Transforming Hematology: The Power of AI

You are listening to the Hematology podcast by Sanofi. The integration of artificial intelligence, AI, into routine clinical practice is poised to revolutionise the diagnosis and treatment of hematologic diseases. AI-based models are already capable of automatically identifying cell populations, including malignant cells, and facilitating early disease detection and prognosis. However, the efficacy of these tools depends on their correct application and interpretation, and hematologists need a basic understanding of machine learning to use AI effectively. In order to exploit the full potential of AI in hematologic diagnostics and to achieve a comprehensive integration of AI-based systems into routine clinical settings, there are challenges and limitations to be solved, including standardisations of the methods, regulatory guidelines, and training-related questions. In this episode, we will dive into the current status and future directions of using AI in hematology. Today's guest, Oskar Bryk, is MD and PhD and leads the Hematoscope Lab at the Helsinki University Hospital and the University of Helsinki. His team combines high-resolution imaging, big data, and machine learning to revolutionise the diagnosis and monitoring of hematological diseases. With a multidisciplinary expertise and approach, the group aims to harness technology to tackle cancer heterogeneity. This is the Hematology Podcast, and I am Mats Mirup. Hello, Oskar. Good to see you. Hello, Mas, and thanks. Thanks for joining me in this discussion. I think this is a difficult topic for many of us, but we all understand that it's probably part of the future. So I hope you can, in a very easy way, explain to me what AI actually is and how it will affect my and our every day in the future. I'll try my best then. And first, thanks for the invitation. I'm very happy to be here. So should we start out just with talking about what AI actually is? So most people are probably very familiar with the term AI. I would maybe start maybe two steps back and to define maybe what algorithms are. So these are a set of instructions or guidelines that transform some kind of input data. So you might have, for example, information on an Excel sheet or images of cells or maybe even multiple data sources. And you want to predict something. It can be from the cell type to the images. What is the cell type? Or if you have prognostic information, you want to know whether the patient, your patient, might be alive in five years or one year or so. So there are multiple methods which can be used. And AI and machine learning are one what we use. And a lot of people use very much as still statistical methods. And I will never forget this because these are really, especially for the clinicians, what they want to know. They want to know as certainly as possible what is the prediction or if you want to know, for example, if you want to know what is the probability of a patient dying after a few years, you want to be as sure as possible and not have a lot of uncertainty there. So statistical methods are actually very good at that. They are exact, they're very transparent, and therefore used a lot. But then more and more we have access to a lot of data. And there maybe these common or traditional statistical methods might have some troubles in really processing that amount of data. And that's why there's been more and more interest in machine learning. So the machine learning is like an automatic or semi-automatic method to detect some patterns in your data or associations in the data without sharing too much instruction study to them. So you might say that I want to see, for example, that from cell images, which of the cells are glass and which aren't. But you're not saying that the cells need to be larger than, say, 15 micrometers in diameter or so, you just give the images. And that's kind of a thing that you give the data and a set of very limited instructions. And you let the machine learning do its task and try to learn the patterns there. And this kind of a machine learning is very typical in the medical field. So you can say that it's a really kind of a narrow machine learning that you have one task, one type of data, and you want to get one result. Yes or no? How many percent of glass? But then when people are talking about artificial intelligence that maybe the definition has been also evolving, but maybe more and more they mean some kind of true artificial intelligence. So that kind of a broad machine learning or what also is used in the field is called AGI, or Artificial General Intelligence, where there is a system, like a computer, that is not doing one prediction task, but rather is able to reason, so use information, process of information, and do some tasks. And also like memorize and learn new tasks without that there is any human intervention. And that is really far away what currently used in medicine, or there is a lot of discussion whether that kind of AGI is going to be achieved or when it's going to be achieved. There is already some evidence that language models are also very interesting and very modern at the moment, whether these already fulfill some definitions of AGI. So they can feel aware of this Turing test where there is a human that is communicating with another human and with a computer. And by just asking questions, tries to guess which of these two is a human, which one is a computer. And like the most recent forms of these language models are already kind of fooling the human so well that we can say that it's kind of a GPT-based language models already have kind of completed this Turing test. So we need to kind of make better tests to see when we have actually achieved this AGI. But this is kind of the broad scope of statistical methods to narrow machine learning, to broad machine learning and AGI. Very