Cancer sucks, and finding it early often means the difference between life and death. That’s why Google has thrown its convolutional neural network (CNN) at digital slides, or images, of tissue samples to detect tumors in breast cancer patients with surprising success, according to a white paper titled Detecting Cancer Metastases on Gigapixel Pathology Images.
The CNN was originally designed for self-driving cars, but when given the chance to detect the spread of tumors on 100,000 x 1000,000-pixel images, the research group at Google reported a 92.4 percent tumor detection rate.
Normally, tissue sample slides are manually reviewed by pathologists slowly and carefully. This can be an arduous and sometimes inaccurate task. It’s been shown that pathologists agree at a rate of under 50 percent, meaning that false positives and false negatives are fairly common. Google’s project throws an artificial intelligence at the cancer prognosis equation as a way to help pathologists along the way.
The technology isn’t meant to be a replacement for human pathologists, mind you. The AI still generates false positives, and isn’t able to find abnormalities that a human would. However, given that a computer can shuffle through thousands of images and doesn’t experience fatigue, the CNN can crunch through thousands of images and lower the amount of time a pathologist would need to spend to look for what it is he or she looks for.
It’s excellent news for pathology and cancer treatment and research, and great news for the future of AI as something that does much more than just get us to the airport and steal our jobs.