So yes, it's really my great pleasure to be here and very, very excited to be on my first trip outside Europe again. Welcome also to the people who are watching this on the live stream. And yes, I'd like to update you on the latest WHO classification for glioma, which was really just recently published. These are my disclosures, which were also shown in the first slide. So this has basically been published in the past few months. So we have a summary from this summer. It's quite a long summary, I have to say. And really just a few weeks ago, the full new classification, the fifth edition of the WHO classification was published online ahead of print. It's probably going to come out in print in March next year. And you can find the URL here. It is unfortunately paid, but it's only 100 euros for the full set of WHO classifications. I'm not involved in this, so I'm not selling this. So what's new in the WHO classification now? So what's very obvious here is that there's going to be now, there's a very clear distinction between adult type and pediatric type diffuse glioma, really acknowledging the fact that these are very, very different tumors. There are 14 new tumor types. This is not just glioma, also the other CNS tumors of which the glioma ones are all the pediatric ones that are new. They've really made this more consistent with other WHO classifications. And there is an even more prominent role of molecular diagnostics. But let's roll back, because that molecular diagnostics, we're talking about that now, but for those of you who are not really familiar with this, so 2016 was the big revolution in the WHO classification for CNS tumors. Because that was the first time that these molecular parameters were added to the classification. And why was that? Well, I think the best way to show it is these survival curves. So if you're looking at the classification based on histopathology, you can see that there's a very clear distinction between outcome in oligodendroglioma, astrocytoma, and glioblastoma. But, sorry, I am going to take my mask off, because I'm finding it difficult to breathe and talk. Sorry, I'll breathe that way. Okay. And you can see that the molecular classification actually gives you a much clearer division or distinction between these groups in terms of survival. So this gives you a much better, I mean, distinction of these tumor types clearly. And this is based on, this is now the 2016, so not the current classification, based on basically two parameters, the 1p19q codilesion and the IDH mutation. So the 1p19q codilesion tumors we know, of course, as the oligodendroglioma with a typical appearance of this very heterogeneous mass, multiple tiny cysts that you can see very indistinct borders, frontal location, and, of course, on CT, the calcifications. Now, in 2021, this has changed a bit, because what's happened on the molecular side is that we now have a clear distinction. This is just the adults. I'm not talking about the pediatric adult patients. We now have a clear distinction between tumors that are IDH wild type and the IDH mutated tumors, which includes this 1p19q codilision tumor, so the oligodendrogliomas. So if we're looking at what else is new, and I'm going to go into these molecular features in a minute, but first the easy parts. So this is what it used to be, histological grading. It used to be with Roman numerals, and now we use Arabic numerals, which is in line with other WHO classifications. So if you want to show that you're up to date, do this. That's already a good start. That's the easy bit. The slightly harder bit is now how these tumors are classified based on their IDH mutation status. So looking at the IDH wild type, and again this is an important point, these tumors, to classify them as glioblastoma, they need to have other mutations as well, such as gain of chromosome 7 or loss of 10 type mutation and EGFR amplification. So if you don't have that or no histological signs of glioblastoma, this is not a glioblastoma. So these are important to be present in IDH wild type tumors. And what we can see here is then that any tumor, irrespective of the grade, will be an IDH wild type glioblastoma. And these are the typical primary glioblastomas as we know them, so the tumors that have the area of central necrosis, irregular ring enhancement, mass effect, hemorrhage, little vessels, and of course very high perfusion. So the typical primary glioblastoma, that's the IDH wild type tumor. So what about the IDH mutated tumor? So I already mentioned the 1p90q co-deleted. So if there's no 1p90q co-deletion, again, there have to be other mutations such as TP53 mutation or ATRX loss. And these are the IDH mutant astrocytoma. I have to admit, this is an old slide, this anaplastic is incorrect. This should not be there. It's just called astrocytoma. But what is more important here is also that we have no more glioblastoma, and I'll come back to that in a minute. So the IDH mutant 1p90q intact tumors, that's the diffuse astrocytoma, which are typically, this is a typical example. So on T2-weighted imaging, we have a very homogeneous signal, very distinct border. And of course, we also know with our esteemed colleagues' work that we can have a loss of the signal on the T2 flare-weighted images, known as the T2 flare mismatch sign, which I'm sure we'll hear about more later this session. So to come back to those grade four tumors, so grade four astrocytoma used to be glioblastoma. That's no longer the case for the IDH mutated tumor. So an IDH mutated tumor cannot be longer a glioblastoma. How do we define a grade four? So again, we used to do that based on histopathological characteristics, such as necrosis and microvascular proliferation. That's still possible, but also we can now use molecular parameters, and this is a homozygous deletion of CDK2NAB that you can see here. So either will give you a grade four. And again, what we see here is what's typically really an IDH mutated astrocytoma, but then we tend to also see some enhancement, where typically the enhancing region is a bit smaller than the, quite small compared to the area that has a high signal. So to come back to