00:00:01

 The next 47th Digipath digest, it's so great to have you here. Uh if you are here live, let me know in the chat uh so that we are on the same page and I know when to start when people join. I did get a few uh error messages that this page is not working, so that's why I'm kind of waiting for your comments before and I start diving into the paper. In the meantime, um Okay, one more time. It says issue, something went wrong, refresh the page. All right, let's go again. Let's try it.

00:00:45

 Um Okay, maybe I can check it on LinkedIn if I'm actually going live on my own. Uh and if I am, you let me know that you hear me and you see me, okay? Uh let's see. No, it's Okay, I am speaking on LinkedIn. Good. Then, you know what? Let's just continue. Let's just continue this, ignore this issue, and whenever you join, just give me a comment. This is a signal for me that everything is working, that you hear me, and that you also see me well and we can do the um and we can do the live stream.

00:01:39

 Okay. Uh so, I was playing with uh different media. So, if you are here live, let me give you a clap. Okay, is this working? Definitely clapping is working. So, congratulations on being here and let's dive into the papers. Um I hope more of you will join as we do it. Um 6:02 in Pennsylvania. Let me know in the chat where you're tuning in from and we're diving in into four papers today. Um Let me put myself in my favorite box. Where is my favorite box? Come on. This is my favorite box.

00:02:31

 Welcome. We are starting with the impact of tissue detection on diagnostic artificial intelligence algorithms in prostate digital pathology. Um so, tissue detection to me is like a no-brainer. How can this be a problem? And apparently it is a problem. So, let's have a look why it is a problem. Uh this particular paper was published in uh scientific reports and the impact factor is 3.9. This uh looks like a European group. I see some Polish names. Um Kordekuwitski, the first author, Bowman. Um so, what they say here

00:03:15

 is that tissue detection is the crucial first step in most digital pathology applications, right? Yes. If you don't detect the tissue, like what are you going to analyze there? Let me make this a little bigger. Um Uh and but applying image segmentation algorithms uh by applying these algorithms, all tissue is delineated and background discarded from a further analysis. So, again, to me it was like, yes, you need to do that and you need to do it well and it should be easy because it's so

00:03:49

 easy visually, right? Uh but apparently details of the segmentation algorithm are rarely reported. And that is the case. I don't remember a single paper where this was like included in the evaluation of of the model. Uh, and there is a lack of studies investigating the downstream effects of a poor segmentation algorithm, which basically like if you don't detect this tissue, then it's going to be everything else is going to be wrong, right? So, uh, disregarding that tissue detection quality will jeopardize

00:04:24

 patient safety in a diagnostically relevant part relevant parts of the specimen. Uh, if the these relevant parts are excluded from analysis. So, uh, if you don't detect well, if you detect only part of it, or anything happens there. So, uh, this study uh, is trying to check whether performance of downstream tasks is sensitive to the tissue detection method. Uh, compare the performance of a classical and AI-based tissue detection. Uh, so, they trained an AI model for Gleason grading on prostate cancer

00:05:01

 on a lot of algo- a lot of images, and they had two tissue detection algorithms, like thresholding classical and U-Net plus plus, uh, which is deep learning AI-based. Um, and they used so many slides, 33,823 whole slide images were scanned on seven digital pathology scanners. So, pretty big study, I have to say. Um, and then the downstream downstream Gleason grading algorithm was trained and tested using 70,000 whole slide images from 13 clinical sites scanned on 13 different scanners. I don't think I

00:05:43

 have seen a study that big. Um, yet that used so many slides, uh, and on the slides where tissue could be detected by both algorithms, there was no difference in overall Gleason grading performance was observed, right? Logical. If you detect the same thing, uh, even with two different methods, your input is the same for whatever comes next. So, uh, but there was a decrease from uh, 118.43 to 23.09 fully undetected tissue samples when switching from thresholding-based tissue detection to AI-based.

