Digital Pathology Podcast

113: DigiPath Digest #16 | PathVisions Recap, Ai in Breast Cancer Diagnostics and Global Health

Aleksandra Zuraw Episode 113

Send us a text

Welcome back to the DigiPath Digest, fresh from PathVision! 

In this episode we will dive into the latest updates from the PathVision conference, covering trends in AI-driven diagnostics, the expansion of digital pathology into primary care, and the exciting new frontier of glassless pathology.

Join me as I recap the highlights of PathVision and the latest updates from the digital pathology literature, including discussions on:

  • AI Integration in Pathology: Learn how AI is advancing breast cancer diagnostics with tools like Ki-67 scoring models and multi-label AI for mammography, aimed at reducing unnecessary biopsies.


  • Global Health & Digital Microscopy: Hear about innovative projects from Sweden and Finland focused on AI-supported digital microscopy in primary healthcare labs, bringing accessible diagnostics to underserved areas.


  • Glassless Pathology with MUSE: Discover how glassless pathology is changing tissue imaging with MUSE (Microscopy with UV Surface Excitation), enabling diagnostics without the need for traditional glass slides. Dr. Zuraw breaks down what this means for future pathology workflows.

Plus, a shout-out to the vendors and partners making these advancements possible, and insights from Dr. Zuraw’s conversations with digital pathology trailblazers from around the globe, including new developments from Asia in digital pathology education and technology.

Timestamps:

  • [0:00] PathVision Highlights & Global Attendees
  • [5:15] AI in Diagnostic Workflows: Dr. Anil Parwani’s “Pathology Train Ride”
  • [12:30] Moving Beyond Narrow AI: Multimodal and Foundational Models
  • [18:45] Glassless Pathology: A New Frontier with MUSE Microscopy
  • [25:10] Integrating Digital Microscopy in Global Health Labs
  • [32:00] Breast Cancer Month: New Advances in AI for Diagnostics
  • [42:00] One Health & AI for Disease Detection in Primary Care
  • [48:30] Special Interviews: Jun Fukuoka and Asian Society of Digital Pathology

Links and Resources:

Publications Discussed Today:

📝
AI-Supported Digital Microscopy Diagnostics in Primary Health Care Laboratories: Protocol for a Scoping Review
🔗https://pubmed.ncbi.nlm.nih.gov/39486020/

📝
Ki-67 evaluation using deep-learning model-assisted digital image analysis in breast cancer
🔗https://pubmed.ncbi.nlm.nih.gov/39478421/

📝
A Multi-label Artificial Intelligence Approach for Improving Breast Cancer Detection With Mammographic Image Analysis
🔗https://pubmed.ncbi.nlm.nih.gov/39477432/

📝
A comprehensive evaluation of an artificial intelligence based digital pathology to monitor large-scale deworming programs against soil-transmitted helminths: A study protocol
🔗 https://pubmed.ncbi.nlm.nih.gov/39466830/


If you enjoyed this episode, please subscribe and leave a review to

Support the show

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

Aleksandra:

We are back with DigiPath Digest fresh from PathVision. In this episode, you're going to get some PathVision updates. Then we're going to talk about advances in AI for breast cancer diagnostics and how digital pathology is making its way outside of pathology departments into primary care and global health initiatives. So let's dive into it.

Intro:

