236: Quality, Teaching, and AI: A Practical Shift in Pathology

Where is AI in pathology actually becoming useful right now? In this episode of DigiPath Digest, I review 4 new PubMed papers across digital pathology, whole slide imaging (WSI), computational pathology, medical education, forensic pathology, and breast cancer AI. We look at a deep learning tool for coronary artery stenosis measurement in forensic autopsies, an AI-powered digital pathology model for renal pathology education, an open-source quality control tool for prostate biopsy whole slide images, and a breast cancer stage prediction model built for resource-constrained settings using low-magnification H&E slides. I also share updates on the upcoming second edition of Digital Pathology 101 and the decision to make AI paper summaries public on the podcast feed to help busy pathology professionals stay current. 

Highlights
 
[01:28] Update on the upcoming second edition of Digital Pathology 101 and the release of public AI paper summaries for faster literature review.

[05:22] Paper 1: Deep learning for coronary artery stenosis evaluation in forensic autopsies using whole slide imaging. Why objective stenosis measurement matters, how the model outperformed visual estimates, and why this could affect adoption in forensic pathology.

[15:18] Paper 2: AI-powered digital pathology with case-based teaching in renal education. A practical discussion on annotated digital slides, flipped classroom learning, and how digital pathology can improve pathology education and diagnostic reasoning.

[21:34] Paper 3: Open-source AI for quantitative quality control in prostate biopsy whole slide images. Why WSI quality control matters, what PathProfiler measures, and how automated QC can support remote pathology workflows.

[32:38] Paper 4: Breast cancer stage prediction from H&E whole slide images in resource-constrained settings. A look at low-magnification AI, vision transformers, and what moderate performance can still mean when access to advanced testing is limited.

[45:06] Closing thoughts, invitation to vote for future AI paper summaries, and a final reminder to download Digital Pathology 101. 

Resources
Paper 1: Development of a deep learning-based tool for coronary artery stenosis evaluation in forensic autopsies using whole slide imaging
PubMed: https://pubmed.ncbi.nlm.nih.gov/41998396/

Paper 2: Integrating AI-Powered Digital Pathology With Case-Based Teaching: A Novel Paradigm for Renal Education in Medical School
PubMed: https://pubmed.ncbi.nlm.nih.gov/41995002/

Paper 3: Application of an open-source AI tool for quantitative quality control in whole slide images of prostate needle core biopsies - a retrospective study
PubMed: https://pubmed.ncbi.nlm.nih.gov/41994924/

Paper 4: Deep-learning-based breast cancer stage prediction from H&E-stained whole-slide images in resource-constrained settings
PubMed: https://pubmed.ncbi.nlm.nih.gov/41993946/

Transcript

00:00:02
Welcome my trailblazers to our weekly dig digipath digest which is our abstract review journal club of what's happening in digital pathology. What's happened actually what was published since last week. It's based on pubmet alerts uh of papers about digital pathology and medical AI. And today I have a few today I have of course I had to deal with some technical difficulties which is mouse is not working even though I exchanged batteries like five times. Tablet is semi working cuz my pen

00:00:43
is not working. Uh but I see you here and this is amazing. So let me just say hi in the chat and we do have four papers for today to discuss. Let me know where you're tuning in from. Let me know what time it is. It is 6:03 in Pennsylvania. So actually I'm 3 minutes over time. I hope I didn't lose not Yeah, three minutes too late. Four minutes too late. I hope I didn't lose anybody. Couple of updates before we jump into the papers. So the book right you know there is a book digital pathology play digital

00:01:28
pathology 101 uh that you can download for free as PDF. So let me give you the code to get the book. So what's happening with the book? I've been talking about updating the book for a long time now and now we're at the end of April and let me tell you the text is already updated. There are also images inside that I still need to work on but the text is already updated. Uh and hopefully by next week you will be able to download the PDF. Currently the PDF is of the previous book, but if you download the

00:02:06
previous book, meaning the first edition, this is the first edition and we're going to have the second edition. If you download the first edition, you will automatically get the PDF of the second one. So, if you're not on the list, if you don't have the book yet, Digital Pathology 101, all you need to know to start and continue your digital pathology journey, there's a code for you on the screen. And we are actually getting so close to the second edition which is going to be so good because so

