Radiomics: The Digital Biopsy

Radiomics is a powerful application of the work of biostatistics, which explores the application of the rational analysis of mathematics to the chaotic and seemingly irrational variability of life. Biostatistics is inherently interdisciplinary work, operating at the crossroads of pure and applied mathematics, statistics, biology, medicine, and artificial intelligence. Lorin Crawford, RGSS Assistant Professor of Biostatistics, says that the need for collaboration is part of what has made his department so successful.

“I wanted to merge these disparate fields, to go to a place that was flat in the sense that, if I wanted access to any kind of community, I could do that. What I love about Brown is not only is it small, but Brown recognizes it’s small, and so its best way to be successful is via collaborations.”

Jón Steingrimsson, assistant professor of biostatistics, agrees. “The problems are a lot more weighted by real world data and collaborations with subject matter experts. You’re much more incentivized to collaborate rather than doing your own pure methodological work.”

Diagnostic Over Therapeutic

Brown SPH has a long and distinguished history as a collaborator in this field. Constantine Gatsonis, professor of biostatistics, joined the American College of Radiology Imaging Network (ACRIN) upon its formation in 1999 as a statistician. Funded by the National Cancer Institute, with an initial grant of $23 million over five years, ACRIN was one of the first national, multi-center research groups focused on diagnostic, rather than therapeutic, medical technologies.

In addition to advances in imaging technology, ACRIN took advantage of other key technological developments, becoming one of the first multi-center studies to collect almost all its data through the new technology of the internet. ACRIN’s interdisciplinary leadership included medical, surgical, and radiation oncologists, patient advocates, and a statistician. Rather than targeting a specific type of cancer, ACRIN investigated the potential of imaging analysis to screen, detect, diagnose and treat cancers across the board.

ACRIN merged with Eastern Cooperative Oncology Group to form ECOG-ACRIN in 2012, the same year that researchers coined the term “radiomics” to describe the extraction of biologically relevant quantitative features from radiological images. It now comprises 1300 member institutions and around 15,000 physicians and research professionals.

Fenghai Duan, associate professor of biostatistics, who joined the Department’s research track in 2008, does most of his research within ECOG-ACRIN. Lung cancer kills more people worldwide than any other cancer. As part of the National Lung Screening Trial (NLST) research team, Duan’s work analyzing the success of screening for lung cancer—which using either X-rays or CT (computed tomography) demonstrated that the sensitivity of CT scans allowed for earlier detection and earlier treatment, leading to a twenty percent relative reduction in lung cancer specific mortality compared to X-rays, leading in 2011 to an immediate change in national screening guidelines.

But both CT scans and X-rays produce a very high number of “false positives,” turning up nodules that more than 90% of the time turn out to be benign and pose little health threat, which is why screening guidelines depend on risk factors—mainly age and smoking history. Public health policymakers must balance the harm of false positives—including anxiety and unnecessary surgery—with the harm of failing to detect cancer until it is too late to treat. Duan and other researchers began to not just look for nodules, but to analyze the features of the nodules they found—size, location, shape, and so on. By connecting these data to patient clinical history and outcomes, Duan and his collaborators developed a pulmonary nodule risk prediction model that could distinguish with much greater accuracy which nodules would develop into full-blown cancer.

But the predictive power of this model was limited by the data on which it was based, and the computing power to process those data. As computer power has increased, so has the quality of radiologic imaging, with increasing resolution producing images of greater and greater granularity. “Rather than just a few features of the nodule, radiologists can now provide hundreds or even thousands of imaging features,” Duan said. “That’s what we call radiomics: massive quantitative features associated with an image of the nodule.” Radiomics harnesses this fount of new data with the new techniques of machine learning, yielding vastly more complex and accurate prognostic tools. In the last two years, Duan has shown that the radiomics-based model he has developed in collaboration with pulmonologists and radiologists is better at distinguishing benign from malignant nodules than the standard of care risk assessment.

This non-invasive, non-surgical “digital biopsy” doesn’t only tell you whether a mass is cancerous. Duan is now exploring whether this approach may be able to better predict outcomes than physical tissue biopsies. Through Deep Learning techniques, Duan thinks it is possible that computers will surpass doctors in their ability to predict medical outcomes.

Predicting Alzheimer’s Disease

Cancers are complex and hard to predict, but, having developed new models to predict lung cancer survival based on tumor heterogeneity, Ani Eloyan, assistant professor of biostatistics, is now tackling an even more complex and enigmatic organ—the human brain. Eloyan’s goal is to better predict the trajectory of disease for people with early-onset Alzheimer’s. Alzheimer’s research has undergone a revolution in the past two decades. Post mortem analyses of the brains of Alzheimer’s patients have shown that they often feature unusual plaques of amyloid proteins, but these plaques were largely not detectable in CT scans of living brains. Advances in Positron Emission Tomography (PET) scan allowed researchers to image brains of living patients with unprecedented detail, even identifying plaques in patients without symptoms. Brain scans would “light up” where amyloid plaques were forming, says Eloyan. Trials of pharmaceutical interventions to slow or prevent the formation of plaques were greeted with tremendous optimism, but ultimately yielded little clinical benefit. Alzheimer’s researchers now suggest that rather than thinking of Alzheimer’s as a single disease, we should think of it as a set of diseases, with multiple pathways to similar clinical outcomes. As Eloyan puts it, “Not just ‘Alzheimer’s or not’ but what type of Alzheimer’s is this—what clinical function is most affected?”

