a nightmare. This freaking thing is

00:56:32
a nightmare. So you stain with PDL1 program death, right? And if you have seen it, it's like a membrane marker in tumor cells, but it also stains macrofasages and it also stains uh other immune cells, lymphosytes. And then you are supposed to visually do this like combined combined positivity and abstract like mentally the immune cells from the tumor cells. Let me check exactly how how you're supposed because I'm of course exaggerating. um it it is for uh pembbrolumab kituda and it calculates the ratio of pdl1 staining

00:57:15
cells tumor cells lymphosytes to the total number of viable tumor cells multiplied by 100. So trailblazers here was a fantastic opportunity to have a computational pathology biomarker and no they are like making pathologists do that it's impossible. And how do I know that it's impossible? There was a one research project where I had to score different ways of PD1 and I tried to be consistent. I don't think I was because it's impossible to be consistent with such a complex visual score. So

00:57:49
um so um yes so we have this this combined and I see comments coming in. I'm going to address them as well uh once we're done with this infographic. Um you have we have the new biomarkers. We have the um mutations and look at this amazing part of the infographic. Why didn't they put like this here and like put the number here and what is the unit IHC provide answer in anyway but even with all these like uh imperfections we know that IHC is faster than nextG sequencing and that is all we

00:58:32
need to know but let's go to and here is uh uh at the bottom we're going to have the advancements in um what happened In IHC we have uh better antibbody engineering, we have synthetic polymer detection, we have better antigen retrieval methods and but then we also have AI. I think I don't know who had this sticker at the conference but it was like a round sticker with gut AI. It was either either uh Vizioform or Indica Labs that had got AI. I of course took the the pin or the sticker or

00:59:19
whatever. So, we've got AI here uh because um and here multiplexing and AI. And here we have a little typo. Multiplicing, but we know it's multipplexing. Um anyway, because so this is something I loved. Let me focus because I'm like rambling and jumping all over the pl place. The brightfield coral color revolution translucent chromogens allow overlapping markers to show green for coexpression. And green is just an example, right? Uh you stain something with yellow, you stain something with blue and then when

01:00:01
they uh are collocicalized or co-expressed, not only Yeah. collocicalized and co-expressed that's the same thing um then they change color and you know that okay I have two now and if you try to do it visually you're going to uh hit a wall uh also another thing multiplexing 23 plus markers on a single slide this depends on the technology it can be like 50 uh it can be I don't know how many even if you do three you are better off with a Um and you can measure the spatial relationships between different cells.

01:00:43
When I started in this field like over a decade ago and I heard this like spatial relationship between tumor micro environment, I thought it's like some fancy science that is just fancy science and people want to quantify it because they are interested in it but no it actually has clinical implications. Um so we have the multiplexing the AI computational AI uh to help quantify and detect these things. Um but obviously another thing is like the explanability uh X AI uh in this particular example it

01:01:23
gener generates spatial heat maps showing exactly which cellular anomalies the AI used. It's kind of it's kind of a proxy. It's like uh okay AI was using an attention mechanism and this is um an where the attention comes from. So an example in real life is for example you example for example anyway you know what I'm going to do I'm going to analyze my transcript and see what are the words that I overuse. I did that I have this voice to text software on my computer. It's called Whisper Flow and every month

01:02:00
it gives you a report and I talk a lot to Claude and to AI for for different things trying to automate different things of my business and you guess which is the most used word that I say to AI. I'll give you a few minutes to do it and then I'll tell you anyway. But um the example I wanted to um give you for this explanability. Okay, there's a image with a dog and if the attention is actually on the dog then you can extrapolate that the information was taken out from the relevant part of the

01:02:36
image. If the attention was on a cloud and still the image was classified as containing a dog, you would doubt your AI because how can you get information from the clouds visually that it's a dog? Uh so this is kind of a proxy for that but um that's pretty good if it comes from the tumor from the uh correct cell then it is good and okay here the reveal what is the most used uh word that Dr. Alex Juraf uses with her AI uh assistant. It's don't. I like review its work and I'm like don't do this, don't

01:03:13
do that. Don't write like a marketer. You're a scientist. I'm a scientist. So anyway, don't is uh the main word that I'm using and let's address our question. Um one why why do you think it takes so much time for hospitals clinics to adopt digital pathology in their workflows? What are the biggest hurdles? So couple of different hurdles. Um so the main one that people mention is money. To me this is like inside the box thinking. H especially after I heard the story that and

01:04:01
it's not that it's not true it's that needs to be approached in a different way and what is the prescription I don't know but a parallel story is okay you're going to say oh the scanner is so expensive it's like $300,000 how much is an MRI machine it's like a million dollars and every hospital has an MRI machine right and there are multiple factors right so but but money is the one cited the other one regulatory uh complexity is another one level of adoption if everybody was uh

01:04:40
digitally native just for reading slides it would be so much easier to adopt these things and so it still requires a lot of change investment and change management and uh actually First comes the change management and the willingness to do it and then uh you figure out the ways to invest in this. So if there was one factor I would think the digital path or like pathology community or medical community not realizing the full value of this technology. So that's kind of like my goal to show it that it

01:05:19
is valuable not just as a nice to have for a digital pathologist that can work from home but actually relevant in quantifying biology that then uh translates into uh predictive information basically an information care for patients. So I do believe that digital pathology is the u that patients have the right to fast diagnosis and the gateway to fast diagnosis is digital pathology not only fast also precise h and personalized. So that's why I talk about digital pathology to healthcare professionals.

01:05:58
Bring the right businesses in front of them so that we can increase access and speed to care together. Thank you so much for joining me, for staying till the end. H if you're watching the recording, let me know in the comments what resonated, what questions you have. If you are not um a subscriber to Digital Pathology Place, you can subscribe and get this book, the PDF version of this book, Digital Pathology 101, All You Need to Know to Start and Continue Your Digital Pathology Journey. And I'm going to put this in the chat

01:06:29
right now and also it's going to be in the description. You can find it right here. I'm so over time. Let me just put it put the website in the chat. digital pathologyplace.com. And it's going to be in the link as well. I need to go cuz I need to get my kids ready for school and I didn't realize that it's 7:10 and you're still here. So, thank you so much for joining me and I talk to you in the next episode.