00:02:34
much has happened uh in the AI space specifically uh that changed since the first edition. Um so that's the first thing and another thing I'm going to keep the uh the the code on the screen. Another thing is um my div and I made a strategic decision to release the AI paper summaries uh to the public. So take it from behind a payw wall. It was supposed to be a well it was for some time um a paid subscription and some of you believed in this and decided this is useful enough to actually subscribe. So thank you for

00:03:18
that. Um, I'll send separate emails to you. But so what are the AI paper summaries? AI paper summaries um are as the name says summaries of the full papers that we're discussing now the abstracts generated with AI. What I'm going to do the only thing standing between me and releasing these AI episodes is me recording an intro for them so that people have context. those who have not subscribed have context why there is an AI generated uh summary on the podcast feed but basically um we

00:03:56
decided it's more valuable like it's a lot of hassle to access them uh because they're not available on your normal podcast feeds uh you have to like subscribe and pay and it just wasn't worth it I decided that it's better for you to have the value and the value the main value of these AI summaries is uh timesaving. Oh my goodness, you don't even know how much time they save um when you don't have the time to read papers. So, um let me share my screen with the first paper. And of course,

00:04:35
there's a little bit of glitch with my PDF reader. So, bear with me. Maybe we can hide the side here. No. How does that work? Well, we're going to make it big anyway. Um, and so the AI paper summaries are going to be released, which is exciting. Let me know if you're excited. Let me know something in the chat because if I I don't see any chats I don't I mean I see like the count of people who's here but just say hi in the chat that would be fantastic uh because it always makes me feel that you are

00:05:22
here with me actually interacting. So uh let's start with our first paper which is developing a deep learning based tool for coronary artery. Let's see if I can make it even bigger. Can I make it slightly bigger? Yes. Okay. Um development of a deep learning based tool for coronary artery stenosis evaluation in forensic autopsies using whole slide imaging. That was a very interesting uh paper. I'm smiling because I see some people commenting. Thank you so much for saying hi. It's it

00:06:01
just makes me smile. Um so development of uh this was an interesting paper. It is international journal of legal medicine which is one of the things that makes it interesting because it's not a specifically like pathology journal and we're going to have our uh pathology journals as well. Um but it's from legal medicine. Um and I had a glimpse into like how legal approaches pathology and in general like medical um expert reviews when I had a guest uh on the podcast. I had a lawyer on the podcast

00:06:40
and um maybe in the show notes I'm going to give you the link to the uh to to this episode where they said like how cumbersome it is uh to work with glass slides and now we have this for uh coronary artery stenosis evaluation. So let's see what happened here. Cardiovascular disease is a leading cause of mortality with coronary artery disease accounting for 60% of case uh in adults. 60% is coronary artery disease and the quantification of the stenosis in forensic autopsies is crucial for

00:07:21
determining causality causality between pathological findings and death. um but it is hindered by subjective visual assessment and interobservability which is always the case when we have human observers right so that's one of the reasons why we develop these AI tools so this particular study wanted to develop an AI tool using whole images for objective stenosis measurement in forensic investigations um so what did they have they had 98 anonymize H& stained autopsy slides and 234 coronary

00:08:04
sections. Then they had 103 highquality region of interests and they split it into training and validation and also test data set which is like the classical way how um you should be developing these models. But um let's see if they're going to be telling us some details. Now let's go through the abstracts and uh then uh I have the images here as well. So they had annotations delineated with data augmentation uh weighted cross entropy loss uh AdamW optimization and um so basically like a lot of um machine

00:08:48
learning, deep learning uh things that you need to do to actually develop a good model. And let's uh check what the validation results looked like. So um agreement with ground truth was was excellent um with uh so this is great. Um and on the test set accuracy improved uh and it outperforms the pathologist visual estimates. Um and the important thing the total inference time was uh 199 seconds uh for seven cases. So it was just 28.44 44 SL seconds per image. Uh so this is huge because it's super fast actually.

