Digging for Gold
By Beth W. Orenstein
Vol. 18 No. 12 P. 16
Radiomics allows radiologists to mine imaging data and enhance diagnoses and treatment decisions.
First came computer-aided detection (CAD) and diagnosis systems. Now radiomics, a more advanced form of CAD, is garnering attention and showing great promise, not only for oncology but also for other diseases and conditions including Alzheimer's, Crohn's, and cardiovascular diseases. Radiologists see great potential for radiomics and believe it will figure into the "value-added" debate about their role in medicine.
Maryellen L. Giger, PhD, a professor of radiology/medical physics at the University of Chicago, says the concept of radiomics has been around for decades. "Radiomics is basically a new term to describe the expanding field of CAD."
What's changed, Giger says, are the major advances in computing and technology. "We now have bigger data sets of medical images and faster computers, and, because we do, we are able to expand the features; we are able to mine more information from images," she says. Radiologists are able to use the quantitative information extracted from MRI, CT, and PET images to make more advanced, personalized prognoses and patient management decisions, she adds.
No Cause for Alarm
Some in radiology are afraid of radiomics and its siblings, deep learning and artificial intelligence (AI), and the impact they will have on radiologists as medical professionals, adds Hugo Aerts, PhD, director of the Computational Imaging and Bioinformatics Laboratory and an associate professor at Harvard University. While radiomics will likely change the role of the radiologist, in Aerts' eyes, it will only make that role more valuable.
"There's always going to be the need for the expert to approve the detection and diagnosis of these AI methods," Aerts says. "Computers can fly planes, but who would get on a plane if it didn't have a pilot on board?"
Years from now, when AI does the majority of the interpretation work, Aerts says radiologists will still be needed to supervise. And with radiomics, radiologists can get even more value out of the images than they ever could on their own, because the technology will allow them to make more accurate diagnoses and prognoses as well as have an even bigger say in what treatments are most likely to succeed for a particular patient, he adds.
The difference between radiomics and CAD, deep learning, or AI is the human factor, says Anant Madabhushi, PhD, a professor in the department of biomedical engineering and director of the Center for Computational Imaging and Personalized Diagnostics at Case Western Reserve University in Cleveland. The radiologist using radiomics is able to apply his intuition to what he sees and what the data sets tell him or her, he says.
"There are folks who are critical of deep learning because of its lack of intuitiveness," Madabhushi says. "Deep learning or AI is based on what the machine thinks. Radiomics combines the best the human reader and the machine can do, which takes it to the next level."
Robert J. Gillies, PhD, the Martin Silbiger Chair of Cancer Imaging and Metabolism and vice chair of radiology at the H. Lee Moffitt Cancer Center and Research Institute in Tampa, Florida, notes another fundamental difference between CAD and radiomics: CAD is usually a standalone diagnostic system, whereas radiomics is designed to be applied on big data in a dynamic way.
"There is a need to share data for radiomics to fulfill its full potential," Gillies says.
Like CAD, radiomics began in oncology. Thanks to support from the National Cancer Institute (NCI) Quantitative Imaging Network and other initiatives from the NCI Cancer Imaging Program, oncology is where radiomics is the most well developed, Gillies says. To date, the majority of the research has been in lung, brain, breast, and prostate cancers and in correlating features extracted from images to aid every step from diagnosis to the prediction of treatment response and monitoring of disease status, Gillies says.
While most of radiomics has been largely in the solid tumor space, "researchers are talking about it playing a role in blood-based tumors such as leukemia," Madabhushi says. "We just haven't gotten there yet."
Assessing Breast Cancer
Much of Giger's research has concentrated on integrating mined imaging data—radiomics—with other "-omic" data including genomics—imaging genomics or radiogenomics—to identify new integrated biomarkers that can serve as diagnostic or prognostic tools for breast cancer patients. Identifying these biomarkers could allow medical imaging to become a noninvasive way of diagnosing and developing treatment plans for these patients, she says. Giger anticipates that radiomics could allow imaging to be used as a virtual digital biopsy, especially when an actual biopsy is not practical, such as in screening or repeated assessments of response to therapy.
Currently, imaging is used to stage breast cancer and, initially, to manage the patient. A follow-up biopsy augments the staging, and treatment is planned in response to the clinical evidence. In one of her earlier studies, Giger and colleagues built models to compare the use of information extracted from the images with pathologic stage and surgically verified lymph node involvement. The extracted information included tumor size and radiomic features that described the "heterogeneity" within the tumors.
"We concluded that computer-extracted MRI phenotypes have promise for predicting breast cancer pathologic stage and lymph node status, and we expect that the integration of these image-based phenotypes with the pathology of the tumor will yield a stronger predictor," she says. The study was published in the Annual Review of Biomedical Engineering in 2013.
In another study, Giger investigated the correlation between quantitative MRI radiomic features and various cancers subtypes such as human epidermal growth factor receptor 2 (HER2). About one out of every five breast cancers is HER2 positive. The subtype is important because each subtype has a different prognosis and responds differently to different treatments, Giger says. The researchers observed statistically significant associations between the observable characteristics of the tumor and receptor status. They found that more aggressive cancers were more likely to be larger and saw greater diversity in their contrast enhancement texture.
"We concluded that computer-extracted MRI phenotypes show promise for high-throughput discrimination of breast cancer subtypes," Giger says. "We may be able to develop a predictive signature for assessing prognosis." The study was published in Breast Cancer in May 2016.
Still another Giger group study looked at radiomic features of invasive breast tumors and multigene profiling procedures. The researchers sought to determine whether the radiomic features and multigene assays could be used to predict breast cancer recurrence. Again, Giger says, the results were promising.
