The Role of Imaging Before and After Diagnosis of Localized Prostate Cancer - Masoom Haider
September 15, 2022
Masoom Haider, MD, Senior Clinician Scientist, Director of the Sinai Health Research MRI, Head of the Radiomics and Machine Learning Lab, Mount Sinai Hospital, Toronto, Ontario
Phillip J. Koo, MD, FACS Division Chief of Diagnostic Imaging at the Banner MD Anderson Cancer Center in Arizona.
Phillip Koo: Hi, this is Philip Koo with the Imaging Center of Excellence at UroToday. We're very fortunate to have with us Dr. Masoom Haider, who is the Director of the Sinai Health Research MRI and Head of the Radiomics and Machine Learning Lab at Mount Sinai Hospital in Ontario. So thank you very much for joining us today.
Masoom Haider: It's my pleasure.
Phillip Koo: So you will be speaking at the 2022 AdMeTech Global Summit for Prostate Cancer and one of the topics you'll be speaking on is radiomics and machine learning. So for a lot of the viewers out there who may not be as familiar with radiomics, can you give us sort of a high level overview of what it is and what potential there is with this field of radiomics?
Masoom Haider: Sure. So radiomics is really predicated on quantifying information in medical images and it's predicated on this idea that there's so much data and information in these images. A prostate MR may have hundreds and hundreds of images in it and we're not really taking full advantage of all the data that's there when we do a visual interpretation. So typically, a number of features are extracted using engineered formulas to derive anywhere between 30 or 40, up to hundreds of imaging features from a typical imaging dataset in prostate MR, let's say. And then there's a variety of machine learning techniques that can be used to derive from that quantitative information, perhaps, something like a better risk profile for clinically significant cancer in prostate cancer patients on active surveillance, let's say, or the likelihood of them harboring a clinically significant cancer. So it's akin to genomics, but it's for images.
Phillip Koo: Great. So I think that's a really great point and I'm a nuclear radiologist and that's one of the areas that we've always or I've focused on is the fact that we have list mode data and all these data sets that we really haven't tapped into. So I think it's wonderful. I didn't realize there was such a wealth of information in MRI as well. What's been holding us back and how do we take that next step forward?
Masoom Haider: Yeah, so I think there's some challenges regarding a classic radiomics approach. And I just want to mention that there are other machine learning approaches that are newer, less than 10 years old, some of them less than five years old, that use convolutional neural networks. And I just want to put that in the category of what's excited people a lot in terms of the AI concept and computer vision. And so there's radiomics, classical radiomics, and then there's these deep convolutional neural networks. And it's really the marriage of these two that may get us to better results.
So the challenge with radiomics itself has been really in getting stability in the quantitation. So this has always been a challenge in medical imaging, where we have different manufacturers, different scanners. These are complex imaging systems, usually. And so the signal and how that signal has converted into digital, minable data by a manufacturer varies. And as a result, it's become difficult to reproduce these results across centers. So if you try, you develop a radiomics model, and you try and use it at another center on another scanner or with some different software there's instability in those measurements. And I think that's been holding back the field from really broad application.
I think the good news, though, is that these newer computer vision approaches in medicine that use deep convolutional neural networks or these artificial neural networks don't necessarily suffer from the same problems in the same way. And I think most people are now trying to use these convolutional neural networks to really surpass this problem and try and get around it. And the rationale there is that our brains are still able to interpret all of these prostate MRs, even though there's variation in quantitative stability from manufacturer to manufacturer. So there is actually enough information in the image for us to come out with pretty good interpretations using our real neural networks, our brains.
And so I think that some of the results are pointing in this direction that we can actually do quite well with these newer convolutional neural networks approach approaches. So I think that's the challenge. I think the future path that we're seeing evolve and some of the results this year are looking good.
I'll say one last thing quickly, not to go on and on too much, but one of the big needs in machine learning, and this is a key point that I want to get across at the AdMeTech meeting this year, is we need very large data sets to get results. And there has been some good progress this year in bringing together large data sets for prostate MR and I think this will be the fuel to really bring good progress.
Phillip Koo: Okay. So do you envision this being an assistive tool to the radiologist and it's sort of a separate sort of output, I guess, as you mentioned, like risk a risk score type of setup?
Masoom Haider: I think so. I think if you look at what's happening with current pieces of software and the FDA and other regulatory agencies, almost all the tools are really to assist the clinician. So can we, for example, get more stable and repeatable interpretations of prostate MR by inexperienced radiologists if we use AI? Can we get better risk calculators and decision-making tools during active surveillance if we use a combination of genomic information or fluidic information and MR and convolutional neural networks and radiomics to decide when to do a biopsy of a patient on active surveillance? And I think these assistive approaches are where we're going. It's interesting, though.
I think that there are what I'll call lower level applications. And these would be things where workflow related.So right now to do a MR fusion biopsy, where we're trying to take an ultrasound and an MR and fuse them together. A radiologist typically has to manually contour the prostate. The urologist has to do a whole bunch of manual contouring and correction on the ultrasound. They have to be fused. There's a need to constantly check for that fusion. And the latest computer vision tools really do a good job at this fully automatically, with very little intervention. So prostate segmentation and image quality assessment are areas where I think the field is there and we're seeing more and more productization of machine learning tools at that level. So we're not making clinical decisions. We're just assisting workflow. And I think that's coming and it's here already in some ways.
Phillip Koo: Great. So it's always tough to look into the future, but when do you predict these radiomics and risk score type setups with prostate MRI will make it to the clinic?
Masoom Haider: So I think we're going to see two things. One is that in radiology, there are a number of... And I'm speculating here so I will say that I don't have inside information or anything. I'm just speculating. So radiology is dominated by a small group of major manufacturers. So where they start to adopt tools and where radiology departments and hospitals and large practices start to buy is typically through a conduit of these major manufacturers. And I think what we're seeing is that some of these manufacturers are starting to put these tools into their image pipelines at various levels. And what that means is the radiologists have, now, tools that are there as part of their clinical infrastructure to do clinical trials. So instead of developing it from scratch or in a research institute, in the near term we're going to see manufacturers come out with interpretive aids, biopsy aids, image quality aids that leverage these new technologies, then I think we'll see clinical trials.
And then... So I think that's, what's going to happen first in the next couple of years is that we'll see clinical trials showing that you can do double reads with these AI tools, like in mammography with AI, or you can get better interpretation from a radiology resident or an inexperienced reader who doesn't do much prostate MR. And we'll see those publications come out and then we'll see risk calculators second. I think we'll then see, "Okay. Well, can we take this tool that has an established, published performance in a multicenter trial and actually build it into PHI or 4k score or genomics assessment on prostate?" And I think... So we're looking at a two to five year window to see some initial stuff, I think. That's my guess. But it's coming. I think it's coming.
Phillip Koo: Your speculative thoughts seem very prophetic and reasonable. So hopefully this all becomes reality soon. I agree. I think it could be a very powerful tool and really help patients, in the end. So thank you very much for joining us. Look forward to your presentation at the AdMeTech meeting.
Masoom Haider: Thank you. It's been a pleasure.