Educational content on VJHemOnc is intended for healthcare professionals only. By visiting this website and accessing this information you confirm that you are a healthcare professional.

The Myelodysplastic Syndromes Channel is supported with funding from Geron (Silver).

The Myeloproliferative Neoplasms Channel is supported with funding from Takeda (Gold) and Kartos Therapeutics, Inc. (Bronze).

VJHemOnc is an independent medical education platform. Supporters, including channel supporters, have no influence over the production of content. The levels of sponsorship listed are reflective of the amount of funding given to support the channel.

Share this video  

The development of AI in hematopathology diagnostics for MPNs and MDS

Sanam Loghavi, MD, The University of Texas MD Anderson Cancer Center, Houston, TX, discusses the development of artificial intelligence (AI) in hematopathology diagnostics for myeloproliferative neoplasms (MPNs) and myelodysplastic syndromes (MDS). Dr Loghavi explains that AI can improve fibrosis assessment and thrombosis risk prediction in MPNs, and support prognostication in MDS. She notes that wider adoption will require collaboration and larger datasets. This presentation took place at the 4th International Workshop on Myelodysplastic Syndromes & Myeloproliferative Neoplasms (iwMDS/iwMPN) 2026, held in Washington, DC.

These works are owned by Magdalen Medical Publishing (MMP) and are protected by copyright laws and treaties around the world. All rights are reserved.

Transcript

So I will be speaking about the development of AI in hematopathology diagnostics, and are we ready to mainline, or mainstream, AI?

So I want to put the acknowledgement slide first. These are all the people that have contributed to this team, and I’m not acknowledging myself; I’m just there as the person that’s part of this team. But essentially, it’s Ground Truth Labs, a spin-off from Oxford, and then our team at MD Anderson is really Rashmi, Kelly and myself...

So I will be speaking about the development of AI in hematopathology diagnostics, and are we ready to mainline, or mainstream, AI?

So I want to put the acknowledgement slide first. These are all the people that have contributed to this team, and I’m not acknowledging myself; I’m just there as the person that’s part of this team. But essentially, it’s Ground Truth Labs, a spin-off from Oxford, and then our team at MD Anderson is really Rashmi, Kelly and myself. And then we have collaborated with other institutions on some of the data and the studies that I’m going to show you.

So what are some of the challenges, right? What is the opportunity here in diagnostic pathology, and why do we need AI? One is that, despite what you may believe, there is actually a shortage of pathologists in this country, and I’m sure in other places in the world. So the workload is definitely higher than, you know, there are pathologists. And then the other is interpretation variability. I’m sure you guys are all familiar with difficult cases that get sent to different pathologists, and you get a different diagnosis. Even within the classification systems, sometimes we have two classification systems. So there’s definitely variability in interpretation.

And then I think most importantly, when we train as pathologists, I think partly subconsciously, but partly they teach us to focus on specific areas that we think or believe are important, and then cancel or filter out everything that we think is noise, right? We believe it is noise, but it may actually not be noise. So when I’m looking at a bone marrow for MDS, I’m really focusing on dysplasia. When I’m looking at a bone marrow for PMF, I’m really focusing on megakaryocytic features and fibrosis. And there are a lot of things that I think we tend to ignore, mostly as a coping mechanism, maybe, because that’s the only way we can do our job.

Okay, so there were really three major advances in computational pathology that have allowed the adoption of AI. And I just want to say that we’re not even close to diagnostically using AI. We don’t really have any FDA-approved platforms. But these are the advances that have been made. First of all, whole slide imaging and scanning in the 1990s. And then the computer vision breakthroughs that happened in the 2000s. And then, of course, the processing, right? GPU computing power that has allowed deep learning tissue analysis.

And then in hematopathology in particular, AI is behind solid tumor pathology. One reason is because the slides that we use, the aspirate smears, are very different from H&E-stained tissue sections. There’s a lot of variability in the cellular architecture. And if you’ve ever seen a pathologist make a diagnosis on a microscope, we tend to look at the aspirate smears at 100x. And to scan something at that resolution, to allow you to look at that magnification, needs a lot of space. So that’s still the major barrier to the adoption of AI in hematopathology.

But, of course, there’s a lot of opportunity if we’re able to adopt AI routinely. And what are some things that we can do with the adoption of digital pathology and diagnostics, which actually is happening at some places? Of course, we’re going to have more digital slides to use for training these models, better scanners, faster scanners, and then clinical trials. I think one of the most important things is leveraging well-annotated data sets to train these models. And then, of course, precision medicine demands.

So for the sake of time, I’m going to skip through this, but I want to show you that it’s taken about 26 years to get to where we are, and we’re definitely not even close to using AI in routine data.

So what are some of the opportunities? This is an MDS/MPN meeting, so I figured I’d focus on MPN. You’ll see these words in our reports, but I kind of wanted to put a face to the name. So when we talk about the staghorn nuclei of essential thrombocythemia, for Dr. Kuykendall, or primary myelofibrosis with the cloud-like nuclei, these are things that we recognize as pathologists, and I think it’s very easy for AI to pick up, right? I mean, MPN diagnosis is very much amenable to AI. It can do it reproducibly and rapidly, and I can show you data.

