NCI Imaging Community Call

12:44 pm - 1:54 pm  Monday, December 6, 2021 | (UTC-05:00) Eastern Time (US & Canada)

Ulli Wagner  Hong Zhao  Jayasree Chakraborty  Ola Adeyemi  Jessica Li  Arwa Fallatah  Saugato Rahman Dhruba  Alex Mays  Alida Palmisano  David Newitt  Janice Knable  Lawrence Tarbox  Walter Bosch  Roshan  Brian Luke  Travis Riggs  Murad Labbad  Bill Longabaugh  Guest  anna k  yogab  Ella Jones  Jurgen Seidel  Me  Jay Ronquillo  Pinyi Lu  Wyatt Tellis UCSF  Kimia Tajik  Sang Ho Lee  Michelle Tacconelli  Brandy Heckman-Stoddard  Joel Saltz  Qinyan Pan  Justin Kirby  Jungwook Shin  tracy  Robert Nordstrom  Yu Fan  Nan Hu  Binsheng Zhao  Steven Reeves  Maria Shkeda  Nolan, Tracy  Li, Fangyu (NIH/NCI)  Ariana Familiar  Keyvan Farahani  John Demchok  John Freymann  Kondal Reddy  Benjamin Bearce  Abhishek  Doori Rose  Grant McConachie  Martin Yaffe  Jack Lee  Maeve Mullooly  


WEBVTT

1
Ulli Wagner 00:00:10.260 --> 00:00:16.465
Come up half of it and I will hand the mic over to Dr. Nordstrom. Thank you.

2
Robert Nordstrom 00:00:17.275 --> 00:00:31.405
Wonderful Thank you. Thank you. And welcome everybody as participants. Continue to come in. We're going to get started. It's my extreme pleasure to introduce our speaker today and Dr Martin Cathy who.

3
Robert Nordstrom 00:01:00.000 --> 00:01:21.145
They deal with range from developing image processing strategies for improving the quality of diagnostic images and methods for quantifying the quality of the medical images themselves to developing and evaluating 3 D image the total synthesis and contrast enhance imaging methods. So you can see that.

4
Robert Nordstrom 00:01:27.225 --> 00:01:36.765
So, um, he has a well over 280 publications in this area and it's my pleasure to turn the microphone over it. Over to you. Now, Martin.

5
Martin Yaffe 00:01:37.695 --> 00:01:47.565
Well, thank you very much. Uh, Robert and, uh, good afternoon. Everybody, it's a, uh, I appreciate being invited to give this presentation. I wish I could be there to see you all.

6
Martin Yaffe 00:01:59.600 --> 00:02:07.275
Patients, and if that works, we're on our way. So please if someone would just confirm that it's happened.

7
Ulli Wagner 00:02:07.755 --> 00:02:08.835
It's perfect, thank you.

8
Martin Yaffe 00:02:08.835 --> 00:02:20.595
Excellent. Okay. Thank you. So, uh, I'm going to talk about the ideas behind precision medicine is a very popular term now. It's mostly applied to therapy.

9
Martin Yaffe 00:02:40.560 --> 00:03:01.705
With people in the industry, and here are my disclosures. Um, but most of what I'll be speaking of today, we'll have little little to do with those interactions. This is sunny. Brooke, it's located at the edge of, uh, central Toronto. It's a large, uh, health sciences center. Uh, it's been around.

10
Martin Yaffe 00:03:01.709 --> 00:03:22.854
Since, uh, early and then 9,009 hundred's and, uh, it's, it's 1 of those very spread out type campuses, which makes it very nice because you can go for a walk in the afternoon and clear your mind or just, or or discuss ideas with colleagues. Um, we have over 1200 beds in the clinical.

11
Martin Yaffe 00:04:00.160 --> 00:04:21.275
But I, I at least give you a sense of what we're doing and if you're interested in any of these areas, I'd be delighted to set up a separate section session with people who really are the experts in those areas in the lab and we can discuss things that much greater depth. So we'll talk a little bit about stratification of breast cancer screening to improve.

12
Martin Yaffe 00:04:21.308 --> 00:04:42.454
Performance I'll just mention which I think you all know about, uh, already it is, uh, NCI, uh, funded, um, again, a bit on micro simulation modeling that we use a lot in our work and a facility that we've created called the biometrics, uh, to make it easier to do.

13
Martin Yaffe 00:04:42.459 --> 00:05:03.574
A multidisciplinary research involving, uh, imaging and particularly oncology, and then I'll end with some exciting work that we're doing and trying to take some of the ideas of imaging and apply them in pathology to make it more quantitative, uh, to, for guidance of, of, of therapy.

14
Martin Yaffe 00:05:03.634 --> 00:05:23.494
To reduce over and under diagnosis and to really move towards precision medicine. And a couple of the things we're focusing on right now are biomarker heterogeneity and creation of a pipeline that combines imaging with more traditional pathologic and genomic markers.

15
Martin Yaffe 00:05:24.934 --> 00:05:45.904
So, the 1st part, uh, stratified breast screening, the question is, we know most, uh, um, uh, breast cancers are now found through, um, mammography, uh, screening. It's quite effective in. Uh, and it's been shown to reduce the, uh, death rate from breast cancer. Uh, when it's done with high quality and women attend regularly.

16
Martin Yaffe 00:05:54.500 --> 00:06:15.615
Is something we call masking? The cancer is essentially hidden in the complex background structure of the fiber glandular tissue and for those women, it may make sense to recommend either alternative or supplemental screening using techniques, which are less susceptible to the degradation that breast density causes in.