good. That was a good start. I think you made me understand here why this is better than a statistical measurement. I put in the data, but it's sort of the machine that can learn that I don't have to ask a question. The machine can identify the question itself in a way. And if I understand it correctly. Yeah, that's correct. So could you say, where are we using AI today in hematology? Are we using it? Maybe not as much as maybe there's a lot of hype in the fields and maybe not as much as one could expect from outside. I think it's mostly used in the diagnostics. Many hospitals, there isn't use digital microscopes for 35 years which can pre-classify cells with the correct classes for human to validate. And this is used because it saves time and probably also can help with the accuracy as well. There is also some software for chromosome band analysis. So humans need to work with physicians and healthcare providers in general when analyzing these chromosome bands. They need to first identify each chromosome, what is the chromosome and then look if there is any sense of any alterations in chromosomes. And for that, there is some software that can detect and classify these chromosome bands quite accurately already. So I would say there are a lot of different new applications. So for example, I wouldn't say necessarily in hematology, but in clinical medicine, for example, it's part of the appointment. There will be like an AI that is supporting in breaking the report after the appointment. And then in some of the modern risk scores that is used to make the prognosis for the methodological patients as well. So we are using it already and we had a podcast earlier about the use in pathology. So that part we have discussed a little bit, but it's good to talk about it again today. If we look again how AI has been used, if we look in research, could you point on some particular advancement that we've made using AI in hematological research so far? Yeah, definitely. And maybe as a disclosure, I'm specifically personally interested in image analysis. So I did follow more that field, but there in the beginning, there's been a lot of research in detecting cells from images and classifying cells. And it started from automatic counting of white blood cells and red blood cells. And we find, especially during the last 10 years, there have been publications on classifying white blood cells correctly. Started off with some initial papers with maybe four classes just to get lymphocytes, neutrophils, monocytes and, last correctly, been advancing more and more. And related to this one kind of the very limiting factor has also been the imaging system. So there are not that many microscopes that you can use to digitize these samples. So in hematology, in comparison to pathology in general, it's not sufficient that you get like a 20x or 40x magnification. The laboratory physicians or metallologists are interested to have the 60 or 100x magnification and oil immersion, so really like a high resolution image of blood cells. And there are not that many systems that can do this, you know, like a high throughput fashion. And I think that's been a really important varonekin in kind of translating these image analysis methods to hematology. But now, luckily, there are a few new systems, and really that's been one with a Chinese morphical system. There is one with an aperio system and one with the German size imaging system. And they all were very promising in the sense that you can not only detect and classify cells, but you can get that actually with quite nice accuracy. And probably one of the most, I would say, important work in the field has been from the Munich leukemia laboratory, which has analyzed a huge data set of from over 100,000 images and almost 1,000 slides to look at this kind of a very accurate that is across different diseases. So how well you can really classify cells. But again, maybe here, I would say, very put another emitting factor is the there's not that much validation done across centers. So how these algorithms we work in different centers, or is it just slides from your hospital? There are some work also done with these plastic cells. So there's one work from Dr. Kimura using perfa blood smears, where they look at if there is any signs of this place. So there was a sample from I think it was hypoplastic MDS and plastic anemia, which is not like always easy to separate cleanly. So if you could use image analysis to not only classify the cells, but also look if there's any signs of dysplasia. So that's helping the clinical differentiation between these two white noses. So in that last case, we could sort of could the AI define other things that could differentiate that we can sort of learn from that? That is sort of a piece of my point. Yes, yes, well, in that particular paper, I feel they pre-classified this plastic pipes first, so that you could really explain how the model is working. I think that's that's a nice way, especially if you want to have a clinical impact. But probably, as I mentioned, there is there might be a lot of subtle changes in the morphology that the oncologist maybe have noticed. See that this didn't occur, but they're not really in the sense they're not in the diagnostic guidelines of dysplasia. And maybe this kind of a subtle changes could be very important, actually, make it more confident. That's a good clinical example. If I now want to learn more about AI, could you give me some ideas of notable public works that highlight the use of AI in hematology? I would say that a really good starting point is the recent review paper by, well, it was actually by the team from the ViniQ chemial laboratory in blood release journal from one or two maybe years ago. I think they did a very good job in describing not only kind of the applications of hematology, but also like really basic concepts about