00:06:28

 I'm sorry. Like, the thresholding did not detect tissue in 118 samples, like, not at all? That is interesting. And then, the AI also did not detect in 24 cases. So, for me as a pathologist, this was like, how? How can that be? Because, you know, you see the tissue, this is the first thing you see, but obviously it's a computer, even, um, when you open a Word document, it can crash on you. So, here, if you don't detect what you're later need to be analyzing, um, then you're not analyzing

00:07:05

 anything. Um, AI model may be more reliable than the classical model for avoiding total failures on slides with unusual appearance. And we're going to, um, see what this unusual appearance can be. Um, and moreover, tissue detection-dependent clinically significant variations in AI grading were observed in 3.5 of malignant slides. Um, that's it for for this abstract. So, let's have a look at our infographic and see if we can um depicted better. Let me know if you have any questions. Let me know if you have

00:07:51

 any comments to this and also let me know where you're tuning in from and what time it is for you. 6:08 for us in Pennsylvania. So, um Let's see. Can I draw here? I need my drawing tool. Come on. Here. Uh the foundation of precision white tissue detection matters in pathology. Well, why does it matter? If you don't detect it, you cannot do anything with it. Uh and it goes in three different levels. Here at level one is the foundation. Then we have the Gleason grading and the different

00:08:30

 algorithms and then we have uh the diagnostic result, right? And like when you take a first glance at the this infographic, you can see that um okay, here's like they they show this is the classical way of detecting and this is the AI way of detecting. Uh and um if this doesn't work, then this doesn't work and you get uh result that this is B9. Uh Pardon my notebook LM. It says D9 instead of B9 [clears throat] and not pignant instead of malignant. I love it. >> [laughter] >> I love how AI is like so smart and so

00:09:17

 stupid at the same time. Makes me smile every time. Anyway, make me smile. Um okay. So, what is happening here is um we have the classical thresholding uh and it struggles with unusual slide appearances. So, unusual slide appearance is going to be um um when it has a color that is not uniform. So, here it shows like white patches and uh pink patches. It's not going to be like this, but there may be a fault, there may be I don't know, some artifacts that basically uh throw off the thresholding algorithm,

00:09:59

 especially if it's adaptive. Whereas uh what they say here that the AI segmentation based on U-Net plus plus uh is doing some kind of color normalization and it looks better. And then obviously, if we have um here 79% reduction in total failures uh when we switch to AI from 118 to 24 samples. I'm still like I'm like I'm sad that you still have 24 samples that were not detected because the tissue was not detected, then it's like what I'm even talking about. Makes me sad. I mean, make

00:10:39

 makes me happy that somebody looked into it in this paper, but then makes me think oh my goodness. Like this workflow is so complex. It has so many parts and you have to check every part. It's like validating any process. No option to have a shortcut when oh, um analysis result because then your your statistics for the result, you don't know where your error comes from. So, yeah, a little bit maybe not discouraging, but a little bit um reality check kind of thing. So, um yeah, for the diagnostic result, uh when

00:11:24

 we have the tissue detected, the concordance between both methods is high. I'm I'm going to check if I'm still on LinkedIn working actually. I don't know. I see I see you. So, let me know which platform you're um viewing on. Okay, just give me a quick comment which platform you're on. I am streaming from a streaming software, so it's kind of platform independent um but sometimes Okay, I think it's working. If it's not, let me know immediately. So, yeah, that is why tissue detection

00:12:05

 is so important. Mhm. And I am Why did we have 3.5 of malignant slide in 3.5 of malignant slides choice choice of tissue detection method led to clinically significant variations in file light final Gleason grade. Okay, yes, so um this is basically what we're seeing here. that uh if you choose one that detects tissue less reliably, then if it's not detected, then you're not going to see any cancer cells there and your uh Gleason grading AI does not have anything to run on. So, obviously it's

00:12:47

 going to say there's no malignancy, so it's benign, which is wrong because it just didn't detect tissue. So, that's that. Let's move um back to our abstracts. And let's look at the next one. The next one is a super cool resource um for bridging the gap between pathology and computer science. Um it's decoding digital histopathology, the building blocks for computational researchers. Um I love this one because uh this is why I started my platform, Digital Pathology Place.