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, my digital pathology trailblazers. let me know that you're here. I see you are coming in. when you are here, let me know where you are tuning in from. I have coffee ready and I have a few updates for you. Let's start with Path Visions. So Path Visions was this week Okay, we have people from Poland And Lebanon. Amazing. Welcome. International trailblazers. Path Visions was in Orlando. What's Path Visions? Path Visions is the annual conference of the Digital Pathology Association, the DPA. I was there and this was amazing. I was there to support my sponsors, two main sponsors this time. That was Epredia and Muse. because of them, I went there and benefited from the whole conference and met many amazing people. Let's talk about the trends I've noticed there. There are three main trends and one we have been seeing in our DigiPath digest already. I've been talking about it, I think, last week or a couple of weeks ago. The first trend is Integration of now a I into the diagnostic workflows. So, the digital pathology journey is long and there was a talk by Dr. Anil Parwani who compared it to a train ride. And it has many stations, It usually starts with okay. What scanner are you going to choose? it goes all the way to AI and diagnostic AI, and I see some more people joining. So feel free and let me know where you are tuning in from. we have from scanner to AI. It's a whole journey and there are different stops on this digital pathology adoption train. for those already on this journey the stop where they have to figure out how to do it, and then. Keep going is to integrate the applications for diagnostics. that will be, an app to, count Ki 67. there are different apps available for diagnostic scoring. That's what's happening now. And we've been seeing that in the literature where people are publishing, okay, we use this algorithm to quantify PD L1 to score colorectal cancer, to count Ki 67, and it shows up across different journals. People are doing this work, figuring out, okay, is it as good as we are? they're figuring out, okay, is it as good as a pathologist? Does it correlate? can we use it? this is the hurdle that the people who have been on this journey for quite some time need to overcome right now. the second trend that I've seen was Integration of less narrow AI. I don't call it general AI I because to me when you talk about artificial intelligence in a more conceptual way, general AI, I would be as good as a human. And there is nothing like that out there. But there is. Less narrow AI that can do text and images. So obviously foundation models we've been talking about and the chatbots, meaning the large language models and the integration of multimodalities. So I see the integration of this into slide viewing platform. So there are different partnerships. For example, path presenter, which is a slide viewing and image management software partnered with Paige.AI not only deploy the narrow algorithms that Paige AI has developed, but also to integrate the new wave off AI into where pathologists already diagnosing into where pathologists already looking at images because the which is fantastic, because The previous trend, the integration of narrow AI, has the problem of Being able to access different apps from different vendors, but not in the same place where you're viewing the slides. It may seem like a first world problem, but if you're going fast through your cases and every second counts, you don't want to be clicking between windows now the vendors are figuring out, how do we partner? How do we make it seamless for the pathologist so they can deliver their diagnosis faster for the patient? there is a third trend, glassless pathology. They've been listening to me, for the last. month and a half, the digital pathology, place sponsor Muse is in that category, glassless pathology. They have a device that images direct, from fresh tissue, There's no need to do a formalin fixation paraffin embedding. We don't need to do, the normal glass slide workflow. We can just image tissue, The, and Muse stands for, microscopy with UV surfaced excitation. Mm. So you ex excite the surface, you image it, and you have an image that is diagnostically as informative as an h and e image, and it actually looks very much like an h and e image. that is the new wave. I was talking to Giovanni Lujan and we called it the New Frontier. 10 years ago. Ah, by the way, Digital Pathology Association is celebrating the 15th anniversary they had special cookies for that. I made a short so check it out on linkedin the next frontier is integration of those new developments in AI into where pathologists are actually looking at slides plus glassless pathology. So I'm super excited about that. It's the 15th anniversary of Digital Pathology Association, right? some of the people who are there, and not only vendors, also the participants were there every single time. it was my second time. I remember the first time I came with a poster about digital pathology and drug development. People had no idea what to do with it, the audience that comes there is mostly, MD pathologists and digital pathology professionals from the clinical space, A little less from drug development. So people didn't really know like how do you use it in drug development? Do you use it for clinical trials? They had no idea that there is something like pre clinical work for drug development my last conference was in 2019 fast forward five years people are a lot more informed about the cross, interactions of different disciplines where digital pathology can redeploy it. shout out to the digital pathology crowd for advancing so rapidly. And another thing that was interesting, there's a lot more collaboration with radiologists. So, there was a talk, under the umbrella of Society for Imaging Informatics in Medicine. SIIM. most of the members of SIIM are radiologists, but now they're inviting more and more pathologists which is fantastic. While I am talking about the cross population of different disciplines in digital pathology, let us go to our papers today. First paper is, um, AI supported digital microscopy diagnostics in primary healthcare laboratories. I want to highlight primary healthcare laboratories. if you have any questions regarding the papers or anything, put