machine learning and how these are trained and evaluated and how they can be used in hematology. So the paper is called the Artificial Intelligencing Hematological Diagnostics Game Changer on Gatshit. I think it's a nice name. And the first author is Wanker Walter. And last authors are Claudia Thorstenhafela. And it was published in the Journal of Blood Reviews in 2022. And I really recommend to read this paper. It's really good. Good. You have a special interest for AI, of course. And how in your own research, how have you used AI and have you particular findings here using AI that you can tell us about? Yeah, I would say that the initial maybe interest came... Well, I was actually doing my PhD in a very nice environment where there were people from different backgrounds. I was really fascinated, especially from the technical applications people were working with them. And especially some colleagues which had a background in image analysis. So my kind of a first, I would say, touch with machine learning was in 2019. When we started this study, where we collected hematoxylene and then we also stained bone marrow slides from MDS of these plastic syndrome patients. And I was really interested to look at whether we could just use these images of these pathology slides and predict whether the patient had a mutation in one of the most common genes. And it was surprised at how reliably we could actually do this task. And we expanded that to all the common mutations, all the common keratin alterations. We went even to see if we could predict the age of the patient and the survival of the patient. And there was a variation in how well the models worked. You could say that the histopathology was definitely associated with the mutations and the keratin changes. And in a sense, MDS is a disease which starts from these genetic changes. And there's definitely some kind of causality with the morphology. So you have that type of mutations or alterations in those chromosomes. So it must affect somehow in a unified way the morphology and the bone marrow composition. So we published that in blood cancer discovery. And it was pretty much one of the main motivation why I want to start my own research group. So we are now kind of all in that field now. So we are not working with the pathology, but with cytomophology. So cell images from bone marrow and airflow blood aspirate samples. Interesting. So we should move now to understand how these machines learn and improve their skill. And how do they actually do it? I mean, for instance, if we take the example with the blasts, do we first tell the machine this is what a blast looks like? And but how does it continue? I mean, blasts can look very differently and it's like teaching a child. They have to stand on their feet first, but then they start running. How do they start running? Yeah, that's a very good question. And you had a nice discussion on that in the other podcast episode on digital pathology. You can broadly divide machine learning to supervised learning and unsupervised learning. And the medical field has really more interest in the supervised learning. So here you then I say this is a blast sort of. Yes, yeah. And that's that's like consuming requires the expert to be really involved there. So Pici, the blasts and all the other cell types that are relevant for the algorithms. It takes a lot of time to collect the images and to annotate those. But the other hand is that you get the really good results. It as good as your data is. So you need to have sufficient number of images or data, high quality annotations. And if you really want to make a change in the field and not only in your hospital or your project, you need also images from different scanners so they have different resolutions. And for the computer is completely different than for the human. You see as a human, you see that OK, these are the same cell. If the pixel size or how many pixels there is, for example, one cell is different. The computer is lost. But also images from multiple clinical sites because professional state to prepare the slides with a big different methods and different staining protocols. So it might be that you make a really good model, but just works for your hospital. And it's a bit of a pity if you cannot really make that more generalizable. And the other kind of a few of these unsupervised learning. But I'm not really aware that there is any clinical method using unsupervised learning. But there, the idea is that the algorithm is looking at the data itself if there is any structure there. So any groups which you can differentiate and optimally not happen. But optimally, if you use like those images and you use the information that is those images, they will separate based on their cell class. So the blast would go to one corner and lymphocytes to another. But it's not always evident from these small images, whether it's a blast or not. So they won't cluster that nicely. So you need this unsupervised learning. Unsupervised learning is just probably more in this kind of a finding outliers. So if you have like some really bad quality samples, you could like automatically identify those and remove those from your analysts. So we've touched on areas where it's AI is actually used in clinical hematology. You have talked about identifying cells, cell morphology, tissue histology, also carrier typing. Are we able today to do complete carrier typing? And where does AI add to it? And is AI better than the cytogeneticists today or how can it be better using my point? Yeah, that's a good question because if you think again from the previous discussion about the supervised learning, if the human is the one that decides what is correct or not, it's very difficult to show that the AI is better because the human is over 100% correct. So the only way to do that kind of a comparison of performance, in my