00:13:33

 And by the way, for those who are here for the first time, I'm Dr. Alex. I'm a veterinary pathologist and a digital pathology educator working in the digital pathology space for the last decade I've been working in the space and on Friday's morning we are doing this interactive journal club called Digital Pathology Digest and this is our 47th time. I have you here for the 47th time. Well, we know how many times have you joined live in the comments. I'd love to hear that. Give me, you know, one, two, three or

00:14:10

 more. That would be amazing to see that. So, um why I started the platform Digital Pathology Place was because Can I make myself bigger? Uh, I don't know. Or maybe I can. Um, just So, I started this platform because I saw that um I was working in a digital pathology company where we were doing image analysis and I was the pathologist working with image analysis scientists and I've noticed very quickly that I don't understand their language nor did they understand my pathology language.

00:14:49

 So, we started building we started learning to communicate better and then I built some trainings and then once I switched jobs I decided, "Okay, if that company needed it, it means the whole um um like discipline needs it needs better communication and that's what the authors here say as well that and the authors are this is a group uh that is representative of European Society of Digital and Integrative Pathology and I have Diana here because I met her in December and at the conference in London Digipath digital

00:15:29

 imaging DBAI digital pathology and AI organized by Global Engage and we actually spoke. I read her papers before and she follows me online as well so we're like online acquaintances but that was the first time we met in person. So what they say here this is published in Plos Digital Health with an impact factor of 7.7 and they say that computational pathology is a novel discipline. Can I make it a little tiny bit bigger? Okay. More or less. So this is a novel discipline at the intersection of pathology

00:16:12

 and computer science and combining the insights from both disciplines remains challenging due to differences in technical background and language between pathologists and engineers. And I have experienced this first hand. I had to learn all the computer vision lingo and how it translates into tissue and the computer scientists needed to speak pathology with me as well. So they acknowledge that literature translating fundamental pathology complex concepts for computer scientists remains limited. And this further

00:16:52

 complicates the understanding of the field especially for those entering the field and that is something I've noticed as well because digital pathology is a pretty niche discipline and those who started in this discipline has have moved a lot during the last decade right? We were starting with validating scanners proving to the world that digital pathology is actually as good as glass and now we have uh 10 scanners cleared by the FDA. Uh there are like templates for how to run these studies and basically this

00:17:28

 tells the world, "Yeah, it is good enough." But, um although those who are doing digital pathology and have been doing this for some time um are like kind of advanced, there are still a lot more people who are not doing digital pathology. And but they are entering the field. So, you know, with the field having moved further away from the beginnings, there is a gap for those people who are entering. And this paper is addressing is one of the resources addressing this gap. Um another resource that I want to share

00:18:03

 with you is my book, Digital Pathology 101, all you need to know to start and continue your digital pathology journey. So, if you're interested in the PDF of this book, I have included it in the chat. I don't know if this chat transfer messages to all the platforms, but um let me just share where you can get it online. Um here. When you go to the link, you can just put your name, email, uh and you're going to get this book into your inbox. And it says what this book is about, what it includes, and it's free. The PDF

00:18:46

 is free. So, if you're starting your digital pathology journey, um I highly recommend my resource and it is probably a fantastic pair uh with the um paper that we are discussing right now. All right? So. Uh So, abstract is is not going to do justice to this paper, so we're going to to the infographic in a second. Uh and I see more of you joining. What's going on? Why am clicking some stuff here. Sorry. I don't want no teleprompter. Okay. Pardon my in uh in unintentional clicking. Okay, so

