TAU MIPLAB at ImageClef 2008

This paper describes the participation of Tel Aviv University Medical Image Processing Laboratory group at the ImageClef 2008 medical retrieval and medical annotation tasks. In both tasks we have used the bag-of-words approach for image representation. We submitted two purely visual automatic runs to the medical retrieval task, which used different normalization in the feature extraction stage. Images were converted to a histogram of visual words, and were compared using L1 distance. Our best run was ranked first among the automatic visual based retrieval systems, with MAP score of 0.042. For the medical annotation task we submitted four runs, all used support-vector-machines trained on the visual word histograms. The runs differ in image resolution, and in the way classifiers of two resolutions were combined. In this task our result was second best among the participating groups, with error scores between 105.75 and 117.17.