160: AI in Medicine: Neuropathology, Renal Disease, Hematology & Cytology

What if the way we quantify pathology is more guesswork than science? In this episode of DigiPath Digest, I take you through the latest research where AI is not just supporting but challenging traditional methods of image analysis in neuropathology, nephrology, hematology, and cytology. From Boston brain banks to Mayo Clinic kidney models, we look at how advanced AI compares to human vision—and where it already outperforms us.

Episode Highlights:

• [00:02:49] Neuropathology image analysis (Boston VA & BU) – Why traditional semiquantitative scoring often fails, and how AI-based density quantification reveals more subtle pathology in CTE.
• [00:13:16] Chronic kidney changes with AI (Mayo Clinic, Cambridge, Emory, Geneva) – A 20-class AI model trained on 20,500 annotations, showing how multiclass segmentation outperforms human guesswork in renal pathology.
• [00:21:09] Digital hematology review (University of Pennsylvania) – Current hurdles in AI for blood and bone marrow evaluation: regulatory oversight, data standardization, and resistance to change.
• [00:25:52] AI in cytology review (Journal of Cytopathology) – From BD FocalPoint to deep learning: two decades of digital cytology, stagnation, and why adoption still lags despite proven benefits.
• [00:32:09] Neuropathology goes digital – Where digital neuropathology is already routine (Ohio State, Mayo Clinic, Leeds, Granada) and why this specialty is crucial for pushing adoption.
• [00:34:19] Personal note – Why I believe learning, sharing, and experimenting with AI tools now will shape the way we practice pathology tomorrow.

Resources from this Episode

• Comparison of quantitative strategies in neuropathologic image analysis – Boston VA / BU Brain Bank study.
• Multiclass AI model for chronic kidney changes – Mayo Clinic, Cambridge, Emory, Georgia Tech, Geneva collaboration.
• Review: Digital hematology in the AI era – International Journal of Laboratory Hematology.
• Review: AI and machine learning in cytology – Journal of the American Society of Cytopathology.
• Digital Pathology 101 (by me, Dr. Aleksandra Zuraw) – Free PDF & Amazon print edition.
• Pathology AI Makeover Course – Practical training for AI in pathology workflows.

Chapters

• 0:00 Introduction to Digital Pathology Episode 28
• 1:15 Neuropathologic Image Analysis Paper Discussion
• 6:05 Computer Vision to Pathology Vision Translation
• 8:30 Multi-class AI Model for Kidney Histology
• 14:05 Hematology in the Digital Era
• 17:54 AI in Cytopathology: Progress and Challenges
• 23:05 Digital Adoption in Neuropathology and Conclusion

Transcript

00:00:01 - 00:01:13

Aleks: Good morning. Good morning, my digital pathology trailblazers. How are you doing today? 6:00 a.m. Friday. Can you believe that this is the 28th time that we are meeting here? Um, I don't think anybody has been here for 28 times, but if you have been here more than three times, let me know in the comments. I'm going to show myself the chat and say hi to everyone. H So when Okay, I see people joining. I always like I have this little icon here when people uh pop up. So this is fantastic. You're

00:00:37 - 00:02:10

tuning in it for you and uh it has been more than three times in I guess in the US you show this three that this edition. So um today we're going to start with the paper about uh neuropathologic image analysis which is so um which is a coincidence because I recently oh and we have we have people joining amazing. I have to finish a sentence first before I like show show comments. Um so coincidence and made me smile because I recently uh was presenting at the uh annual meeting of the society of toxicologic pathology

00:01:26 - 00:02:44

and the topic of my talk was okay digital pathology in neuropathology for toxicologic pathology and I did some research right and obviously I was focusing on image analysis because this is like the added value of digital pathology not just the reading but the reading and makes everything so much more connected and there is study material or we call it study material or diagnostic material or slides uh that then can be analyzed with image analysis. Um and thank you so much for the kind words that I'm awesome. This makes my day

00:02:05 - 00:03:24

because I wake up here at 6 6:03 and I um and then I see these comments up live. This is Thank you so much our papers. If you're joining right here more than once and if you've been here three times, I'm going to give you a special discount. just put it in the for um whatever is in the shop. Oh, so I have the QR code for pathology AI makeover because we just officially launched this course. I launched it to my uh email list and some people are already going through it learning AI in

00:02:49 - 00:04:28

pathology. So if you're interested um you have this little code on the left for me on the screen in the top left um and you can check it out. And in the meantime, we will start talking about our paper. Everything working? I checked my pen and everything. So, comparison of multiple quantitative strategies for neuropathologic image analysis. um very close to my heart image analysis because you know what is my opinion on trying to quantify IHC other than you know positive or negative um IHC I feel

