Latest News

 

Improving Cancer Treatment with Frontera

Published on September 3, 2019 by Aaron Dubrow



Results for a real tumor taken from the BraTS'18 TCIA dataset. The images show the tumor core (enhancing and necrotic tumor cells) indicated as a gray wireframe with the reconstructed initial condition (magenta volume) and parts of the patient brain geometry.

Gliomas are a particularly nasty form of brain tumor — fast-growing, deadly, and difficult to root out. But there is hope among physicians and researchers that even this most challenging foe can be managed more effectively.

George Biros, professor at The University of Texas at Austin, with joint appointments at the Oden Institute for Computational Engineering and Sciences and the Walker Department of Mechanical Engineering, is leading an effort to apply massive, high-speed computers, machine learning, and biophysical models of cells to the problem of diagnosing and treating gliomas.

"Despite 30 years of research on chemotherapy, radiotherapy, and surgery, these cancers are still very hard to treat," Biros said. "The goal is to combine machine learning and biophysical models to characterize tumors and help create hypotheses of how tumors evolve, with the ultimate goals of creating clinical decision support tools for diagnosis, prognosis, and treatment planning."

The project is a collaboration with professors Andreas Mang of the University of Houston, and Christos Davatzikos of the University of Pennsylvania.

Training a machine learning system is difficult and time-consuming in the best of circumstances. But when the results of the training have life-or-death consequences, the accuracy and confidence must be much higher than it would be for an image search or movie recommendation.

In Biros' work, machine learning is used for many aspects of the analysis from cleaning up images, to automatically identifying different structures in the brain, to extracting features that correlate with clinical outcomes.

George Biros, professor at the Oden Institute at The University of Texas at Austin, and an early user of Frontera.

As an early user of Frontera — the fastest academic supercomputer in the world — Biros is working to build bio-physical models of brain tumor development that include more factors than ever before, and train automated medical image processing systems to detect the extent of cancers beyond the main tumor growth, which must be removed during surgery to prevent the cancer from returning.

"We're focusing mostly on trying to augment this process with prior information," he said. "We know that as tumors grow, they interact mechanically with the surrounding healthy brain tissue. One hypothesis is that quantifying this interaction may give clues on specific mutations that drive the cancer. Another hypothesis is that if we can figure out where exactly the tumor started this will also give us information on specific mutations."

Knowing the specific mutations of a tumor can lead to better, patient-specific, chemotherapy and treatment plans. "With the help of the biophysical model we can also project forward in time to try to see what is the most likely infiltration area so we can be a little bit more aggressive with surgery," he said.

Working with partners at the University of Pennsylvania, Biros's predictive system is being used in critical trials to guide surgeons to take tissue from a prespecified location in the patient brain so it can be compared with the models.

On Frontera, Biros and his team are trying to train more complex models than have ever been created, containing parameters that capture how new blood vessels form, and how diverse types of cells within a tumor interact. Doing so means incorporating data from many patients.

"We can easily come up with models that have hundreds of parameters. But with these models, even to test out basic hypotheses, we need to conduct simulations on a big machine," Biros said. "The algorithm and application development and training need a big resource capable of a quick turnaround. Without Frontera, and the support we have received from the TACC staff, it would be impossible."

Frontera's unique design is critical to the future success of Biros' work.

"You need state-of-the-art resources to do science," he explained. "With Frontera, everything is integrated in the system — GPUs, CPUs, visualization, analysis, common file systems. That's exceptional, especially at this scale."

"You need state of the art resources to do science. With Frontera, everything is integrated in the system — GPUs, CPUs, visualization, analysis, common file systems. That's exceptional, especially at this scale."
George Biros, The University of Texas at Austin.

Despite growing interest in machine learning, training models across multiple or hundreds of nodes, instead of just a single computer, is rare and challenging.

"There are only a handful of research groups that are doing multi-node deep learning and machine learning research," Biros said. "For clinical image analysis, it's really necessary. Unless you explore the parallelism, you'll never be able to train different models to look at different parameters to test different ideas. Parallelizing the biophysical models and scaling them to multiple node machines is absolutely critical."

Biros also intends to use Frontera to explore the shape and dynamics of red blood cells so they can be more realistically incorporated into studies of vascular diseases, enabling researchers to develop better artificial valves and stents.

Any one of these problems would exceed the computing capacity found at most universities, even the most elite. But Biros is fortunate to work at UT Austin, which, for over the past decade, has been at the forefront of advanced computing.

Growing and evolving in tandem with the resources at TACC, Biros has been able to dream big, twice leading teams that won the ACM Gordon Bell Prize (a high distinction in supercomputing), and tackling problems that would be unthinkable without massive computing.

"We are extremely fortunate to be able to use Frontera and the other resources that TACC provide," said Biros. "We're grateful to TACC staff, The University of Texas at Austin, the National Institutes of Health, and the National Science Foundation for making this research possible."


This research is supported by the NSF Division of Computing and Communication Foundations, within the NSF Directorate for Computer and Information Science and Engineering: Award #1817048: SHF: Small: Algorithms and Software for Scalable Kernel Methods


Contact

Faith Singer-Villalobos

Communications Manager
faith@tacc.utexas.edu | 512-232-5771

Aaron Dubrow

Science And Technology Writer
aarondubrow@tacc.utexas.edu

Jorge Salazar

Technical Writer/Editor
jorge@tacc.utexas.edu | 512-475-9411