00:09:49
Um and what they say is hostled imagebased transform pipeline enables rapid auditable and reproducible coronary stenosis measurement reducing interobserver variability and supporting standardized interpretation in forensic investigation. Um, I don't think this abstract actually does uh this paper a justice because and let's see if we can maybe get some information from the image here. Yes, I love this because uh they did like how this uh annotation um and everything looked like. So, first

00:10:31
of all, um the I don't know what I wanted to say. I had a thought. That's a good image. So, um they where okay so, so here annotating process of a coronary with actual stenosis 74.15%. So the stenosis ah yeah so um I know what I wanted to say that um okay when you have this um acute cardiac death or uh an infuction uh you are expecting to see changes in the heart but the changes in the heart only appear after so so what's happening there is a um lack of blood flow to the heart muscle and then

00:11:22
you have changes in the heart muscle when there is a heart attack. It's eskemia, lack of uh oxygen and the muscles die. But if it's like super acute, you can only see the changes may still in the heart muscle may still not be visible. And this is when they are uh checking if there was actually a stenosis and if there was a stenosis of uh 70% or greater then it is considered enough evidence to actually uh say yes this was the likely cause of that and you know I'm not a witness in this case and they

00:12:03
probably brief witnesses how to speak about it and uh uh how to like qualify this but basically Um the message here okay 70% or more even without cardiac changes uh could be a likely cause of death and so here is already uh a problem right 70% visually how are you going to do this uh what are you going to do you're going to guesstimate and that's basically what pathologist did and in this paper uh they like show how big the discrepancies between the pathologists were and they bigger than the normal

00:12:39
like 70% agreement or 30 30% disagreement that we are expecting right um and here if you would like the AI summary of this particular paper uh after the live stream we're going to give a poll uh and there's going to be the four abstracts that we are discussing today you can vote uh which one is going to end up as an AI summary so if this is interesting for you You can already put in the chat that you want this one as an AI summary. So, we're going to go by these here in the live stream. It's going to be uh

00:13:15
paper one, two, three, and four. We have four uh for today. And this one is the first one. Um so, if you like to know more about this one, and they are open source. You can download them. So, if you like reading, by all means, go read. But if you want an AI summary on the podcast feed, let me know. just put a uh put a one in the chat. Uh so anyway, and they developed this model and the model was super accurate. Uh and also defensible in court. Uh at the end of this paper, they are saying, hey, how is

00:13:52
now the visual assessment of this going to be defensible if you have an AI tool uh that is actually accurate? uh matches the matches or exceeds the ground truth of pathologists. Um so we are entering an interesting um for me this is an interesting time where now it is not justifiable anymore to not to not use a tool in a case like this. So if this is something you're interested in having as an AI summary, let me know. I like I already have all the AI summaries. It's now uh up to you which ones are going to end up uh on the

00:14:34
feed. Um so if that's interesting to you, let me know and we're going to move on to our next paper. If you have any questions, if you have any comments, uh if you have any thoughts like oh how can it be now illegal not to use such a tool, let me know in the chat because it's like semicontroversial statement I would say. Uh which for me means it's going to be driving digital pathology adoption. Um, now let's change what we're sharing here and go to our next paper in our classical view.

00:15:18
Okay. Integration of AI powered digital pathology with case-based teaching and novel parting for renal education in medical school. Um let me tell you a a story about this not about this paper but in general about uh using host images for medical education. So I was doing my uh PhD and residency in Germany. I'm from Poland and I studied in Poland you know on glass slides classical way and in Germany we also would teach on glass slides but in addition to that they developed this whole slide image portal with

00:16:00
annotations for histo I think definitely for hisystopathology I saw the version for histopathology where they had all the lesions of all the different um changes that the students were supposed to learn in pathology annotated I was so jealous was I was like why didn't I have this as a student? I was like really literally jealous that the students uh had access to this and I didn't when I was learning h and then I thought like well I have access to it now why don't I just go through it which is what I did

00:16:34
so um that's how AI can power um education right so let's see what they say here integrating a integrating AI powered digital pathology with case-based teaching and novel paring for renal education in medical school. So medical students, veterary students, any kind of u healthcare related uh thing that uh teaches anything on glass slides, they struggle with understanding um different things including renal pathology and traditional teaching approaches that rely on the didactic lectures and static microscopy images

00:17:16
frequently fail. Let's see if I can highlight as I do go through it. Um, yes, they do what I remember also learning hisystologology where we would get these box of slides and then we would draw them like we were actually uh obligated to buy colored pencils and draw them and the professor would sign off on our drawings. But uh basically it was not really based on understanding. It was based on learning these slides by heart because we knew that then one of these slides is going to be on the exam