00:03:40 - 00:05:07

deeply feel sorry for moving this basically gueststimation so like quant Obviously, as I already said, the traditional semiquantitative scoring for neuropathologic assessment is prone to variability among assessors. This is what I'm referring to. I have this thinner. Can I have this thinner somehow? I should be able to. Okay. Okay. So um it's prone to the semi-quantitative is always going to be prone to uh variability between people right variability with the same person like I'm not going to be as good in the

00:04:30 - 00:06:12

morning as I'm going to be in the evening you know depending on uh what the like prime hours or mental best mental capacity hours are for the people. Um so then and also uh it does not capture full spectrum of pathological changes nor does AI but you know nor do we right so to address these limitations um they decided and and by they I mean and the group from Boston everything Boston VA Boston healthcare system Um yeah, everybody from Boston uh they um strategic quantification or advanced artificial intelligence techniques. I

00:05:25 - 00:06:53

love this distinction. Okay, you have like the very basic positive pixel quantification or everything else is very advanced. I love it. Sweet. Um, anyway, let me just I'm going to bring the QR code back later, but I just don't want to obscure our view right now. But if you're interested in the pathology AI makeover, uh, where I break down everything um, AI for pathology, you can also comment. So, they decided to do the advanced AI or positive pixel quantification, right? And um then they did uh so however

00:06:09 - 00:07:43

comprehensive comparison of these measures have never been performed. So they they had a uh they had 100 1,412 cases from Boston University brain banks and then human-driven semiquantitive scoring was compared with the computer-driven percentage area stained measures and AIdriven density quantification of tow pathology lateral frontal cortex. The advanced uh things that they did was basically which uh are not that advanced when it comes to computer but definitely a lot more um or quantitative than what what we do

00:07:00 - 00:08:20

visually, right? So um they observed general agreement between measures. So they compared each measure to themselves. So the um positive pixels then the area stained and then the cell densities and they had a good correlation general agreement between measures which is good especially if it's a cellular staining but um they have one caveat for the pixel quantification um and also because this data set include included a large range of different neuropathologist uh they decided to um restrict the cases um and

00:07:40 - 00:09:14

uh they decided to do the SAB analysis uh in cases with the neurodeenerative disease chronic traumatic encphylopathy um and examine correlations with clinical and neuropathologic variables. So we have more data now. We have uh you know the staining um for the to staining for the CTE. We have clinical and neuropathological data human um assessment of the and also computer assist uh computer assessment of the staining. So all methods demonstrated significant ability to pure pathology. uh they had a problem with the pixel

00:08:28 - 00:09:41

count. Does not surprise me because when you just count positive pixels, you may have inconsistent background, non-cellular elements staining and artifacts increasing variability of the positive pixel method. Uh but the AIdriven method was better at identifying pathological changes associated with spares pathology. So overall their results demonstrate important differences among neuropathological assessment techniques and highlight the need for careful consideration when selecting analysis method and I

00:09:03 - 00:10:35

totally agree with that. I think it is I guess it's general like prop property of human brain. we tend to use what we know of like also in image analysis and let me put my code for the book and because this is something I describe in the book right and by the way I started uh writing the new version so if you don't have the book yet the PDF of the digital pathology 101 is free but I want to show you something in this book let me make myself big here. Okay. So, and now with the transformers and the

00:09:50 - 00:11:18

the ways of approaching image analysis, um this consider even more options. But I have this picture here. Oh, here of the dress. So, I should be able to see them. Is it? Come on camera. Anyway, so on page 77. You can download it right now. You will see a picture of giraffes. I took this picture at Disney and it basically explains computer vision concepts. Um, and I call it computer vision to pathology vision translation. And you have three giraffes. And what you can do with these giraffes, depending on what you want to know about

00:10:34 - 00:11:45

these giraffes, you can either just like put a sorry, put one box around them and say these are giraffes and that's it. If you want to know the area where the giraffes are, you can put a box around each of them and say this is giraffe one to three. But I don't need to delineate them. Then you can delineate all of them and say this is the area on the picture where the giraffes are. Or you can be super fancy and say, "Oh, this is giraffe one." Exactly. Delineated. Giraffe two as well. Um

00:11:13 - 00:12:35

to see these like all together in one big herd. Um with computer vision, you don't need to do all this fancy stuff. If you only need the information where all the giraffes are, right? Uh and then like you know the complexity depends on what kind of information you need and this is the question that you need to ask yourself uh when you do these image analysis and they did that and they emphasize it in this paper that um hey okay where is my code? Oh sorry I I don't know what happened with the code. Um, but if you're

