Primary brain tumors encompass a wide range of tumors depending on the cell type, aggressiveness and stage of the tumor. Characterizing a tumor quickly and accurately is a very important part of treatment planning. Currently, the identification process is reserved for trained radiologists, but in the future, computing and high-performance computing will play a supportive role in diagnosis.
George Biros, professor of mechanical engineering and leader of the ICES Parallel Algorithms for Data Analysis and Simulation Group at the University of Texas at Austin, has worked for a decade to create accurate and efficient computing algorithms that can characterize gliomas, a common and aggressive type of primary brain tumor.
The Biros’ team tested their new method in the Multimodal Brain Tumor Segmentation Challenge 2017 (BRaTS’17).
Their system scored in the top 25 percent in the challenge and was near the top for whole tumor segmentation.
"The competition is related to the characterization of abnormal tissue on patients who suffer from glioma tumors, the most prevalent form of primary brain tumor," Biros said. "Our goal is to take an image and delineate it automatically and identify different types of abnormal tissue – edema, enhancing tumor (areas with very aggressive tumors), and necrotic tissue. It's similar to taking a picture of one's family and doing facial recognition to identify each member, but here you do tissue recognition, and all this has to be done automatically."
For the challenge, Biros and his team were provided with 300 sets of brain images on which all teams calibrated their methods.
In the final part of the challenge, groups were given data from 140 patients. They had to identify the location of the tumors and segment them into different tissue types over two days.
The image processing analysis and prediction pipeline the team used had two main steps: a supervised machine learning step where the computer creates a probability map for the target classes, and in the second step they combined these probabilities with the biophysical model that represents how tumors grow in mathematical terms. This imposes limits on the analyses and helps find correlations.
TACC computing resources allowed the Biros team to use large-scale nearest neighbor classifiers. For every voxel or 3D pixel in an MR brain image, the system tries to find similar voxels in the brains it has already seen to determine if the area represents a tumor or a non-tumor.
With 1.5 million voxels per brain and 300 brains to assess, the computer had to look at half a billion voxels for every new voxel of the 140 unknown brains that it analyzes, deciding for each of the voxels is a tumor or healthy tissue.
"We used fast algorithms and approximations to make this possible, but we still needed supercomputers," Biros said.
Each of the steps in the analysis pipeline uses separate TACC computing systems. The nearest neighbor machine learning classification component simultaneously used 60 nodes on Stampede2, TACC’s latest supercomputer. The team used Lonestar 5 to run biophysical models and Maverick to combine the segmentations.
Most teams had to limit the training data they used or apply more simplified classifier algorithms on the whole training set, but priority access to TACC’s ecosystem of supercomputers allowed Biros’ team to explore more complex methods.
"George came to us before the BRaTS Challenge and asked if they could get priority access to Stampede2, Lonestar5 and Maverick to ensure that their jobs got through in time to complete the challenge," said Bill Barth, TACC's Director of High-Performance Computing. "We decided that just increasing their priority probably wouldn't cut it, so we decided to give them a reservation on each system to cover their needs for the 48 hours of the challenge."
Biros and his team were able to run their analysis pipeline on 140 brains in less than 4 hours and correctly characterized the testing data with almost 90 percent accuracy.
Their method is fully automatic and needed only a small number of initial algorithmic parameters to assess the image data and classify tumors without hands-on effort.
The team’s scalable, biophysics-based image analysis system was the culmination of 10 years of research into a variety of computational problems.
"In our group and our collaborators' groups, we have multiple research threads on image analysis, scalable machine learning and numerical algorithms," he explained. "But this was the first time we put everything together for an application to make our method work for a really challenging problem. It's not easy, but it's very fulfilling."
The BRaTS competition represents a turning point in the research.
"We have all the tools and basic ideas, now we polish it and see how we can improve it," Biros said.
The image segmentation classifier will be deployed at the University of Pennsylvania by the end of 2017, in partnership with Christos Davatzikos, a collaborator of the project and director of the Center for Biomedical Image Computing and Analytics and Professor of Radiology. It won’t be a substitute for radiologists and surgeons, but it will improve the reproducibility of assessments and could speed up the diagnosis process.
The methods the team developed go beyond brain tumor identification. They can be applied to many problems in medicine, as well as physics.