00:19:38

 as I said, the um European Society of Digital and Integrative Pathology, just double-checking the abbreviation here. Um uh their mission is to promote education and interdisciplinary collaboration in digital and computational uh pathology. Is it computational? No, integrative. Um so um that's what they did in this paper. They aim to provide a comprehensive but accessible guide to pathology for computational scientists and other researchers. So if you are starting um the digital pathology journey as a

00:20:18

 computer scientist, this is a fantastic resource. Um Okay. And let's move to the infographic. We have an infographic on this one. Don't get intimidated by this huge tower of the infographic. We can do it. Uh we're going to start with the bottom. Um and what they uh what they explain here is something very very uh important but often overlooked uh that what happens with the tissue when it actually gets collected and converted into a slide, right? I love this tissue looks like steak. Um

00:21:12

 but steak is tissue, right? Muscle tissue. Anyway, so this steak um is collected, then is fixed, and you know what? It does This is not real, right? It's just an infographic, but basically what's happening in the lab, you collect the tissue, then you fixate the tissue in formalin, you cut it to the shape you want to put it in a cassette where it it gets embedded, then you cut it into thin slices, um you stain it. Staining, come on, notebook. You cannot distinguish A from O. And then you cover

00:21:55

 slip, sometimes you put a tape or different thing, but then this is how tissue becomes a slide. And then what happens? It gets digitized. Uh the digital layer, whole slide imaging. And whole slide imaging, again, it's not like a snapshot picture, it's going to be >> [clears throat] >> uh taken as multiple pictures. It can be tiles, so this should be tiles, or it can be lines that are then later stitched together uh into one whole slide image. And like each of these levels introduces some

00:22:32

 kind of variability, some some kind of complexity, some kind of error source. Uh we call the bottom part this lab journey the pre-analytics, and then we start with the analytics. So the first problem um out of many that can happen is going to be how do we create this image, right? They are a lot bigger size than the radiology images and other medical images, so even if people are coming from other medical imaging disciplines, they're going to be um confronted with a different way of working when they do whole slide images.

00:23:09

 And then uh we have the interpretative layer, the diagnostic logic. Okay, how do we do it? We don't jump super deep immediately. We start assess the architecture boundaries at low magnification, then we assess the uh cells, the nuclear uh the the nuclear shape, the nuclear features, the cellular features at high magnification. But first, we go low magnification and find where to even look. It doesn't make sense to like look at every single cell of this huge um image because then you're going to

00:23:49

 like what there is the saying um about like trees and forest. I don't remember the saying. Anyway, uh that's not how you do it. You identify a region by pattern recognition, by seeing like that there is a disrupted uh architecture of tissue, and um then you perform your diagnostic uh part. And um for this, you need the necessary clinical context. And then, if you want to go one step further, then you're going to do analysis of this tissue. We just talked about segmentation, uh locating the cells and tissue. And if

00:24:26

 you don't locate the tissue, your thing is not going to work. Uh you can have a diagnostic uh algorithm, you can do some kind of discovery. For example, here they're showing a DNA molecule uh prediction of uh molecular properties from tissue is an area in this discipline that is heavily worked on. Um so, that is what is happening with the tissue in computational pathology. From tissue that looks like steak to uh some kind of for example molecular prediction from hematoxylin and eosin slides. So,

00:25:08

 if you're just starting, this is a good resource. And now let me address a comment. We have a comment from Indonesia. AI model has completed the data source from histopathology. Yes, so they train these models on different data sets. There are different data sets gathered, a lot of them apparently, but the one that is most famous is the TCGA, the Cancer Genome Atlas, and that for a long time was like the only one that people had access to with a lot of slides, and this is where the models learn histopathology. But now

00:25:53

 different institutions are collecting images of different tissues for different purposes both in human medicine, veterinary medicine, the clinical applications, preclinical applications, research applications, and drug development. So, everybody who has who makes glass slides now, they have an incentive to actually digitize them, collect them, and do some work on them. So, I hope this answers the question. If not, feel free to ask a clarifying question, and we're going to move on to our next abstract.