17
Martin Yaffe 00:06:16.154 --> 00:06:36.734
Mammography ultrasound contrast, enhanced mammography with exogenous contrast agents, uh, nuclear medicine approaches, et cetera and the upper, um, uh, slide part of the slide. Just shows 4 different classic categories of breast density ranging from the very fatty.

18
Martin Yaffe 00:06:56.300 --> 00:07:17.445
The increases at a given age. Women are at a higher background risk of developing breast cancer. But, as I mentioned, also at a higher risk of that cancer, not being detected on the mammogram and you can see this is work from a from wonders in the Netherlands. There they looked at the sensitivity of mammography and you can see it monotonically fall.

19
Martin Yaffe 00:07:34.460 --> 00:07:55.605
Upper 2 categories be stratified to receive some sort of supplemental screening to improve accuracy. But that would mean 50% of women would have to have additional examinations and I should point out overall. And this is from, uh, Canadian data that approximately 25% of breast cancers are not detected.

20
Martin Yaffe 00:07:55.609 --> 00:08:16.754
In mammography screening, and if I can extrapolate those numbers to the U. S where there's 37Million mammograms done each year, um, that would equate to 52,000 MIS cancers and, of course, some cancer deaths, because of the late detection of those cancers. So, it's really not practical to offer.

21
Martin Yaffe 00:08:16.759 --> 00:08:37.904
Elemental screening to half of the women can we come up with a more sensible approach that might actually be practical in the future? So, what we did is we developed using the signal detection theory and looking at the structural complexity of the.

22
Martin Yaffe 00:08:37.934 --> 00:08:59.054
Caused by density as a form of noise, we looked really at the signal to noise ratio of detection of a lesion that was electronically or digitally implanted in a mammogram at various locations evaluated. The, the chances that that cancer would be missed and came up with kind of a global masking metric.

23
Martin Yaffe 00:08:59.084 --> 00:09:20.204
Index of masking, uh, based on various radio mic features and, uh, which may also incorporate other features such as the woman's age. And the idea is to evaluate, um, on the mammogram for individual women, whether it was likely that if she.

24
Martin Yaffe 00:09:32.160 --> 00:09:52.585
To develop the model, and you also have to test it and validate it using a large you require a large data bank of, of of images where, you know, which cancers were screen detected found on mammography and which were found, not found on mammography, non screen detected. So, we use.

25
Martin Yaffe 00:09:53.454 --> 00:10:14.454
Of structure that we've created, talk about, in a minute call the biometrics, uh, which is just a big data warehouse, uh, repository for clinical images and clinical reports as our source of data for the model building and validation we do. And because there's so many reports. 1 has to go through we've also developed some.

26
Martin Yaffe 00:10:14.459 --> 00:10:35.604
Natural language processing techniques that we use to try and go through the records to determine the mode of detection screen detected, not screen detected, et cetera. So, in the most recent study, we looked at 115, almost 116,000 records on patients to try to develop our set for.

27
Martin Yaffe 00:10:35.609 --> 00:10:56.754
Model building and our set for model testing, and I won't get go into detail on this but and Martell my colleague who's a machine language person in our Institute, and her student great cooling developed a natural language processing algorithm. That would help us go through these records to identify.

28
Martin Yaffe 00:10:56.784 --> 00:11:17.904
By the motive detection, and I think more and more of these kind of tools are going to be useful in the sort of research that we do and maybe you're using them already. Um, in some after we developed our initial model. Uh, we tested it initially using, uh, through a collaboration that we had.

29
Martin Yaffe 00:11:17.909 --> 00:11:39.054
With the University of Cambridge and doctors Fiona Gilbert and Sarah Hickman there and they had cases that they described as being subtle, difficult to find cancers and ones where the cancers were more obvious and we applied the masking index to those cases. And we found that in the case of in the subtle cases, the.

30
Martin Yaffe 00:11:39.084 --> 00:12:00.204
Index was on average, 32 and a half percent higher. So it gave us some suggestion that what we were measuring within this index was related to what the radiologist was facing and trying to find cancers in mammograms. We tested the performance further, using a cohort of images from the University of.

31
Martin Yaffe 00:12:00.209 --> 00:12:21.354
Virginia and our colleague who was there at the time she's now at Rochester Dr Jennifer Harvey and this is like an curve where what we're looking at is the on the ordinate is the fraction of the cancers, which are the interval catches the ones that would be missed on screening that are actually picked up.

32
Martin Yaffe 00:12:21.359 --> 00:12:41.874
Up and identified in this, so called masked group and would be sent off to supplemental screening versus the overall fraction of the examinations that would have to be sent over. So, it's kind of an efficiency. Ideally, you would capture.

33
Martin Yaffe 00:12:42.864 --> 00:13:03.654
You would only be sending women over whose cancers would be missed on a mammography screening and you can see in this example at this point, on the operating curve, we can catch half of the interval cancers and send them for, say, supplemental screening with ultrasound by.

34
Martin Yaffe 00:13:03.659 --> 00:13:24.504
Brooding 20%, rather than the 50% needed in the Carla model. So it's, it's more efficient by least a factor of 2. we're hoping to do better than that in the future. Of course, we can change the operating point to catch more of these MIS cancers, but at the cost of sending more women.

35
Martin Yaffe 00:13:24.894 --> 00:13:45.954
Uh, who don't wouldn't have MIS cancers over into the other group and the by reds, uh, performance using the buyer ads categories are are shown, uh, comparatively just by these black, uh, indicators. And if there's, uh, we have a continuous scale, whereas by reds, because we actually use a, um.