00:11:55 - 00:13:04

interested in the book, you can grab it through the code below and PDF totally free and the giraffes are on page 77 in this version. I assume on PDF it's the same version. Um, so it is important to ask yourself, okay, do I and and these all delineated drafts, this is called instance segmentation, and it's like the let's call it the most fancy way of uh delineating on an image. And like our brain gravitates to yeah like we would like to have that. Do you really need to have that? Maybe you can just need can

00:12:29 - 00:13:47

have the boxes or h in this case maybe you can just have the pixels which um is also valid, right? And if it correlates and if it brings you the same information and is such an emphasize come okay but I'll bring it up uh if anybody is interested on and of course always if you need something from me any course any any book anything that you see on the screen that I later take down you can always send me uh you always put it in the comments so let's move to our next candidate today we don't have too many so bear with me

00:13:16 - 00:15:05

come Come on. Okay, the next one is an interesting one. It's chronic. What's with this code? Give me one second. All right. Okay. Chat clicking here between you and all the codes and everything. Um, okay. So chronic changes on kidney hisystologology by a multiclass artificial intelligent intelligence model and Julio who is co-author of this Andrew Janick who is also the author of um histo opensource software and we did a webinar together um I think you can find this webinar on my website as well and of course actually

00:14:18 - 00:15:24

you can go you can go to the store digital pathology play store and you can see all the all the courses and everything and there is stuff that's for free there it's um paid and also my earrings but I'm going to show them uh show the versions that we have to you in a second now let's focus on the science okay enough of the promo let's talk about the chronic changes on kidney hisytologology by a multiclass artificial intelligence model. I want to highlight multiclass. So this is a group from Mayo Clinic,

00:14:51 - 00:16:07

Cambridge, Emory University, Georgia Tech and also Genev Geneva, Switzerland. I think Andrea is in Geneva. Um and what they did they uh did they say the chronic ch uh changes in kidney hisystologology are often approximated by using human vision with limited accuracy. Have we heard that before? Human vision with limited accuracy. Yes, probably every single paper starts like that and every single live stream uh or you know episode I start like that that our vision is basically guesstimation when it comes to like counting

00:15:29 - 00:16:49

something. So, and they know they know because they've been doing digital um and train um trained an artificial intelligence model for segmentation struct uh segmenting structure images of kidney tissue and 20,500 annotations. um trained the model with 20 classes. Oh my goodness, I say that's a lot of classes for one kidney model. Can you imagine? Uh 20 classes of structures and they're going to tell us some of the classes including separate detection of cortex from medulla uh because there are

00:16:14 - 00:17:23

some dependent measurements that rely on this and they compared the AI model detection with humanbased annotations in an independent validation set. So this is very much like the classical way of evaluating uh the label is the human annotation and then um that your prediction is the prediction right and you compare it um to the annotation. Let's see what they actually annotated. So a model was then um applied. So they trained it, they designed it, they did it. Why is my camera not working? Hey

00:16:48 - 00:18:24

camera move with me. Okay. Sometimes it moves with me and sometimes it doesn't. And now it doesn't want to move with me. Hey camera. And while I like fumble with my camera that doesn't want to with me, let me know if you've been here more than once or for three times. You're getting a special discount code. All right. Now move. Okay. So uh then they a um applied the uh this to this AI model to uh 1,400 donors and 1,600 patients with renal tumor to calculate chronic changes as

00:17:36 - 00:19:12

defined by measures of nephron size. So here I'm already like get triggered like you visually tell people to like measure nefron size. I hope not. Maybe I hope not. Um but they included it right. Uh glumber volume cortex volume per chromolus. So here is this cortex volume per glomeulus and mean tubular areas and nephrosclerosis. So globally sclerotic glomeili increased interestium. So like this um connective tissue increased tubular atrophy and arterolar hyalinosis um and also artery lumininal stenosis

00:18:24 - 00:20:06

from intimal thickening. So the inner layer of the um of the artery uh gets thickened and then you have test I need to order my lines. And the results were that were uh oh, what happened? I switched myself. Okay. Uh the AI detection and human annotations, the agreement was similar between human pairs. um except that the AI model showed less agreement for age and ah was this arterular h highinosis um arterial arter so these are very tiny arteries h would I be surprised with that probably not maybe it was confusing

00:19:29 - 00:20:56

stuff for sure was confusing stuff um I don't know exactly okay so anybody here the second time third time let me know in the chat. Uh so okay the results a detection and human annotations was similar all good except for this one class out of 20 classes. I think that's pretty good still. uh and uh chronic changes calculated uh solely from AI based detections associated with low glomeular filtration rate during follow-up after kidney donation and with kidney after radical nefrectomy based was calculated from cortex per

