Google published a scientific article which explains how it has managed to identify cancer through artificial intelligence more accurately than physicians experienced in the field of pathology. The company is using a version of its intelligent image recognition system originally developed for its standalone car design.
A “raw” or pristine version of the same system used in cars has been trained to identify cancer in high-resolution images of the patient’s body. Without any personalization, the system has already achieved good results. However, a customization that made the software analyze each pixel broadly and also with reduced zoom greatly improved the performance of the machine.
After further customization, including network training to examine images at different magnifications, in much the same way a pathologist works, Google shows that it is possible to train a model that compares or exceeds the performance of a physician who has not had the time limit to examine the same images.
At present, the company’s artificial intelligence can identify any type of cancer in digital images with 89% accuracy, when medical experts rate the accuracy of 73% by analyzing the same content without any time limit.
While performing better than doctors, Google’s system will not replace their work. The idea is for the software to do a preliminary analysis and apply markings to suspect points, giving directions to where the doctors need to look and work.
After that, the professional is the one who will validate this information and evaluate all the findings of the system, which should make it more and more precise.