them in the chat because then I can answer them. I love answering your questions. primary care, primary health collaboratories. this is a protocol. not a full publication. I was reading this abstract and I'm like, where are the results? it's a protocol. So there are no results yet. I'm excited because it's a primary healthcare setting. Basically digital microscopy It was called digital microscopy because it's actually microscopy you can use not only for pathology. pathology is the science of disease any disease for me is pathology, but in the more classical way, it's more like cancer diagnostics, primary health care laboratories. I like it because it's a one off health approach where, so One Health is where you combine human medicine, veterinary medicine, and environmental sciences to improve health care. I am a veterinary health professional, so that's my part of One Health, I gave a talk one time in Africa about that. these are, research protocols, impact factor 1. 4, basically telling us what study they're going to be doing. This is a Department of Global Public Health in Sweden, and we have Institute for Molecular Medicine in Finland. they're combining digital microscopy and AI, and it is being implemented in advanced laboratory settings. Yes, it's pretty advanced, pretty specialized. it could be especially advantageous in primary healthcare settings, You could use it for different things, looking at, fresh tissue in the future, looking at blood. this would improve access to diagnostics via automation and lead to a decreased need for experts on site. Yes, decreased need for experts on site. There are not enough experts to have them every time on site. this is a scoping review. a specific type of publication where You map, different things, how to do it. they mapped peer reviewed studies on AI supported digital microscopy in primary health care laboratories. and the method was a systematic search, databases PubMed, Web of Science, MBASE, and, IEEE. This publication is now this month last month probably October November November 1st It was published. So I am actually surprised when I look at these New publications, there are no AI literature, research tools implemented yet. um, I can make a video on them, but it's still, PubMed, Web of Science, and MBASE,, and there are different ones, right? But no AI tools, just the databases. They will get there. they only use peer reviewed articles in English. No limit of publication year, will be applied. studies that have applied AI supported digital microscopy with the aim of achieving a diagnosis. So this is important. It's going to be for diagnosis, on the subject level. These studies must have been performed at primary health care laboratories, and the criteria was they did not have a pathologist on site, and they were using simple sample preparation. simple sample preparation is going to be cytology, smears, blood smears. It can be fecal smears. there is a super cool publication that also talks about. One health that we're going to discuss in a second. And the question is, will there be a recording available? Yes, the recordings are going to be on YouTube and whichever platform you're watching it on. But also if you're on my mailing list. you're going to get the recording to your inbox. So if you're not yet, then go to digitalpathologyplace.com and download the book. then you're on my list and you're going to get the recording every time we have this session. they didn't have a pathologist on site. that was the requirement. I am very curious, what they're gonna come up with. So, they anticipate to identify the diseases where these novel methods could be advantageous. these were the simple sample preparation methods, smears, cytology, but there is going to be the new wave of imaging of tissue, and this is going to be the direct to digital from tissue. This is coming. So whenever it's out, I want to be the first to let you know. digital microscopy in primary health care settings. last month was breast cancer awareness month. So we do have a lot of breast cancer papers. from October. In histopathology, impact factor 3. 9, and what are we talking About implementing NarrowAI again. So, Ki 67 evaluation using deep learning model assisted digital image analysis in breast cancer. This is a publication from Japan, What did they do? They checked if Ki 67 by AI is good enough. And they said yes. they checked the efficacy of AI assisted Ki 67 digital image analysis in invasive breast carcinoma. And I'm going to tell you why I want to, talk about it, in one second. they, did quantitative assessment of AI model performance. Standard thing, you have a model, you check if it's good enough, and then, you decide what to do with it, and we have seen that most of those papers says, Oh, it was good enough, but we do, are not ready to do anything with it. for this particular study, we had 494 cases of Ki 67 slide images from invasive, breast carcinoma, and needle biopsies. it wasn't resection, Like the initial diagnostic settings where they take a biopsy from your tumor. they had two steps. they developed an, image analysis algorithm, deep learning model from scratch. And they did it using Halo. uh, Halo involvement here is The reason why I wanted to, bring in this paper because it's something that is commercially available. You can get this, software and develop this algorithm. On your own as well. you don't have to have any fancy integration. You don't have to buy the algorithm. You can just develop it on your samples. You're gonna be sure you do your own validation. You're gonna, sure, you're gonna be sure it's gonna work on your lab because you did it, your team did it. And that is the advantage of doing it with commercially available software. there are, disadvantages. and advantages to everything If you want to do like a homegrown on your own There are platforms that are going to help you do it and halo is one of them they divided it in two steps first tissue classifier so they found the tissue with the classifier and then Nuclear detection model because for Ki 67 evaluation, you need to have it in the tumor detect all nuclei because you need to count the positive for Ki 67, which is a proliferation marker and the negative, and then calculate an index. they had almost 32, 000 annotations, which is annotating just. Single cells, in 300 Ki 67 digital slide images, and it was compared with the annotations of ground truth at the pixel level. they probably did semantic segmentation rather than object detection. it was performed on 194 luminal type cases, and the correlation with manual counting and clinical outcome were investigated to confirm the accuracy and prognostic potential of this image analysis, They liked it. The performance was excellent. the numbers were, precision 0. 