00:26:38

 And of course, let me know where you're tuning in from and what time it is for you today. Well, right now. And you know what? If you're watching the recording, you can tell me what time you're watching it as well. I'm just going to be smiling later reading the comments. The next one is self-refining segment anything model SAM, segment anything. Well, apparently it's not segment anything because it doesn't do so well with nuclei. So, self-refining self segment anything model for nuclear

00:27:14

 segmentation as contrastical contrastive learning approach to label efficient pathological imaging. Um the problem with labels is is a bottleneck. It's always a bottleneck, especially if a person needs to provide this label. So, you have a person in the loop. And if this person is a pathologist, they have other stuff to do on a daily basis. Um and also, if they don't, they're going to get a pretty bored and tired with these annotations very fast. So, let's see what the authors say. Um

00:27:52

 or the This is Journal of Diagnostics from Basel, 3.3 impact factor. the authors Nam and Park. And they say that precise nuclear instant segmentation is a prerequisite for reliable digital pathology. Yet, the scarcity of pixel level annotations remains a significant bottleneck for deep learning models. So, here I'm like thinking did we not solve this problem yet? Like nuclear segmentation. And I think there are like models that do that. Some Stardust this this was a model like an open-source model. But I'm like in every

00:28:29

 slide you will have to segment nuclei. We still have a problem with that and we still need to annotate them. Well, apparently we do. And it's good to talk about it as we just realized that there is a problem with tissue detection something sometimes that can cause diagnostic problems. So, let's look at the um nuclear segmentation situation. They propose a self-evolving framework for robust nuclei nuclei segmentation that uses only sparse point annotations extending the segment anything model

00:29:05

 SAM. So, why does it not segment nuclei so well? It's a zero-shot model that recognizes all natural images. This one is from Meta. So, I bet our Facebook images anything we posted on Facebook since we started posting on Facebook is there. But, only a fraction of the Facebook users post >> [snorts] >> histology images and it's not the descriptions. So, there there is the problem, right? The the data the scarcity of this data in the training of the model. So, >> [clears throat]

00:29:43

 >> and like Okay. Let me show you another just to like just to um visualize like what it would be to annotate these images, right? So, >> [gasps] >> and I did that. I did that for years on different things and every now and then I need need to keep doing it. Uh but I want to I want to make this bigger. How can I make this bigger? Of course, it's not making the right screen bigger, but that's okay. So, you have these are nuclear right here. Um It's AI-generated image, right?

00:30:39

 Uh whatever. Anyway, you would have to like go in and annotate every single freaking nucleus like that in the specific field of view, not miss a single one, and be super precise. And after 10, I would be done. I would say, "You know what? No, I'm not doing that again." But, now what they're asking me is to click a dot. 1 2 3 4 5 6 7 8 Um and I should have a different color, right? Let's make it a dot. 1 2 That's not writing. Anyway, you get the point, right? Uh you just point at these nuclei um instead of

00:31:23

 drawing around them. And here I can say, "You know what? I can give you 15 minutes of my time to click on the nuclei as I, you know, open my files, listen to a podcast, or something. I can do that. It's okay." Um So, going back to the abstract, and their method introduces a self-evolving labeling strategy via exponential moving average. The uh the papers today are pretty computationally heavy. Um so, they integrate instance-aware contrastive learning using point prompts as spatial anchors.

00:32:09

 So, point prompts is going to be the clicking. Um and spatial anchors, then they can like grow these points or like see what pixels are around um and implement a consensus-based filtering mechanism between prompt-guided and prompt-feed decoders. And uh they did an evaluation on different data sets. So, I didn't even know these data sets. On CPM 17, MoNuSeg, some nuclear segmentation, and the challenging Co- ConCept ConCept, some nuclear segmentation data sets demonstrated that our framework achieves