36
Martin Yaffe 00:13:49.460 --> 00:14:10.605
Of the of the measure as opposed to by reds, which is a 4 point scale. So in a, uh, screening trial assimilated, uh, screening trial, it might look something like this, where, 20 or 25% of the women would be identified as having a high probability of masking. They would get.

37
Martin Yaffe 00:14:10.609 --> 00:14:31.724
You know, uh, imaging using 1 of these techniques. You might find you might drop the interval cancer rate from 1.4 per 1002.7, per 1000, uh, in in this group. And then, of course, there's going to be some cancers. That you're not going to catch in the, uh, who will be put into the low masking.

38
Martin Yaffe 00:14:31.759 --> 00:14:52.874
Ability but overall, at the end, you find that you've dropped the interval cancer rate by, by, about a factor of 2 and improve the detect ability of cancers this by the way is that a 2 year screening interval, which is done typically done here in Canada. Although why you could easily adapt this to a an annual.

39
Martin Yaffe 00:14:52.909 --> 00:15:14.054
Screening, so the next step for us, we're just beginning this work. We're going to do kind of the next step of testing in an intervention trial called ribs in Italy in parallel with. She has a large screening cohort. There their standard is 2 year screening. They will apply a breast cancer risk.

40
Martin Yaffe 00:15:14.059 --> 00:15:35.204
And also, as their standard measure, volumetric breast density, we're going to add the masking risk index measurement to their trial. So, we're just kind of stepping into their trial with, uh, with this 1 additional measure. The plan is their plan is that women who have elevated risk.

41
Martin Yaffe 00:15:35.235 --> 00:15:48.495
Determined by the, um, breast cancer risk model will go from biannual to annual screening. Women with elevated breast density will get supplemental breast ultrasound and women with both will get annual.

42
Wyatt Tellis UCSF 00:15:48.495 --> 00:15:49.125
Screening.

43
Martin Yaffe 00:15:49.155 --> 00:15:56.355
Plus the ultrasound and what we'll do is to compare the stratification recommendation that would occur.

44
Martin Yaffe 00:15:56.594 --> 00:16:16.604
If you just use density, or if you used our masking algorithm, I'll just mentioned, uh, uh, very briefly. Uh, because I think it is an important trial. It's an attempt to put some science into evaluation of a new technology. Typically, many of these technologies just get thrown out there.

45
Martin Yaffe 00:16:17.564 --> 00:16:38.654
And, um, don't really know how much impact they're going to have this case. We're doing a randomized trial to try to see and the question's a little different. It's not just can we find cancers, but can the use of this new modality reduce the number of advanced or aggressive cancers in a screening population compared to to the.

46
Martin Yaffe 00:16:38.684 --> 00:16:59.804
2 dimensional mammography, uh, the number involved has just been reduced to 130,000. it was originally 165,000 women 45 to 74 who are randomized between the 2 dimensional or Thomas synthesis 5 screens with follow up. But I think of interest is the fact that there is collection.

47
Martin Yaffe 00:16:59.809 --> 00:17:20.803
Of tissue in anyone that gets biopsy an optional provision of tissue, blood and buckle cells for biomarker studies some of, which are to be determined but of course, would have to be all have to be approved. And I think this offer is something that goes beyond my discussion here today but.

48
Martin Yaffe 00:17:21.135 --> 00:17:42.105
Something I've always thought would be valuable, which is to have kind of a Pre screen, using a, um, some sort of a, a blood or saliva based a measure in this case, perhaps a multi cancer detection algorithm. That's based on selfie DNA and perhaps a methylation profile as a potential for.

49
Martin Yaffe 00:18:16.960 --> 00:18:38.105
So that any difference between them is really resulted as a result of their inherent performance differences and not just the variability of how how well they're being used in in the field. I'll just mention the biometrics very quickly because it may be interesting to you. It's essentially just a large data warehouse.

50
Martin Yaffe 00:18:59.065 --> 00:19:18.745
So, we have a large center. We have a lot of data. Our vision is that anyone who walks into our institution should have the opportunity to become a research partner and make, uh, if they want make their data available to be used for approved research studies. And, of course, in some cases that.

51
Martin Yaffe 00:19:18.749 --> 00:19:39.894
Well, it will always require informed consent. There's some uses of data that don't, but we try to get prospective consent, right up front from these individuals. But the, the problem that we're trying to address is the fact that in many health institutions, the data that would be.

52
Martin Yaffe 00:19:39.899 --> 00:20:01.044
Useful for researchers start on different computer systems that don't talk to each other and researches often, get trapped into using Excel spreadsheets to store their data. It's not a very safe, reliable or a consistent way of doing things. So, we're trying to create a more uniform platform to facilitate research.

53
Martin Yaffe 00:20:01.049 --> 00:20:22.194
In this way, and essentially what we have there, and we've been doing this for some years is images, diagnostic reports, tissue samples and slides information on treatment and information on outcomes and patient demographics. So our researches have found is very useful and pulling together.

54
Martin Yaffe 00:20:22.224 --> 00:20:43.254
Various pieces of information, they need to do a number of studies and health services research in looking at the effectiveness of various therapies. And in my case in looking at imaging type interventions. So I won't spend a lot of time here. These, these.