the IDH wild type, so what if it's not a grade four histopathologically? So these are the patients that we used to see and were very confusing because they had, both on conventional imaging and histopathology, an appearance that was low grade, but these patients actually did not do very well. In fact, these tumors behave more like glioblastoma. So these are now, looking back, the tumors that are the IDH wild type tumors with the additional mutations, but not yet the features that we associate typically with the glioblastoma. So here we have an example of a 57-year-old female presenting with seizures and a progressive deficit. She had these multiple regions of hyperintensity on T2-weighted imaging, no enhancement, but we did have an area where there was some increased perfusion. So in a way, we did have an indication that this tumor was maybe not as low grade as it looked, even just, you could even wonder whether this is a tumor looking at this. This was biopsied and the histology came back, as estrocytoma grade two. So histopathology and the imaging appearance on conventional imaging does look fairly benign, but if we then look at the mutational status, so the next generation sequencing, we see that this is an IDH wild type tumor with a P10 mutation and EGFR amplification. So this is a glioblastoma, and this is also the course that this wild patient went, unfortunately. So the big question, of course, while we're all sort of taking this in and all of these mutations, should we care? We always have to have, of course, an opportunity to put Dr. House in our presentations as well. I think we do. I mean, the big argument against this, well, you need to do surgery anyway. You need to do biopsy anyway, and that is absolutely true. But of course, we as radiologists, and I think that's what the whole purpose of this session is as well, is that we do have a very big value here, and I just want to give you an example from our recent tumor board, where this was a relatively young patient in her 40s who presented with this very big lesion with a cystic portion that was enhancement, but you can also see that there's quite a large area of non-enhancing tumor, which is very homogeneous and looks fairly solid. This tumor was treated like it, the surgery was done as in thinking that this was potentially a glioblastoma, so the enhancing portion, as you can see, it's beautifully removed. This was a gross total resection. Surgically and also on imaging, we don't have any enhancing portion anymore, but then the histopathology and the molecular status came back, and this is an IDH-mutated osteocytoma, where we know that this large portion of non-enhancing abnormality is, in fact, really a solid tumor, like you would see in lower-grade tumors. So there is actually a very large residual tumor in this patient, which means that if we had known beforehand that maybe the resection could have been bigger, this patient went for a re-resection, or maybe if the tumor is so large that you can't obtain a gross total resection or a near-total resection, you might not go for that and go for biopsy instead. So what I've tried to show you in this very short period of time is that this glioma classification, and I'm only talking about the adults now, so pediatrics is even more complicated, is a very rapidly changing landscape where we have a very rapid succession of updates, and this is also stated to be a transitional state. It's very likely that this current classification will be updated again with more focus, probably, on molecular classification. The IDH well-type tumors, so importantly, they are considered now a very separate entity and it defines glioblastoma, but in the presence of other mutations. So not all IDH well-type tumors are glioblastoma. Low-grade tumors are now called lower-grade. The term anaplastic has disappeared, and the lower-grade IDH-mutated tumors, grades two and three, have a fairly similar outcome. The grade four IDH-mutated tumors are no longer glioblastoma, but astrocytoma grade four, and there is a clear distinction between adult-type and pediatric-type diffused glioma. And what I've also hoped to convince you of already is that tissue diagnosis, of course, is important, but we really do matter, and it's important to know this stuff. And with that, I'd like to thank you for your attention. Good morning, everyone. Nice to see everyone in person again. So we just heard about the WHO, the latest WHO classification, and in the next 12 minutes or so, we're gonna talk about the role of the radiologist, all of us, in the diagnosis of brain tumors. So in the preoperative setting, not surprisingly, our role is mainly to determine whether or not a lesion is a tumor, tumor or not tumor, but more importantly is to determine whether or not a patient should be going to the OR. So I'm gonna show you seven cases, and I want you to decide for yourself, tumor or not tumor, and should this patient be going to the OR? So this is case number one. I have the sequences labeled below. I'm just gonna show it to you for a couple seconds. Case number one. Case number two, tumor or not tumor, should this patient go to the OR? Case number four, here's an arrow to help direct your eyes. Case number five, here's a box to help direct your eyes. And the image enlarged. Case number six, again a box to help direct your eyes. Image enlarged. And finally case number seven. So again in the preoperative setting, our role is to determine whether or not a lesion is tumor or not tumor. But really whether or not the patient should be going to the OR. So how often do you see this history? Brain tumor, brain lab MR? Usually when the patient's already on route to the OR. But just to take a step back, you know, just a few years ago, the wonderful Dr. Sunmi Cha and her group actually looked at almost 1,000 cases that were diagnosed as brain tumor. And it turned out that 12% of the time, the history was incorrect. They were misdiagnosed as brain tumor. So it's always important to take a step back and ask yourself, is this really a tumor? Should this patient actually be going to the OR? So we're going to run through the cases. This was case number