85, recall 0. 88 almost, and F1 score 0. 86 high Ki 67 index cutoff, 20%, showed significantly worse recurrence free survival and breast cancer specific survival. above this cutoff, the survival was worse. I like it. this plays into our trend of validation and using narrow AI, we have a publication from radiology, about a multi label artificial intelligence approach for improving breast cancer detection with mammographic images. What is multi label? Usually when you do classical supervised deep learning, you assign just one label to a specific class, but here they assign multiple labels and that help them actually. This is from In Vivo Impact Factor 1. 8 and the group is from Republic of Korea and these are radiation oncology researchers and also AI researchers. so what did they do? In this study, they aimed to develop a deep learning based artificial intelligence model that predicts the malignancy of mammographic lesions and reduces unnecessary biopsies in patients with breast cancer. how can you be certain? That this is good enough in terms of regulatory, does it need to be an FDA, cleared product if you actually want to use it, as such screening tool, regardless, of the implementation, let's look at the result of the research their deep learning based AI, was used to predict whether lesions in mammographic images are malignant. So if it would be, not multi label, but one label, you would have the mammography image, and then the label malignant or benign. that would be, the pairing of labels and the training of the model. But here they had this image. multi label training similar to the diagnostic process of a radiologist. a radiologist does not just say your cancer is malignant. That's it. No, they like, look how big it is, where it is. Is it in the lymph nodes? And I see more people joining. So let me know where you're tuning in from. Say hi in the chat and let me know. a doctor is not just going to say, malignant, benign, goodbye. They will give you a full report. they assign it multiple labels, different types and different things. let's look at the labels used in this particular, research setting. they used curated breast imaging subset of digital database for screening mammography. And, the data sets includes annotations for mass lesions, and the multi label classification approach enabled the model to recognize the malignancy, so they have the information whether it's malignant or benign, and lesion attributes. The multi label classification model trained on both lesion shape and margin. here are the lesions attributes they used and it was compared with models trained solely on malignancy. we added this additional information. what is the shape of the lesion? What is the margin of the lesion? And the information about malignancy. we had a better predictor. by considering margin and shape, the model assigned higher importance to border areas and analyzed pixels more uniformly when classifying malignant lesions. this approach showed improved diagnostic accuracy, particularly in challenging cases. The challenging cases, such as for every disease, there is a specific diagnostic system or diagnostic guidelines. the American College of Radiology, breast imaging, reporting and data system categories three and four are the ones that are challenging The breast density exceeds 50 percent there and the conclusion is potential of AI is improving the diagnosis of breast cancer I always laugh when that's the conclusion of a paper, the potential, congratulations to everybody who is waking up or staying up late. What I wanted to say about people waking up at 6 a. m. and me hosting this at 6 a. m. one day, I think it was the third day of PathVisions, and I meet a friend, whom I already talked to several times, and she looks at me and like, You look tired. I'm like, okay, I guess I am tired. I looked at myself in this camera with you and I look tired, but that's okay we show up tired and we do it anyway by the way, have you noticed that my glasses don't glare anymore? You don't see everything i'm looking at on my screen because I got specific, glare free What are they called? Anti glare, lenses specifically for streaming live so that you don't have to look at everything that I'm looking. the next step, I need to check the metaglasses with camera. A friend, Nick Best from Pathology News had them on, Pathology Visions. And I didn't believe it because I saw it on a billboard and I'm like, no way they already launched and people can buy it. Yeah, you can buy it and they have cameras and virtual reality is the next thing coming up in digital pathology, let's go back to our papers. if you have any questions or just joined, let me know where you're joining from and your question. This is another paper and it has the one health component, but actually it's a Study Protocol so, they leave me excited, but then I'm like, where are the results? Well, there are no results because they didn't do the study yet, what they're going to do, ties into the first publication that we discussed, the one about, using digital microscopy in primary health here, they want comprehensive evaluation of an artificial intelligence based digital pathology to monitor large scale deworming programs against soil transmitted helminths. This is a study protocol published in PLOS ONE, I like PLOS ONE because it's an open source, journal and the papers are published under Creative Common License, so, You can just download them, when I wanted to publish at some point, under a creative common license or open source science, I realized that you actually have to pay to publish it. you probably have to pay in every journal, but that was my first attempt in publishing something. I'm writing something and you charge me to publish it. Yes, they do. I was naive at that time. That was maybe 15 years ago. what about those deworming programs? This is, the, the group, public