This article describes the participation of the Geneva University Hospitals and the University of Geneva at the 2008 ImageCLEF image retrieval benchmark. We concentrated on the two tasks concerning medical imaging. The visual information analysis is based on the GNU Image Finding Tool (GIFT). Other information such as textual information and aspect ratio are integrated to improve the results. The main techniques are the same as in past years, with a little tuning to slightly improve results. For the visual tasks it becomes clear that the baseline GIFT runs do not have the same performance as more sophisticated modern techniques do. GIFT can be seen as a baseline for the visual retrieval as it has been used for the past ve years in ImageCLEF. Due to time constraints no optimizations could be performed and no relevance feedback was used, usually one of the strong points of GIFT.
The Geneva University Hospitals and the University of Geneva contribute regularly to the
ImageCLEF1 campaign. The domains of interest are medical image retrieval and medical image
annotation [
The text retrieval approach used in 2008 is detailed in a paper of the text retrieval group of the Geneva University Hospitals. It is very similar to approaches in pas years, where queries and documents are tranlated into MeSH (Medical Subject Headings) terms. 2.2
The technology used for the visual retrieval is mainly taken from the Viper 2 project [
Local color features at di erent scales by partitioning the images successively into four equally sized regions (four times) and taking the mode color of each region as a descriptor; global color features in the form of a color histogram, compared by a simple histogram intersection; local texture features by partitioning the image and applying Gabor lters in various scales and directions, quantized into 10 strengths; global texture features represented as a simple histogram of responses of the local Gabor lters in various directions and scales.
A particularity of GIFT is that it uses many techniques well{known from text retrieval. Visual
features are quantized and the feature space is similar to the distribution of words in texts. A
simple tf/idf weighting is used and the query weights are normalized by the results of the query
itself. The histogram features are compared based on a histogram intersection [
In this section the results and technical details for the two medical tasks of ImageCLEF 2008 are detailed. 3.1
Results for the medical retrieval task are shown in Table 1 highlighting the most important performance measures such as MAP, Bpref, and early precision. 3 purely visual retrieval runs using GIFT with 4 gray levels (GIFT4 ), 8 gray levels (GIFT8 ), 16 gray levels (GIFT16 ) were submitted for evaluation. Using GIFT with 8 gray levels gives the best result for purely visual retrieval. Increasing the number of gray levels decreases basically all performance measures.
Purely visual retrieval results proved to be little robust [
Mixing two runs with varying weights based on the topics (EN GIFT8 mix ) gives second best early precision (P30), and third best MAP among the 5 runs. The best MAP is given by simply combining textual and visual runs with equal weight (GIFT8 EN0.5 ). Favoring the textual run (GIFT8 EN0.9 ) gives best early precision, but surprisingly poor MAP. Compared to the original text runs, the combination with our visual run improves early precision slightly, but reduces MAP signi cantly. 3.2
For the medical image annotation task, the basic GIFT system was used for the feature extraction.
The work of this year followed work performed in 2007 [
The annotation is based on the known labels of similar images retrieved by the GIFT system.
In [
One problem of classifying images with similar images with known labels is that the classi cation strategy favors large classes in the training data and punishes small ones, as images of large classes have a higher chance to be selected. The frequency of each class in the training data is analyzed to avoid this bias. Such a dynamic kNN approach is then used instead of a standard kNN approach to give a di erent k value for each class. As a result, the disadvantages of small classes are reduced.
Another useful information is the hierarchy information inside each code axis (there are four in the IRMA code). The output of classi cation per axis is usually an entire axis or a wildcard for the entire axis. Another possibility is to chop only the lowest level (the last letter) of each axis. The remainder can then be used for a second round of classi cation. This additional step gives the possibility to use less wildcards in the classi cation process and thus can potentially improve the score.
The results of our submitted runs and the best overall system are presented in Table 2. Three
submitted runs use the kNN approach with classi cation for the entire code (kNN ), classi cation
per axis (akNN ), and dynamic kNN classi cation per axis (adkNN ). Dynamic kNN gives the
best results in our tests. Three submitted runs use a voting{based approach as described in [
For the medical retrieval task the use of text alone is still better than our combinations with visual retrieval, which means that combination techniques still need signi cant work to preform reasonably well and stable. Only early precision can be improved through the combination of textual runs with visual runs. The visual baseline seems to be of insu cient quality for really improving the combined runs. A small number of colors still gives best results. For a signi cant improvement in visual retrieval quality new visual features seem necessary.
For the classi cation of images the di erence between our runs and the best techniques is reduced compared to previous years. The voting{based approaches perform generally better than the simple kNN approaches. Classifying each axis separately with a suitable threshold gives always good results. When the threshold cannot be reached in the rst step, chopping the lowest level and redoing the classi cation for the remaining levels can further improve the result signi cantly. The advantage of the "chopping" strategy is that the classi cation is redone iteratively, thus high threshold values increase the con dence without totally blocking the classi cation. This study was partially supported by the Swiss National Science Foundation (Grants 200020{ 118638/1), the HES SO with the BeMeVIS project, and the European Union in the 6th Framework Program through the KnowARC project (Grant IST 032691).