one. You see this large T2 hyper intense lesion, non-enhancing, with a sign that we alluded to earlier, the T2 flare mismatch. Where you have this dropout of signal centrally and this rim of T2 hyper intensity. For those of you that do 2HGMR spectroscopy, this is an example where you can see a small peak. So it's positive. And this was pathology showing an IDH1 mutation. So this was a tumor, an IDH mutated astrocytoma. In these cases, the surgeon is aiming for a gross total resection. And the T2 flare mismatch, again, is a highly specific sign for IDH mutated astrocytoma. And it differentiates it from oligodendroglioma, which is IDH mutated, but also has the 1p19q co-deletion. It was first described a few years ago by our wonderful co-speaker, Dr. Jane, and his group. And it was recently validated in a large meta-analysis of over 1,000 patients, which showed a pool sensitivity of 42%. And again, a very high specificity of 100%. In cases where you can't completely resect the tumor, imaging, including advanced imaging, can help you find the highest grade portion of the tumor. Because that can actually help direct the surgeon as to which part they're going to work the hardest to resect. So moving on to case number two. You can see this large enhancing mass with broad-based contact with the tentorium, compatible with an extra axial lesion. This is how it looked on pathology, mesenchymal spindle cells. And this was a solitary fibrous tumor, previously known as hemangiopericytoma. In the differential would be other extra axial masses, meningioma, for example. It's notable that SFTs are less likely to have that long dural tail that we see with meningiomas. Dural metastasis would be another consideration. These can be WHO grades one through three. And they're often associated with a specific gene fusion. Moving on to case number three. You can see enhancing periventricular nodules, which are diffusion restricting. This is how it looked on pathology, large nuclei with multiple chromocenters and scant cytoplasm. And this was primary CNS lymphoma, commonly diffuse large B cell. This one was EBV negative. These are pretty characteristic findings on MR. On CT, they can be hyperattenuating because of the hypercellularity. They can involve the corpus callosum, extend along ependymal margins. It's important to note that these are treated medically. So in cases that you need it for diagnosis, you would biopsy it. But there shouldn't be a complete resection. Case number four. You can see a small, round, non-enhancing lesion in the inferior aspect of the fourth ventricle. This is how it looked on pathology. We usually don't have pathology for these patients. Microcystic spaces with large posticellular areas. And this was a subependymoma. So technically a tumor, a WHO grade one. They're typically in the fourth ventricle. The second most common site are the lateral ventricles. Non-enhancing, slow growing. I've seen these be stable for over a decade. They can calcify and they're often incidental in older individuals. So because they really don't grow very much and they're often incidental, they're usually not resected despite the fact that they're a tumor. Case number five. So the enlarged image was to help you see these small subcortical bubbles. These little T2 hyper intense round foci. Which are very typical for this entity. Multinodular evacuating neuronal tumor, MVNT. So again by WHO it's considered a grade one tumor. But they are often incidental. And occasionally they can have seizures. But because they're often incidental, they're usually not resected. Case number six. We probably get one of these every month or so. Prepontine ovoid T2 hyper intense lesion. The non-contrast T1 is to show you the marrow associated with this little osseous stalk that's leading to this lesion. Which is very characteristic for E. Chordosis Physoliphora. So not a tumor, but a notochordal remnant. So these should not be resected. They're commonly in this location, retroclival prepontine, and they're non-enhancing, which differentiates it from chordoma, which usually are resected. And finally, case number seven. So dural based mass with parenchymal vasogenic edema. And notably I put mass in quotation marks. On both the CT and the MR, you can see that it's difficult to discern if there's truly a solid mass there, or if it's more of an infiltrative lesion involving the cortex. The differential diagnosis is broad. So again, you would include extra axial lesions, including metastasis, meningioma, lymphoma potentially. You think about inflammatory conditions, such as sarcoid. And infection would also be in the differential here. And this actually did turn out to be infection. This was a case of neurosyphilis, or gamma. So not a tumor, but infection by trypanema pallidum. It's a manifestation of tertiary syphilis. We don't commonly see it nowadays in the antibiotic era. Occasionally we see it associated with individuals with HIV. In this case, as in many cases, the diagnosis was a bit of a conundrum. And so they did end up going to biopsy. But notably, not full resection. So just biopsy, just to clarify the diagnosis, but not full resection. I just want to acknowledge my funding. And I want to thank my colleague, Dr. Pasapia, and Pathology for his slides. So thank you very much for your attention. Great, thank you so much. Appreciate everybody turning out. It's a good-sized crowd. So I get extra time, I guess. Yep. Awesome, great. More time to opine and more questions. So I'm going to kind of give what I feel is a nuts and bolts for surveillance imaging of glioblastoma. And I think this probably best pertains to IDH wild-type glioblastoma, as we just learned the classifications and the diagnoses. And hopefully that will provide a framework from which this can be used to assess your institution. So we'll talk about the history of response assessment. I think knowing the history is really important and kind of helps us understand how we got to where we're at today with current recommendations. Really broad overview of the modified RANO criteria, the response assessment for neuro-oncology working group. There's been lots of these papers