55
Martin Yaffe 00:20:43.349 --> 00:21:04.464
Diagrams all tend to look the same, but basically many data sources we've just done by the way, made it a collaboration with Microsoft who's going to help in terms of facilitating the extraction of data from individual hospital sources. So, That'll take a lot of the load off our shoulders. It's web based and it has.

56
Martin Yaffe 00:21:04.974 --> 00:21:25.644
Um, uh, tools to try to help ensure accuracy of data, entry and consistency, privacy and security features, which have been reviewed externally. Uh, we have a group who work to make sure that the data definitions are consistent, uh, which is very important. And I think importantly we have a tool that lets you I.

57
Martin Yaffe 00:21:36.800 --> 00:21:57.495
Our work is done on de identified data from what we call pseudo patients, they have names, but they're not the correct names. They have birth dates that have been altered enough. So, you can identify the individual, we built a tissue specimen tracking system so that we can link information in the biometrics to, um.

58
Martin Yaffe 00:21:57.949 --> 00:22:19.094
Slides and specimens in our bio bank, uh, and particularly digitized history pathology slides and I'm going to skip, I think, just in the interest of time, and I'd be happy if any of you are interested in this, we can talk about it in a future session. Perhaps, um, so I'm just going to jump to.

59
Martin Yaffe 00:22:19.124 --> 00:22:40.244
Quickly just mentioned the fact that that in my lab, we use micro simulation modeling extensively to inform our research, and also to try to understand issues around policy and particularly the cancer screening policies. This is just something that we did using an sponsored model. This is.

60
Martin Yaffe 00:22:40.274 --> 00:23:01.394
The Wisconsin, Harvard model to look at different intervals times to start screening intervals for screening, et cetera on breast cancer mortality. The assistant group has published much work in this area, but we find these models very useful. Uh, more recently. There's been a model developed in Canada with the help of the assistant.

61
Martin Yaffe 00:23:01.399 --> 00:23:22.544
Model W, people called and is nice, and that it actually provides some additional flexibility and you can consider things like breast cancer subtypes and other, um, other interventions. So it's fairly flexible in terms of modifying the conditions under, which.

62
Martin Yaffe 00:23:22.549 --> 00:23:43.694
You're watching the evolution of breast cancer and the effect of an intervention, like screening or a therapy. This is just from just showing the effect of changing the interval and the ages associated with screening. You may have seen these kind of graphs before. But 1 of the things that we did recently is we applied the model.

63
Martin Yaffe 00:23:43.699 --> 00:24:04.844
And this was the, uh, model to try to understand the effective covet and the interruption of screening that occurred during Colvin in Canada was approximately 4 months. But then, even now, we're not screening at the volumes that we did Pre pandemic and what we're seeing, what we're predicting is an.

64
Martin Yaffe 00:24:04.849 --> 00:24:25.994
Increasing the number of women who die from breast cancer, but in addition a change in the in the stage at which breast cancer is detected as as shown here and that distribution really depends a lot on the interruption that takes place. And the rent, the profile of the ramping back to normal screening.

65
Martin Yaffe 00:24:25.999 --> 00:24:47.144
This is for breast cancer, and we did this in collaboration with a group interested in colorectal cancer. There's also an colorectal model and the lung model and a few others that are available. I'll just skip over that 1 and move into the last part of my top. Because I see time is.

66
Martin Yaffe 00:24:47.564 --> 00:25:08.084
Flying here, but I wanted to give you a flavor of some exciting work we're doing in, uh, we've created a pathology lab. I'm a physicist by training. I really knew nothing about pathology and probably still don't know much about it but I have good people who do molecular biologist and pathologists in the lab to, um, to straighten me out.

67
Martin Yaffe 00:25:08.594 --> 00:25:29.204
But the idea behind it is to take the tools from imaging, which, uh, are quantitative, and we found them useful in characterizing images and to apply them to make pathology more quantitative. 1 of the, our concept is really summarize in this slide where we go from in.

68
Martin Yaffe 00:25:29.594 --> 00:25:50.594
Imaging, which is low, special, poor, special resolution, relatively speaking, but gives you good sampling and a fairly complete coverage, spatially to histone pathology, fusing information between the 2 realms. If we, when we can. And then, uh.

69
Martin Yaffe 00:25:50.599 --> 00:26:11.714
Looking at individual areas, by using multiplex protein analysis of of biomarkers and in some cases doing a genomic type analysis on specific regions of interest in my lab we've developed, uh, techniques for doing very large 5 by 7 inch.

70
Martin Yaffe 00:26:11.775 --> 00:26:32.895
Type, um, pathology slides. This is H. E of a whole section through the breast. You can see the breast. You can see where the cancer is, and then we have a Google Earth type, uh, display system that allows us to blow these images up and to look at them in microscopic detail. So you can go from the macroscopic millimeter type resolution.

71
Martin Yaffe 00:26:32.899 --> 00:26:53.714
And to, uh, you know, Micron, uh, level resolution, uh, very quickly and seamlessly. Uh, this is the idea in terms of fusion. This is the work of James, main prize in my lab who's, uh, working with, uh, images of the breast and basically, um.

72
Martin Yaffe 00:26:54.824 --> 00:27:15.194
Combining them fusing information between that, and our historic pathology platform. So the color components are, this is a invasive cancer here in the breast abnormal breast tissue and just some skin for, for reference. And these are, I didn't want to risk showing a movie, but you can rotate.

73
Martin Yaffe 00:27:15.199 --> 00:27:33.494
These things and and basically get a good 3 dimensional impression of the juxtaposition of disease and normal tissue structures. So this is the direction we're going in. And right now we're actually beginning a big project in prostate cancer.