00:17:54
and then we will have to talk about what hisystologology that is and describe this tissue. Uh but it definitely was not based on understanding. It was very much based on memorization which we did very well. Um but I suspect anybody who did not end up in pathology already forgot all these tissue architectures. So um here what they're proposing a concurrently casebased learning uh and flip classroom strategies are gaining traction for fostering active clinically relevant learning. So in this paper they

00:18:27
discuss how AI assisted whole sled imaging platforms can support interactive exploration of renal lesions and simulate diagnostic reasoning. Um, they also present a conceptual framework for a case-based flip classroom approach that leverages annotated slides, clinical cases, and active discussions. Um, and this hybrid model has the potential to improve student engagement, diagnostic accuracy, and readiness for modern digital pathology practice while also allowing with competency based medical education principles. Um and honestly I

00:19:08
did love this hybrid model that we did in Berlin because uh people came prepared to class. Uh so they had the access to the uh virtual version online. They had to like do homework actually like do some tests. So they they had to do it. It wasn't really optional. And then they would sit on the microscopes and us as assistants would go around and help them find the different features, find different cells on the microscope. So you know there was the microscopy component hands on but they had the

00:19:41
virtual option as well. And the case-based learning is also not really something super new because I remember learning dermatology like that. So here we have renal pathology but I remember dermatology like that where it was like um this diagnostic casebased workflow where you had a patient presenting with something h and then you had to make a decision. If uh if it was the right decision a good thing happened. If it was the bad decision then it was worsen worsening. So then you had to make another decision to like reverse your

00:20:19
decision. At some point you would get uh the microscopic images of what was there and you were able to diagnose on that and I really remember this as the most impactful learning experience that was reasoning based like diagnostic uh way of thinking was introduced for the first time and this thing I did in Spain when I was doing my Arasmos exchange in Europe. you have these um exchange programs for students where you can go and study abroad for a semester or two semesters. I went for two semesters to Spain and that's where I

00:20:55
got introduced to the case-based learning and I'm glad this is happening in the medical field as well. So the other ones are not marked because of some technical difficulties that I had today. So bear with me. But I did listen to uh and of course if you're interested in detail how the renal pathology is taught in this hybrid model give me a comment in the chat that you want the second paper as an AI summary as well and now and just let me know where you're tuning in from. Uh I see you guys are are

00:21:34
tuning in. So just give me a few comments. What do you think? Why are you even coming here so early in the morning? I appreciate you so much. Okay, so let's this one. Application of an open-source AI tool for quantitative quality control in whole slide images and prostate needle core of prostate needle core biopsies. A retrospective study. So this study talks about a kind of a trend in medical AI and the trend is quality control for a long for the longest time and it's still important the trend was like

00:22:17
diagnostic support computer aided diagnostics showing the pathologist where they are supposed to look they already know where they're supposed to look but you who hates looking at stuff that um is super boring and I would hate the same thing too. The historicians who are now in the digital pathology implementation time tasked with manual QC. Oh my goodness. Can you imagine? Just imagine or maybe you are a historician listening to this. Imagine, I don't know how many cases they they I'm going to give you my example from

00:22:59
from my work. So, let's say I have a study that has 1,000 cases. Oh, sorry, 1,000 slides. And manual quality control the person. So, so there are two steps of quality control. The uh first step is actually when you're making the slide and checking that everything is there, no faults, and that it's a diagnostic slide. And then you scan this slide and you can introduce uh well you can you are introducing a new source of variables because you're creating a digital image from the glass slides. So

00:23:31
what can happen? It can be blurry. Uh the folds I mean the folds if they were excessive um they they were probably detected but some are not detected right. Uh what else can happen mostly out of focus and some um hisystologology processing like folds uh torn tissue or whatever right and now you go on the screen and can look at them for eight hours. I would quit like within a week. I would quit within a week. And um the thing is that people are quitting basically uh or some labs are not going digital because

00:24:12
they are facing the threat of these people who are so valuable to the whole hisystologology process not just like this digitization piece which I love very much but it's just one piece of the workflow. They like please who would want to do that? So the trend is or like the the new application of digital pathology is quality control and that's what they did here in this case uh they did it for prostate needle core biopsies right and this was a retrospective study where was it published do we have it I