published. And I'll just go through the most recent one, which is the modified criteria. And then we'll discuss briefly about clinical context and how I think that's really important. And then just a brief summary of my experience. So these are the learning objectives that were on the website. I won't go through them. So the history of glioma diagnosis and resection and treatment and imaging is actually really fascinating. The famous Dr. Verkow first described the word glioma classification based off gross histology in 1865. His methods were rather crude at the time, but he was able to define in patients who had died of neurologic disease, foci within the brain that appeared neoplastic or appeared different from the normal brain tissue. I have a little bit of a slide there with what we would see today. Seeing his, I should probably include his original drawings, they're really interesting. But that gave us the basis for the diagnosis of glioma and scientists moving forward. A few years later, Dr. Godley was the first clinician to report in the literature the resection of a glioma. And it's a very brief report, which I have here on the left. The patient unfortunately ultimately died, it sounds like, from brain herniation. But it was the first, quote, successful resection of a glioma at the time. So moving forward from the late 1800s, the standard of care did over time become resection and radiation therapy that became established predominantly in the 1970s, where that became more the standard of care. And then towards the end of the century, there was more use of conformal radiation and stereotactic radiotherapy rather than whole brain radiotherapy to treat gliomas. And also the introduction of more types of chemotherapies. And as we know it today, the current standard of care, the STOOP protocol, was first reported in the early 2000s, where Dr. STOOP added temozolomide with concurrent radiation therapy and found a survival benefit of about a month and a half for these patients with glioblastoma. And so that's where we're at today, where patients, when diagnosed with presumed glioblastoma, come in as discussed. We'll get gross total resection of the enhancing components as defined by the MRI. They'll then get radiation therapy, so this is a little timeline of the treatment. They'll get radiation therapy over six weeks or so, five days a week. The patients get the weekend off, fortunately. I don't know if that's for the patients or for the radiation oncologist, but that's the way it goes. And then the patients will concurrently get oral temozolomide. And then they'll follow that treatment with continued oral temozolomide for about six months. And it depends based on the oncologist how long they'll treat with oral temozolomide. Some will treat indefinitely, some will treat with a shorter course. And so through this time, as we all know, we do MRIs about every three months or so to see how the course of therapy is going. And so that prompted the development of and the need for response criteria. And those response criteria were first published by Dr. McDonald in 1990, and at the time was pretty innovative. And because it integrated the clinical criteria from which people were treating the patients and seeing their patients with the imaging criteria. And it really emphasized four categories of the complete response where you get rid of all the enhancements. Partial response where enhancement volume decreases or at the time the size of the enhancement decreases. So called stable disease or progressive disease. But it's recognized that those criteria have some limitations. It relies on the use of gadolin enhancement, which we know is enhancement doesn't equal tumor. As we talked about in the first two publications, you can have a large extent of non-enhancing tumor, which is frank tumor. I think to some degree, it's kind of commonplace for a lot of physicians to think that the T2 component, not necessarily radiologists, but other physicians, that the T2 component is more edema, not so much tumor, but that's certainly not the case. And it plays a really important component, and so that's why we should consider that. Enhancement, gadolinium enhancement on MRI is really only an assessment of the integrity of the neurovascular unit. So the more disruption of the blood-brain barrier you have, the more enhancement you're gonna have, and the more avidly it's gonna enhance. And there's a lot of factors that affect the integrity of the neurovascular unit. One of them is chemotherapy and radiation therapy. So the RAINO, or the Response Assessment Neuro-Oncology Working Group, in the early 2000s published a paper to describe kind of a systematic approach to looking at these tumors through the course of therapy. And it was gauged at trying to improve upon, or build upon, the McDonald criteria, and address some known issues at the time, which was pseudoresponse. And pseudoresponse really was in the context of bevacizumab or Avastin therapy, VEGF antibody therapy, that when these were being developed, was thought to improve outcomes for patients with glioma, because the enhancement went away, as we see in this figure, it would decrease. Well, all it turns out doing is normalizing the neurovascular unit and decreasing enhancement. And it hasn't actually been shown to prolong survival, but at the time, it was an important thing to recognize and understand the enhancing characteristics and the pseudoresponse phenomenon. The other important context in which we have to understand what's happening through the course of therapy is so-called pseudoprogression, or when the patient's enhancement will increase and then miraculously will decrease. And it's an important concept to know because if you see the enhancement getting larger, I think it's easy to call this progressing disease, treatment failure, patient needs to go the OR, get it resected, and or change therapy, but that's not always the case. And that's a big implication for the patient, because what happens if treatment failure is called early, if you call pseudoprogression growing tumor, is you really alter the course of the patient's therapy, right? You have a therapy