Digital Pathology Podcast

119: DigitPath Digest #19 | Cytology's Digital Revolution, Prostate cancer tsunami + Live AI Demo

• Aleksandra Zuraw • Episode 119

Share episode

Send us a text

In this episode of the Digital Pathology Podcast, you will learn about cytology's entrance into the digital pathology space, including successful AI and scanner implementations.

We cover AI's role in rapid on-site evaluation for lung cancer and share insights on a looming prostate cancer surge and how digital pathology and AI can help. I

You will also listen to a live demo of me using an AI assistant to decode a scientific paper in real-time. Tune in to stay on top of the digital pathology research in 2025!

00:00 Welcome to DigiPath Digest
00:53 Introduction and New Year Greetings
01:41 Diving into DigiPath Digest
01:44 AI in Respiratory Cytology
06:11 The Role of AI in Pathology
09:49 Multi-Omics and AI
11:28 Radiomics and Pathomics
14:44 Live Q&A and Future Plans
20:09 Prostate Cancer Tsunami
22:34 Thyroid Cytology and Live AI-Assistant demo
31:07 Conclusion and the option to send texts :)

Links and Resources:


Publications Discussed Today:

📝 Evaluation of an enhanced ResNet-18 classification model for rapid On-site diagnosis in respiratory cytology

📝 Advancing precision medicine: the transformative role of artificial intelligence in immunogenomics, radiomics, and pathomics for biomarker discovery and immunotherapy optimization

📝 A machine learning approach to predict HPV positivity of oropharyngeal squamous cell carcinoma

📝 The uropathologist of the future: getting ready with intelligence for the prostate cancer tsunami

📝 Artificial Intelligence and Whole Slide Imaging Assist in Thyroid Indeterminate Cytology: A Systematic Review

Support the show

Become a Digital Pathology Trailblazer get the "Digital Pathology 101" FREE E-book and join us!

Hello, my digital pathology trailblazers! Today's DigiPath Digest is packed. We're talking about cytology finally getting its digital moment. I'll share which scanners are actually working for this. AI matching pathologists in rapid diagnosis rose. And heads up, there is a prostate cancer tsunami coming in the next two decades. Plus, listen to me do a live demo using an AI assistant to decode a paper I was confused about in real time. Learn about the newest digital pathology trends in science and industry. Meet the most interesting people in the niche, and gain insights relevant to your own projects. Here is where pathology meets computer science. You are listening to the Digital Pathology Podcast with your host, Dr. Aleksandr Zuraw.

Aleksandra:

