AI beat oncologists again
The neural network is better than living doctors saw signs of the re-formation of tumors
Polina Gershberg, Naked Science
In an article published in Nature Communications (Yamamoto et al., Automated acquisition of explicable knowledge from unannoted histopathology images), scientists from the RIKEN Center for Advanced Intelligence in Japan talked about new opportunities for medicine, which were demonstrated by the artificial intelligence they created.
The technology developed by Japanese scientists in collaboration with a number of university hospitals was able to successfully detect relevant features on the images of tissue samples of cancer patients without annotations that human doctors could understand. In addition, the AI found signs correlating with the risk of cancer recurrence, which were not previously noted by pathologists. As a result, the forecast of the computer program turned out to be more accurate than the diagnosis based on the conclusion of doctors. The best results were shown by forecasts that combined the predictions of AI with the predictions of doctors.
Figure from the press release Artificial intelligence identifies previously unknown features associated with cancer recurrence – VM.
These works, according to the authors, can contribute to understanding how AI can be safely used in medicine, helping to solve the problem of the "black box". As long as people teach AI themselves, it is impossible to get knowledge from it beyond what we know today. The authors of the work used the technology "teaching without a teacher". Instead of "teaching" AI medical knowledge, they used uncontrolled deep neural networks known as autoencoders. Researchers have developed a method for converting the functions detected by AI – initially only numbers – into high-resolution images that can be understood by humans.
To do this, the team took 13188 images of prostate tissue preparations at Nippon Medical School Hospital (NMSH). The complete data was very large. When divided into fragments for deep neural networks, about 86 billion patches were obtained. Array processing was performed on a powerful AIP RAIDEN supercomputer. Without diagnostic annotation, AI has learned to use pathological images from 11 million patches.
The signs detected by AI included diagnostic criteria for cancer on the Gleason scale, used all over the world. In addition, he found features in the stroma – a special connective tissue that supports internal organs – in areas not associated with cancer. Experts did not know about these specific signs of relapse before, and yet, from the point of view of prognosis, this turned out to be an important and useful diagnostic criterion, even more accurate than the actual diagnosis on the Gleason scale.
After testing the AI on the data obtained at NMSH, the scientists tested their findings on images of patients from other hospitals to exclude the specifics of the technique or protocols – and the results were confirmed.
According to the first author of the work, Yoichiro Yamamoto, this technology can make a great contribution to personalized medicine. It can help to make highly accurate predictions of cancer recurrence based on images of samples. At least it is suitable in the case of prostate cancer, on which it was tested, but probably for many others.
AI can help to find new features of diseases that have not yet been recorded by humans. In addition to improving the accuracy of diagnostics and improving the quality of medical care, artificial intelligence based on such technologies can be used to search for new knowledge in other areas outside of medicine.
Portal "Eternal youth" http://vechnayamolodost.ru