opinion, is that you should look at the objectivity. So you have two humans and now suddenly the humans are not agreeing on everything. So is there two truths or is it so that in the case is what they disagree? The correct answer is something in between. And that kind of objectivity should be, in my opinion, the first metric for AI evaluation. So again, regarding the carrier type banding, so one is the objectivity. But I think there might be more benefit also in the speed. So how fast a human can detect these carrier type bands and classify them correctly. But the machine could do that maybe faster. And there is some evidence already on that day that they could do that faster. So the human can verify those results and have them making more time for a cup of coffee or so. So that speed is one, objectivity is one, and maybe the full benefit is their combination. You sort of get me to understand that at the level we are right now, it seems that AI can sort of standardize different morphology of cells from different hospitals, the interpretation of different cytogeneticists. But it doesn't really lift yet to sort of increase the knowledge far away from the human specialist at this point. Is it correct to say it like that? Yes, I think these are two different tasks. If you really want to affect clinical work, whether it's a lot of discussion on that, what should you do, what should you do? You should do the same things as the humans are doing or something, what humans are not doing at the moment. So if you want to do something to reduce the workload of humans, then you should do the things that they're doing now, and especially the ones that take time or are somehow really tedious to do. For example, calculating the number of glass, I think it's a very typical example. But then again, if you want to go beyond the current spectrum of what physicians are working with at the moment, then you need to ask different questions. So it could be that, can you predict, can it be that you have some risk scores that use genetics and laboratory values? But if you add morphology there or AI in this hope that you look at how, for example, the lab values have been progressing with time, you get different results. And you get maybe some assistance to the clinical treatment decisions that current prognosis scores are not needed. So it really depends on what you want to do. You cannot solve everything with one model, so you have one model, maybe for the simple tasks and then all the current tasks that take a lot of time for the humans, and then the other models are doing these things that go beyond. We, hematologists, we love to classify diseases, and I guess that's the way how we have become better. We understand which diseases or which patients belong to the same group that we give a certain disease name. And that has changed during the time. Can AI, you sort of touched on that, could AI help us here? Might be. I say might be because I'm not aware of any example where AI has done the kind of disease etiological classification. So these are typically based on expert opinions. So I think this is kind of probably the top level of truth you can get. So an opinion from multiple experts in the field. Well, especially now, lately, when there's been two kind of teams of expert opinions, well, you know, the WHO and ICC. So there might be that AI could be now the first, hopefully not the third, but maybe that could be one way to maybe unify this. So if there is some disagreement between experts, so maybe solve like this with AI, if you cannot do that with discussion. But for the risk scores, so this has to be different than again. So there is evidence that AI is or has been used to make scoring models. And I would say, especially for MDS, there is now from a few years ago in the England Journal of Anderson. There was a publication on this IP SSM where the authors combined first a huge amount of patients and data from from these patients. As the information was the gene mutations and the previous information was the carrier type and laboratory values. And really try to make a model that is that recombined these all together. I think they really made up the great job there. There is a nice portal how to use that you don't need any expertise in AI to use this this model. I think that's that's a really nice use case on how to do AI for the clinical or for the clinical issues. There are everyday issues. That's interesting. Could AI help me to identify my patients with a really poor prognosis and maybe suggest a better treatment for that patient? Could that be possible with AI? If you ask it like that, I would say yes, but we have very good prognostics scores at the moment. And maybe that's kind of one issue of this time that we have in CML we have the ELTS score, which is based on prognosis. We have in AML we have the ELN scoring use, which is based on prognosis. And also I mentioned this IP SSM and the previous IP SSR in MBS, they're all based on prognosis and they don't really touch upon the treatment sensitivity. So you might have in certain diseases an option to decide between two different treatments. And currently this course, look at if you have a poor risk, you should be treated more intensively and you kind of add on that. But they don't look that whether OK, you have this type of mutations and this type of mutation or this type of morphology in the SAS could actually be associated with a higher sensitivity to a drug, even though it would be high risk or low risk. And maybe that's something that could be interesting to study with these computational methods. I think we haven't done too much progress on that in the morphology yet. OK, one area where we have been doing a lot of research during the year, that is how if we can predict the sensitivity to cytostatics and drugs for patients, individual leukemic cells or some other malignant