Visual Feature Selection/Dynamic Visualization

by falk for Scaffold Hunter

Current state of the art:           Scaffold Hunter (SH) was designed as a Visual Analytics tool for chemical space. Recently, a plugin was developed that leverages SH for the visual analysis of content-based medical image retrieval (VAMIR). Different from traditional approaches relying on textual annotations and queries, content-based image retrieval (CBIR) exploits visual features that in the medical domain may comprise segmented anatomical or pathological regions, their spatial relationships, volume and texture as well as Wavelet- and Fourier-transformations. A typical CBIR framework presents the user a result list sorted according to some similarity metric that accumulates all features used for image comparison. This way, no insight is given on which and how features have contributed to the ranking. In a novel approach, the VAMIR plugin allowed SH to visualize the outcome of a CBIR algorithm in a way that enables the user to make judgements about a query’s similarity to other database entries with respect to selected features.           Motivation for this project:         The tool’s performance is promising, but VAMIR is still at a prototype stage. A current shortcoming consists in the static nature of the feature set that is visualized, which needs to be specified in a config file before execution. Also, no hint is given as to which feature set is likely to profit from a visual analysis (with respect to the Visual Analytics core mantra of “discovering the unexpected”). Implementation of methods to dynamically select a set of features to be visualized as well as computerized reasoning behind a meaningful starting set of features would mean an essential step away from VAMIR’s prototype stage towards a deployable application.