Good morning. Happy New Year, my digital pathology trailblazers. How are you today? I'm super excited. This is the first live stream in 2025. It's already the 10th of January, but that's okay because I was feeling so overwhelmed. At the beginning of the year, because everybody's like making resolutions and getting ready for a new year and, all those goals. I'm still working through the backlog from December. I was feeling behind, but apparently I've heard that like some energy fields say that we're gonna start better off on the 21st. I don't know if I believe it or not, but it made me feel better. Let's dive into DigiPath Digest. we are starting with Evaluation of Enhanced ResNet 18 Classification for Rapid On Site Evaluation in Respiratory Cytology. What's happening here? Of course, we have AI. Of course, we have, digital pathology, but digital pathology for cytology. Let's pay attention to this because that's going to be a theme in a few more papers. And there was a video I posted last year and the, use of digital pathology in cytology. I listed it as, a drawback that it's not that, Available for cytology, but in today's DigiPath Digest, we'll see that's not the case anymore. In January 2025, cytology is, entering the digital pathology space on par with anatomic pathology with, formal and fixed paraffin embedded sections. So let's see what they did with this ResNet classification. This is a group from China It's interesting what's happening in China, that can be extrapolated to other places that are different than, the Western world and different than the U. S. The objective here was rapid on site evaluation, ROSE. This is something that takes less than five minutes, this is the, rapid on site evaluation of, in this case, respiratory cytology specimen. And this is a critical technique for timely diagnosis of lung cancer. it's either, through, endoscopy transbronchial or transcutaneous. You sample a mass, do a cytology, and then you stain it with something, called Difquik. It's a special staining, Difquik. I remember this from my veterinary practicing days. it was super quick. You just had three jars of different stains. You put in one for one minute, two minutes or whatever. And in less than five minutes, you have a stain specimen that you can evaluate. there is in China, there's a limited familiarity with difficult staining method and a shortage of trained cytopathologists. And this hampers the utilization of ROSE. Is that specific to China? No, it can be found in different places in the world. So definitely we can Take the learnings from China and apply it somewhere else They wanted to involve deep learning because then you have an AI assistant for whatever you're doing, if you're not familiar with it, it's good to have an assistant, even if it's an AI agent or robot or whatever. And by robot, I mean software here. they, analyzed 116 digital images of diff quick stain cytology samples. And a whole slide scans. I was using my phone to find what kind of scanner they use. Let me see if I can check what kind of scanner they used. No, I don't see it. We would have to go into the paper. there's another paper that talks about different scanners for cytology. if you're interested in what scanners can be used for cytology, stay on the live stream and give me questions. Let's finish this paper. So they included six diagnostic categories, carcinoid, normal cells, adenocarcinoma, squamous cell, carcinoma, non small cell carcinoma, and Small cell carcinoma. the thing with cytology is that it's less specific than, than anatomic pathology. All malignant diagnosis were confirmed by histopathology and immunohistochemistry. And then the test that was presented to. three cytopathologists different hospitals and also different varying levels of experience and an artificial intelligence system and they say diagnostic accuracy of the Cytopathologist correlated with their years of practice and hospital setting I would assume the longer you practice the better you are at it. So I hope it was a linear correlation correlating better, accuracy with longer experience, but who knows? And the AI model demonstrated proficiency comparable to humans, which is good because that's what we want. we do want an assistant that is not worse than us. If an assistant is worse than us, then it's not a good assistant in this particular pathology diagnostic settings. And this has been shown before already, combinations of AI assistants and human cytopathologists increase diagnostic efficiency to varying degrees. if we have an AI assistant, we're faster here. So we have a promising capability as an aid for on site diagnosis of respiratory cytology samples and human expertise remains essential. So I see it with, consumer AI. with your knowledge you can sort through the answers i've been using Perplexity on my phone with voice commands for over a month and I ask it very specific questions. I ask it to search literature and specific journals It comes back, with answers, not all are, accurate, but I immediately know that this is a little bit aside from what I was asking for, and with, my knowledge, my partial knowledge of what I'm looking for example, sometimes I know the author of the paper is, sometimes I know one main message of a paper, So me combined with this tool is a lot more efficient than me going and searching through the classical methods that were keyword based PubMed, Google, whatever. even though my assistant, perplexity in this case is, not always like it does not always deliver me a hundred percent of the stuff that I want. I want, I know how to guide it. this was image based, but, it was a similar thing. Another paper that we have here, we have a lot of not really digital pathology specific is. And I think it's like a beginning of the year. Let's set the stage for, the situation that we're going to be encountering moving forward, which is, the digital pathology and AI will be helping us more and more. Because we have a few reviews in this particular PubMed alert is that, reviews, setting the stage for AI is going to help us. Digital pathology is going to help us and also cytology is part of the game now. Don't feel, if you are a cytotechnologist, cytopathologist, somebody working with cytology sample, don't feel underserved by the digital pathology community because today in those papers, the