June 4, 2018
In this #WheatonExperts post, Associate Professor of Physics Dr. Heather Whitney shares her recent research.
My training and scholarly expertise is in medical physics, which is the study of the interaction of energy with the body and how we can make use of that to diagnose or treat medical issues or protect humans from the energy as needed. Even more specifically, my graduate training combined radiation therapy and magnetic resonance imaging methods. When I began to teach at Wheaton, my focus shifted to nuclear magnetic resonance, which is the fundamental measurement behind magnetic resonance imaging. This was a better fit for the liberal arts environment, in terms of resources.
When it came time to plan for my sabbatical in 2017, I wondered if it might be possible to get back into medical imaging through a collaboration. I missed it quite a bit! Long story short, a friend from graduate school connected me with my current collaborator, Dr. Maryellen L. Giger, A.N. Pritzker Professor of Radiology at the University of Chicago, who is one of my co-authors for our most recent study published in Academic Radiology: “Additive Benefit of Radiomics Over Size Alone in the Distinction Between Benign Lesions and Luminal A Cancers on a Large Clinical Breast MRI Dataset.”
I officially started working on this project in January 2017. I was anticipating officially joining the Giger Lab as a Visiting Scholar in Summer 2017, and wanted to take on a small project so I could learn the methods used in the lab. I also had an excellent student, Nathan Taylor ’19, who was interested in working on a project. Working together with Dr. Giger and her associate, Dr. Karen Drukker, we decided to focus on ways that radiomics can distinguish between benign lesions and luminal A breast cancers, as luminal A breast cancers are the most frequently diagnosed (about three-fourths of breast cancers) in the U.S.
Nathan and I worked together with members of the Giger Lab on the research from January to April 2017. In April we submitted an abstract of our initial findings to the meeting of the Radiological Society of North America (RSNA). In July 2017 we got word that our abstract had been accepted. I wrote the manuscript during the fall 2017 semester along with the co-authors and presented the research at the RSNA meeting the week after Thanksgiving. Along the way I submitted the paper to Academic Radiology and it was accepted pending revisions in late December. We addressed the revisions and got the good news about acceptance in April 2018.
Radiomics is a field in which you extract quantitative information from medical images and look for associations with disease (or lack of disease). It might surprise many, but sometimes and for a long while, medical images can be and have been used to provide a qualitative assessment: to identify if something is there or not, roughly how big it is, etc. However, there is so much more to be known! In radiomics, you can make all sorts of measurements, including how spherical a lesion is, or how rough it is, or how quickly it takes up contrast agent. The Giger Lab has developed 38 radiomic features that can be determined for breast lesions imaged with magnetic resonance. Then you can use classification methods with a large set of measurements and compare how different algorithms predict what the state of the lesion is versus what it was determined to be, later, from biopsy. You can also include additional information in classification, such as how patients responded to therapy, such as chemotherapy. The goal is to produce a set of predictions from retrospective data that can be used for future predictions on new images.
So, ideally you might be able to make radiomic measurements of a lesion and provide a physician with a quantitative probability, a prediction about how, based upon a large collection of previous results, the patient might benefit from one therapy versus another. Essentially, taking these measurements from images could act as a digital biopsy. Our work determined that certain radiomic measurements are useful in distinguishing between luminal A cancers and benign lesions, but that it’s not just a single radiomic feature that does this best. We first compared classification performance using just the linear size of the lesions against if you use the full set of 38 features, and found that using the set improved performance, and that the feature of irregularity was used the most in our algorithm. But it didn’t stand alone: a few texture features were always helpful too. We also found that when you removed size information from the set of features used for classification, the classification performance was equivalent, which suggests that size doesn’t add anything to the classification process.
This finding somewhat surprised us. You always want to make sure you make strong predictions based upon knowledge of the physical system you are working with, so it makes sense that more information that corresponds to the physiology of the lesion helps the classification; we know from surgical removal of lesions that benign lesions are smoother than cancerous ones, for example. But it is interesting to see how you don’t really have any difference in the classification when you don’t use size features, especially as that can be one of the first bits of information a radiologist thinks about when looking at an image.
I think the role of medical physics is always centered around providing more information to clinicians, who make the decision about including it in clinical care. To be truly a part of standard of care, findings like ours would have to be vetted and approved by the Food & Drug Administration, which takes a long time. I will note that this research involved magnetic resonance (MR) images of breast lesions, which in the U.S. is typically done for patients at high-risk for breast cancer or with a history of breast cancer, or for pre-surgical planning, or for follow-up after treatment. (At the University of Chicago, MRI is used as part of diagnostic workup in breast imaging, not just for high-risk patients.) The lab I am working in has also done many separate studies of radiomics from mammograms and ultrasound. There is an ongoing discussion in the medical field about the risks and benefits of mammograms vs. MR vs. ultrasound for breast imaging, with discussion about cost (MR is markedly higher in cost), length of time needed to do the exam, and the trade off in radiation dose from mammograms vs. dose of contrast agent in MRI. So hopefully our findings will be a small piece of the puzzle in larger, long-term clinical discussions about if MRI should be part of standard of care in breast imaging.
All in all, here are three major takeaways from my research:
First, medical physics is a really fantastic field and it is very unusual to have researchers in the field on faculty at small liberal arts colleges. I’m really excited to be able to talk about it frequently with students and include them in my research.
Second, there is a lot going on behind the scenes for a paper like this. This project might seem fairly small, but it involved dedicated work by many people, including researchers whose job it is to collect the images and make the radiomic feature measurements. While it was a relatively short period of work for me on this paper, it was possible only because of the over 30 year effort of my collaborator, Dr. Giger, and her associates to build up these methods. Federal funding has been very crucial in supporting the work over the years. It’s also important to be patient as research findings come out and as the field of medicine works through whether or not to bring them to clinical care. If you get a mammogram right now, you might see a note on the report that the radiologist’s interpretation was supported by “computer aided detection.” The research papers on that were beginning to come out almost 30 years ago, and Dr. Giger is a pioneer in the field.
Last, I think it is really important that women know how important it is to ask your doctor questions about your own breast cancer risk. I think many women know to think through family history, but there are other risk factors, such as breast density, which can be determined only through screening. So it’s great to start talking with your doctor at around age 35 about screening possibilities.
As far as what’s next for me, I am continuing to collaborate with the Giger Lab. The University of Chicago just appointed me for a second year as a Visiting Scholar, which is exciting. I don’t want to give too much away about our specific initiatives, as getting scooped is always a concern! But we’re working on some interesting things with classification of other molecular subtypes of breast cancer and deep learning, making use of images that are already collected as part of standard of care, and seeing what additional information we can get out of them.
Dr. Heather Whitney is associate professor of physics and a Visiting Scholar at the University of Chicago. Her research is in medical physics, and specifically how the principles of physics can be used to improve the measurements made by technologies such as ultrasound, magnetic resonance imaging (MRI), and other medical imaging tools. Her current work focuses on radiomics of cancer imaging.