 Good morning. I am a medical physicist imaging scientist. I work in extracting information from images and my the mouse and my collaborator Juan G is a biostatistician working in the field of genomics We're working with the TCGA breast phenotype research group on extracting information from the MR images Unfortunately, Juan G came down ill today. However, Yitanzu is in the audience and we have three posters on this So this is the Breast phenotype research group it includes various image analysis and Genomic folks NCI as well as radiologists from around the country that have worked with these images and our purpose is to demonstrate using this data set the role of quantitative Radiomics and characterizing the molecular subtypes of breast cancer and associating that with genomic data We come from a history of Research in computer aided diagnosis where we've used computer vision techniques to extract out the information and then based on clinical information Create the knowledge to aid in diagnosis or prognosis of the patient and now using it in data mining a Few definitions so radiomics is the high throughput conversion of images to minable data and Radiogenomics or imaging genomics is the association of these features with genomics and other omics data So we're interested in asking questions about the relationship seen in these clinical medical images that are obtained Routinely in practice and their relationship with the biology of cancer We want to know which correlate which compliment to potentially lead to personalized screening and personalized treatment Most of you are in this area and and what we're interested in is in incorporating this Radiomics part into the association and predictive model development So this is our Computer aided diagnostic workstation, which is now our high throughput MRI phenotyping system once the lesion is identified the computer Automatically segments the lesion and pulls out the various phenotypes This is our data set we work with the breast cancer cases from the TCGA where clinical histopathology and genomic data is downloaded using TCGA assembler and Molecular subtyping and risk recurrent values were obtained from the Peru lab at UNC In this data set there were only at the time 91 cases that had imaging data more are being collected This is the distribution of those cases many our Well our ER positive PR positive her to negative and most are not triple negative less Most of them are luminal a and the cancer molecular subtype and From these cases we downloaded the breast MRI images and do the computational image analysis phenotyping Let me take a few moments to talk about breast MRI Since tumors have the increased blood vessels and different micro vascular density and vessel permeability We can look at the we can obtain MR images over time after injection of a gadolinium contrast agent So here we have a pre contrast a post contrast and a subtraction and you can see the tumour is highlighted We do this analysis in 4d that is across as we inject the contrasts We continue to image and we can obtain data in 3d so 4d analysis Most of the figures I'm showing will be a slice, but the analysis is done in 4d What we're interested in is looking at this uptake So if it's high uptake is quick and the washout is quick chances are it's a cancerous lesion Whereas other ones might be more persistent or plateau doesn't mean they're not cancerous, but likely not The first step we do is to segment the tumor Here examples are two cases and then read out the computer segmented lesions to better give you an idea what this is This is a 3d breast MRI. This is a clinically obtained breast image you can see right here is the tumor and So this is we indicate that tumor to the computer the computer then does the segmentation within that segmented region We are looking at Characteristics of the tumor both around the margin as well as distribution of uptake within the tumor And we're calling this a virtual biopsy note here. We can analyze the entire tumor Relative to real biopsy. We're doing it not and not invasively so we can repeat it over time And we can spatially know where we're looking at as opposed to a Surgical biopsy. We're not quite sure where the needle is going in And so we hope to correlate between the two Once we have that lesion we then can extract out multiple phenotypes so we can look at Size volume maximum linear size surface area We look at the shape of the tumor the sphericity or how irregular the shape is we look at the sharpness of the margin of the tumor and other features and Based on the uptake because this is dynamically obtained contrast enhance MRI We look at the enhancement heterogeneity and kinetics of the uptake and washout For example here this is a slice through one of the tumors and each at each of those voxels we can look at the Uptake over time and you see that there was curves on the right very greatly And that's the heterogeneity of the uptake within the tumor So what we do is we look at at one point in time How that differs and we're interested in relating the heterogeneity that we're seeing in the imaging from the angiogenesis To the heterogeneity that you are seeing in that analysis of cancer in the genomics at