researchers are discovering that this is doable cytology is part of the game as well, and, that's what this advancing precision medicine, the transformative role of artificial intelligence in immunogenomics, radiomics, and pathomics for biomarker discovery and immunotherapy optimization. Yeah, I've seen this pathomics word. It's a relatively new word. That the most narrow definition of this would be predicting molecular, because omics is the omics from genomics, right? So some kind of, molecular properties, deriving or predicting molecular properties in pathology images or some kind of biomarkers, this is where, cancer biology med and this is a group from China and the USA. So we have a collaboration, China, USA and AI, is advancing precision, medicine in the fields of immunogenomics, radiomics, and pathomics. Immunogenomics, AI can process vast amounts of genomic and multiomic data to identify biomarkers associated with immunotherapy responses and disease prognosis. So let's see if I can. Highlight something super important. Multi omics is what I want to highlight. multi omics is where everybody wants to go and by everybody I mean the medical and scientific community bioscientific community life science community If we can integrate all these different information, then, we're so much better than just basing diagnosis on one type of information. We can detect things faster. We can predict, responses to treatment better. And we need assistance of computational pathology or computational medicine tools because so many. Pieces of data at the same time, it's not physiologically, possible for us to process, right? I'm always referring to, immunofluorescence, three markers at a time, but to evaluate, how many different, Lymphoid cells are there you have to switch those markers off you can only evaluate one at a time and that takes three times as much time to evaluate a triplex then Just a single marker, right and then you know the adjustment of time to do it In, a fluorescent image, in contrast to a bright field image that we are more used to. So we are very good in logic and everything, but processing a lot of data at the same time is too much. Multiomics is the keyword. If you don't remember anything from this paper, remember multiomics needs assistance of AI. this, supports personalized treatments. In radiomics, AI can analyze high dimensional features from tomography, MRI and PET scans as well. like 3D analysis of shapes and everything. this is very important for discovering so called imaging biomarkers, not pathology imaging biomarkers, but, radiology imaging biomarkers. Two more heterogeneity treatment responses and disease progression because you can see how this mass grows, where it grows, what kind of, shape it takes. if this, is important, then you have this data and you can analyze this data and analyze if there is a treatment response based on this. the cool thing about radiology in contrast to pathology, is It's non invasive, or let's say less invasive in real time. Although ROSE would be real time, but it's a little bit more invasive, depending. Anyway, radiology, usually less invasive, even though you have to go into a machine or get some kind of radiation. But a real time assessment, for personalized therapy. And our friend Pathomics leverages AI for deep analysis of digital pathology images. that can uncover subtle changes in tissue microenvironments. specifically in tumor microenvironment and for immune oncology. If you have so many of those markers that you want to analyze, this is a task for AI and by AI here, image analysis. pathology images whenever we have like multiple markers to analyze subtle changes also multitude of changes there was a group from Mayo Clinic, that used software Aiforia to detect a lot of different visually distinguishable features of colorectal cancer, and then they combined these, features and, created a prognostic model. So you can see visually each of those features. They are explainable, understandable, and they were loosely or not numerically associated with prognosis, but when they analyze them with image analysis, it was a totally different, piece of information, totally different piece of value for the care of that patient. Even though it was previously visible by human eyes, the combination of data together was something different. Also cellular characteristic, morphological features and, Again, not that the pathologist doesn't see it and the trend is, to use those that pathologists actually see for predictions because then we have the explainability covered, so to say, we do want to have this explainable, but combining them together. Into insights or extracting insights from the combination that are actionable for the patient is something that we have not been able to do without the assistance of AI or digital pathology or computational tools. computers are helping us. unique insights into immunotherapy response prediction and biomarker what do we gain? speed, accuracy, robustness of biomarker discovery, improved precision, personalization and effectiveness of clinical treatments. This is the paper setting the stage for, hey guys, embrace AI and let's leverage it for better patient care. Let me know if you have any questions. I don't see any questions in the chat. Can be related to the paper, can be unrelated, whatever comes to your mind, because there is something I want to do in the near future, which is a live Q& A. We'll see when and how often. I ambitiously am thinking about doing it once a week. Let's see about the bandwidth. But definitely once a month where I gather all your questions from LinkedIn, other social media platforms, emails, responses to emails, and in those, DigiPath Digest specifically, because this is where you tend to have the most questions. And I'm going to be taking those questions. And answering them, if I don't, manage to answer them in the particular live stream or piece of content, I will take them and I will have a Q& A session every regular time, let's say every month for now. If you want to put it in the chat, put it in the chat. If you want to send me an email or a LinkedIn message with a question about digital pathology and AI in pathology, let me know. And I will be answering those questions on a regular basis. And now back to machine learning approach to predict. HPV positivity of oropharyngeal squamous cell carcinoma. This is Pathologica group from Italy and the impact factor is 4. 