cells. Could AI assist us here in determining the sensitivity of treatments in hematology? Yes, there are some ideas here. So I could maybe advertise the bit work that has been done in Helsinki and a bit of similar work has been done in Vienna and Austria and in Oregon in the USA. So there have been a couple of research teams looking at how cells. So you take a sample from the patient and you treat that sample with a lot of different drugs and all these drugs are in a separate cell culture. And I mentioned that there are some multiple teams that have been looking at that and how these cells respond and how that is associated with the genetics of the patient and also the true treatment response of the patient. And feed one of the best examples from that is for venetoklax, so this cytotoxic drug that affect the apoptosis pathway and has been approved now in CLL and AML. And based on those results, there were a team in Helsinki, whereby megalosis called Nika Kontro, who's been running a trial, a perfected trial to use that kind of a differentiation. So how well the cells respond, ex vivo in cell culture. So based on that, I'll locate patients to receive, at the site that implies venetoklax treatment or not. There's no AI there in itself, but maybe the AI there is a very similar method has been used in Vienna and in Switzerland by one of the biases, but in Snyder, looking at the how cells or the morphology of the cells is associated with the treatment response or and there they use these kind of mother image analysis methods. Employing AI and then from a clinical point of view, there is also a study by us in Nature Genetics by Boris Gerstun, maybe eight years ago, where they looked at data from clinical trials of AML patients and really tried to make this kind of a knowledge banks they call that have mutations and what is the association with the clinical phenotype. So the outcome is some whether the patient will have a rich remission or have a relapse and that I think is kind of a really valuable for the clinical practice as well, because they really are based on clinical data and have a really easy to interpret as well. You have really mentioned that we can't quite far, but I still get the feeling that we have a long way to go. What are the main challenges and obstacles in adopting AI software and hematology? Yeah, that's one of my favorite points to discuss. So so again, as a disclosure, so we have developed image analysis algorithms to analyze the cell images and now also develop the software for clinical use. And we have really, I would say, have three challenges in kind of make this clinical translation possible because there is a lot of regulation, which is a good thing to have regulations, but they're really demanding to really make your innovation into a medical device. So, especially in the European Union, we have regulations that are quite strict for the medical devices. We have the MBR, so medical device regulation and IVDR, so the in vitro device regulation, which are quite similar, but a bit for different medical devices and then data privacy regulations are also quite strict. The purpose is always, I mean, it's obvious, it's very good that we have these regulations. But the kind of amount of work you need to first document everything takes a lot of time and resources. So to give an example, so first, you need to have quality management system. So an entity, for example, a company has a quality management system means that they adhere to certain protocols on, for example, that you have documented from how completely as accurate as possible, how your device works, you need to have a performance evaluation of your device, which adheres to certain guidelines. When it is news, you need to have post-market surveillance and all of these are very good to have. It would probably anyways, you would like to develop these, but be just surprising the amount of time you need to develop these so that they meet the regulatory requirements. Another thing is the data privacy, which again, is really important. I see that from a finish point of view, where we have another regulation that dictates quite much on where you can build this type of model, this AI models. And it's a very expensive to build and to maintain and any development requires also a lot of money from the organizations, for example, for the hospitals. And also, if you want to do some machine learning there, it becomes quite expensive. And that's maybe the kind of the biggest issue there, that when you want data privacy, the cost is that it's actually so expensive to develop. Sometimes this is that models that are useful to patients that they probably would just reduce that that of development in these countries. And I'm a bit worried about that kind of a tendency, if that will actually happen. But it seems that there would be a more and more understanding or more dialogue between these kind of researchers and developers and then the regulators to understand the issues from those. And so you're sort of saying that the regulation of AI is done in different, there's no particular authority that regulates this in Europe or is it the national level or how does it work? Yeah, so for the medical devices, there is they usually there's a national or like a medical device authority, and then there are independent companies that audit the medical devices themselves. If you have a medical device that is classified as high risk, meaning that it is somehow related to the diagnostics, or if you want to do any treatment response predictions, or they are all classified as high risk. The low risk are them if you, for example, classify images of cells, these are considered low risk. If the device requires that the human is always part of the assessment, so only the high risk devices require like an external audit. And the audit thing is again, it's super expensive, especially