This time, the research group uses deep learning and non-hierarchical clustering, which are machine learning methods, to automatically acquire the characteristics of cancer without being taught by humans and output it as information that humans can understand. Was developed.

　When this technology was applied to prostate pathological images without doctor's diagnostic information (about 11 billion divided images for AI learning), AI diagnoses cancer used all over the world. In addition to the criteria, we also found a previously unknown lesion called "changes in the stroma outside the oncology area."

　More than 15,000 prostate pathological images (for AI learning) of three university hospitals, Nippon Medical School Hospital, St. Marianna University Hospital, and Aichi Medical School Hospital, to confirm whether these factors found by AI are useful for predicting recurrence. It was verified using about 960 billion images as divided images.As a result, it was found that recurrence can be predicted with higher accuracy than the diagnostic criteria for prostate cancer currently used all over the world, and the combination of the features found by AI and the diagnosis of a pathologist may further improve the prediction accuracy. Do you get it.

　The results of this research can be used as a highly accurate method for predicting cancer recurrence, and the developed technology will be useful as an automatic analysis method for acquiring new knowledge from images.In addition, AI is sometimes called a black box, but since this technology can extract information that can be understood by humans, it can be expected to contribute to the realization of AI that can be used safely in medical treatment.

Paper information:[Nature Communications] Automated acquisition of explainable knowledge from unannotated histopathology images