Unknown Features Related to Cancer Recurrence Identified by Artificial Intelligence

2019-12-30

The artificial intelligence technology developed by the Advanced Intelligence Project (AIP) at the RIKEN Center in Japan has successfully identified unknown features from pathological images of human cancer patients, without annotations, but understandable to human doctors. In addition, the AI has discovered some features related to cancer prognosis that were previously unnoticed by pathologists, resulting in higher accuracy in predicting prostate cancer recurrence compared to pathological diagnosis. By combining the predictions made by AI with those made by human pathologists, even higher accuracy can be achieved. According to Yoichiro Yamamoto, MD, the first author of this study published in the journal Nature Communications, "This technology contributes to personalized medicine by enabling highly accurate prediction of cancer recurrence through the acquisition of new knowledge from images. It also helps address the 'black box' problem of artificial intelligence, thus facilitating the safe use of AI in medicine." The research team collaborated with several university hospitals in Japan and employed an unsupervised learning method called "autoencoder" for the AI to learn without being "taught" medical knowledge. The researchers developed a method to convert the features discovered by AI (initially represented as numbers) into high-resolution images understandable to humans. To accomplish this feat, the team obtained 13,188 complete prostate pathology slide images from the National Medical School Hospital (NMSH) in Japan, resulting in a massive dataset equivalent to approximately 86 billion image patches (sub-images of the deep neural network). The calculations were performed on the powerful RAIDEN supercomputer at AIP. The AI learned from 11 million image patches using pathology images without diagnostic annotations. The discovered features included cancer diagnostic criteria used worldwide, as well as the stroma (connective tissue supporting organs) in non-cancerous areas that experts were unaware of. To evaluate these discovered features, the research group validated the performance of recurrence prediction using the remaining cases from NMSH. The team found that the predictions made by AI based on the discovered features (AUC=0.820) were more accurate than those made by pathologists using the human-established cancer standard Gleason score (AUC=0.744). Furthermore, combining the features discovered by AI with the human-established standard resulted in even higher accuracy in predicting cancer recurrence compared to using either method alone (AUC=0.842).

RECENT AI NEWS

RECENT AI TOOLS