Computers were more successful than doctors in the detection of lung cancer based on the results of the scanner in the new research and offered a new insight into the future of artificial intelligence in medicine.
Identifying patterns and interpreting images – the same skill with which people read the results of the scanner – is a very promising area when it comes to artificial intelligence.
By storing huge amounts of data in systems called artificial neural networks, researchers have been able to train computers to recognize patterns associated with a particular situation.
This system monitors the algorithm and, in the meantime, learns. The more data is stored, the more accurate it is. This process is called deep learning and is already used to enable computers to understand speech and register objects – so the autonomous car recognizes the stop sign.
Google has created systems that help pathologists to read microscopic slides, and ophthalmologists in diagnosing eye diseases.
Researchers from Google and several other medical centers in the new research have applied artificial intelligence to the results of a scanner recorded for the diagnosis of lung cancer, which has only claimed 1.7 million lives in the past year.
In addition to detecting cancer, the scanner also identifies spots that can develop into cancer later, so radiologists can classify patients in risk groups and assessments in which patients need biopsy or more frequent control. The scanner, however, can override a tumor or benign spot replacements for malignant, and different radiologists who examine the same image can have completely different opinions.
The researchers thought that computers might be better able to do a job than a doctor. They created a neural network and gave her photos of patients whose diagnoses had already been determined.
“The whole experimental process seems to be like teaching a student at school,” says Dr. Daniel Ce from Google, the author of the research article. “We use huge data sets to train the machine, give us lessons and quizzes so that we can begin to learn what cancer is and what it wants, and what will not develop in cancer in the future.”
The system was tested in 6,716 cases in which the diagnosis is already known and in 94 percent of cases it has been diagnosed with accurate diagnosis.
Compared to six extraordinary radiologists, the deep learning system had less false positives and negative diagnoses.
Dr. Erik Topol, who writes about artificial intelligence in medicine, but did not participate in the study, said: “I am pretty sure this discovery will be useful, but it must be proven first.”
A radiologist who misinterprets the recording can harm one patient, while the artificial intelligence system that does not work can perfectly harm many, warns Dr. Topol. Before this system is used in practice, he adds, it should be tested in real terms.
“We work with institutions around the world to find out which is the most productive way to implement this technology in clinical practice,” says Dr. Ce. “We do not want to run in front of the ore.”