00:24:46
think it's journal of histo technology but uh the quality of images is a determining determining Sometimes these words have interesting pronunciation determining factor for proper diagnosis and prognosis. Um and for enhanced performance in of digital and computational pathology and this is crucial especially if you want to like run algorithms or on these slides uh they should be as artifact free as possible. Uh so um now uh in the context where diagnosis are increasingly quantitative well that's a

00:25:29
statement I would argue some are some are not uh I see this quantitative trend more in the biomarker discovery space than actually diagnostic space but let's assume that's the case and sorry nobody's telling me that I have this small text as you cannot read this when it's so small. Let me know in the chat. Um okay. So um if you want to do something with these digital images in terms of AI, automated, precise, effective and rapid quality control is a of paramount importance. And the uh nice

00:26:07
thing here is that uh they developed a tool path profiler. Uh this is a deep learning based software trained on prostatic tissue that provides a usability score uh of host imaging evaluating its suitability for diagnosis. So they develop some kind of score and they describe what the score is. But this path profiler is open source if I understand correctly. Do we have it? Yes. Yes. Ha. Why did I not highlight this? Let's do it a different color. opensource AI tool. So no excuses again like the court said in the first

00:26:46
paper, no excuses, no more guesstimation here. No excuses for those who those who say hey well we don't have uh money for uh QC software now you don't need money because open source is you can download it for free. Um so here um what they did at the central de anatomia pathologica Germano de susa um they have 200 sorry 200 prostate biopsies a year um and they are distributed to pathologists remotely. So this is also an important fact that hey digital supports remote work for pathologists which is not just nice to

00:27:36
have for pathologists if you prefer working remotely h after I don't know how many years of working remotely I still wouldn't change it for anything but every now and then I need to go to a coffee shop to focus better uh on whatever cognitive task I don't read slides at the coffee shop but I do write things write pieces and work on um like written pieces. Anyway, so uh but what I'm getting at is remote meaning digital powers remote. Uh why do we need remote for pathology and probably for other

00:28:09
medical specialties as well is because there's a shortage of specialists. So you want to tap into the pool of uh global um global resources, right? Okay. So uh it was crucial to investigate path for fire's viability for automated and quantitative hostled image quality control and its monitoring for diagnostic purposes. uh and what they did in the last 3 months of 2024 um they used these 226 uh H& host images from prostate needle core biopsies uh where they retros uh they were retrospectively

00:28:52
analyzed by path profiler and they had this usability score focus um and h& quality was registered numerically what they also say in the full paper I don't know if they tell it to us in the abstract is that usually uh so whenever there is like a scanning u artifact a problem with the slide that comes from the scanner usually it's blur um but you can have other things um but then you have to rescan it right uh if not re um remake the slide like recut it uh But if it just comes from the scanner, you

00:29:37
rescan it and you want to keep your rescan rate under uh 5%. Right? So if it's over 5% uh you have a problem with the scanner or with whatever, right? Um they want to keep it under 5%. But oh my goodness, if you have to look at it visually without uh an assisted tool anyway. Uh and here we have a list of artifacts. Um the artifacts such as mounting media, dust and folded tissue where the major artifacts detected. Uh so that's the CL these are the classics right? Uh a combination of um but they don't say they don't say

00:30:24
um anything about blurry but blurry as well. And then they also had extra prostatic tissue uh was recognized as other artifacts which is not really an artifact per se but what they used it they used it as a proxy for checking the quality of the surgery because um so let's see what they say here patrol profiler is a valuable tool in the automatic quality control for of prostate biopsies, allowing quick evaluation and identification of cases requiring review before being handed over to a pathologist. Um, and promotes

00:31:08
recognition of opportunities to improve laboratory and clinical quality. They actually don't say it in the abstract, but because I listened to the AI paper summary. Uh, so is is this our sec? This is our third paper. So if you want this as AI power paper summary, uh, put three in the comment. um was that okay there was extra prosthetic tissue meaning it came from somewhere it came from the surgery right so they also used it as a proxy metric for the quality of the surgery um okay so that was this one let

00:31:46
me know if you have any comments any any anything you want to tell me you can tell me on the comments Um, okay. And now we have another interesting one. But let me make it smaller so that we see the full title. Can I do it make it smaller? Of course I can. I just need to know how. What do we have here? Okay. So this is a novel way of approaching things when it comes to diagnostics because uh we are talking about deep learning based breast cancer stage prediction from H& stained halllight images