health from Belgium. We have group from Sweden, and, we have, Groups from Ethiopia and Uganda. I visited Ethiopia and that was where I gave About one health so super cool to see this paper. what are they gonna do? They are gonna Evaluate something that was evaluated manually or visually through the microscope with AI manual screenings of Kato-Katz thick stool smear and Remains the current standard to monitor the impact of large scale deworming programs against soil transmitted helminths. So, what is this? This is a way of preparing stool samples to see the parasite eggs. If you can see those eggs with your microscope then you can develop AI to see them as well. they designed an artificial intelligence digital pathology system, AIDP, for Digital image capture and analysis of those smears. I'm interested with how, how are they capturing the images? Probably not going to be scanned. It's probably going to be capture of digital images with a camera. This is the way that this is being done in developing countries we recently looked at the publication from Pakistan. once this, is out, because this is a protocol, we're going to learn that. they want a comprehensive evaluation of this technology as a cost efficient and end to end diagnostic to inform The soil transmitted helmet control programs against the target product profiles of the, um, WHO. This is the next step of their validation. what are these target product profiles when you have, something to, use against a disease, a medicine or something, it has a profile, what it works for. And this is the target product profiles. you're going to be checking, does it actually work for the parasites that it's supposed to work? our methods are the comprehensive evaluation based on diagnostic performance, repeatability and reproducibility, time to result, super crucial time to result, cost efficiency, and to inform large scale deworming programs, and usability in both lab and field settings. Field settings, in this particular case, this is super important because maybe not everybody has access to a lab. It needs to be doable in the field. I'm super curious. Maybe they're going to figure out something like how to do it on the smartphone or something. those researchers are creative. they will conduct, in two soil transmitted health helmet endemic countries with national deworming programs. that's going to be Ethiopia and Uganda in school aged children only. this is an official clinical trial. Trial registration was registered in September 29. 2023 a year ago. A year and a month. Clinicaltrials. gov. I am super excited about seeing the results of this particular trial. These are our publications for today. I'm going to give you a little bit more update coming out on the content side from Path Visions. I wanted to give the shout out to the vendors who go there it's intense. it's like a boot camp because you come in and set up the booth. those people have been there for 15 years joining conferences and trying to convince people to go digital with whichever product, right? It doesn't matter where, you enter on this train ride of digitization. Maybe you're gonna, you're gonna enter in the future where the current, the previous stations are like going to be abandoned already. Wherever you're going to enter it, there is going to be somebody enabling you with a commercially available device. these are the vendors. shout out to them. I had the honor to interview, Jun Fukuoka, president of the Asian Society of Digital Pathology. This is a recently, founded, society, and, he gave me, And, overview why we need a specific Asian one, how diverse the Asian society is even showed me a super cool AI tools. So I was interviewing him. He's Japanese. He speaks great English, perfect English. but he had this AI tool on his tablet that was Transcribing me live and then translating on the other, side. he Had this, and he says, This is a mandatory tool for all the members when we meet on our society meetings to use, because there's no way you're going to have the same level of English and one language, to communicate at the same level. So we need to, leverage AI and the translation to do it. I talked to him and it's amazing what's happening in Asia and how diverse this is. you will have, the adoption of digital pathology, from 50 percent in the Arab peninsula countries to less than 1 percent or 0 percent in some countries. for comparison, in the U. S. people say between 5 to 10%. I do not have a publication to cite that actually counted this, but this is more or less the number they say less than 10 percent in the U. S. where a lot of development is happening, then I also met with, two other, educators in the field of digital pathology. one was Nick Best from Pathology News. he approaches digital pathology from, okay, what tools are there? Can you compare them? what is the best informed decision you can reach on your own before you actually reach out to your preferred vendors? They have the same. special technology buyer's guide where you can compare different solutions I talked with him and Imogen Fitt. She is from Signify Research and she has this monthly recap. Oh, what's up in digital pathology space? Who partnered with him from the business perspective? What's going on? What are the trends? So we had the podcast together. Then there's another podcast coming out with, Allae Kwam, she is in the content creation committee of the digital pathology association they had a Podcast set up there as well. I was a guest so stay tuned for that There's going to be a vlog on YouTube. There are shorts, uh, on different platforms on YouTube as well. And by the way, an interesting thing I've noticed, there's a lot of trailblazers coming back to the YouTube channel, watching videos over and over again. Like 70 percent of you are coming back, but are not subscribed. So please hit the subscribe button. It's free. And if you're coming back over and over again, then just do it. Hit the subscribe button. I would love that. It's not only going to me, um, it's not only going to help you stay on top of the things, but it will give a signal to the algorithms to show it to more people who might be interested in digital pathology. And this is the point of this podcast. So thank you for that. So thank you so much for subscribing and I talk to you in the next episode.