Improving content-based retrieval of medical images through dynamic distance on relevance feedback

Content-based image retrieval approaches rely on automatic features extracted from images to perform similarity queries. The major drawback is that such features often do not satisfactorily represent what the users understand and expect from them, e.g. when searching for similar images. In order to deal with the gap between the user semantic interpretation of the images and what the system can automatically provide, relevance feedback techniques have been employed. However, it has been used without prior analysis about the distance function that best suits the user intention in each relevance feedback cycle and leading to the increase of such gap. Hence, in the present paper we employ user profiling in conjunction to content-based image retrieval and relevance feedback techniques to exploit the user intentions and to reach the best configuration according to the user intention in each relevance feedback cycle. To do so, we introduce a novel approach and a mediator architecture to enhance this process through user feedback and profiling, allowing to dynamically modify the distance function in each feedback cycle choosing the best one for each cycle according to the user expectation. Experiments have shown that the proposed method outperformed the traditional static distance approach, improving in up to 74% the precision of similarity search of medical images.