different locations within the tumor So we can look at the texture and we do various Computer vision texture analysis methods. We look at the uptake and washout and we look at the variance of that and those Are given as our various phenotypes of our non-invasive virtual biopsy So in the in the TCIA breast group. We're relating these computer extracted phenotypes to various tasks for example For a clinical tumor status here We show the surface area extracted by the computer relative to stage and we see that it is predictive of breast cancer tumor stage We can look at it relative to molecular classification and cancer type subtype for example We've shown that e on this limited data set that e are negative breast cancers tended to have larger size a more irregular shape And we're more heterogeneous in terms of the contrast enhancements For triple negative they tended to have a more irregular shape and more heterogeneous in terms of the contrast enhancement What happened? Oh Not sure what happened, but okay We also compare this to the molecular subtype and here we're showing that the enhancement texture of the tumor heterogenea Appears to be predictive a molecular subtype and here at the bottom We show normal like luminal a luminal be her to enrich and basil life breast cancers and As you go across the graphic it's the prognosis is worse And this was a statistical trend now in our analyses and we realized as the tumor size is very important So we try to control for that and we still see this enhancement texture of the tumor heterogeneity being predictive of the molecular subtype As shown here where the plot on the left there for very small tumors and the one our right is two centimeters to five centimeters We've also related these to the risk of recurrence from multi gene assays From Chuck Bruce lab at UNC Here's an example of two cancers The one on the left is one of a good prognostic case and one on the poor is a poor prognostic case these were analyzed with The research on cotyped DX mammal print in the pan 50 tests as well as Our radio mix looking at tumor size tumor shape irregularity and the heterogeneity of the uptake And I know you can't read this but I put this up only again to show that We are finding statistically significant correlation and the same types of features are Being shown again tumor size irregularity and uptake heterogeneity I'm not sure We can also present this in terms of our OC analysis in looking at good prognosis versus poor prognosis for these various research gene assays And the black curve here on the upper right of the ROC analysis is the performance of these of these computer extracted tumor signatures in terms of predicting poor and not so poor prognostic and Then ultimately we can relate these phenotypes to genomic pathways One thing I want to point out is that while we're we're pushing Imaging to see as much as we can what what we can see from just the imaging alone It's not that we would just use imaging ultimately we want to see if we can merge imaging with genomics and other features To come up with a better prognostic signature Also, we want to ultimately correlate our Imaging signatures our virtual biopsy signatures with actual biopsies so that we could Non-invasively follow a patient over time during therapy and also spatially correlate that information So moving on to the genomic pathways here The data was extracted from the TCGA using the TCGA assembler Software again, and we use the imaging data from the MRIs We looked at genetic pathways associated with the radiomic Phenotypes regression analysis was done as was clustering analysis so in the exploratory cluster analysis of the MR phenotypes significant associations were found between these radiomic features and the clinical outcomes as shown earlier and However, this was done using the typical genomic clustering methods treating the MR phenotype as another phenotype There various associations were shown with the MR Phenotypes and various pathway Transcriptional activities and I will list the posters where this is shown in more detail at the end of the talk Here red is pos high positive correlation and dark green is And light green is the high negative associations We also identified significant associations in particularly the size phenotype very strong related to various gene expressions of pathways and This enhancement texture heterogeneity phenotypes were related again to the MRI RNA Expression so the same phenotypes based looking at them from different ways keep appearing so in summary Computational quantitative MRI analysis shows promise as a means for high throughput image-based phenotyping and appears to predict breast cancer molecular subtypes Radiomics of tumor size and enhancement heterogeneity appear as dominant phenotypes and classifying these subtypes and risk of recurrence and Significant associations were identified between the phenotypes And molecular features involved in multiple regulation layers a main limitation is that there are only 91 breast MRI cases even though there's a thousand breast cancers in the TCGA TCIA is collecting additional images and we're also organizing outside of that a multi institutional Radiomics network to to collect cases for