74
Martin Yaffe 00:27:37.214 --> 00:27:57.494
Alison Chung, who's a molecular biologist in my lab has got very interested in looking at phenotypes of single cells and I'll show you how that works and looking at this, some of the spatial characteristics of those cells, uh, using a protein, uh, multiplexing, uh, strategy. So, um, 1 of the questions, we're very.

75
Martin Yaffe 00:27:57.499 --> 00:28:18.644
Interested in is heterogeneity and heterogeneity comes in part from, you know, clonal, evolution in cancer. And when you look at cancers, when a lot of the measurements basically give you 1 number characterizing the cancer. But when you look at a cancer, you find there's a lot of spatial variability in the cancer.

76
Martin Yaffe 00:28:18.649 --> 00:28:39.524
And that will be important in terms of its response to traditional chemo and hormonal therapies. And also to immunotherapy, which is an increasing interest in my lab, just going to move a little bit more quickly here. So the questions really relate which patients will respond to.

77
Martin Yaffe 00:28:39.944 --> 00:29:00.854
Particular type of immunotherapy can you predict who's going to be responsive or non responsive where there's going to be resistance by looking at the information that is shown in terms of these spatial biomarker Maps? Uh, the platform we use is, uh.

78
Martin Yaffe 00:29:00.950 --> 00:29:22.095
Called it was developed by the research group, uh, in upstate, New York, their global research labs and essentially what involved it's a cyclical process. Where you take unstained tissue, you measure it's auto flourescence and correct for that. You remove auto flourescence effects and then you label with antibody.

79
Martin Yaffe 00:29:22.099 --> 00:29:42.944
Bodies that have attached to them, uh, Florida fours, which you can image with fluorescence. Then you bleach you, you, uh, acquired digital images, you bleached the Florida for, to inactivate it and then repeat with additional, um, additional antibodies.

80
Martin Yaffe 00:29:43.394 --> 00:30:04.394
So, you can, uh, image as many as 40 or 50 different antibodies on the very same tissue section, and therefore look at Co location and you can do single cell counting. You can do Boolean operations. You can measure distances. So you can now you now have images with, uh, very rich.

81
Martin Yaffe 00:30:04.399 --> 00:30:25.544
Information in them that are amenable to quantitative analysis, we developed a segmentation program that uses segment, various segmentation markers for the nuclei for the cell membrane for the cancer cells and using machine learning to segment these, uh, any kind of tissue.

82
Martin Yaffe 00:30:25.574 --> 00:30:46.664
Section image into the individual cell components, and then we can insert the fluorescent measurements of the intensity of the various biomarkers into and attribute them to that specific component. So, this just kind of illustrates that process here. This is actually worked from the G.

83
Martin Yaffe 00:30:46.725 --> 00:31:07.845
But done done in in our lab, and in our lab, we've developed the techniques of, um, of basically, uh, doing the conjugation of the antibodies specifically onto, um, onto the Florida for. So, we have the capability of if somebody is interested in a particular marker, we can usually, uh, create a.

84
Martin Yaffe 00:31:07.849 --> 00:31:28.784
Present version that we can use for imaging in the lab so you could also create from the fluorescent images. You can create a pseudo R. H E, virtual image, which is useful because most pathologists are pretty uncomfortable if they don't have an H. D image in front of them when you present some of the more.

85
Martin Yaffe 00:31:29.324 --> 00:31:50.144
Sophisticated, uh, other marker images, so we're interested in, uh, heterogeneity of the traditional biomarkers in breasts. Some of them are shown here and you can see, for example, key 67, which is a proliferation marker is a present, but it's certainly not homogeneous fully present in the tumor area.

86
Martin Yaffe 00:31:50.414 --> 00:32:11.294
You can see the same thing with, uh, the yellow, uh, um, uh, the, uh, marker and green, uh, et cetera. So it gives us a picture of really the spatial, uh, relationship really on a molecular basis. Um, almost of what's going on in within the cancer.

87
Martin Yaffe 00:32:12.284 --> 00:32:32.444
And we can look at different, uh, we see a lot of heterogeneity in the quantity and the intensity of markers in these samples. Sorry I didn't realize this was animated, but very interestingly. Um, it's very common to use immuno history chemistry to characterize.

88
Martin Yaffe 00:32:32.504 --> 00:32:53.594
Samples as a surrogate for doing molecular analysis, which is more costly and time consuming and with what we discovered is within in our lab is within the history, chemical surrogates for the breast cancer subtypes. We see a lot of heterogeneity and we've actually.

89
Martin Yaffe 00:32:53.624 --> 00:33:14.744
Found that in in our hands, uh, it might make sense to subdivide to produce additional granularity. Um, in terms of the, uh, the subtypes of breast cancer, using the multiplex Emil immuno fluorescence data. So, here, we have actually got 16 different.

90
Martin Yaffe 00:33:14.804 --> 00:33:35.864
Categories breast cancer, based on the measurements of the estrogen receptor sort of, the estrogen receptor, the projector in her 2 and whether key 67, the periphery, the proliferation marker is plus or minus and you can see, um, basically the.

91
Martin Yaffe 00:33:35.899 --> 00:33:55.994
Variability these are the amino histone chemist, chemical groups of a number of different cancer samples, and you can see the triple negative. They tend to be fairly homogeneous lead, not showing the markers, but you can see that for the, these different characteristic.