that's working, it's causing inflammation, it's causing tumor cells to die, and that's gonna be stopped, at least to go to the OR before they can get the diagnosis. But I think helping to try to avoid that context with understanding some of the imaging features can be useful to our patients. So like I said, the original RANO criteria were published in the early 2000s and 2010, and it was an attempt to build on the McDonald criteria to provide this framework. It included T2-weighted imaging, particularly to understand the pseudoresponse phenomenon that was happening with the Aston. And it provided a framework for differentiating pseudoresponse and pseudoprogression from growing tumor. So since 2010, I guess what, we're 11 years later, there's been multiple publications. These are four of the publications that pertain to glioblastoma or glioma imaging in general. They now include PET imaging, which I think is a really important foundation for the future. But again, we provide this framework for our interpretation of the imaging. And so a broad overview of this, it's the modified RANO criteria, which were published in 2017, really are aimed towards clinical trials. But I think it can be applicable to clinical care. One emphasis is to use BTIP imaging or standard compliant sequences kind of throughout the clinical trial world, so we can have homogeneous imaging. We try to have a standard approach to measuring these tumors, right, either using volume, calculating the volume of the enhancement, or using a sum product diameter. In the modified RANO criteria, we have what's a redefinition of the baseline imaging, which is now post-radiation rather than pre-radiation, and that's an important point. And then also, it's interesting cuz we've kind of come full circle with the modified RANO criteria where the enhancing component is really emphasized. And so the flow diagram here is really scary. And I'll go through a couple of examples. But this is a good read if you have some time. Like I said, the new baseline, if you wanna use this criteria, would be the post-radiation MRI. There's many reasons for that. And then that would be followed up with the post-treatment MRI. Sorry, so the post-radiation MRI is gonna be your baseline. And then, depending on how the patient's doing, anywhere from three to six months later, after the first round of temozolomide's been completed, that's gonna be your first so-called post-treat imaging time point. And you kinda have at that point three options, right? You can either have preliminarily progressive disease, which is an increase in the sum product diameter of enhancement by 25% or 40% if you're doing volumetric analysis. So if you see that at that six month time point or so, I would caution and say be very wary of calling progressive disease cuz we know there's this phenomenon of inflammation or pseudoprogression which occurs. And so you're gonna wait another one to three months to get that post time point follow up. If the volume continues to increase or the sum product diameter continues to increase, that's gonna be coined progression of disease. Patient's likely gonna go to the OR, get resected if they can, and then start on a new type of therapy if that's confirmed by histology. But we hope that that doesn't happen, right? So if you have anything other than these two measurement criteria, you get the diagnosis of pseudoprogression at that follow-up time point. And that will kind of reset the clock and you can progress forward to continue to follow the patient. But if the enhancement again, changes as the criteria suggests. You don't get another presumptive diagnosis. You get the diagnosis of tumor progression. So I think that's the most important point to hit on. We'll kind of go through some criteria here. This is kind of a graphical example of that, where we have a patient with their baseline here of post-treatment. I'm sorry, of post-radiation. You see the enhancing component. Six months later, it enlarges. This is the presumed progression time point. If you do a follow-up one to three months later, if it enlarges again by that 40% by volume or 25% by some product diameter, that's going to be imaging, the definition of imaging progression. Conversely, using the same patient, different imaging time point, the enhancement goes away. That would be coined pseudoprogression. And so you would continue to follow this patient with MRI, and hopefully that tumor will stay not enhancing. Pseudoresponse and the new criteria is kind of, the modified criteria is kind of built into this framework, where if, you know, the enhancement goes away, you're going to continue to follow it because you're going to want to see that progressive enhancement rather than really relying on the T2. I think the imaging, the T2 imaging features are, I think, pretty important, and I use them clinically, but that's the way this modified radial criteria is described. So clinical context is really important. You know, I tell the trainees, I really expect them to know the history, especially for these brain tumor patients because it profoundly affects how I think of what's going to happen with the imaging and what the imaging means. You know, some patients may be more likely to get pseudoprogression, and GMT molecular status is an important, could be an important feature of that. It also knows where, it's important to know where they're at in their disease treatment. Like I said, if they're within that first six months post-radiation, you got to convince me to call it disease progression because there's just so much happening in that tumor bed because of the radiation and the temozolomide that it can really have ugly imaging and not necessarily be growing tumor. And so I include just a picture from a meta-analysis here which just shows the nonspecificity and sensitivity of enhancement at best. We're getting it wrong 20% of the time using only enhancement. So like I said, these features provide a framework, but I think a lot of clinical judgment can go into that as well, and especially physiologic imaging. So kind of a summary of my experience, which I learned from the great Sumi Cha as well, is