SHAPE-BASED RETRIEVAL IN MEDICAL IMAGE DATABASES

点击翻译按钮获取中文摘要

DESCRIPTION (adapted from the Abstract): Retrieval in Medical Image Databases is greatly facilitated by queries which use visual and graphical examples. In such queries, the user points to features of interest in an example image and the database is expected to retrieve other images which have similar features. This proposal seeks to create the capacity for retrieval by shape similarity in an echocardiographic image database. The proposal has three aims. The first aim is to create robust shape features. Much of the previous work on shape analysis uses shape features which are invariant to certain image transformations. Unfortunately these features are very susceptible to noise and are not practically useful. A primary aim of this proposal is to seek alternate and more robust shape features. Preliminary work by the investigators and others suggests that there are non-invariant features which satisfy these demands, and the proposal seeks to investigate these in more detail. The second aim is to create novel structures and algorithms for indexing these features. Classical indexing techniques, such as indexing trees, are applicable for indexing invariant features. The proposal seeks to extend these techniques substantially for indexing non-invariant shape features. The extension is based on the application of the theory of interval valued arithmetic to indexing. Finally, the proposal seeks to validate the mechanism of retrieval by shape similarity in echocardiography. First, the robustness of the new shape features will be compared to the robustness of old shape features by adding noise to the data. Second, the efficiency of retrieval with the new indexing mechanisms will be measured by simulations and by evaluating the performance for the echocardiographic image database. Lastly, the utility of shape similarity retrieval will be evaluated by measuring the performance of the database to that of expert cardiac radiologists.