Editor’s note: This is one of five profiles of finalists for NVIDIA’s 2017 Global Impact Award, which provides $150,000 to researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.
When it comes to brain tumors, it’s all in the genes.
Cancer begins when genetic mutations make cells grow uncontrollably. Spotting certain mutations can point the way to the most effective treatments for the disease.
Doctors now do this by testing tissue samples collected during surgery. But Dr. Bradley Erickson, a Mayo Clinic neuroradiologist, taps the power of AI to predict brain tumor genomics using MRIs.
Tuning in to Radiogenomics
His method could give doctors easier access to invaluable genetic information, which can help predict how fast a tumor will progress, and if it will respond to specific drugs and other treatments.
The work, called radiogenomics, “reflects the almost unthinkable thought that in the appearance of images, we can figure out the genomic properties of tumors,” Erickson said.
This achievement has placed Erickson and a team of Mayo Clinic researchers among five finalists for NVIDIA’s 2017 Global Impact Award. We award annual grants totaling $150,000 to two teams of researchers using NVIDIA technology for groundbreaking work that addresses social, humanitarian and environmental problems.
MRIs Beat Tissue Test
Although surgery is still required to remove a brain tumor, Erickson’s GPU-accelerated deep learning technique could lead to an earlier and more accurate way to diagnose and treat brain tumors. It’s also a way to track the tumor’s progress or its response to treatment without surgery. (Erickson will speak at our GPU Technology Conference, May 8-11, in Silicon Valley.)
“You’d think the tissue test would be 100 percent correct, but it’s not,” Erickson said. According to a study by Johns Hopkins Medicine, about half of patients with tumors examined with a genome test get results that are potentially misleading.
Sometimes tissue tests show chromosomal damage when it’s not there. Testing for some mutations requires a large amount of suitable tissue, which often isn’t available after the biopsy. Even if doctors have enough tissue, there’s no way to guarantee the mutation will show up in that particular sample.
AI to Predict Brain Tumor Genomics
In one set of experiments, researchers identified an alteration in a gene connected with glioblastoma multiforme – the most common and deadliest type of brain tumor – that interferes with DNA repair. Cancers with this change to the MGMT gene (called methylation) usually respond better to chemotherapy and radiation than to radiation alone, Erickson said. If the tumor doesn’t have this variation, doctors can opt for a treatment with fewer side effects.
The team also identified a type of chromosomal damage that’s important for predicting how well patients with a low-grade type of brain tumor will respond to chemotherapy and radiation.
Erickson’s team trained their neural network using MRIs of tumors with and without the genetic mutations. They used the CUDA parallel computing platform and a range NVIDIA GPUs with cuDNN. They deployed their algorithm using our Tesla P40 GPU accelerators and other GPUs.
“It’s always been a dream to use computers to help with tumor image interpretation, but the technology wasn’t there,” Erickson said.