Indexing Large Collections of Tumor-Like Shapes

We investigated the problem of retrieving similar shapes from a large medical database of tumor shapes (‘find tumors that are similar to a given pattern’). We used a natural similarity function for shape matching based on state-of-the-art concepts from Mathematical Morphology, and showed how the function can be lower bounded by a set of features extracted from the shapes, thus leading to “correct” output (i.e., no false dismissals), a key requirement for medical applications. These features can be organized in a spatial access method, leading to fast indexing for range queries (‘Find objects within distance e of the given object.’) and nearest neighbor queries (‘Find the first k closest objects to the query object.’). In addition to the lower-bounding, our second contribution is the design of a fast algorithm for nearest neighbor search, which achieves significant speedup while provably guaranteeing correctness. Our experiments demonstrate up to 27 times better performance for the proposed method compared to sequential scanning. We also verified that the similarity function matches human perception of shape similarity, with experiments on human subjects obtaining 80% precision for up to 100% recall.