So look at this. On the left-hand side is without AI. We’re focusing on the exact same region. And even to the trained eye of a hematopathologist, I really can’t identify all of the megakaryocytes in that field that are being identified by the AI. So the greens are all megakaryocytes. If you look, some of those are actually difficult to identify by the human eye. So it’s much better than the human eye.

So there have been studies. This is actually by the Oxford group, where they’ve looked at using AI in quantifying fibrosis, so what they call Continuous Indexing of Fibrosis, to discriminate between ET and prefibrotic MF, and also to determine the risk of thrombosis for these patients. So there’s for sure an improved C-index when you combine the AI with the human. The inter-pathologist agreement with this tool increased from 66% to 80%, and then it identified patients that were at risk of disease progression.

The other question is, can we predict features like vascular events, fibrotic transformation, AML transformation and overall survival? And in large data sets that have been used, the short answer is yes. One of the most important things is that if you look at the revised IPSET score, and this is a study looking at several patients with ET, 36% of the patients who actually had thrombotic events down the line were missed using this risk score, and were predicted to be very low risk for having a thrombotic event. The major reason for that was that these patients were young, and so they didn’t score as high on the IPSET score. But if you use the AI model, it actually rescued 16 of 27 of those patients. So it predicted correctly that they were at high risk of having a thrombotic event.

And then you’ll say, well, how much better does AI do compared to a human pathologist? So this is a study; this is not fibrosis, this is cell identification in the bone marrow in general. And if you compare expert hematopathologists at top academic institutions with the AI model, the AI model doesn’t outperform the hematopathologist, but it does just as well. If you look at the middle line, that’s where there was agreement on every cell line. So on one axis, you have the human, and on the other, on the top one, you have the AI. And there’s really good agreement between the two.

So using a similar model, we did a collaborative study, actually with Dr. Hasserjian, sitting in the room, and with our NYU cohorts. And we looked at MDS, normal bone marrow, and MDS mimics, so patients that presented with cytopenias and did not have MDS. And then we really tried to see how the AI performs compared to humans. So there were expert hematopathologists that were blinded to the history and everything. And I have to say, we were placed at somewhat of a disadvantage because they only showed us the core biopsy, and some of the core biopsies were very bad in quality. But nevertheless, the AI had the same; it didn’t have the aspirate smear. And then we tried to see how the AI performs compared to the pathologists.

So there’s a lot that goes into the algorithm, including the cellular marrow, the bone, the periosteum and the fat. And the cohort was enriched in features identified by AI that are known to humans to be associated with MDS, like dysplasia and hypercellularity. And then what was depleted in MDS was essentially low cellularity, so that predicted a negative prediction. But then there were also things that, as I told you, we ignore as humans, but were flagged by the AI as being associated with MDS, like loss of the normal trabecular architecture or how the marrow was spatially distributed. So that was informative.

And then if you look at the performance of the model, the area under the curve for MDS versus negative controls was 0.8; MDS versus cytopenic mimics that were not MDS, 0.8; and then MDS versus all controls, 0.7. So this is not perfect. I think it still needs work. But again, the input was not perfect. It was really the model and us; we were all put at a disadvantage. But nevertheless, it was good. And then when you compare the model with the hematopathologists, the sensitivity was 68% and specificity 62%. But I think as we train these models and as we put more data into the models, I’m sure that they’re going to get better with time.

And then what about as a prognostic model in MDS? So this is a study that we’ve done in collaboration with our colleagues at Moffitt and at MD Anderson. So we’ve looked at cohorts of MDS. Some AML patients were included, just because we wanted to have the end of the spectrum, and then we had one CMML here. These were patients that were all treated on trial with HMA-based therapies. So again, similar to the last model that I showed you, the AI looks at similar features to predict risk and also classify the disease.

So if you combine the features that are identified by AI with our traditional prognostic risk scoring systems, like IPSS-R and IPSS-M, it actually outperforms them. But it’s a combination. So the model by itself doesn’t actually outperform, but the combination does. And then it also predicts response to HMA therapy, and again, the combination does it better than just the traditional prognostic scoring systems.

So in summary, I think there is some positivity. But if you talk to pathologists, there’s still some hesitation because the AI isn’t performing perfectly. But I do think that it’s going to get better and better, and as the models improve, the adoption is likely going to be higher. We are concerned that we’re going to be out of a job, but we’ll see how that plays out.

And then what do we need for the next step to adoption? I think really collaboration and large data cohorts. So if anyone here is interested in contributing cases or collaborating with us to develop these models, please come talk to us. We’re here. And I think this is a nice schematic from the MLL laboratory. They have actually adopted AI in their diagnostic algorithms and are using it, and I hear that they’re very happy with the performance. But again, just to put a plug in here that we’re very receptive to collaborations and receiving additional cases to try to improve these models, to hopefully help us become better as pathologists. That’s it.

AI-generated transcript. Please refer to the video as the authoritative source.

Read more...

Choose your cookie settings

VJHemOnc uses necessary cookies to make the site work. With your permission we'd also like to use analytics, preference and marketing cookies to understand how the site is used and improve it.

Choose an option to continue. You can change it at any time on our Cookie Policy page, linked at the bottom of every page.