AI Achieves Near-Human Detection of Breast Cancer
Pathologists still do the bulk of their diagnosis of metastatic cancer cells in tissue and lymph nodes by hand, putting slides under a microscope and looking for signature irregularities they’re trained to see. Recent advances in computer technology, however, particularly in artificial intelligence (AI), have begun to teach machines to do this kind of detection with growing rates of improvement.
Now, a research team from Beth Israel Deaconess Medical Center (BIDMC) and Harvard Medical School have developed a form of AI that can interpret these pathology images with accuracy levels of 92.5 percent. That's not too far below the human detection rate of 97 percent. Moreover, when the two are used in combination, the detection rate approaches 100 percent (approximately 99.5 percent).
Their AI method is a form of deep learning, in which the system attempts to replicate the activity of the human neocortex through artificial neural networks. The goal was to teach the machine to interpret patterns and structures. Andrew Beck, director of bioinformatics at the Cancer Research Institute at BIDMC and an associate professor at Harvard Medical School, is co-author of the technical report describing these findings, recently uploaded to arXiv.org [PDF], an open access archive. He tells mental_floss, “We use a subset of AI where you are trying to train the computer to do something in a data-driven way to learn model parameters and to make predictions on new examples.”
To teach and test the AI, they input 400 whole slide images—270 for teaching and 130 for testing. Some of the slides contained metastatic breast cancer lymph node tissue, and some healthy tissue. The team was able to identify which slides the computer was more prone to making mistakes about—primarily by flagging false positives—and used those examples to re-train the computer to improve its performance.
They submitted their system to the International Symposium of Biomedical Imaging (ISBI), where they placed first in two categories in the ISBI's Camelyon Grand Challenge 2016, up against private companies and academic research institutions from around the world. According to ISBI’s website, the goal of this challenge is "to evaluate new and existing algorithms for automated detection of metastases … in stained whole-slide images of lymph node sections."
Beck was surprised at how efficient the system turned out to be. “I was impressed at how well the computer did, because it is really a complicated visual task," he says. "The cancer can take a bunch of different appearances and the normal lymph nodes, too. To think that a single model in a purely data-driven way could accurately make this classification was surprising.”
It did a much more accurate job of detecting cancer than a recent study that reported that pigeons had an 85 percent accuracy rate of detecting breast cancer individually; when the scores of a flock of four were combined, they had a 99 percent accuracy rate. Beck feels that associating the two studies is like comparing apples to oranges because his study wasn’t diagnosing breast cancer, but breast cancer in lymph nodes, he explains. "It wasn’t trying to separate normal breasts from pre-invasive breast lesions and breast cancers."
Moreover, he says, "I think you can imagine computers being worked into the workflow much more simply than pigeons.”
One especially positive application of this kind of AI is its ability to remove some of the burden of detection from the pathologist, who can then focus more on treatment plans and patient health. “You can imagine that in the future the computer will keep getting better. I can see things evolving where pathologists move away from the more tedious, lower-level tasks because there are higher-level, more integrated things that humans are way better at than computers,” says Beck. For instance, the computer could count all of the individual cells.
It may also help solve diagnostic errors by improving accuracy in combination with the manual method. Further research by his team will continue to test the system by expanding the types of cancer used, and increasing the number of slides. “This could be integrated into existing workflows to make the process faster, more accurate and hopefully more cost effective, ranging from the clinic, to research in pharmaceutical companies, to global health,” says Beck.