ImageCLEF 2004: Combining Image and Multi-lingual Search for Medical Image Retrieval

This article describes the technologies used for the various runs submitted by the University of Geneva in the context of the 2004 ImageCLEF competition. As our expertise is mainly in the field of medical image retrieval, most of our effort was concentrated on the medical image retrieval task. Described are the runs that were submitted including technical details for each of the single runs and a short explication of the obtained results compared with the results of submissions from other groups. We describe the problems encountered with respect to optimising the system and with respect to finding a balance between weighting textual and visual features for retrieval. A better balance seems possible when using training data for optimisation and with relevance judgements being available for a control of the retrieval quality.

The results show that relevance feedback is extremely important for optimal results. Query expansion with visual features only gives minimal changes in result quality. If textual features are added in the automatic query expansion, the results improve significantly. Visual and textual results combined deliver the best performance.