September 9, 2019 — Screening mammography is an important tool against breast cancer. But it has its limitations. A “normal” screening mammogram means that a woman doesn’t have cancer now. But doctors wonder whether “normal” images contain clues about a woman’s risk of developing breast cancer in the future. Also, most women “recalled” for more tests when their mammograms show suspicious masses don’t have cancer. With the help of XSEDE’s Extended Collaborative Support Service and Novel and Innovative Projects, scientists at the University of Pittsburgh Medical Center are using the XSEDE-allocated Bridges-AI supercomputer at the Pittsburgh Supercomputing Center to run artificial intelligence programs meant to determine the risk of developing breast cancer and to prevent false recalls.
Why It’s Important
Despite a lot of progress in improving survival and quality of life for women with breast cancer, the disease remains a major threat to women’s health. It’s the most common cancer in women and is either the first or second most common cause of cancer death for women in the largest racial and ethnic groups, accounting for 41,000 deaths in 2016 alone.
Screening mammography is an important tool for getting early warning, when the disease is easiest to treat. But it’s not perfect. For women whose scans show no signs of breast cancer, doctors wonder whether that scan may contain information they could use to predict future risk. More than 10 percent of women who get mammograms are “recalled” for further testing. But nearly 90 percent of the time it’s a false alarm. That’s something like 3 million women in the U.S. who go through the stress of unnecessary recall each year.
“We collected a large set of images from UPMC’s digital screening mammography … We wanted to see if ‘normal,’ or negative, digital mammography images in screening were predictive for the risk of breast cancer in the future. There’s also the very critical breast-cancer screening issue that when a lot of women come for mammography and when radiologists visually assess their images, it’s not certain in many images whether cancer is present or not. This creates a difficult decision-making process on whether to ask these women to come back for additional workup,” said Shandong Wu, University of Pittsburgh Medical Center
Expert radiologists can tell a lot from a modern digital mammogram. But Shandong Wu and his colleagues at the University of Pittsburgh Medical Center (UPMC) wondered if artificial intelligence (AI) could detect subtle hints in mammograms that the human eye can’t see. They tested their “deep learning” AIs on digital mammograms from UPMC patients whose status was already known, running the programs on the XSEDE-allocated Bridges and Bridges-AI systems at the Pittsburgh Supercomputing Center.
How XSEDE Helped
The task for the UPMC scientists’ deep-learning software was a big one. Each digital mammogram is more than 2,000 by 3,000 pixels large—that’s dozens of megabytes of data for each image. And to do their study, they needed to “train” and then test their AIs on thousands of images. Their AIs are also pretrained on large datasets with tens of thousands of images. The size of the data for AI modeling was enormous, making the computations slow on the computers available to the researchers at their own laboratory.
“[Roberto] was really helpful. He converted a Matlab container to Singularity, and he wrote a wrapper to run Matlab R2019a on the DGX-2. Sergiu Sanielevici [leader of XSEDE’s Novel and Innovative Projects] has been very helpful in supporting our research and he regularly inquires our progress and needs, making sure our issues are properly addressed. Tom Maiden, Rick Costa and Bryan Learn also assisted us in solving problems. Without the support of XSEDE, I don’t think we would have been able to do this work.”—Shandong Wu, University of Pittsburgh Medical Center
Working with experts from XSEDE’s Novel and Innovative Projects program and Extended Collaborative Support Service (ECSS), including Roberto Gomez and other ECSS staff at PSC, the team used the graphics processing unit (GPU) nodes of Bridges-AI to train and run their AIs. Deep learning, which works by building up layers of different kinds of information and then pruning connections between the layers that don’t produce the desired result, tends to work best on GPUs. The new NVIDIA “Volta” GPUs in Bridges-AI contain accelerators, called “tensor cores,” specifically designed for deep learning. Bridges-AI’s GPU nodes combine eight to 16 GPUs each for up to 512 gigabytes of extremely fast GPU memory. The large memory available to PSC’s GPU nodes was central to the success of the AIs, bringing the computation time down from weeks to hours. The NVIDIA DGX-2 node, deployed in Bridges-AI as a first for open research, and its massive memory were particularly useful.
When an AI is designed to produce a binary result—yes or no, positive or negative—scientists often report that result as a graph of true positives versus false positives. The larger the “area under the curve,” or AUC, the better the AI’s accuracy. AUC can range from zero to one, where zero means that the classifier has no predictive value and one implies perfect prediction. Versions of the screening AIs had an AUC of 0.73 when predicting whether a woman with a negative mammogram would develop cancer over the next 1.5 years. Better, the recall AIs could tell the difference between women with cancer and those who would have been recalled even though they didn’t have cancer with an AUC of 0.80. These results are promising, and could have value for clinical use after further evaluation.
“In this kind of work—medical images—we deal with larger-scale and sometime 3D volumetric data. All those images are high-resolution images … and we have more than 10 thousand [of them]. Our local GPUs did not have enough memory to accommodate such a scale of data for AI modeling. It could take weeks to run one experiment without the support of powerful GPUs. Using the GPUs from XSEDE, with larger memory, reduced that to a couple of hours,” said Shandong Wu, University of Pittsburgh Medical Center.
With XSEDE support, Dr. Wu’s lab is developing several other AIs to improve breast cancer diagnosis. One would pre-read digital breast tomosynthesis—a kind of 3D breast imaging method—to reduce the time it takes radiologists to read the scans. Another is designed to automatically identify and correct mistakes in the labeling in a dataset for AI learning. Finally, the scientists are also working on AIs to predict breast cancer pathology test markers and the recurrence risk for women who’ve already been diagnosed with breast cancer.
Further work beyond these preliminary results will compare the benefits of improved AIs against the current methods used by doctors. The aim is to improve care and lower cost in real-world clinical practice. The UPMC team reported their results in five oral presentations at the Radiological Society of North America (RSNA) Annual Meeting in Chicago last year, three presentations at the Society of Photo-Optical Instrumentation Engineers Medical Imaging conference in San Diego this year, and several upcoming journal manuscripts.
Source: The Extreme Science and Engineering Discovery Environment (XSEDE)