 Good morning, and thank you for the opportunity to present my work this morning. My presentation is on the imaging genomic analysis of TCGA and TCIA diffuse lower-grade gliomas by molecular subtype. And this work is a collaborative effort with the Cancer Imaging Archive TCGA glioma phenotype research group. So since we're a diverse group of cancer researchers and we each have our own tumor specialties, I thought I'd start with a few fast facts on brain tumors. So there are over 120 types of brain tumors, but today we'll focus on a subset known as glioma. This year it is expected that over 66,000 cases of primary brain tumors will be diagnosed, and about 30% of those will be glioma, so that's just over 19,000 cases. Glioma makes up about 80% of all malignant tumors, and as it indicates here there's only four FDA treatments approved. And treatment effect varies, as we've seen with all of our cancers, and here we can see survival will actually range between less than two years, for some cases over 15. So glioma is a cancer of the glial cells, so not your neurons that's transmitting the signals in the brain, but those cells that support the brain. So if you have a cancer of the astrocyte, this is astrocytoma. If you have a cancer of the oligodendrocyte, this is oligodendroglioma, and there's actually a mixed form currently being recognized as oligoastrocytoma. Now in the lower grade glioma groups we focused on who grades two and three, which are infiltrative and moderately diffused or proliferative disease, but these may progress into glioblastoma, which is who grade four disease. And those have a high mitotic rate in our prone to necrosis. So in the past five or so years we've learned that there are specific subsets of glioma with genomic alterations, namely about 80% of grade two and grade three gliomas harbor a mutation in either the IDH1 or IDH2, that is isocitris dehydrogenase genes, and this leads to alteration in methylation and genome instability. In contrast, only about 10% of glioblastomas happen to harbor this IDH mutation. Furthermore, in about 30% of gliomas, we have observed mostly in oligodendrogliomas a loss of both chromosome 1P and 19Q arms. So we see this co-deletion about 30% at the time. And nearly 100% of those that have co-deletion also harbor an IDH1 or 2 mutation. Both the IDH mutation and the 1P19Q co-deletion are associated with better prognosis. So though we have preconceived notions, as I mentioned, we tend to see that co-deletion going with oligodendrogliomas, about which types of glioma will harbor which types of genomic alteration, we found that there's actually not a strong concordance between the IDH mutation and co-deletion and those that histology and grade subtypes, as shown here in this cord diagram, where you can follow the IDH mutation with co-deletion across to the oligodendrogliomas of type 2 and 3, as expected, but also find those purple cords going into oligoastrocytoma in other groups. So in particular, we notice that IDH wild type here goes into all of the six histology and grade categories. And actually the IDH wild type follows right along with a GBM profile on a survival curve. So in fact, the analysis working group for Loa-Greig glioma went on to show that the distinct molecular profiles shown here, and you'll notice this color coding kind of following throughout, these three groups, IDH wild type, IDH mutant with co-deletion and IDH mutant without co-deletion, are following this molecular pattern as shown by the distinct groups here, DNA methylation, copy number, and then the cluster of clusters across the group specifically, but also with these molecular profiles. And of note, again here, these IDH wild types are really mimicking GBMs showing EGFR, P10, CDNK, 2A, a lot of those same genomic alterations that we would expect to see. But notably, it's not following histology or grade, so that's kind of the punchline there. But in parallel, the Cancer Imaging Archive, which is in the Informatics Division of the NCI's Cancer Imaging Program, was collecting preoperative radiological imaging sets for the glioma cases as profiled by TCGA. And this is something that they've done for over two dozen, I believe, now, different cancer types. So this includes T1-weighted MRI with and without contrast enhancement, T2-weighted MRI, fluid attenuated inversion recovery, or FLARE MRI, and the DWI-80C maps for our particular cancer. And so you'll see a lot of these screenshot images here as we go forward, and I use the xnatview.org, I'll give a little plug for you, David, that this is a great way for those of us who are not wanting to download all the images to be able to go through and view them. So unlike the study we saw yesterday by Mary Ellen Geiger, which was doing this by computer automated searches, we are actually using a set of neuro radiologists. There were nine neuro radiologists that read each sample in triplicates. So we had three people per sample, they were blinded to the diagnosis or molecular group of the tumor, and they looked for 26, I keep hitting the wrong button, 26 features of the tumor. And these are features that were defined to be related to what you would expect to see if you were looking as a radiologist attempting to diagnose this tumor in the clinic. Okay. So our question for the study today is, are there neuro imaging features that are associated with that IDH co-deletion-defined molecular subtype that we found from the LGG Working Group that we can see here in the radiology images? And one thing in particular that I was really fascinated to see is, do we see that IDH wild type really showing up as a glioblastoma? So just to get a sense, we have 70 cases here in our first round of reviewer assessed images. You can see that this pattern here of IDH wild type mutant non-codel and mutant codel is following just what we saw. It's really mixed with the different subtypes. We see that this subset also has that characteristic drop in survival for the IDH wild type. So these 70 is a good representation of the 293 that we used in the full study for LGG. So some of the things we found, first thing we found was that IDH mutant codel tumors were likely to be centered in the frontal lobe. In fact, 75% of those tumors were found in the frontal lobe compared to other locations. Similarly, the IDH mutant non-codel tumors were split between the frontal lobe 41% and the temporal lobe 41%. So 80% were found