Imperial College Uses GPUs to Spot Brain Damage
Editor’s note: This is one of five posts profiling finalists for NVIDIA’s 2016 Global Impact Award, which provides $150,000 to researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.
Bumps, blows, blasts all cause traumatic brain injuries, which are fast becoming a top public health concern.
To help diagnose damage to our most complex organ, researchers at Imperial College London are deploying sophisticated image analysis tools, powered by GPUs and deep learning.
Computational methods are at the core of work by a team led by Ben Glocker, a lecturer at Imperial’s computing department. Taking advantage of the rich data in biomedical scans, their system provides automated, image-based assessments of traumatic brain injuries at speeds other systems can’t match.
The work has placed Glocker and team among five finalists for NVIDIA’s 2016 Global Impact Award. Our annual grant of $150,000 is given to researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.
Fast Response Time
While millions sustain traumatic brain injuries each year, most aren’t life threatening. But anything disrupting the brain’s normal function can lead to mental disabilities or emotional problems, such as impaired reasoning or depression, with unknown lifelong consequences.
The variety of such injuries can make them hard to identify and treat. And a fast response during medical emergencies is critical, something Glocker saw firsthand working as an ambulance driver in Germany while fulfilling his civil service requirements.
“Doctors need to see what’s happening to the organs and brain, and are making decisions based on what’s in emergency room images,” Glocker said. “What we’re doing with computing technology is helping doctors make better informed decisions.”
Looking for Patterns
Ph.D. student Konstantinos Kamnitsas and others on Glocker’s team developed their image analysis technique using one of the most successful types of deep learning techniques for computer vision — convolutional neural networks, which apply thousands of filters to an image sequentially to look for patterns.
The Imperial team applied these GPU-accelerated computational tools to 3D medical scans. This computationally intense brain lesion segmentation method was developed using a cuDNN-capable Python library called Theano, which allows CUDA-accelerated computations and the full use of NVIDIA’s latest technology.
Using the extended memory of various GeForce GTX graphics cards let the team experiment with bigger networks. A recent deployment of a cluster of 16 NVIDIA Tesla K80 GPU accelerators in the research lab gave a further boost.
“GPUs enable fast, efficient algorithms and have the right hardware infrastructure for this type of image analysis — we couldn’t do this without them,” Glocker said. “Analyzing a single brain scan can be done in two minutes using a GPU. Without one, it can take hours and hours and hours.”
The result is a pipeline for automatic detection and segmentation of brain lesions tied to traumatic brain injury that outperforms other state-of-the-art segmentation systems. The team is collaborating with research groups and hospitals because their work could accelerate clinical research in medical imaging and have an immediate impact on diagnosis and therapy.
The accomplishments of Glocker and his team “contribute to a paradigm shift in the use of MRI for developing and delivering treatments for traumatic brain injury,” said Professor David Menon, head of the Division of Anesthesia at the University of Cambridge.
The team found their method worked well on stroke victims and patients with brain tumors. It can also be used for cardiac and fetal imaging.
The winner of the 2016 Global Impact Award will be announced at the GPU Technology Conference, April 4-7, in Silicon Valley.
More Global Impact Award 2016 nominees
Check out the work of last year’s NVIDIA Global Impact Award winner.