4. I didn't do impact factor for the other ones. I'll do it next time. HPV status is an important prognostic factor in oropharyngeal squamous cell carcinoma. I love those abbreviations, O P S C C. Oropharyngeal squamous carcinoma, and HPV is the human papillomavirus, positive tumors are associated with better overall survival. To determine the HPV status, there is a lot you need to do. to find out. So now it's IHC immunohistochemical investigation for expression of the P16 INK4A protein. And this must be associated with molecular investigation for the presence of viral DNA. So we have two, we have IHC and molecular DNA detection. They want to do image analysis for this instead. Do machine learning image analysis, Machine learning would be a tool used for image analysis to predict HPV status from H& E stains from hematoxylin and eosin. And if you are joining a little later, let me know in the chat. Oh, we have somebody from Munich. Hello. I spent some time in Munich. I worked there three years, two years, for two years. It's beautiful there. I bet it's cold, winter now. For this HPV status, we extracted a pool of 41 morphometric and colorimetric features. Colorimetric, I was like, oh, and if you have staining variability, why are you doing colorimetric? But maybe they had a good solution to that. But This is another proof point of what I just said about the explainability situation. they didn't, just train a classifier with examples here. They, actually extracted features. So they actually defined those features. those morphometric and colorimetric features, and they were working with the TCGA, the Cancer Genome Atlas, samples and archives of pathological anatomy of Federico II of Naples. From their, Italian, Naples, Pathology Institute, they built a random forest classifier. their model showed 90 percent accuracy. there is a debate whether accuracy is a good metric or not, but we're not going to engage in this debate right now, because we don't have time for that. we would need to invite experts on, Image analysis, model performance metrics, I used to have a presentation on that, but this is a topic where you forget at the moment after you finish the presentation, unless you do it every day, but I'm digressing. They Studied the variable importance to define a criterion useful for the explainability of the model. So here we have this word explainability again. In medicine, we do want to have explainable models. features extracted, explainability, the model was accurate. They build a classifier capable of anticipating the result of p16 immunohistochemistry and molecular tests to assess the HPV status of squamous carcinomas of the oropharynx by analyzing H&E. So this would be something that I would call pathomics as well because we're anticipating or predicting a status of a molecular test just from the images. thank you so much for the hearts. If you can show me some likes and love. In form of likes that's going to take this live stream and show it to more people who might benefit from this content and learn more about digital pathology so thank you so much for the hearts and likes and if You are a uropathologist You need to prepare for a tsunami of prostate cancer. this was also Published in Pathologica, the Uropathologist of the future getting ready with intelligence for the prostate cancer tsunami. This is a group from Italy, also USA, and Spain, Ireland. And they We're talking about that we need to embrace digital pathology and AI for prostate cancer. So another paper I classified here as setting the stage for digital pathology and AI readiness. there was a paper by the Lancet Commission on prostate cancer and they project that, prostate cancer, will raise from 1. 4 million in 2020. That was already four times four years ago to 2. 9 millions in 2014. And I'm like, why would it raise so much? It, is. Connected to lifestyle, earlier detection that influences the, prevalence, And also not the prevalence, but the detection of it and also developing country being more and more developed. entering this, different lifestyle stage that, Western countries already have and have those, factors, contribute to prostate cancer. They say that late diagnosis of prostate cancer, it's widespread worldwide, but especially in low income and middle income countries. the best way to cope with the harm due to increase in case numbers is to develop systems for earlier diagnosis. But obviously systems, it's a catch 22. Systems for earlier diagnosis are going to detect more cases faster, which is what we want, but it means there is going to be this tsunami. We will need to leverage AI and digital pathology diagnostics, In this particular paper, they describe how digital pathology and AI can help pathologists for the prostate cancer tsunami about to come. So basically, we know we have prostate cancer algorithms available commercially from different companies, PAIGE AI having the prostate one being the most famous that was cleared by the FDA. this paper goes through what we are talking about, Every time we discuss paper or digital pathology it should be faster. It should give pathologists Confidence AI plus pathologists are more efficient and more accurate than AI or pathologists alone. So this tandem of AI, and pathologists, enabled by digital pathology is going to be a must if we want to provide the level of care that our patients deserve The last paper today is talking about cytology, my friends. Again, cytology. Everybody who is a cytology person, now you also have access to digital pathology. And this one actually talks about all the different scanners. if you want to know which scanners to use for digital pathology, this paper is for you because it mentions them. And this is for thyroid cytology assist So how whole slide imaging can assist in thyroid in the term in the terminate cytology This is a systematic review So it goes through all the different reviews that they found in literature on PubMed or wherever and talks about okay What's going on who used what so Thyroid cytology, particularly in cases of atypia of undetermined significance, follicular lesions of undetermined significance, and I don't know if I forgot to look for my tumor slides when I was in Poland. Shoot. Anyway, but what I wanted to tell you is that, when I had my, thyroid tumor diagnosed, I had cytology three times and it was always inconclusive. at some point I had to go to surgery, but, I experienced this undetermined significance of a cytological, evaluation the problem is suboptimal sensitivity and specificity challenges. this systematic review looks at papers between 2020 and 2023. They used different stains. They found 176 studies. 