in this region. In contrast, we saw that the IDH wild type could be found in any region of the brain and they did not have any specific localization. And this is something that we did note in the LGG marker paper, but it was nice to have that confirmation that we were finding on the same page. So other IDH mutated associated features. IDH mutant codel tumors were more likely to have a T1 flare signal across the midline. We'll see a little bit more on what the T1 flare signal is looking like. They have presence of hemorrhage and they have presence of cysts, more prevalent than other tumor types. The IDH mutant non-codel tumors were the least likely to have the presence of satellites. So less likely to have little bits in other regions of the brain. So these IDH wild type tumors, we found that they tended to be more infiltrative. So here's that T1 versus flare. And in fact, what we were looking at in this case is the T1 region and the flare region and were they of similar dimension or did the T1 flare, the T2 flare region have a more expansive molecular footprint than the T1 weighted region. And so for the mutant codel cases and the mutant non-codel cases, we tended to have more of them where the T1 and the flare regions were similar, showing a similar expanse here. So not showing that infiltration, whereas in the IDH wild type, and you can kind of see this guy here, but you can see it's much more expansive in the flare region. And so we're showing that infiltrative pattern. IDH wild type were also less likely to have a not well-defined non-enhancing region. So the enhancing region is that part that you scoop out when you're actually doing the resection. It's that part defined by the blood-brain barrier being broken down. But the non-enhancing region is that expansive region beyond it is showing that edema, that growth, that infiltration. And so if you were in IDH wild type, you were more likely to have this type of a fuzzy, just broad, expansive infiltration, whereas you would have a better or more circumscribed tumor if you were in an IDH mutant case. The IDH wild type tumors also tended to be smaller. So here we see an example of an IDH mutant codeleted tumor that has 57.6 centimeters squared area in its largest region. And you can see this big guy here in the frontal lobe as expected for an IDH mutant codel. And then the IDH wild type here is only this one is 21.5 centimeters squared area in its maximal range. So they tended to be smaller. And the hypothesis there is kind of twofold. So one might be that it's more aggressive, it's expanding quickly, and so we're catching it in that rapid growth phase, whereas these guys are kind of growing a little more slowly and we're not catching them. Also, it could be that being in the frontal lobe, you're less likely to get into some of those eloquent regions as quickly. And so we're not catching them by symptomatic diagnosis as early until they've grown to a larger state. So even though we've seen some things that are related to IDH wild type tumors and aggressivity, there are two key features that we see in GBM that we did not actually see more common in IDH wild type. So this is a GBM. So you can see that real bright enhancing region showing that breakdown and that hemorrhage of the blood brain barrier and that hemorrhage. We've got a necrotic region here, which is key to the who grade four disease that necrosis prone. Now in the lower grade glioma, the grades two and three, we actually saw about 66% of our tumors had an enhancing component and it was not more prevalent in the wild type than the other groups. It was more prevalent in grade three, but we did see some groups with enhancing tumor in grade two disease as well. Likewise, the necrotic component was found in about 26% of our cases. So these are lower grade gliomas. Grades two and three did show necrosis in 26% of our cases, but it was not isolated to wild type. And in fact, five of those cases are about 30% were found in the grade two disease. So we tried to go one step further to see if maybe we could come up with some classifications, some clustering algorithms, and this is one of our attempts at it. We used a bi-clustering metric because this allows for the ordinal and the nominal types of data that we have because really the only thing that we have that is continuous is the dimension. And even though we found some great clusters here, none of them are really aligning except for this one cluster which has five IDH wild type grade three tumors and shows this series of features in common. So in conclusion, there are several MR imaging features that are associated with LGG molecular classes. IDH wild type does seem to show an aggressive phenotype, but it does not seem to be that it's simply an underdiagnosed GBM. Imaging patterns are present, but we're still working on exactly what those patterns mean and how we can use them for clinical impact or whether they have other genomic associations. So this work could not be done without the support of our large group here in the TCGA glioma phenotype research group at the Cancer Imaging Archive, without the support of the Hermelin Brain Tumor Center for my time and our colleagues in the lower grade glioma analysis working group. I'm happy to take questions now. You're welcome to email me at any time or I'll be wandering around my poster number 98 later today. Thank you. Great presentation. And I applaud you for using nine different observers. But the question that comes in is what was the concordance between the nine observers and what was the Kappa value? So it actually depended on each of the 26. So concordance was higher for some than others. We ended up in some cases we would have five categories like really strong, strong, neutral and so forth where we had to collapse some of the highest or the too high and too low and have a high none and low to gain concordance. So we did go through a training phase. I unfortunately I can't give you the Kappas for all of them. We went through a training phase with each of our readers ahead of time to make sure that everybody understood this 26 features and how to score them. So in all in all it was fairly strong. In the end we did a voting and they had somebody who went through and did a check to make sure that it was best representative. So there was a fairly rigorous amount of data checking and quality behind that. But I don't have the exact Kappas for each one.