Researchers at Emory University are reporting a significant improvement in their ability to analyze and understand changes of cancer tumors over time thanks to HPC work done on a Keeneland Project supercomputer. Analysis of high resolution cancer tumor images that used to take weeks can now be completed in a matter of minutes on the hybrid GPU-CPU system.

Emory University researchers are leading the project, which also includes experts in pathology, microscopy imaging, image analysis, and HPC. The work involves analyzing, integrating, and correlating macro radiology, microscopy imaging, and genomic data, with the goal of bolstering the understanding of what is going on in brain tumors at the molecular, micro-anatomic, and macro-anatomic levels. Researchers hope to make progress in the general knowledge of cancer tumor morphology, the development of new treatments, and the use of HPC applications in this highly specialized field.

Thanks to a decrease in the price of high resolution scanners, researchers are amassing large sets of high resolution medical images. In this case, Emory University researchers are scanning gliomas, which are thin cross sections of tumors. An image of a glioma that is 50,000 pixels by 50,000 pixels can now be captured in a few minutes.

While acquiring the images is more cost effective than ever, the actual analysis of the images remained a challenge, according to a story by The National Institute for Computational Sciences at the University of Tennessee at Knoxville on the project. To overcome this challenge, the team of researchers developed methods that optimize the utilization of the Keeneland supercomputer to perform the analysis.

Keeneland is a National Science Foundation-funded partnership between Georgia Tech, NICS, Oak Ridge National Laboratory, the UT, NVIDIA, and Hewlett-Packard. The project has several HPC resources, including the Keeneland Full Scale System, which is a 615 teraflops HP Proliant SL250-based supercomputer with 264 nodes, where each node contains two Intel Sandy Bridge processors, three NVIDIA M2090 GPU accelerators, 32 GB of host memory, and a Mellanox InfiniBand FDR interconnection network. 

The techniques used by the Emory University team include a combination of image-analysis and artificial intelligence classification methods. According to the NICS story, the team is using the Keeneland system to run several analyses on the images, including: identification of the boundaries of the tumors’ nuclei; extraction of the shape and texture of each segmented nucleus; summarizing the values of the nuclei; clustering the images to form groups; and finally, running “association mining” routines that compare and correlate information from the images with genomic data.

The power of Keeneland has been instrumental in the team’s ability to get the work done in a timely manner, team member Tahsin Kurc of the Center for Comprehensive Informatics (CCI) program at Emory University tells NICS. “By carefully scheduling the operations to CPU cores and GPUs and distributing computation across multiple nodes, we are able to analyze one hundred and fifty 4,000 pixels by 4,000 pixels image tiles per second on 100 nodes of Keeneland,” Kurc said.

The team of researchers is led by Joel Saltz, professor and chair of Biomedical Informatics and director of the CIC program at Emory University.

Related Articles

‘Sandbox for Geeks’ Powers Open Medical Research

Intermolecular Lends Genomics Data to Materials Project

Oxford University Unveils Big Medical Data Mining Facility