Baidu has presented a new artificial intelligence, a deep learning algorithm capable of tumor detection much better than a human. In fact, the algorithm, which is also open source so that it can be used by others, will help human pathologists to pinpoint tumors more easily.
Baidu presented its deep learning algorithm for tumor detection
The artificial intelligence system was trained with 400 images, so large that they can occupy several GB of size, divided into grids of tens of thousands of smaller images. Then, the scientists randomly selected 200,000 of those smaller images and tested the algorithm.
The level of analysis accuracy of the deep learning algorithm is much more reliable, so it detects more tumors than any specialists, just by looking at some images.
It’s very similar to Google’s Artificial Intelligence microscope which also detects tumors much faster (and more effectively) than a professional human pathologist with many years of experience. In the case of Google, however, the artificial intelligence is combined with human expertise.
The Baidu deep learning algorithm identifies cell by cell and finds out if it is a cancer cell or if it is a standard cell. And, through spatial correlation, it can make accurate predictions, as explained by Yi Li from the Baidu Research Artificial Intelligence Laboratory, in Silicon Valley.
The Baidu’s artificial intelligence outscored any human pathologist
In the image above we see in (a) the original slide, in (b) the pathologists’ notes, where the white areas are tumors, in (c) the tumor regions identified by an Artificial Intelligence that does not consider neighboring patches, and in (d) the result that the Baidu deep learning artificial intelligence algorithm considers.
In the FROC (free-response receiver operating characteristic) curve, which measures average tumor detection sensitivity, Baidu AI algorithm scored 80.9 out of 100, higher than the highest human pathologist’s score of 72.4 out of 100.