92
Martin Yaffe 00:33:57.049 --> 00:34:18.194
We call them categories are represented within a single immuno history, chemical group. So I think getting that granularity may explain to some extent why, in some cases therapies aren't working, because you're really not dealing with a.

93
Martin Yaffe 00:34:18.224 --> 00:34:39.344
Continuously responding set of cells, just more examples here just looking at different blocks from the same tumor analyzing them and we can do measurements using something called the diversity index. We can actually measure the amount of diversity in representing.

94
Martin Yaffe 00:34:39.350 --> 00:35:00.435
Information of of important markers, like PR and PR in these, these cancers we're very interested in immunotherapy. We're collaborating with a really good group downtown Toronto at Princess Margaret hospital who are working on various clinical studies. So we're trying to use these.

95
Martin Yaffe 00:35:00.884 --> 00:35:21.194
The same approach to characterize these markers you can see listed, which are important markers of the immune environment. And I can see, I'm almost out of time here so I'm just gonna kind of give you a, uh, maybe another minute of, um, perhaps a bit of a description of, of what we're doing here.

96
Martin Yaffe 00:35:21.649 --> 00:35:42.734
But, for example, we can look at immune subsets, we can look at the presence of the important markers like, and how it's represented as a percent of the immune cells, uh, the CD 8, positive immune cells that you can see that, um, that many of these measurements differ.

97
Martin Yaffe 00:35:42.915 --> 00:36:03.525
Substantially these, by the way are not breast cancer. This is a variance sample, uh, cancers of samples that we're getting from Dr. lab at princess, Margaret hospital, and you can see huge variability in the representation of these important immune markers in these samples as a percentage of the.

98
Martin Yaffe 00:36:04.004 --> 00:36:25.034
General population of immune cells so, uh, this is probably going to tell us important things about the responsiveness of these cancers, these different cancers, which from classic characterization pretty much look similar to the use of, uh, of the immune therapy.

99
Martin Yaffe 00:36:25.485 --> 00:36:46.215
I'll just mention quickly that we are using a cluster analysis techniques, like, uh, uh, et cetera to look at the clustering, uh, characteristics of the markers within examples of cancer. We can see that they're quite different. And we think this is going to be helpful in understanding, uh, which cancers will respond to, which.

100
Martin Yaffe 00:36:46.249 --> 00:37:07.394
Therapies I think I'm, this may be a good, um, I think maybe this is the last slide I'm going to show you, which is just looking at before and after treatment and this is part of the inspire trial, which is a trial being done by our colleagues at princess, Margaret hospital, but embedded on that trial.

101
Martin Yaffe 00:37:07.399 --> 00:37:28.544
Our measurements that we're making in my lab of these markers and looking at the, basically the representation of important immune markers, and seeing how they vary between the Pre treatment and post treatment situations and how that correlate.

102
Martin Yaffe 00:37:28.550 --> 00:37:49.665
It's with the actual responsiveness of these cancers to to therapy. So these are the kind of tools I realize this has been a fairly superficial presentation, but I hope this is giving you some insight into how you can use imaging tools. I hope in a valuable way.

103
Martin Yaffe 00:37:50.054 --> 00:38:10.844
Uh, to answering important questions in the treatment of of cancer and with that, I'd like to thank you for your attention. Um, acknowledge a large group of people of color people in my lab and collaborators at other institutions, uh, with whom I work. Um, thank you for your attention and, uh, just.

104
Martin Yaffe 00:38:10.875 --> 00:38:15.375
And that we have a lot of things that we have to think about very carefully. Thanks very much.

105
Robert Nordstrom 00:38:17.115 --> 00:38:27.405
Wonderful wonderful. Thank you so much. Um, it's obvious that you have a lot of interest and a lot of research going on. Do we have questions that we want to ask now?

106
Ulli Wagner 00:38:34.124 --> 00:38:37.334
If you have a question, please unmute your line and ask the question.

107
Robert Nordstrom 00:38:37.634 --> 00:38:38.024
Well, you.

108
Robert Nordstrom 00:38:41.265 --> 00:38:52.875
While you're unmuting, let me ask this. Um, I'm very interested and curious in the fact that you're able to collect so many images from a single slide. 30 or 40. you said.

109
Martin Yaffe 00:38:53.685 --> 00:38:56.385
By biomarkers from a single slide yes. Right.

110
Robert Nordstrom 00:38:56.415 --> 00:39:00.195
Right and now what kind of I'm just curious what kind of time frame.

111
Robert Nordstrom 00:39:00.224 --> 00:39:04.664
Does that take I mean, how long is the slide? Good for it. This is what I'm trying to ask.

112
Martin Yaffe 00:39:04.874 --> 00:39:21.344
Well, you know, DNA is very stable, so it, if you're if you're looking at, um, and the proteins, sorry, I should mention, in this case, we're looking at at at proteins, and they tend to be very stable and you can process the slides. This is not my area of.

113
Martin Yaffe 00:39:21.375 --> 00:39:28.755
But the people in in the lab are very confident that they can process slides even the ones that have been around for a few years.

114
Robert Nordstrom 00:39:29.025 --> 00:39:29.235
Oh.

115
Martin Yaffe 00:39:29.235 --> 00:39:42.495
Retrieve the antigens so that you can actually measure them using these, these labeled antibodies. So I'd say years, I'm sure they degrade over time, but so far that we haven't really.

