The first tests of Google’s AI in cancer diagnosis have had excellent results without changing anything.
When it comes to cancer detection and diagnosis, time is of the essence; early treatment can make the difference between the life and death of the patient. That is why the pressure on oncologists is even greater than that of other specialists.
And it is not a simple job; in many occasions it consists of the detailed study of scanner images; With 10 gigapixel images at 40x magnification, it is quite a methodical task in which it is easy to skip a pixel that does not look as it should. As a demonstration of how difficult this is, some studies claim that when faced with the same images, any two experts only interpret them 48% of the time.
Using machine learning to diagnose cancer
Google believes it can help with this time-consuming and resource-consuming task; your idea is to use GoogLeNet, the same machine learning system that you have developed for other projects, like your autonomous car. In effect, the same AI that detects if a rider is approaching, can detect tumors.
In fact, the striking thing is that Google has not had to change anything. As it stands, the AI has been able to analyze patient scanner images; With machine learning, an AI is capable of learning by itself from thousands of samples, without having to be programmed to perform a certain action.
From these images, the AI generated heat maps, areas that clearly differed from healthy patients and where there could be tumors; They were not perfect, though, as they contained too much noise and therefore it was sometimes difficult to discern where the tumor had been detected.
The surprise of Google’s IA in the diagnosis of cancer
Sure, he didn’t stay here. The engineers modified the system to adapt it to the analysis of this type of images; It was then that the results were more than surprising. The system got a FROC accuracy score of 89%, greater than 73% that a pathologist with unlimited time can achieve.
Google is quick to calm hopes; We cannot expect a perfect system that detects tumors instantly based on just a few scans. To use it in real applications, it would be necessary to make more adaptations; and you will always need the support and knowledge of an expert in the field. By itself it will not be able to function as well.
In any case, the results allow optimism; Google’s AI has gotten a higher score than other systems with the same images. And better yet, it seems to adapt very well to different types of images and detections.