Deep Lesion Graphs in the Wild: Relationship Learning and Organization of Significant Radiology Image Findings in a Diverse Large-scale Lesion Database

摘要: Radiologists in their daily work routinely find and annotate significant abnormalities on a large number of radiology images. Such abnormalities, or lesions, have collected over years stored hospitals' picture archiving communication systems. However, they are basically unsorted lack semantic annotations like type location. In this paper, we aim to organize explore them by learning deep feature representation for each lesion. A large-scale comprehensive dataset, DeepLesion, is introduced task. DeepLesion contains bounding boxes size measurements 32K lesions. To model similarity relationship, leverage multiple supervision information including types, self-supervised location coordinates sizes. They require little manual annotation effort but describe useful attributes the Then, triplet network utilized learn lesion embeddings with sequential sampling strategy depict hierarchical structure. Experiments show promising qualitative quantitative results retrieval, clustering, classification. The learned can be further employed build graph various clinically applications. We propose algorithms intra-patient matching missing mining. Experimental validate effectiveness.