00:20:19 - 00:21:55

glomeus percent glomemerous pronouncing these things is um pretty challenging and then uh tubular atrophy fosi density as well as mean area of arterial hyalinosis and these lesions showed good prognosis discriminating for kidney failure. So basically the conclusion is it's good multiclass AI model can help automate quantification of chronic changes on whole slide images of kidney hisystologology. So I was impressed with how many classes um they trained it for and uh yeah if you can like somehow have a computer

00:21:09 - 00:22:40

aided assessment of so many things um that's that's really golden. What about hematology practice in the digital era? What has changed? It's a review and the um internal internal journal of laboratory hematology international journal of laboratory hematology. Uh so these um revview abstracts are a little bit challenging to uh to to at the journal club but uh and this of Pennsylvania and I'm in Pennsylvania. So hi to University of Pennsylvania and are uh pretty short and not that positive

00:21:56 - 00:23:11

because they say obviously hematopathology is complex but I mean the kidney evaluation was complex too but um he pathology is complex as well because they include numerous data points necessary for guiding further testing h diagnosis and patient management. What do they mean by numerous data points very multimodel because we need complete blood counts uh blood cell counts and then we have subsequent morphologic evaluation of peripheral blood and bone marrow and uh AI and image analysis digital pathology

00:22:33 - 00:24:02

as well could potentially revolutionize the um peripheral blood and bone marrow assessment and did it revolutionize let's have a look at the rest of this abstract Um well we could you know implement artificial intelligence for assisted and automated evaluation but there remain major hurdles like and I have a star here because I don't really like believe in these hurdles anymore and anybody who has been with me during that even with for any of series which is a module in um pathology a makeover. So if you want to make over

00:23:21 - 00:24:26

this is a paid course um and you're going to see um like all there's a full module on all the QR code on the screen right now. Um, so there's a full module based on the seven live streams that we had and all the live stream fluff is edited out. So you don't have to like re no rewind, wind forward, what's the word for like accelerate, speed up or like uh go through the hours of live streams. It's actually uh all the fluff is chopped out and you can and and it's a full module

00:23:54 - 00:25:06

with all the papers attached. So if you're interested in this, go ahead and uh scan the code. Um and here in this uh our current paper, what do they say? They say, "Oh, oh, so many hurdles. For example, lack of uh regulatory oversight." And if you remember the paper about regulatory oversight from the uh seven part series, oh my goodness, I didn't know how much regulatory guidelines are there in the world uh of digital pathology. You know, everybody thinks just especially in the

00:24:29 - 00:25:49

US FDA and oh, they lag. Oh, they like look at it very closely and there's a lot of uh a lot of guidance regulatory uh well, but definitely guidance and awareness from the reg. So my friends this one go to the seven part series and you will know that regulatory oversight is there. You just need to want to find uh what you know what they say. Then also data standardization. Yeah, this is a chronic problem in the digital pathology space because we still like we want to have diccom but we don't have diccom

00:25:14 - 00:26:29

everywhere and different formats um and in general different data standard data standardization is going to be probably a problem in the whole world um and also then insufficient knowledge and training uh well I don't know you know I told you that the uh the abstracts like this um are pretty difficult for journal club we would need to like go in into the paper and um and read it. But one thing that I very much agree with is resistant to change among others. Um and this article reviews current state of digitization

00:25:52 - 00:27:20

and hematopathology practices and uh recent research using machine learning models, automated specimen analysis and the offer intelligence driven clinical workflows for efficient and comp workup. Um so yeah there are hurdles but we can overcome the hurdles like we can right us let me take the field of view let's have a look at our last paper machine learning deep learning and artificial intelligence as applied to the field of cytoathology a comprehensive review so we have like two similar well two similar everything is

00:26:38 - 00:28:07

based on um not on tissue Right. The hammopathology and cytology and this one it is a comprehensive review journal of American of cytoathology. And this is a little bit more optimistic. Thank you for being a little bit more optimistic because um and it it highlights something that I was asking myself about and I keep asking myself about because um the the this digitization and the implementation of image analysis uh in pathology actually started in cytoathology, right? Um and it was a long time before whole slide im

00:27:23 - 00:28:42

well I don't know maybe they used whole slide imaging but anyway they were ahead of uh of the of our anatomic pathology times right so let's look at the paper um imaging informatics h have play in gynecologic cytologology since the introduction of BD focal point and the thin prep imaging system and that was at the dawn of the new millennium. So that was super early for digital pathology. That's the beginning. And the development of these systems nearly 20 years ago led to increased diagnostic