00:32:38
uh in resource constraint settings. Yes, the title doesn't tell us too much about it. Uh it just tells us that we can predict it and we can apply it in research uh constraint settings. But the methodology is pretty different from what I have seen so far. So let's have a look. uh we know that AI shows promise for evaluating primary breast cancer including nodal status and molecular subtype. Um we know that AI there there is uh models being developed that can predict molecular status from the H&

00:33:30
nothing yet in routine diagnostics. But this is also a trend that uh could potentially benefit uh low resource uh places, low resource um environments because uh molecular testing is more expensive than uh the computation that goes into this molecular prediction. Let's just park the discussion. Okay, you would need a um digital infrastructure for this. But there are also approaches being developed on static images and um people are working on it, right? And here um they present a reusaw aware deep learning pipeline that

00:34:08
combine com combines a vision transformer feature extractor with an annotationbased multiple instance learning aggregator to predict predict pathological tumor note uh metastasis stage from hemattoxiline and using hallite images. Nothing crazy here. Good combination of methods, right? We have a vision transformer being leveraged. So this is a new newer architecture new. I don't say new because uh it probably started coming out. Well, they probably worked on it before uh I published the book. I

00:34:45
don't have it in I don't have it described in the first edition from um 2023. Let me just put the book code for you on the screen one more time. Oh, and I I see some votes. uh which papers uh for AI summary uh I see vote for paper three. So we can actually do two papers. So you can put two preferences. I'm not going to be doing more than two because then that's going to be AI podcast instead of Dr. Alex podcast. Uh but I think it's valuable enough. Let me give you the QR code for the book. Whoever

00:35:24
does not have the book yet. Um, but why do am I showing the book? Because it has the date 2023. The next edition is going to be 2026. It will have vision transformers. Here I'm like not fitting into my camera. Anyway, so new edition will have vision transformers and uh what they did in this uh particular paper, they have a vision transformer. they have an uh with an attentionbased multiple instantless learning aggregator uh to predict tumor node metastasis. So um multi multiple instance learning was

00:36:04
also an interesting architecture that um powered the I think believe page AI initial prostate algorithm. Uh and in the AI powered summary they also explain how this all works. So if you're interested, so the AI powered summaries um are pretty good for understanding the methods underlying uh these u computer vision models, right? So it's going to tell you uh how a transformer works in contrast to multiple instance learning, how they make these models uh these methods combined. So, if you're interested in

00:36:41
that, let me know in the chat that you want paper three to be as No, sorry, four. This is for this is our last one. Paper four to be uh in on the podcast as an AI powered summary. Okay. But the suspense, I hope nobody left because they tried to make the suspense why this paper is actually interesting. Um so, uh they have these methods, right? methods are cutting edge but actually not new anymore. What is new is that they operate at look at this 2.5 magnification. Like what kind of malignancy cytologic

00:37:28
self features are you going to see at 2.5 magnification? Not too many. But uh because this is well below the uh 20 to 40x uh typically used in computational pathology and um originally it was 20x now uh 40x or a combination of these depending like what validation studies you ran but definitely not like oh just use your two 2.5x for diagnostics. uh but that's what they did and that is where the resource saving component comes into play. So um here let's see for feature extraction they evaluated

00:38:10
three backbones the uni foundation model. So another new thing foundation model uh uni fine-tuned on the breast carcinoma subtyping bra data set. So we have foundation model that what is a foundation model? A model that uh for pathology in this case that already recognizes some features, understands the image. Uh so it could like for example uh if it was recreating parts of the image, it would be able to recreate the correct part in the of the image. So, for example, if um you would uh give it a cardiac muscle missing a patch, it

00:38:58
wouldn't put uh fat tissue in there. Uh right, it would put if it was supposed to like draw uh finish the drawing, it would put the correct texture there. So, it like understands obviously in quotation marks uh what the tissue architecture is, the underlying components. Maybe it can already like understand um not only like the edges and the normal visual features of an image but also the pathology specific but for a particular use case in this case uh breast cancer subtyping it was fine-tuned on a breast cancer data set