116
Martin Yaffe 00:39:42.499 --> 00:39:46.664
He had a problem with with samples that were kind of expired.

117
Robert Nordstrom 00:39:47.564 --> 00:39:49.184
That's very interesting. Thank you.

118
Joel Saltz 00:39:52.755 --> 00:40:13.395
I have a question now. This is great. Talk really fantastic stuff. Martin. So I'm interested in what you were talking about red path, uh, integrative analysis, slash correlation. Um, and I wasn't sure are you, are you actually looking at tactic?

119
Joel Saltz 00:40:13.519 --> 00:40:23.714
Scenarios or are you looking at the deep learning algorithms that use both to make predictions or you talk a little bit more about how you're relating those to modalities?

120
Martin Yaffe 00:40:24.914 --> 00:40:27.104
What the imaging you mean and the.

121
Joel Saltz 00:40:27.584 --> 00:40:29.324
And the pathology yeah.

122
Martin Yaffe 00:40:29.594 --> 00:40:34.664
Well, uh, you know, currently it's really a matter of, of, uh, of trying to do.

123
Martin Yaffe 00:40:34.669 --> 00:40:44.594
Confusion with, uh, traditional information and and institutional information is it's challenging. It's Dr, salt.

124
Joel Saltz 00:40:45.134 --> 00:40:48.704
Yes, hi. Yeah, yeah, yeah, yeah, yeah. Yeah.

125
Martin Yaffe 00:40:48.704 --> 00:40:55.814
It's, it's, it's rather challenging. I mean, we use whatever we can, but is 1 of the things we.

126
Martin Yaffe 00:40:55.820 --> 00:41:16.965
We certainly do we use some, uh, fairly sophisticated computer techniques for trying to do alignment of images in different domains, bearing in mind that the, the resolution scale is. So so very different that the advantage of the whole mount is that we can look at the tissue in a pathology.

127
Martin Yaffe 00:41:16.969 --> 00:41:26.234
That are starting to look a little bit like the mammograms and the images so that you can, you know, create reference points for alignment. But it is a challenge.

128
Joel Saltz 00:41:27.314 --> 00:41:36.344
I think, okay, that's so so so then you, you actually are trying to do a comparable spatial registration.

129
Martin Yaffe 00:41:36.494 --> 00:41:38.114
When we can and we've.

130
Martin Yaffe 00:41:38.120 --> 00:41:59.265
Develop some interesting techniques for block face correlation. We, we use, believe it or not. Uh, it's like a squid or octopus. Inc. pasta is a really good. Um, uh, it's something that's very good to to use as a way of going between a pathology block.

131
Martin Yaffe 00:41:59.954 --> 00:42:20.414
And, uh, and some of the images you can, uh, basically put that into, um, into tissue. That's been freshly recepted. It will show up on it will show up on on, on various kinds of imaging and give you sort of, uh, external that lets you align structures. You know, at the, at the macroscopic.

132
Martin Yaffe 00:42:20.444 --> 00:42:20.864
So, anyway.

133
Joel Saltz 00:42:22.004 --> 00:42:34.094
Cool and if you have you thought about, or I've tried to to use Super resolution approaches to try to train the interpretation of the radiology, using the pathology.

134
Martin Yaffe 00:42:34.604 --> 00:42:38.954
Uh, we haven't got into that yet, but that's a very good suggestion. I mean.

135
Joel Saltz 00:42:40.154 --> 00:42:41.294
Because we've got a group.

136
Joel Saltz 00:42:41.655 --> 00:42:52.365
That particular work with, uh, we have a team of deep learning folks and Super resolution is 1 of our particular interest. So that could be that could be a fun discussion.

137
Martin Yaffe 00:42:52.455 --> 00:42:54.915
Well, that's a great idea. I'd love to talk with you about that.

138
Robert Nordstrom 00:42:57.284 --> 00:43:03.434
Super super good things. Um, anything else anymore any more questions.

139
Robert Nordstrom 00:43:10.304 --> 00:43:26.534
Well, hearing none, I guess it's appropriate to thank you very much for your presentation. Um, we've learned quite a bit and I'm sure people will go over the recording from time to time to.

140
Robert Nordstrom 00:43:26.955 --> 00:43:27.615
To learn more.

141
Martin Yaffe 00:43:28.215 --> 00:43:42.675
Yeah, and please if any if any of you, I realize that I could only give a very quick, superficial picture of things. So if any of you are interested in any of these areas, we can certainly get into it with, uh, much more detail. Um, if you want to set something up afterwards.

142
Robert Nordstrom 00:43:43.875 --> 00:43:45.345
That's a good opportunity. Thank you.

143
Martin Yaffe 00:43:45.705 --> 00:43:46.185
Thank you.

144
Robert Nordstrom 00:43:46.245 --> 00:43:46.605
Thank you.

145
Robert Nordstrom 00:43:48.465 --> 00:43:50.085
Okay, Julia, I guess this back to you.

146
Ulli Wagner 00:43:51.255 --> 00:43:51.795
Thank you.

147
Ulli Wagner 00:43:57.284 --> 00:44:02.144
Please, let me know if you can see the agenda slide not anymore.

148
Robert Nordstrom 00:44:05.744 --> 00:44:06.254
That's it.

149
Ulli Wagner 00:44:06.794 --> 00:44:24.464
Okay, um, next is an update, a brief update on, uh, new things from the imaging data comments. Uh, 1 of the most exciting updates that I would like to share with you, is that the imaging data comments data, which.

150
Ulli Wagner 00:44:24.524 --> 00:44:45.644
Originally came from, are now available through the Google public data set program, and therefore they have become available to an even larger audience. The Google public data set program in itself contains more than 200 public data sets and.