"We found some highly specific imaging-genomic associations," she says. "And they have great potential to make image-based diagnoses that can predict recurrence and drive the best treatments." This study was published in Radiology in November 2016.
Researchers at Case Western are also looking at whether pretreatment MRI scans of breast cancer patients can tell how patients will respond. "We looked at 116 patients who all had been treated with neoadjuvant chemotherapy prior to surgery," Madabhushi says. "We looked at their MRI scans prior to their chemotherapy and were able to find features on their pretreatment scans that could tell us whether they were going to respond to the chemotherapy or not. We were able to predict pathologic complete response. It's very, very exciting." The findings were published in Breast Cancer Research in May 2017.
Lung, Prostate, and Brain Cancers
The Case Western researchers have also done a number of studies involving lung cancer and radiomics. One study presented at the 2016 American Society of Clinical Oncology (ASCO) meeting concluded that texture and shape features extracted from within and around the lung nodule on CT images could identify nonsquamous nonsmall cell lung cancer (NSCLC) patients who could potentially benefit from pemetrexed-based chemotherapy, which is the standard. The group also presented a study of NSCLC patients at the ASCO meeting in May 2017 that found certain radiomic texture features between baseline imaging and scans taken two weeks following chemotherapy—nivolumab—could identify early clinical response to the treatment. The researchers were able to identify six features that most significantly changed between baseline and posttreatment scans and could determine which of these features were most significant in identifying responders vs nonresponders.
In a collaborative project, researchers at the Moffitt Cancer Center and Dana-Farber Cancer Institute in Boston analyzed CT image features from 262 North American patients and 89 European patients with NSCLC to guide their cancer treatment and predict their response to therapy. The researchers identified associations between the image features, molecular markers, biological pathways, and clinical outcomes. Among other things, they were able to determine certain sets of image features that could predict the stage of the tumor. They also found that the presence of certain biological and genetic markers drove tumor growth.
"We demonstrated that, with radiomics, it is possible to increase prognostic power," Gillies says. The study was published in the July 21, 2017, issue of eLife, a novel emerging journal in biomedicine founded by National Academy members and Nobel Prize winners.
In another study published in Scientific Reports in February 2017, Case Western researchers used MRI to help diagnose prostate cancer. The researchers found that the shape of the prostate and a compartment within the gland, called the transitional zone, was indicative of prostate cancer. They are doing follow-up research to determine whether features in the peripheral and transitional zones that are discernable on MRI suggest whether the cancer is aggressive or slow moving, which is critical when making treatment decisions.
Also, a team of researchers led by Pallavi Tiwari, PhD, a professor of biomedical engineering at Case Western, published a study in September 2016 in the American Journal of Neuroradiology in which radiomic features beat radiologists in determining whether changes seen on MRI were dead brain cells caused by radiation—radiation necrosis—or whether the patient's brain cancer had returned.
Researchers are turning to radiomics outside of oncology as well. Case Western researchers have developed an algorithm based on MRI measurements of the brain that appears to be able to predict Alzheimer's disease before patients have symptoms. They published their findings in a September 2017 study in Scientific Reports. The researchers developed an algorithm that integrates features of the hippocampus, glucose metabolism rates in the brain, proteomics (the study of proteins), genomics, mild cognitive impairment, and other parameters.
Researchers at Case Western are also studying features captured from MR enterography to develop a treatment score that could be used to predict patient response to therapy for Crohn's disease. "Satish Viswanath, PhD, an assistant professor of biomedical engineering at Case Western, and his team are looking to develop a radiographic enterographic treatment score that can help us distinguish between mild and severe Crohn's," Madabhushi says. "Knowing which Crohn's a patient has can tell us who might need more aggressive treatment and who might need less aggressive treatment." Future work in this area will include validation through a prospective multicenter study, Madabhushi says. Also, some researchers are looking at features on cardiac images to determine whether they are useful in predicting cardiovascular disease progression and prognosis.
Gillies believes the implementation of radiomics in clinical practice faces some substantial challenges. Although radiomics can be performed with as few as 100 patients, it takes time to capture and curate larger data sets, he says. Also, radiomics is a young discipline, and physicians, especially radiologists, are slow to change practice, he notes. Still, Gillies believes the conversion of digital images to mineable data will eventually become routine practice.
In Gillies' vision of the radiology reading room of the future, "Radiologists will add their knowledge to curate data at the point of care, identifying volumes of interest, and overseeing the extraction of radiomic data to automatically contribute to databases." In this way, Gillies says, "We can go beyond the current hundreds of patients in our cohorts to thousands or tens of thousands, which will be key to increase the power and potential of this approach."
Being able to determine who will respond to a particular treatment is important for a number of reasons, the radiologists say. One is time: There is no point in spending time, often weeks, on a treatment that a patient is not likely to respond to. Another is cost: Chemotherapy and immunotherapy treatments can be quite costly. If patients aren't likely to respond, it's better to spend the money on a treatment that is more likely to be effective. If radiologists can provide this information to clinicians by extracting it from the images they see, it can be extremely valuable to not only the patients but also the health care system, Madabhushi says.
"Nothing is 100% accurate," Giger says. "Not computers, and not humans." Giger says she's extremely optimistic that radiomics will have a positive impact on radiology and the role of radiologists, especially as radiomics is incorporated in the assessment of an increasing number of diseases. At the same time, she is very cautious.
"There is currently much hype about radiomics and deep learning," Giger says. "And thus, we have to be careful because people tend to mainly publish their positive results."
— Beth W. Orenstein of Northampton, Pennsylvania, is a freelance medical writer and regular contributor to Radiology Today.