example at my institution We've collected 800 breast cancer cases with MRI and now we're trying to and we're collecting the pathology And if possible the genomics on them to validate these promising features Identification of radio mix of molecular subtypes of breast tumors is expected to allow for these virtual biopsies to augment actual biopsies And ongoing research involves relating and merging the MR phenotypes with genomic data to develop improved predictive Models, I like to end with these questions Is it possible to decide targeted therapy based on imaging and genomics association results? Can imaging features inform important genomic features not in this work in prior work that we've done We've in in in the assessment of future risk of breast cancer. We've used imaging and other genomic association results to Direct the analysis as opposed to going to GWAS to improve the statistics Can integration of imaging and genomic features lead to a higher power and prediction and can imaging serve as a virtual biopsy? Being non-invasive covering the complete tumor and being repeatable. I like to thank you and Reference a workshop of the TCIA here on imaging resources for the TCGA That's today at four and five and posters 91 79 and 105. Thank you Thank you for excellent talk and just two notes one and Radiogenomic right now is used by two groups effect of radiation on genome and genes and also the Radiomic and genomic Combination it needs just somehow clarification when we we are using one term and we are coining actually one term It's better to be more precise. I totally agree with your usage. I in my opinion and radiogenomic more Implied radiomic and genomic, but there are other opinion also out there and the second one is that Radiomic in particular is able to capture the heterogeneity in the tumor and What we have at the moment at least is just one snapshot of the genomic many times when you are going to Integrate the heterogeneity captured by radiomic with just one Sample and the genomic sequencing There will be a mismatch and how you're going to handle that we don't have that ability at the moment at least Well, first off, I completely agree with you on both points one. I actually the radiomics I Agree with that term radio genomics actually I prefer saying imaging genomics because of that confusion because there are the other folks where Radiogenomics is looking at more of the changes based on the radiation So I completely agree that name is sticking But I will go back to pushing for imaging genomics and I also agree on your second point And that's exactly why we want to take the imaging data where we can get complete coverage of the tumor We realize that the genomics is based on maybe one or two biopsies only and that the genomic Heterogeneity is slightly different to the imaging Heterogeneity that we're seeing but we need to relate them. So I have a slide I didn't put it in It's it's my own proposal which I'm Pressing for is that We we can do whole-mount pathology So instead of the small slide when there's a mastectomy we have a slide this big and it's kind of like a slice through the breast Where you have the prankama and the tumor and what we want to do is spatially orientated multi-omics Relationships because at that point. Yes, we can do our we can capture the entire tumor with imaging right now You can't do that. It's a partial picture with the genomics, but with the whole-mouth comparison And what we will do is then you could do Complete genomics through the tumor and you can then spatially relate it to the MRI in fact I almost put that slide in I I did not ask him to ask these two questions But those are the perfect two questions I would want someone to ask me and that will help answer that problem of the because once we Understand for the cases that we do get the entire breast not only we can we look at the tumor We can do virtual biopsy on the stroma around it and that's very and we want and that's very important So we're setting up this whole-mount pathology lab to Precisely do that and once we can understand that relationship then for cases where we don't get the complete tumor We can hopefully rely on what we're seeing in the images and then do it repeatedly over time during treatment I just wanted to add that right now by by the facilities by image guided Biopsies and intervention we have the ability to have the tissue sample from the exact location which we Know in the image and if you want to just correlate the texture or whatever of that special location To the genes it's possible and interventional oncology can can really provide a very good platform For this endeavor. Thank you Yes, we've done it with the breast cancer and also prostate cancer is what we're looking at that. Thank you So if you have images of some of your cases and you'd like to collaborate, please let me know So thank you to all the speakers from this morning session I do have one request if you are a speaker in the afternoon session if you can please come up here Right after I finish speaking. We'd like to talk to you for a second Also, we will have lunch now on your own posters start at 1 o'clock and we'll start back promptly at 2 o'clock with session 2 Thank you everybody