that, you know, I think a lot of times, especially in tumor board, we're asked to be clairvoyant and predict the future. I just, I can never do it. I just, you know, go by the imaging features. I think of aggressive biology as well. I didn't talk about perfusion-weighted imaging or diffusion-weighted imaging or PET imaging features that I think are important and I use clinically to try to understand what's happening with these tumors. The biology is really important, and it's very different, depending on what's happening. But it's often an admixture. And so if the surgeons are going to take the patient to the OR, I really do encourage them to use perfusion-weighted imaging to help guide the surgery and to biopsy what I feel is going to be the most aggressive biology that I can see. The RANDO criteria provide a framework. And I think overall, we're working on this, but we really do need biologically specific imaging metrics to help these patients better. Thank you so much. What I'm going to talk about in the end is talk about a little bit about future, which things which we are not using right now currently in clinical practice, most of it is in research. And that's what I'm going to talk about is radiomics and a little bit about what the future holds for glioma imagers like me. So let me start. You know, Hummingbird effect, it's an innovation or cluster of innovations in one field which end up triggering changes in domains which it is not connected with, right? And that's what is happening in medicine right now with computing power increasing. You know, we are seeing changes happening in medicine, in healthcare, in radiology. And then many of these inventions, they kind of occur in clusters, right? We know that. And this graph, you guys might have seen this many a times. It's very fascinating to me that the computing power of, you know, is going to be almost reaching one human brain computing power by 2023. Maybe it's already there. And by 2045, in 25 years, it's going to be like all humans on the earth. And that's what is exciting to me. You know, I think the change what we're going to see in the next 15, 20 years, if some of us continue practicing, is this is going to be the slope we are going to be facing, right? And this is what we, as glioma imagers, we are facing too. So radiomics, what is radiomics? Identifying important characteristics from an image was described from aerial photographs as early as 1955. And then from medical imaging by Harlech in 1973, using computable texture features, it's basically high throughput data extracted in large amount of quantitative imaging features that are generated from medical images using algorithms and computers. And the potential to see, to uncover disease characteristics which human eye, a naked eye radiologist cannot really see. And I'll show you a few examples of that. Now, in tumor world, radiomic features could be divided into four different categories. I'm not going to talk too much about that. There are first order statistics, texture features, wavelet features, and then shape features. The first three are kind of, radiologists don't understand these. I don't understand because these are not part of the radiological lexicon. So they're kind of agnostic features to me. And the ones which I understand a little bit are the shape features where, you know, your roundedness of the tumor, borders, flatness, compactness, those are some of the things I probably can understand as a radiologist. But, you know, beyond that, you know, the first three, as I said, radiologists typically don't understand. The typical pipeline for a post-processing workflow in radiomics imaging is basically 3D segmentation, which has, you know, now has achieved great success with, you know, federated learning run by UPenn group, Spiros and other people, Brad Challenge. We know that the 3D segmentation of the tumors is now doable in 2021, followed by intensity normalization and then feature extraction, right? And as I said, you know, these multiple different feature sets, there are hundreds and thousands of feature sets you can generate, and then you're putting them through some kind of analysis, machine learning tools, and then at the end, hopefully integrating them with molecular and clinical features, right? That's what gives us some more exciting results. One of the important things, you know, people learn very quickly is that feature reduction is an important aspect because there are so many features, right, hundreds and thousands of these features. And if the features are more than the clinical subjects or, you know, the subjects you have in a study, there is going to be overfitting of the data, right? So that's a big issue with radiomic studies, which are being published, you know, left and right. You have to look at that. So feature reduction is an important aspect, which people have now started using in radiomic studies. This is a review we wrote a couple of years back, how far we are from the clinical practice, you know? So difficulties with clinical implementation of radiomic studies, the big risk, as I said, is risk of overfitting because you have hundreds and thousands of these features you are generating. And if your subject population is 100 patients, it's going to be, the data is going to be, there's going to be heavily overfitted, right? So validation is an important thing. Majority studies you will see are single center studies published, right? And that's an important thing. You need to validate some of these radiomic studies with multi-institute data, multiple data sets and external data sets and things like that. And then, you know, of course, it comes back to the quality of these studies, right? So that's a big issue. This is a very nice article. You know, I think the table is not going to look really well over here because of the size, but if you can read this article, you know, it talks about the scoring system. So as a reviewer, if I'm reviewing these radiomic studies, you know, it becomes really challenging. The quality of these studies is really poor and I'll show you why. And one of the major reasons, as I discussed, is, you know, these number of