Investigation on continual training of computer-aided diagnosis systems by semi-supervised learning

Medical image analysis systems can help radiologists in reading images accurately and promptly. The systems are expected to be robust to images obtained with any imaging systems by different vendors and at different facilities. However, because of the prevalence of diseases and difficulty of collecting samples with various characteristics, the systems may not apply well to all the images. Artificial intelligence-powered systems are expected to learn from experience and be improved continuously. In this study, we investigated whether continual training with local samples can improve system performance using mammography and lung CT databases. Different training strategies using unlabeled data are compared.