Radiologists’ work may soon become a lot more efficient. With inroads in computer vision and deep learning, they will someday use computer algorithms to more efficiently analyze cancer tissue. The field is proceeding slowly but surely.
What’s driving the field forward are new capabilities in data storage that, until a few years ago, were unimaginable. “Computers could only scale up so far,” Dr. Chris Pal, associate professor of computer software and engineering, Polytechnique Montreal, the engineering school of the University of Montreal, told MedicalExpo e-magazine.
“Today it is a lot easier to collect data; hard drives are cheap; and you can put more processors on a chip,” he added. “We will soon be able to reproduce human images on a number of benchmark problems, within the realm of usability.”
Interest in building better clinical decision support tools led Dr. Ronald Summers to pursue deep learning, which he thinks will add value to medical imaging in reducing errors. Summers is a senior investigator in the imaging biomarkers and CAD Laboratory, at the National Institutes of Health Clinical Center in Bethesda, Maryland. He is a key player in research efforts at the federal level in the United States.
Recognizing Patterns Better Than a Human
Deep learning constructs many layers of abstraction to help map machine inputs to progressively higher-level representations. Sometimes mocked as the stuff of science fiction, it is making incremental advances in medical imaging and seems poised for broader acceptance.
Machine learning relies on big data and the bigger the database, the better the pattern recognition becomes.
In medical imaging, the ultimate goal of machine learning is to recognize patterns better and faster than humans can, which ratchets up accuracy and productivity. Artificial intelligence uses deep neural networks, a form of machine learning, to train medical imaging machines. Machine learning relies on big data and the bigger the database, the better the pattern recognition becomes.
“So far, deep learning is accurate for such tasks as non-medical image analysis and speech recognition,” he said. “Right now, although there have been large improvements, deep learning is not as good as the experts for most medical image diagnosis tasks.”
In a recent review of the current status of the field, Dr. William M. Wells III, of Harvard Medical School and Brigham and Women’s Hospital, Boston, identified some of the most “powerful capabilities emerging” in the areas of “segmentation and registration of medical images, and registration of medical images, and representations of shapes of anatomical structures in individuals and populations, to name a few.”
Uses for Detecting Cancer
Only a small number of peer-reviewed studies with machine learning and cancer have been performed, but early results are laying the foundation for the field. Dr. Summers described results by his group, which show impressive improvements in sensitivity for three areas of oncology.
In detecting polyps of the colon, sensitivity rose from 58% to 75%; for cancer of the spine, sensitivity rose from 57% to 70%; and for lymph node analysis, improvement in sensitivity was especially striking, rising from 43% to 77%.
Dr. Igor Barani, CEO of the deep-learning start-up Enlitic, San Francisco, told MedicalExpo that Enlitic analyzed lung CT scans with its deep-learning system, finding it was 50% better at classifying malignant tumors, with a false-negative rate of zero, compared with 7% for humans.
In separate research using MR images to evaluate glioblastoma tumor segmentation with deep neural networks by Mohammad Havaei, Chris Pal, and others, the authors report superior architecture depiction and speed. “It not only offers far more detail and accuracy than state-of-the-art studies, it is also 30 times faster,” said Pal, associate professor at Polytechnique Montreal.
Pal sometimes teams up with the Montreal deep-learning start-up Imagia, which is devoted to working on cancer and deep learning. Alexandre Le Bouthillier, CEO of Imagia, told MedicalExpo: “Deep learning is key to the future of medical image processing because of its ability to merge large and diverse data sources to make more accurate predictions. We long ago reached our limits with conventional radiology.”