00:28:04 - 00:29:20

performance with potential to address issues relating to speed, throughput, and reliability of psychologic assessment. So like yeah let's it's increased um increased speed increased reliability and everything like better so let's take it and run with it right no wrong uh I mean however rather than ushering it in an age of increasing automation what I would be very uh interested in and you know I would want to do and uh quantitative analysis algorithmic advances in psychology and in pathology in general

00:28:42 - 00:30:16

stagnated for approximately a decade after its introduction. H then in the 2010s 201 how do you say that? Let me put the chat. Maybe you have the right word for that. Um oh sorry confused with my screens again. Okay, there was Alex Net and what a coincidence. My name is Alex. I had nothing to do with Alex Net coupled with plummeting storage and due cost open-source development h and also more affordable high throughput digital slide scanners accelerated this progress. Um so while the widespread

00:29:33 - 00:30:38

adoption of largecale consortium level databases for example the TCGA the cancer genome atlas uh spur developments of models algorithmic modeling of cytologology specimens has lagged behind. So here we have the same problem actually that in the previous one it has lacked defi behind due to the varied nature of cytologology preparations and assessment. And you know what? I don't want to be overly enthusiastic, which I already am. Uh anyway, but cytology is challenging, okay? Don't don't

00:30:06 - 00:31:14

understand me. Uh don't don't get me wrong. It's more challenging to image, a lot more challenging to image than um anatomic pathology where you have like a super thin, nicely prepped slide. And if the hisystologology lab knows it's going to be for digital, they're going to make it even thinner and more beautiful and all that stuff without any folds and run automated QC tools. H you cannot do that with cytology. you what you have you preparation still it's not going to be

00:30:41 - 00:32:03

flat to struggle with imaging but there are devices right however increasing attention uh to cytoathology uh with a common sure increase in publications yeah in this uh journal club we already have two publications and uh commercial applications in this area. Still applications of AI and machine learning to cytologology and the herculean effort to implement fully fully digital pathology services remains at its nent stages. This review explored the current state of research and commercial development in digital cytoathology

00:31:25 - 00:32:54

and with the focus on AI and machine learning technologies and um yeah I mean we're going to be reading about that a lot in many papers that the um adoption is lagging and it is because um I mentioned this um this presentation I gave at the beginning uh at the um annual meeting of society of toxicologic pathology about uh neuropathology uh about the use of digital pathology in neuropathology and when I was doing research uh for like okay in which um medical institutions not veterary or pharmaceutical um companies but medical

00:32:09 - 00:33:18

diagnostic institutions using publications publications I wanted that um there were publications saying yes we do it for everything but then I had to actually on LinkedIn ask the people and say hey do you also do it in in neuropath and they do so for example I did reach out to Giovanni Luhan at Ohio State and they do neuropath digitally I think for five years they've been doing that and this was one of the first specialties where they implemented And I don't know who else I I checked Maya Clinic some digital neuropathology

00:32:56 - 00:34:38

service and and several right but I didn't find like bunch of them I found like four or maybe six and also um uh Great Britain Leadeds Hospital and also Granada Hospital in Spain. internationally a few more but I didn't find um references specific references on more than 10 in your pathology so that kind of uh confirms that uh it's lagging it is lagging so we need to be active and we need to help this to not lag by learning what we can do uh is to learn. So if you're interested in H

00:33:46 - 00:34:51

started version, we need to update the couple of chapters because a lot has happened in the digital pathology world. So I'm already working on the new version. If you have this this PDF version and I know that you have it because you're in my database and you will get the new version without anything additional. So, if you don't have the book or if you want a new version of this later, scan the code that's on the screen right now. Um, and if you are proud of being a pathologist,

00:34:19 - 00:35:43

being a digital pathology trailblazer or anyone who lo looks under the microscope, who loves life sciences and laboratory sciences, uh you may be interested in wearing earrings like that. So, I have uh them in the store as well. Let me show you the designs that we have. So, let me put the store on the screen as well. Can I put all the codes? No. Um, so the store, if you're interested in the earrings, getting it for yourself or for someone, this is in the upper corner right now. And my favorite are the Malt Queen giant

00:35:01 - 00:36:16

cells. These I wear them at conferences and people stop and ask me about them. So now they can stop and ask you the book is something I write for anybody digital or already having some expertise in one area of digital pathology but wanting to be this book is a fast read and it's going to bring you up to speed. Thank you so much for joining me. Uh if you're viewing the replay and you want any of the uh things that I'm um offering here, just leave a comment and I'm going to get back to