variables you have, number of features you have in a study, if they're more than number of variables and the subjects in that particular data set, there is overfitting, right? So you can see the scoring system actually gives you minus points if you don't have done, you know, the feature reduction. Another important thing, as I mentioned, is validation, right? So again, you know, minus five points if there is no validation. And then there are points, you know, going, if there is a single set validation, multi-institute validation, you know, you get maximum points like plus five points or something like that. And then another important thing which radiology, RSNA, and everybody were pushing is, you know, open science, right? You have to put your data sets out and Dr. Flanders is here. He's a big proponent of that. You know, you have to put your data sets on some kind of open platform so that people can really redo things, redo analysis in their own way and validate some of the stuff, right? So this is a very nice systematic review done by Ascent Medical Center looking at the quality of radiomics studies in neuro-oncologic imaging. And you can see the quality of the studies published is rather poor, you know, the level of evidence provided is not that great, right? And that goes back to how we score them. So that's very important, right? So that's a problem we face. The reason the radiology implementation is gonna be a little bit delayed of some of these tools is because of that. But there are very good studies too, right? This study from Heidelberg looking at, you know, the radiomics and prediction errors decreased by almost 36% for progression-free survival or overall survival, you know, when they were added beyond the molecular and clinical features, right? So definitely helping in these scientific studies if they are done properly, right? This is a study from Case Western looking at radiomics features and creating these hypoxia enrichment score for these gliomas, right? And showing tumors which have high hypoxia enrichment score, you know, they're doing poorly compared to the ones which have low hypoxia enrichment score, right? So again, these are generated from these radiomic features and very well done studies, right? Another problem which Ramon touched very nicely about surveillance imaging, right? We are asked to read these, you know, brain tumor studies every, almost every day where we're trying to figure out, you know, these are four different patients post-treatment, you know, whether this is recurrent tumor or radiation necrosis, right? And it's, there is no way to say that based on just morphologic imaging. We're using perfusion imaging. A lot of people are using, you know, PETMR and other things to really differentiate that. But can radiomics help you with that, right? So these are two recurrent tumors. These are two radiation necrosis. How does that, and we know visual inspection by any expert is rather poor, right? It's like flipping a coin. And then these radiomic studies come, you know, this is one of the first order features, entropy within the localized gradient orientation of the lesion. So this is one example, you know, from Pallavi Tiwari's work at Case Western, again, they published it, that you look at this entropy of the localized, you know, the gradient field, and it's an indicative of the degree of disorder in the pathology, right? The more disorder the pathology has, this is the recurrent tumor. This is radiation necrosis, right? So there is, there are visual, there are differences on radiomic studies, on radiomic feature sets, which are there, which just the human naked eye examination cannot really help with that, right? So this is an example over here. Again, the same thing, you know, recurrent tumor on the left over here, with the higher degree of disorder within the pathology, and they actually went one step further, and they actually published, you know, kind of a grading system for these also. You can see, you know, the degree of entropy kind of going higher in pure tumor cases versus, you know, predominant tumor, predominant radiation necrosis, or pure radiation necrosis, right? So there is, there are, there is information available in these image sets, which is just not, you know, we cannot really appreciate that with naked human, you know, naked eye examination, just looking at the thing. And then the big thing is, you know, how do we assess tumor response? You know, response after therapy, and we know resist and McDonald criteria, Ramon again went much more in detail, I'm not going to talk, volumetric assessments are much more reliable than 2D measurements, we know that, but they are very labor intensive, time consuming, and complex, right? So we all know that it's not doable in the actual clinical practice when you're reading 50 or 60 scans every day, right? So 3D segmentation tools are there now, again, with BRATS challenge, RSNA has a now challenge with the BRATS, and federated learning, as I said, run by Spiros Bakas is, you know, making very good progress. And then I saw this paper in 2019, again, from Heidelberg, you know, again, the tables are small, but this is a great study, again, you know, looking at, you know, a test set, which was from the Heidelberg data set, and then testing sets, which were from the Heidelberg, another set, you know, another independent set of patients, plus another testing set from one of the European trials, so, right, so this is a great study, in that sense, you know, they have enough test and, sorry, the validation and the testing data sets, right? And what they've shown over here is that the tumor segmentation agreement was very high, when you look at the contrast enhancing and the non-enhancing tumor, and the tumor volume agreement was even higher, right? So with some kind of volumetric assessment done in a semi-automated manner, and when I saw this paper, I usually, you know, will send something like that to the tumor board, colleagues, and I wrote, this paper gives me a shiver down my spine, because, you know, this is what radiologists like me, you know, thrive on in the tumor board, we say, oh, this is recurrent tumor