With this more nuanced and complex understanding of the disease, Eloyan is working to build models that correlate patterns revealed by brain scans with specific patterns of impairment. Her work involves deep collaboration with physicians, radiologists, and neurologists. “I think the best research comes when we work with collaborators who understand more about the biology.”

“Ten years ago, a lot of the Alzheimer’s studies were very small. So you could have a finding and then it wouldn’t hold in the future.” The Longitudinal Early-Onset Alzheimer’s Disease Study (LEADS) is funded by the National Institute on Aging in the National Institutes of Health, and aims to follow 500 cognitively impaired participants and 100 cognitively normal participants at multiple sites around the country. The Alzheimer’s Disease Neuroimaging Initiative—another multisite study—has been running for longer than LEADS and Eloyan says has collected thousands of images. “There’s a lot of data, and so hopefully some of what we are finding will hold up better.”

For the moment, however, Eloyan is wary of discussing how her work might translate into treatments for Alzheimer’s, but her work is paving the way to better understand new treatments and assess how well they work with much greater precision. “We’re not comparing therapies,” she says. “We are trying to identify what are the main markers we should look at in clinical trials. If you want to show that a certain treatment is helpful, then you have to show effects on this marker, that it’s changing in this way over the course of treatment.”

The research is also limited by the expense and accessibility of PET scans. So far, Eloyan’s data has been drawn entirely from American study participants, something that Eloyan, who grew up in Armenia, is acutely aware of. “I just found out four months ago that they acquired the first PET scanner in the country. Granted it’s a small country. But until four months ago, anybody that wanted to get a PET scan would have to travel internationally to do it.” International research is also limited by clinical challenges. “If you run the same test, but in a different language, things may be different. This is another aspect of biostatistics—How can we deal with this? What kind of data can we collect that would be universal, and what are the data that would have to be analyzed differently or incorporated differently because of language and differences across cultures. There’s a lot that goes into these discussions.”

A Biostatistics Omnivore

Perhaps no-one is working harder to expand the scope of the field of biostatistics than Lorin Crawford. Crawford is the RGSS Assistant Professor of Biostatistics affiliated with the Center for Computational Molecular Biology, as well as a senior researcher at Microsoft New England, and in the last two years has received both the Alfred P. Sloan Research Fellowship and the Packard Fellowship. Crawford’s approach is omnivorous and wide-ranging, concerned more with pushing the methodological envelope than with any specific problem or application. “We use shape variation as a way to explain genotypic or phenotypic variation,” says Crawford. Crawford’s work on glioblastoma parallels Duan’s work on lung cancer, demonstrating that structural patterns of tumors better predict key outcomes than gene expression.

Crawford is now working on an analytical tool that could parse the key topological features of almost any object, from teeth to proteins. Called SINATRA (for Sub-Image Analysis using Topological Summary Statistics), the tool aims to “bridge the gap between the new conceptual technologies and new physical technologies we have,” says Crawford. “I love the name SINATRA. It’s an acronym, it stands for something, but it’s also a project that takes everything we do in my lab—the genetics, the statistics, and the really hard math—and kind of mixes it all together into one cohesive ‘sound.’” The model was trained on data from primate teeth, collaborating with anthropologists to analyze the structural features that differentiated the teeth of carnivores and herbivores—which have a more limited range of variation—what Crawford calls “low intra-class heterogeneity”—than tumors. But Crawford has his sights set on some of the hardest challenges—and complex shapes—known to science. Working with Was Shing Tang, a physics graduate student, and other members of his lab, Crawford is developing SINATRA Pro to analyze the highly flexible structure of proteins.

Crawford says that Brown is uniquely conducive to this sort of innovative, interdisciplinary work. “The thing I love about people at Brown is other people also mostly work at the intersection of different fields. You’ll have someone like Brenda Rubenstein [Joukowsky Family Assistant Professor of Chemistry] who’s in chemistry, but also knows a lot of physics, and a lot about statistics and Gaussian processes and stuff. At Brown, you have a department, but if you look beneath the surface people have all these different skill sets. So it’s easy to meet someone and say ‘These are the things I’m interested in,’ and someone else would be like ‘I’m in EEB [Department of Ecology and Evolutionary Biology] and I’m interested in skeletal movement,’ and you’re like ‘Oh, wow, this would be really cool!’”

Crawford believes we are only just beginning to tap the potential of radiomics. “As technology gets better, I think there are going to be more opportunities to think about noninvasive approaches to studying complex diseases.” The development of three dimensional and textural imaging techniques have opened new ways of understanding diseases, which can be harnessed using innovative analytical techniques—statistics, data science, and machine learning. The key, in Crawford’s view, is collaboration. “As people continue to explore and use tools from seemingly disparate disciplines, I think we’re going to see more and more innovative solutions to these problems. People don’t work in silos anymore, which is fantastic. I tell my group, you have to speak multiple academic languages.”