Posted: January 9, 2012
CASE STUDY: Using New Maps to Navigate Cancer Treatment
The Cancer Genome Atlas (TCGA) has been described as a map of the molecular landscape in cancer. Unlike a traditional map, which describes appearance, the TCGA map describes much more—it shows what cancer cells are doing with their genes and proteins and how this differs from what happens in normal, healthy cells. Having such a map is beginning to change the way doctors understand diseases like glioblastoma and how they search for better cancer treatments.
Meanwhile, every day in hundreds of clinics around the country, radiologists use another kind of map—the magnetic resonance imaging (MRI) scan—to gather images of tumors like glioblastoma. These images show the size, number and location of tumors, so that doctors can determine the best course of treatment for their patients. The images also can show whether tumors are responding to treatment.
Now, with support from the National Cancer Institute's (NCI) Cancer Imaging Program through the TCGA Radiology Initiative, researchers have developed a way to overlay these two types of maps to improve cancer care. Analyses have revealed imaging markers in glioblastoma MRI scans that appear to match up with some of the key molecular properties identified by TCGA researchers. This discovery could be a first plank in an important bridge between the genomics lab and the clinic, revealing what glioblastoma is doing in its earliest stages, when it could be more treatable.
"The TCGA Radiology Initiative is providing genomics researchers a way to move their science into the clinic, so that doctors can use molecular and imaging markers together to benefit their patients," explains Carl Jaffe, M.D., who ran the Clinical Trials Branch of NCI’s extramural Cancer Imaging Program from 2001 to 2008. “We’re trying to help build virtual research communities that involve experts from very different backgrounds to achieve a common goal,” he said.
This convergence of two different approaches could have a dramatic impact on cancer medicine. Currently, surgeons usually perform a biopsy to help diagnose a patient’s cancer. Researchers take a sample of tumor tissue to see the molecular detail of how a cancer is progressing or responding to treatment. But if a series of MRI scans uncovered an “imaging signature” linked to those genetic and molecular markers, it could be a tremendous step forward in cancer diagnosis and prognosis.
“Sometimes you need to step back and look at the big picture,” explained Scott Hwang, M.D., Ph.D., a neuroradiologist at Emory University, who is focusing on glioblastoma as part of the TCGA Glioma Phenotype Research Group. “That’s what MRI provides. If it turns out that certain features we see on pictures of the brain match up with specific patterns of gene expression, we might be better able to diagnose and predict that cancer’s behavior better than if we were using either of those methods alone.”
However, the ability to see the big picture objectively can be affected by the viewer. Subjectivity can be introduced to cancer imaging not only through the visual interpretation of the person who is looking, but also through his or her memory and experience viewing similar pictures. An additional factor is whether the person looking at the image can explain to others what he or she has seen, so that everyone understands and agrees about what the picture shows.
Controlling the Variables
The seeds of the TCGA Radiology Initiative were sown in 2003, when NCI invited Adam Flanders, M.D., from Thomas Jefferson University to explore a project called the Repository of Molecular Brain Neoplasia Data, or Rembrandt, for short. NCI researchers had collected and characterized a number of tumor specimens from patients who had glioblastoma or a related brain neoplasm called low-grade glioma. The patients also had numerous MRI brain scans during their treatment. Dr. Flanders, a neuroradiologist, was asked to figure out how to classify the information derived from those images to compliment the rest of the Rembrandt data.
During this process, he developed a tool called the VASARI framework, which today provides a controlled, descriptive vocabulary set for MRI readers across the TCGA Glioma Phenotype Research Group, a network that includes radiologists and researchers at six sites. This vocabulary reduces the errors that can come from subjectivity.
Along with the digital MR images, VASARI is loaded onto a computer workstation at each network site. At least three network neuroradiologists score the same set of images according to a list of 26 different features. The radiologists receive instructions and examples to guide their responses for each VASARI feature.
The views and features that VASARI provides are not new––they include the most common ones that any neuroradiologist would notice when looking at a brain MRI. But each image is preselected to give the best view of that feature, and the readers must confine their interpretation to answering a number of specific multiple-choice questions about that particular feature.
The radiologists are not surrendering their professional judgment, of course. Many of the features require measurement, judgment or estimation. But those elements are confined to selection from multiple-choice lists that are built into VASARI, rather than coming from the unstructured language that doctors might normally use in conversation to describe their expert opinion.
“This approach gives us a common language and scale that we can all use to be consistent,” said Chad Holder, M.D., also a neuroradiologist, and colleague of Dr. Hwang’s at Emory University. “We need to have a way to score these features objectively.”
“Not only do we develop regularized data that can be shared and corroborated across the sites,” said Dr. Hwang, “but we also use the computer to do quantitative imaging, which allows us to characterize more precisely aspects of the tissue that can be measured, the things your eyes can’t do.” To accomplish this, biostatisticians work with raw data from the neuroradiologists to employ statistical tools that look for patterns and correlations across the features described in VASARI.
For example, when examining a set of MR images from a 45-year-old man with glioblastoma, VASARI feature f21 asks the radiologists, “Does the brain tumor extend into the deep white matter of the brain, and if so, where?” The radiologists look at all of the images necessary to answer this query, and then choose from among the structured responses, which include “no”, as well as a list of locations within the deep white matter where tumor may be visible: the brainstem, the corpus callosum, or the internal capsule.
Results presented at major meetings, but not yet published, are promising: the structured responses for six of the 26 features have a statistically significant association with how long patients survive. One of those responses predicted survival in 72 percent of patients, and the accuracy increased to 82 percent when information about four specific genes was included.
With these early successes, the TCGA Radiology Initiative is expanding from glioblastoma to other TCGA cancer types. The next cancer selected for study will be breast cancer, and the experience with VASARI will be re-adapted to the features for which mammography radiologists normally look. Again, the institutions that contributed the breast tumor tissue samples to TCGA will provide the MR images of those same patients. As these images are collected by the Cancer Imaging Program they are being made freely available on The Cancer Imaging Archive.
“Doctors and patients will have a clearer picture of what's really happening inside,” said Dr. Jaffe. That picture will show not only where and how big tumors are, but also how aggressive they are; whether the patient can get more-or-less toxic treatment; and whether the tumors are responding to the treatments that have been given, he explained. “These are all factors that can make a huge difference in a patient's quality of life.”