00:39:39
and um they also used a third one which is a ResNet 50 tuned on the same breast assets sorry the data set and ResNet is an uh deep learning architecture, an older architecture that still performing well. Uh deep learning works pretty well for uh different pathology tasks. Um and the embeddings from the best performing uni fine-tune network were used as input for the multiple instance learning model. So here this is where the vision transformer powered mill multiple instance learning comes into play

00:40:20
because uh you so so the multiple instance learning my pathologist understanding of this is right like you you divide the image into a bunch of patches and then like try to predict uh which one is important h and it works but it's not that efficient h And here this foundation model is actually telling the multiple instance learning model, hey I already like went through these patches and I understand the tissue enough to tell you that this one this patch is more important than the other patch. So here take my information

00:41:00
the weights not the weights the importance of the patches for diagnosis. Here you have this additional information and go work faster. And this is also where the resource savings comes into play. Um and the performance was assessed on three test sets. Internal hold from the uh semlice cohort 82 whole site images from 72 patients. Then um they also had a curated sub set of the nighting gale high-risk breast cancer prediction data sets bunch of slides 9,489 hostlide images from almost 600 patients

00:41:44
and we have our um the hero of many many papers TCGA uh BRCA data set h and this was 731 hostlide images um and 678 patients and they will all so let's not forget we're still operating at the 2.5x not at 40x. Uh so the pipeline achieved area under the receiver operating characteristic curves values of 0.6 663 uh 0.672 and 0.632 632 which uh 0.5 is a random like coin toss h and one is perfect. So we are well above 0.5 um so so these different numbers were for the different data sets h and

00:42:45
they definitely acknowledge that operating at 2.x X may limit access to fine cellar cues. Their results indicate that stage relevant information can still still be captured at this resolution. Um this study provides a transparent compute efficient whole slide image only baseline for uh PTNM stage prediction from wholeside images supporting feasibility in resource resource constrainted environments. So, um it's sad but true that hey 0.66 or even 0.67 you would want more for diagnostics. But then if you have nothing you're

00:43:34
going to take this over nothing. And this is the it's sad but it's true. It's the same uh kind of principle with molecular prediction from H& you you predict something. It's not like 100% we did PCR or next gene sequencing or whatever other method and we like amplified your gene and this is what it is. But if you don't have access to that, you're probably going to choose access to something that rather than access uh to nothing. Um and that's why the caveat uh resource constraint but also like in

00:44:15
which way uh other than just leveraging all these different um new methods together to make them even uh more efficient. this um option to do it at 2.5x slides uh means that you can have a whole slide scanner that only scans and at this magnification, right? So that drastically reduces the cost, increases the speed of uh acquiring these images. Um, and that's what uh makes the the resource requirements a lot lower. So, going back to the chat, let me know if you have votes for this fourth paper

00:45:06
about the um breast cancer stage prediction from H& stained host images at 2.5x. Uh if you're interested in that, give me a four in the chat. Um and if you don't have the book yet, get it right now, my friends. Get the book. Uh the PDF is free. You're going to get it automatically. Uh when you um check the code, uh you're going to be taken to a page where you uh have to give me your email address and then you will get emails from me. So, be prepared uh for that. But you can unsubscribe

00:45:46
anytime as well. Just get the book and read it. But uh if you're here, you probably want to show up more times. Uh and hopefully next week, let let's let's make it a challenge for me. I don't know how many ch of these challenges I already did. Let me tell you, even if I didn't like fulfill the whole challenge, didn't like meet whatever the goal was, I did get closer to the goal of updating uh digital pathology oneonone and including all the new AI information uh for you in a

00:46:23
accessible clear language clear format if you're just starting in digital pathology or if you have expertise in one area of digital pathology but would like to learn about uh different aspects of dig pathology, this book is for you. And uh thank you so much for joining me today uh and for bearing with all my my technical difficulties. Um I'd like to think that uh because it's a live stream and all these things happen. Um obviously like 15 minutes before I want to go live and some I can fix and some I

00:46:58
cannot fix. It makes it a little bit more relatable, a little bit more fun for you. Thank you so much for joining me. Uh if you're watching the recording, vote for which paper you would like to have as an AI powered summary. And if you have no idea what these summaries are and if you even like to listen to them, I'm going to release a bunch of them today in the podcast feed a digital pathology podcast. You just Google that if you're not listening yet and have a listen and see um if it resonates. And

00:47:31
if it does, then I hope to see you next time. Thank you so much and I talk to you in the next episode.