00:32:50

 state-of-the-art performance across various backbones, and they included different vision transformers, B H M transition from general-purpose foundation models to specialized histopathology experts. So, they made the SAM segment anything into a specialized pathology expert. And this self-refining approach delivers a highly efficient, accurate solution for automated diagnostic workflows in clinical settings. So, you know what? I'm going to tell you something what I think about this like a redoing

00:33:25

 annotations over and over again for the same freaking thing as nuclear segmentation that you're going to need probably in every single slide. Like, can't we just make a model that everybody can use this use or like this model to be part of whatever other models are downstream? Like, why do we need to reinvent the wheel? Um so, also like so many annotations have been made for training, and these nuclei look the same. So, you know, if you are computer vision expert, uh please argue me on that. And let's continue this

00:34:06

 discussion uh between pathologists and computer scientists, where a pathologist says, "Like, it's obvious that it's a nucleus." And computer scientist says, "No, we need more annotations." Uh so, you know, prove me wrong from the computer vision standpoint. Um and let's look at our infographic. Uh if you have any questions, any comments, let me know. Um even if they're not that related to what we're talking about, just anything digital pathology. Um okay, so here.

00:34:40

 The high cost of ground truth. I already like wind and complained enough about this. Uh the cost is high uh because the fully supervised models require thousands of pixel-perfect manual annotations. And pixel-perfect like we used to sit together with different pathologists and like argue, oh, are we too off to the left of this nucleus? Or mostly it was about invasive margin of the tumor with that requires some interpretation. It's like, oh, how detailed do you need to be? Can you just like circle it? Or do

00:35:18

 you need to be super detailed? And I was always super detailed and was at least in this task. I'm not always super detailed in everything, but in this like I wanted to delineate every cell. And then I was like, if you think this is not relevant, prove me wrong. Do like artificial intel- artificial annotations that are sloppy, and let's do correlations. Obviously, nobody has time to do these studies. And if you're a pathologist, maybe that should be That's a good idea for a paper. Like how

00:35:47

 granular do the annotations need to be? But maybe it's irrelevant now because we have um the weekly supervised um methods where you can do point prompts. Um like on this side of our infographic with supervision, instead of drawing boundaries, pathologists provide simple clicks at the center of the nuclei. So here again, at the center. If I go more to the left or more to the right, then are they going to penalize me? Probably it's relevant, so probably they will. Um at the center of nucleus nuclei reducing

00:36:19

 annotation time by orders of magnitude. Yes, it does reduce. And and the problem with the SAM uh segment anything zero shot SAM models targets with H&E stained images because they were trained on natural scenes not microscopic cellular structures. So what they did they like froze some image encoder and they trained it trained like a portion of it and then what happened is they had the EMA refined pseudo labeling and self correction loop. And then they were able to segment nuclei. They used the point prompt as special

00:37:09

 anchors. They had hard negative mining and they result the dense clusters. So when you have a cells like at the top of this image like they're very dense especially in lymphoid structures you cannot really say where one nucleus ends and the other one begins. So they somehow did resolve that and here this is the result of the segmentation right all the nuclei are segmented and we didn't have to annotate so much. So thank you so much for this work. We have how many do we have? We have one

00:37:49

 more. So stay with me. We have now one about virtual immunohistochemistry. Virtual immunohistochemistry for HER2. Okay. So HER2 score our virtual immunohistochemistry via non-contrastive multitask translation. So again pretty computationally heavy. This is also in Journal of Diagnostics and Basel and um what they tell us here is that HER2 human epidermal growth factor receptor 2 IHC is pivotal for breast cancer management, but it relies on additional tissue processing beyond the routine H&E

00:38:45

 staining. And it remains a clinical burden. That's basically a statement, "Hey, you're using up tissue and you are taking time to do IHC, which is necessary for the full diagnosis and later for treatment." And that is the case, and if we could solve this, that's fantastic and that's what they're trying to do. And they they say that virtual staining offers a potential solution. Current methods often fail to explicitly account for HER2 score specific expression patterns. So, they developed

