Scaling Up Image Analysis
Automated tools enable scientists to evaluate vast numbers of images in a fraction of the time, yielding new insight into brain cancer diagnoses and outcomes
The In Silico Brain Tumor Research Center (ISBTRC) at Emory University is leveraging imaging standards and algorithms in a variety of ways to gain new understanding of the factors that drive tumor progression in human brain cancer (glioma). As an In Silico Research Center for Excellence, Emory University is one of six caBIG®-designated centers that are demonstrating the effectiveness of in silico research methods in speeding research.
Comparing "Apples to Apples"
In many cases, a patient will progress from a relatively slow-growing tumor (called an astrocytoma) to a more aggressive and lethal form of cancer called glioblastoma. This progression is usually marked by specific, characteristic patterns that can be seen in MRIs of tumors.
"One of the things we're trying to understand is what's going in with a tumor's progression and how that correlates to observable changes in nuclear morphology, the molecular and genetic classifications, and the gene expression profiles of the tumor," explains Joel Saltz, M.D., Ph.D., the program director for the Emory ISBTRC. "Our interest is the integrative analysis of a number of complementary types of information, ranging from "omics" data that is captured by TCGA (The Cancer Genome Atlas) to imaging data."
To accomplish this, Saltz and his team have heavily leveraged caBIG® resources to combine these disparate types of data, as well as to ensure consistent analysis and markup of radiology images.
"By using these standards one has a systematic way to describe 3D elements of the tumor and mark up the area, rather than rely on a verbal description of what a radiologist sees," explains Saltz. "So, you're able to quantify and analyze that information in concert with other relevant data."
The ISBTRC is using publicly available data sources that are funded by the NCI and hosted by caBIG®, including TCGA, REMBRANDT, and VASARI to access a rich collection of information on tumor pathology, genetic mutations, gene expression patterns and radiology images.
Adam Flanders, M.D., a neuroradiologist at Thomas Jefferson University, has been deeply involved with the REMBRANDT and VASARI projects. He's now working with the ISBTRC group to develop a set of standard image "features" and a controlled vocabulary to describe those features that can be used by radiologists anywhere via a simple web-based form. This process helps to ensure uniformity of annotation and diagnosis, and can provide objective criteria to measure tumor progression. "Medical imaging has a lot of value not only in diagnostics, but in finding associations of image features with outcomes, genomics, histology, and response to treatments. It's a big piece of doing personalized genetics for patients with cancer."
"The resources developed by the ISBTRC and our collaborators allow us to ask and answer critical scientific questions about tumor progression that advance the field in ways that haven't been possible before now," states Dan Brat, M.D., Ph.D., the lead neuropathologist for the TCGA project, and the scientific lead for the ISBTRC.
Automating Image Analysis
The use of these standards is yielding a similar benefit when applied in pathology. Traditionally, pathology slides are manually viewed and analyzed by a human using a microscope, which is an accurate but time-consuming process since a single slide may contain 100,000 to one million cela. In the case of tumor samples, these cells may be a mixture of normal cells and tumor cells at various stages of progression, making it essential to examine large numbers of samples.
"We have developed a number of different, complementary algorithms, each of which check each other and look at slightly different aspects of the cells. In this way we can analyze 90 million cell nuclei in a few hours, something no human pathologist could do," explains Brat. "We're finding particular nuclear features that are correlated with specific mutations and outcomes. Using these new tools, which can quantify up to 28 different morphological characteristics of cell nuclei, we've been able to classify these cases of brain tumors very accurately—our results match those from a panel of six neuropathologists from TCGA."
A small hospital typically deals with "tens of thousands" of slides a year, making it virtually impossible for pathologists to perform comprehensive analyses of the requisite number of tumor slides manually. As the technology for creating high resolution digital images of pathology slides ("virtual slides") has improved, the need for algorithms to quickly classify and annotate these images has increased. These advances in hardware and software are creating accurate disease models for tumor progression and driving the migration to digital data acquisition for pathology. The Emory team is working with computer scientists from the Oak Ridge National Laboratory to scale up these tools for use on supercomputers, allowing for even faster analysis of these vast digital pathology data collections.
"Pathologists certainly won't go away, but we can make their jobs easier and more integrative by adding the ability to look at vast quantities of raw data and correlate the pathology information with radiology and molecular data in ways that have never been possible," says Saltz.
Seeing the Benefits
Even at this early stage, the work done by the ISBTRC has already generated tangible benefits to clinical researchers and those treating patients with brain tumors.
For example, tools are now available to help physicians differentiate between two forms of brain cancer—astrocytomas and oligodendroglioma—which are each associated with very different outcomes. The team from Emory has developed algorithms that help clinicians more accurately distinguish between these two forms of cancer based on a computer analysis of cellular morphology on digitized slides, and on specific gene expression patterns. These tools are expected to be made broadly available to pathologists within six months. In addition to providing immediate benefits to patients, these tools can be used to help educate radiologists and oncologists to more accurately recognize the tumors in the real world.
"The number of questions we can ask and answer is actually limitless, and we're just starting to discover the capabilities of the datasets and our own skills," says Brat.
Additional work examining gene expression patterns that differentiate low grade tumors from more aggressive forms of cancer has yielded methods to identify the precursors of these characteristic patterns in tumors at earlier stages, providing valuable insight about patient outcomes to their physicians. This work has recently been submitted to the online journal PLoS ONE.
In addition to the image analysis tools described previously, the ISBTRC makes use of a wide variety of caBIG® tools including caB2B (for gene annotation), caIntegrator (to create translational web portals for data presentation), NBIA (to manage radiology images), caMicroscope (to view digitized pathology slides), AIME (caGrid service for image annotation, developed at Emory), XIP/AVT (an image analysis toolkit), and multiple images viewers including OsiriX iPAD and ClearCanvas. These tools are connected to caGrid, the services-oriented information architecture developed by caBIG® to facilitate data sharing, and the tools and data developed as part of their research will be made publicly available.
