Efficient Similarity Search on 3D Bounding Box Annotations

Searching for similar image regions in medical databases yields valuable information for diagnosis. However, most of the current approaches are restricted to special cases or they are only available for rather small data stores. In this paper, we propose a fast query pipeline for 3D similarity queries on large databases of computed tomography (CT) scans consisting of minimum bounding box annotations. As these box annotations also contain background information which is not part of the item that was actually annotated, we employ approximate segmentation approaches for distinguishing between within-object texture and background texture in order to correctly describe the annotated objects. Our method allows a compact form of object description. In our framework, we exploit this advantage for enabling very fast query times. We have validated our method on data sets of 111 and 1293 bounding box lesion annotations within the liver and other organs. Our experiments show a significant performance improvement over previous approaches in both runtime and precision.