00:39:16

 a score-aware framework designed um for the precise generation of virtual H&E uh sorry, virtual HER2 IHC images. Uh and here they use non-contrastive multitask framework. Before they used contrastive learning in the previous paper, here they used non-contrastive. Um and it integrates negative free patch alignment, style content constraints, and auxiliary HER2 score supervision for high-fidelity H&E to IHC translation. Uh and the model was validated on the BC I data set utilizing an official split

00:39:53

 of 3,896 training, 977 test images, and uh but they had like images, patches, or image patches from um 51 whole slide images. Wow. You could have taken some more, please. Um but what they demonstrated is superior virtual staining performance. Um and in downstream HER2 scoring tasks, the virtual IHC alone yielded only accuracy of 83 83.01%. The fusion of H&E and virtual IHC further elevated it to 97.85% accuracy and 98.23 F1 score. Um this is pretty good. Um what they say that their model enables H HER2 score aware virtual IHC

00:40:56

 generation from H&E and can serve as a complementary tool for HER2 assessment in digital pathology. So, um definitely complementary. So, this this um kind of is in the category of molecular predictions from H&E, right? And so, so, IHC is visualizing the proteins. Um actually, like pro- protein prediction, but um it is a molecular method, right? So, um this is this this space of predicting something that you cannot see from um H&E. So, what how this is being used currently is as something that is like a screening or

00:41:50

 could be used because there are few applications so far. There there was a case report and there are commercial image analysis solution that can predict different mutations. Um but it is always as kind of a triage tool because then you have to confirm because this is prediction. This is not a chemical way of showing that uh, in this case, HER2 would be there. So, um, they did predict this. Um, and they had H&E morphology, and then they had the conventional IHC with antibodies. Um, and they had they did this H&E to

00:42:35

 virtual IHC translation. Um, and they had layer one, layer two, and layer three what they did. Uh, layer one was patch-wise, um, patch-wise. This is not seems here. I'm lost. This is, um, non-contrastive alignment, uh, that learns strictly from positive patch pairs. So, uh, this is like I don't know what, uh, notebook LM, uh, did here, but this is non-contrastive alignment that learns strictly from positive patch pairs, uh, and it avoids false negatives and ensures structural consistency.

00:43:23

 Um, then um, we want to balance the, uh, preservation of tissue morphology and the positivity of the target IHC stain, and, uh, then we have a score our multitask learn. So, so there is, um, the scoring rate zero, where there is no positivity, one plus when there is very little, two plus when there is more, and three plus when there is a lot. So, so that's like a label, like a categorical variable. And, um, they created this virtual IHC image tool. Um, I do, like, I'm always pretty skeptical

00:44:06

 about these IHC virtual IHCs. Maybe because I just don't understand them enough and it's like I don't know. I don't know even how to describe my attitude towards virtual IHC and why do I feel differently about that than predicting molecular properties which like oh if I like feel totally okay with predicting a mutation status from H&E and I feel like funny about predicting IHC maybe because I see IHC on the slide and I never have seen a mutation on a slide. I don't know it's just like personal

00:44:59

 perception of these two things whereas they're both like in the same bucket of molecular predictions from H&E and they like stand in between the H&E and the molecular method itself so they're like you still can do IHC or should if you have a specific prediction that the the staining is going to be there right although no it's just my personal perception and it's so funny because it's just my personal bias based on understanding on of different methods right IHC is something that is like

00:45:40

 tangible visual molecular pathology mutations this is something that is less implemented in veterinary pathology so this was not really part of my training so to me it's like abstract out there so let's just predict it from H&E whereas IHC I have in my mind these two pairs of H&E slide and IHC to and like I want this to be a perfect match. So, anyway, that's just my personal feelings. Let me know if you have any questions. Let you Let me know if you have any comments and join me next Friday in the next

00:46:17

 episode and let's let's give us some claps for uh here in here. Uh if you are watching the recording, leave me some comments and that always shows that it's valuable and I talk to you in the next episode.