검색 상세

Deep Learning for image guide using multi object classification and localization in US

초록/요약

As technology advances and societal demands increase, demand for POC (Point of Care) is increasing. Particularly, portable ultrasound system which can be carried is not only increasing the market shares in the ultrasound machine market, but also the market share of the entire medical appliance market is continuously increasing. Accordingly, more and more inexperienced users are using ultrasound device. For non-experts, it is difficult to detect ultrasound images like experts. Therefore, ultrasound image guide for unskilled users is required. In this paper, we propose a method to assist non-experts in detecting objects in ultrasound image. In thyroid ultrasound image, we propose multi-label classification to detect several thyroid objects, in addition to provide a localization of the thyroid objects without region-level annotation using CNN (Convolutional Neural Network). A novel method is that the network learns to localize multi objects using saliency maps without locations using guided backpropagation. The network architecture is designed to make possible multi-label classification with providing multi localization tasks. We represent the task results based on dataset consisting of images and video recordings of various view of thyroid.

more