13 met the inclusion criteria and the data sets ranged from 145 to 964 whole slide images. we had 494, A, U. S. cases. Ranging so the overall number and then ranging from 8 to 254 and 5 studies used convolutional neural networks. Two used artificial neural networks, which is interesting because CNN is like a subset of ANN, but, okay. And then, okay, so we had five studies, two studies, that's seven studies. I'm getting a little confused, in those numbers, but I bet They have it well calculated, but the methods for preparation included Romanovsky stained smears, Papa Nicolaus stained, or H& E, and liquid based cytology, thin prep. let me tell you the scanners that they used. The scanners that were used in studies for cytology include, drumroll please, Leica Aperio, Alyuda Neurointelligence Cupertino, and PNNORAMIC Desk II Scanner. Let me show you the names. Here, Leica,/Aperio, Alyuda Neurointelligence Cupertino, and PNNORAMIC Desk II Scanner. then there were different classifiers as well, Feed Forward Neural Networks, Two Layer Feed Forward Neural Networks, all the different classifiers that were used, maybe in combination we just said that they had, 5 studies with CNNs and 2 studies with ANNs, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10. I am confused by the numbers in this paper, by the different combinations. what I would do now, I would take my perplexity, Maybe I can do it live. Let's see if I can. Maybe I'll give you a demo, Okay, this is our last paper, so I can do a demo. I have this perplexity installed on my phone, and I can talk to it. let's see. In the paper, Artificial Intelligence of Whole Slide Imaging Assists in Thyroid Indeterminate Cytology, a systematic review, published in Acta Cytologica, how many studies used artificial intelligence? Let's see. The paper, Artificial Intelligence of Whole Slide Imaging Assists in Thyroid Indeterminate Cytology, a, a systematic review, published in Acta Cytologica, reviewed 13 studies that utilized artificial intelligence in this context. Okay, that makes sense because we have 13 that met our inclusion criteria. What was my confusion? The data sets ranged from number ranging from eight, overall number, okay, ranging from eight. Oh, so this is per study. Let's see. How many overall cases of AUS were evaluated in this study? The study, artificial intelligence and whole slide imaging, assist in thyroid indeterminate. Cytology, a systematic review, evaluated a total of 494 cases of atypia of undetermined significance. AUS Across the included studies with individual data sets ranging from eight to 254 cases. Thank you perplexity. Now, individual dataset from eight to 254. And then, let me see, what was I confused about? five studies used convolutional neural networks and two used artificial neural networks. I see that five studies used CNNs and two studies used ANNs, artificial neural networks. What was used in the remaining studies? In the remaining studies reviewed in the paper, various machine learning models and classifiers were used. These included Feed Forward Neural Networks, or FFNN, Two Layer Feed Forward Neural Networks, 2L FFNN, Gradient Boosting Trees, or GBT, Extra Trees Classifier, ETC, Visual Geometry Group 11, VGG11, YOLOV, EfficientNet V2 L, Backprop Okay, it stopped on me. It sometimes does that but what do you think let me know in the chat What do you think about this? so I do it quite a lot for like Things like that when I get confused and I like understand part of the information and then I ask a question to clarify so there was like, you know a lot of data probably if I was reading it on my own not talking to you Not you Trying, to, ask you for comments and all these things, I will understand, but, because I was doing different things, I didn't. So I asked my AI assistant, and we have one question, what is my opinion about adoption and direction of hyperspectral imaging in digital pathology? I do have an, opinion on this. To me. This is a discovery method. How I see everything in pathology and by proxy, everything in science, there is the high effort, high cost, high information methods, which is, what I would classify the hyperspectral imaging as that give you some insights that other methods wouldn't give you. often, in the diagnostic space, you, try to generate insights. once you have this insight, you figure out how can I, Reverse engineer this to get the same insight consistently with the simpler method. So I see it as a funnel, like in drug development. You have this discovery phase where you have so many compounds and then you narrow it down to the one compound that makes it, to the pharmacies, In pathology and everywhere, you use different methods hyperspectral imaging is a lot of information, but also I consider, for example, immunofluorescence a higher in the funnel method than, for example, brightfield. it requires more effort, it can give you more information. So I hope that helps. And thank you so much for this question. I might include it in my Q& A session so that more people can get an answer to this question. If anybody has any other questions, let me know in the chat. I'm going to be monitoring them and collecting them into a separate document that I can later Use for my Q& A. For now. Thank you so much for joining and I talk to you next Friday Bye

By the way, there is a feature available only for, the audio podcast listeners. And it, uh, when you go into the audio listening app, whatever you choose, in the podcast description, there should be a link called send us a text. And you can send me a text with your question. So if you're listening to this on your phone, you can do it through the send as a text link and I'll get your question directly to my inbox. This is for audio listeners only. I did get the text after last episode, so go ahead, try this out and I'm gonna have your question in the next Q& A live stream. Talk to you in the next episode.