Baidu, the Chinese search giant, is sometimes hailed as China’s Google. And just like Google, it has a bunch of innovative side projects which go way beyond straightforward search. We recently covered Baidu’s amazing handheld universal translator. Now the company has announced its latest feat: Using artificial intelligence algorithms to help pathologists better diagnose cancer.
The company has developed an A.I. that is capable of analyzing slides containing biopsied tissue. Reviewing these slides can be difficult, even for experienced pathologists, but Baidu’s deep learning technology is able to look for tiny tumor cells faster and with greater accuracy than previous approaches. In tests, the algorithm was able to outperform both a professional pathologist and the previous winner of the so-called Camelyon16 challenge, a competition intended to evaluate algorithms for automated detection of cancer metastasis in lymph node tissue sections.
“Using A.I. to analyze pathology images is a very challenging task,” Yi Li’s, a machine learning research scientist at Baidu, told Digital Trends. “A digitized pathology slide at 40x magnification often contains billions of pixels, which is too large for a neural network to process. As a result, the mega-image is divided into tens of thousands of smaller individual images so that a neural network can analyze each of them separately. What’s unique about our neural conditional random field (NCRF) algorithm is that it can look at multiple images — including the potentially cancerous region and its surroundings, simultaneously. This new capability significantly reduces the number of false positives [in the form of] misclassified normal cells.”
To its credit, Baidu isn’t keeping this technology to itself. Instead, it is making it available to the medical research community via open source in the hopes that it can help as many people as possible. (And, you know, help raise Baidu’s name value in the process!)
“We hope this open-sourced algorithm can serve as a high-quality baseline for future research in this area,” Li said. “The algorithm is only evaluated on a limited number of public datasets at this stage. However, the algorithm needs to be further assessed using much more clinically relevant data to prove it still maintains higher accuracy than experienced pathologists. Our team will continue improving the algorithm and collaborating with researchers with whom we can share new datasets.”
Li notes that the goal isn’t to replace doctors in performing this valuable task, but rather improving pathologists’ efficiency in their daily work. Should this algorithm work as well as hoped, doctors in the future will no longer need to spend hours looking at every slide from a biopsy, but just focus on the affected areas as identified by the algorithm.