Deep Learning for Pulmonary Nodule CT Image Retrieval — An Online Assistance System for Novice Radiologists

摘要

Lung cancer is one of the most common types of cancer originated from malignant lung nodules. Early detection of lung nodule is key in prevention of lung cancer. In this paper, we developed an online content-based image retrieval (CBIR) system to assist novice radiologists in identifying lung nodules. The system takes advantages of cloud computing and deep learning to retrieve similar lung nodules from a large database, which contains rich diagnostic information generated by experienced radiologists, to help novice radiologists diagnose lung nodules. The cloud computing platform provides a PC or Smartphone accessible interface and deep learning extracts semantic rich features for the retrieval. We utilized dynamic time warping (DTW), Euclidean and Manhattan distance measures to compute similarity between nodules. We evaluated the developed system on the large Lung Image Database Consortium (LIDC) and Image Database Resource Initiative (IDRI) databases, and compared the deep learning features with other hand-crafted featuresforlungnoduleretrieval. Oursystemwasabletoretrieve the most similar nodules in about 0.14 seconds with a best precision of 71.43% (when one nodule was retrieved) in terms of the five malignancy levels given by experienced radiologists. The improvement margin of the deep learning features over handcrafted features is in the range of [4.3% - 20.3%]. Overall, the proposed system offers an innovative online education tool for novice radiologists.