for smaller companies. And it might be that it will drive more and more that the bigger companies have a bigger market share. But then for the AI, there is a separate AI act, which is really recent, and they regulate, again, just the high risk medical devices. And there the demands is that you have, first of all, again, the quality management system, but also that you described the training data, how you have evaluated your device. Is there any possible bias in the patient selection? Can you share some information on the confidence of your predictions? And again, these are very good points, but this also means that you should have all of these ready before you can even use the medical device. And sometimes it might be more useful that you get some user feedback before having the kind of a final version ready. So it might just delay a bit also the development of these medical devices. What AI applications are currently approved for use in hematology? Again, I'm a bit biased, but I think it's related to the cell classification. There is one from Cellavision Cysemex, one from Midrake. Probably there are others well that are used for the peripheral blood smears. So how to automatically or semi-automatically classify cells? I think this is the most commonly used. So we are approaching the end here and we should look ahead. I think it seems to me talking to you that it's a fantastic tool, but it seems that we're still at the beginning. And where do you think this will take us in the future? Will I still have my job or will I be replaced by AI in the future? So the second question is easier because I again refer a bit to the regulations. So the kind of a guide now, the company is to make devices to where the humans are always necessary. So I wouldn't be afraid of AI reflux in anybody's work. It will definitely make or hopefully will make the work, the tedious work more interesting again and maybe more efficient. I really hope it will be more objective. So especially those tasks where there might be multiple opinions. Hopefully, again, as you mentioned, there's been a lot of discussion on AI recent years. And hope we can really meet this height and have nice devices in applicable work. But that remains to be seen, hopefully in the recent years. So you think I will keep my job, but where apart from what we've said so far, do you think AI can advance in the future? I mean, to me, what you've talked about is a lot of pathology. Some research, could it come closer to me in the clinic with a patient? Could it be applied more clearly in the future, do you think? Do you mean dreaming the appointment room or more in the kind of a development? Well, in the relationship with the patient, the treatment, the diagnosis, the procedure, maybe incorporating all the data or something. Yeah, I think that there might be. So there is no limit in that sense that, I mean, your imagination is the limit in it. I really hope that it's kind of the, for example, as a hematologist, you use a lot of the computer as well. And maybe if we can reduce that time and use that time of the patient, I think everybody would be super happy finding new drugs, definitely. So I guess there's a lot of data from research groups and the pharma companies that they have already collected from previously. But maybe these can be re-analyzed and find some targets, better targets with fewer side effects in terms of scoring and diagnosis. There's a lot of sequencing data, and we can classify a lot of the variants correctly at the moment. But there is still a significant proportion of variants that are called variants of unknown significance. And there's a lot of interest at the moment. Can we somehow classify those better? And for example, my team is working also on the screening. So it's if a lot of different solid tumors have some kind of a screening program available. But hematology, there is no screen available. And we just try to find whether there could be some changes in the blood few years before blood cancers emerge and could be somehow applied. So first, can we do that reliably? I think the next thing that would open a completely new field. So not only can we do that reliably, but also can we find the optimal treatments if that is effectively okay? And then if that is, can we delay or even stop the pathogenesis of some blood cancer? I think that would be like a wonderful advancement as well. Okay, and now if we get AI to help us with all these routine monotonous work looking at cells, will we lose our skills then? I don't think so. It's difficult to say, but how the kind of overview of your work as a hematologist or work as a morphologist or hematologist will change. But I doubt that it will at least, I mean, the field itself will have the same, at least same level, probably better even if we can get more information. And again, the medical devices will not replace the physicians. So I guess it will just change their kind of work a bit, but the same information is needed to make the diagnosis. The same information is needed to make the correct treatment decisions to really confirm the AI's predictions and so on. So I don't think that will be a realistic issue. At least, I don't know how far in the future we can see, but at least for the next 5, 10 years, I would say. It seems a bit impossible. Well, that's relieving. Thank you, Oskar. I think this has been a very good discussion. You have sort of made me understand where AI is today, and we have touched upon also the complicated aspects, the regulatory aspects and maybe a little bit about what we can expect in the future. I think it's been very interesting. It's quite a difficult subject, but I feel I've picked up quite a lot here. So thank you so much, Oskar. I appreciate your knowledge. Thanks, Mats. I really enjoyed the discussion. Thank you. You have just listened to the Hematology podcast by Sanofi.