or not, and one of my friends, Dr. Kunzeoka, a neurosurgeon, he wrote back immediately, and he's a great guy, he says, you know, and the best thing was, you know, he said, your biggest fan, right? So that's what validation means to me, not the studies. But anyway, so let's move on to what we're doing with that now. This is a study from UPenn, and they looked at, you know, this machine learning, again, for early recurrence versus no recurrence, and what they noticed is that, if you look at signature coming from one imaging data set, you know, T1, post-cat, FLIR, the signature is there, but it's very small, right? If you combine that, putting through some kind of machine learning tool, this signature difference is pretty impressive, right? So really, you know, by putting multi-parametric imaging in some of these machine learning tools, you can really do a great job in differentiating early recurrence from no recurrence. Another study from the same group, you know, looking at infiltration, right? Again, a radiomic study, done really well, showing you the infiltration maps. You can really generate that. Are we going to be there? Hopefully very soon, right? You know, we should be trying to see if these can be implemented in clinical practice. But the problem for radiologists like me, we need explanation for some of these tools, right? We don't believe in these tools because we don't have explanation, right? And this is what we're working on. We're trying to create some explainers for some of the radiomic tools. People who are pure scientists, they don't care. They just say, you know, how do we care? You know, we have the results, great results, we can implement it. But people like me, we want to know what these tools are looking at. And that's what we're trying to create. This is a work done in collaboration with the University of Colorado we're doing. We're looking at these shapely additive explanations, SHAP, looking at these radiomic features and looking at their weightage, right? And trying to see how the model output changes by changing some of these variables, right? So that's going to help me explain what these features mean and which features is one of the most important ones coming out. For example, you know, we're just looking at one feature and we're trying to figure out what that feature, we can actually get these quantitative numbers for each features. These are IDH mutated tumors versus IDH wild type. So we can actually get the numbers like that. And we can see, we can put some of the examples up. It's very crude work, crude way to do these things. But looking at that, you know, what does that feature mean on the imaging side so that we can come back and say, hey, this is what you should be looking for. Again, this is still work in progress, but what we do is we can actually do, you know, these weightage on the computer software program and we can change it and we can drag things around and we can play with it a little bit, right? Another thing we did a few years back is looking at EGFR signal, EGFR expression signal. These maps are upside down, but you know, you can see that these glioblastomas where that EGFR signal is coming based on the radiomic features, right? So that's, these are saliency maps looking at that. And that, there's another great study by Yupeng Group and Suresh Mohan, one of the coauthors is sitting over here, which they looked at, you know, some of these tools, deep learning tools and looking at differential diagnosis, you know, just looking at clinical differential diagnosis for 35 different entities in deep gray matter location lesions, right? And they did a great study showing that when you put all these clinical features and the imaging features and very simple binary way and put that through, you know, the AI tools actually do much better than or almost similar to an attending neuroradiologist, right? Much better than a trainee neuroradiologist or a radiology trainee, but almost similar to an attending neuroradiologist, right? But that was not the beauty of this study. To me, the beauty of the study was this table. You're looking at how each feature, each parameter, each variable they put it in is changing your decision-making, how it's affecting the actual output, right? And for example, I'm just going to very quickly show you chronicity, just putting chronicity of the lesion. If you take that away, the top differential diagnosis or top three differential diagnosis, you lose, you know, 11 or 12%, something like that, right? So if you put all the clinical features away, you know, you lose almost 20, 25% of the accuracy of this kind of analysis, all signal features, imaging features, you know, again, 20, 20. And then one of the top features, which was coming out on the T1-weighted signal, right? So if you take that away, so the idea I'm trying to say is that, you know, we need some of these explainers trying to figure out what they're doing, what the scientists are doing, what these researchers are doing and how they're putting it together. And that could be an important aspect going forward, implementation of some of these tools. And, you know, as radiologists, clinical radiologists, we having more belief in these, we have more faith and more, you know, a little more confidence in saying that this is working, right? So again, you know, worrying about AI is, I don't think we should be doing that and I don't think I'm the first person to say that, right? And this is what future looks for me, you know, I'm going to probably have a good looking boat like that, you know, this is at Seoul airport a few years back. And I think this is what is going to be accompanying me in the reading room, hopefully in the five or 10 years, my own personal assistant, right? So I hope Adam is listening and working on that, so he's going to get me one. So with that, thank you very much. This is a work of, you know, collaborative effort with multiple institutes, individuals I've been working for many years. So thank you.

gliomas

WHO guidelines

molecular diagnostics

radiological techniques

tumor classification

pseudoprogression

radiomics

artificial intelligence

tumor assessment