151
Ulli Wagner 00:44:45.674 --> 00:45:05.444
They are listed on the Google cloud marketplace and for your convenience, I have put the URLs here, but I will also put the URL to the imaging data comments in the chat as well as the link to the Google public data sets program.

152
Ulli Wagner 00:45:07.245 --> 00:45:27.735
The big advantage of having the imaging data comments data in the Google public dataset program is that they are now, uh, that users no longer have to set up a billing account to work with the data. So we encourage everybody to go and just try it out. And if you have.

153
Ulli Wagner 00:45:28.664 --> 00:45:41.324
Uh, you can put them into the chat. I know 1 of our team leads from the imaging data team, belong about is on the call. So if you have very specific questions, he might be able to answer them.

154
Ulli Wagner 00:45:45.165 --> 00:45:56.865
Thank you and, uh, the, the next topic on the agenda is an update from the cancer imaging archive. Justin. Would you like to share your screen?

155
Justin Kirby 00:45:59.294 --> 00:46:00.224
Yeah, thanks.

156
Justin Kirby 00:46:10.245 --> 00:46:29.895
Okay, uh, so a couple of quick updates from our side, um, I think we had mentioned previously, the National long screening trial data sets now available. This is a totally open access, 11, terabytes of imaging and pathology data from this large trial of approximately 26,000 subjects.

157
Justin Kirby 00:46:33.074 --> 00:46:51.134
Um, then another update on the software side of things, the MBA software that we use to manage all of our dot com data, um, had a new update recently and the, the most interesting thing from the user side of it.

158
Justin Kirby 00:46:51.164 --> 00:47:12.284
Is the data retriever software that you use to download the data is now provides a feature where you can do a check some verification to make sure that all the data that you download matches exactly. With what we have on the server side, and there's a screenshot here just showing you can find it under the file menu to turn that option on. If you.

159
Justin Kirby 00:47:12.314 --> 00:47:12.884
To use it.

160
Justin Kirby 00:47:16.724 --> 00:47:35.714
Um, then we also wanted to highlight, uh, the new data set that was published as a pathology collection called nascent prostate cancer, heterogeneity, drives evolution and resistance to intense hormonal therapy with the nickname of prostate. So, this, this was just released.

161
Justin Kirby 00:47:36.259 --> 00:47:41.384
In the middle of November and has a publication that goes along with it.

162
Justin Kirby 00:47:46.605 --> 00:48:05.865
Um, then another thing on the software side, um, we've been collaborating with some folks that project, and they put together a Python notebook tutorial that describes how to quickly and easily load data into the Monet.

163
Justin Kirby 00:48:05.985 --> 00:48:07.695
To do model development and deployment.

164
Justin Kirby 00:48:13.994 --> 00:48:31.064
Um, and then the last thing we had, uh, actually 3 different sessions that we were, um, that either had featured in, or or were discussing data sets. Um, Jonathan and, uh, prior did a presentation about, uh.

165
Justin Kirby 00:48:32.024 --> 00:48:52.364
An optimization and data privacy. Um, and then there were 2 deep learning lab classes that were part of that series this year. There were more of a hands on type of thing where you're working directly with, uh, with Python notebooks and so 1 of them was the collaboration with the imaging data.

166
Justin Kirby 00:48:52.400 --> 00:49:13.035
Comments Andre, and I presented a number of examples about how to import data from into a deep learning model. Um, and then there was another 1 that was focused specifically on the cancer genome Atlas Glioblastoma dataset to integrate both the genomic data along with the imaging data.

167
Justin Kirby 00:49:13.605 --> 00:49:30.495
Um, with the genomic data, coming from the NCI genomic data comments, and, uh, if, if you follow this link, it will actually take you to, uh, a number of different notebooks that we're all prepared and posted on GitHub for, for the deep learning lab sessions.

168
Justin Kirby 00:49:34.184 --> 00:49:35.714
Happy to answer any questions if there are any.

169
Ulli Wagner 00:49:43.574 --> 00:50:00.074
Thank you Justin, and that leaves just a few announcements for today's call, uh, as always invitations to the webinar are distributed using the and.

170
Ulli Wagner 00:50:00.105 --> 00:50:21.225
The email list, and are announced on the seabot website uh, you can also always find the information on the NCI wiki and here is the link to it on the wiki. You can also find the link in to the Webex. Uh, if presenters agree.

171
Ulli Wagner 00:50:21.254 --> 00:50:42.194
We have the presentation recorded you can also find the recordings of previous webinars on this wiki page. Our next call would have been January 3rd 2022 but because this is more or less the 1st day, uh, the 1st work day after the.

172
Ulli Wagner 00:50:42.404 --> 00:51:03.404
New year, we decided to cancel it. So, uh, we hope to see you again on February 7th for our next call, and we will announce the agenda in advance. So, you know, what's coming and this leaves me.

173
Ulli Wagner 00:51:03.554 --> 00:51:11.714
Everybody to wish you happy holidays and I'll hand the mic back over to Dr Nordstrom.

174
Robert Nordstrom 00:51:12.824 --> 00:51:24.404
Thank Julie and happy holidays to you. It's a very nice slide. Uh, happy holidays to everybody. Thank you for your attendance today. Um, it was a very informative discussion.

175
Robert Nordstrom 00:51:24.679 --> 00:51:25.844
Well, see you in February.

176
Ulli Wagner 00:51:28.334 --> 00:51:29.684
Thank you. Bye.