This paper describes the participation of MIRACLE research consortium at the ImageCLEF Medical Image Annotation task of ImageCLEF 2008. A lot of effort was invested this year to develop our own image analysis system, based on MATLAB, to be used in our experiments. This system extracts a variety of global and local features including histogram, image statistics, Gabor features, fractal dimension, DCT and DWT coefficients, Tamura features and coocurrency matrix statistics. Then a k-Nearest Neighbour algorithm analyzes the extracted image feature vectors to determine the IRMA code associated to a given image. The focus of our experiments is mainly to test and evaluate this system in-depth and to make a comparison among diverse configuration parameters such as number of images for the relevance feedback to use in the classification module.
MIRACLE team is a research consortium formed by research groups of three different universities in Madrid (Universidad Politécnica de Madrid, Universidad Autónoma de Madrid and Universidad Carlos III de Madrid) along with DAEDALUS, a small/medium size enterprise (SME) founded in 1998 as a spin-off of two of these groups and a leading company in the field of linguistic technologies in Spain. MIRACLE has taken part in CLEF since 2003 in many different tracks and tasks, including the main bilingual, monolingual and cross lingual tasks as well as in ImageCLEF, Question Answering, WebCLEF, GeoCLEF and VideoCLEF (VID2RSS) tracks.
This paper describes our participation in the Medical Image Annotation task of ImageCLEF 2008 [
While in previous participations [
The architecture of our system is composed of different functional blocks: • •
Feature extraction module: in charge of the calculation and extraction of a variety of features of each image, both the training set used to build the classifier and the test set that is to be actually classified. This module has been entirely developed using MATLAB and extracts vectors with a total of 3,741 features for each image.
Images are converted to gray-scale, rescaled to 256x256 pixels and the following features are extracted: o o
Global features: gray histogram (128 levels of gray), image statistics (mean, median, variance, maximum singular value, skewness and kurtosis ), Gabor features (4 scales, 6 filter orientations), fractal dimension, Discrete Cosine Transform (DCT) coefficients, Discrete Wavelet Transform (DWT) coefficients, Tamura features (coarseness, contrast, directionality), and coocurrency matrix statistics (energy, entropy, contrast, homogeneity, correlation) Local features: images are cut up into 64x64 pixel blocks and then the previous features are extracted for each block.
Classifier: determines the IRMA code associated to a given image, from its feature vector and the feature matrix of the training set. The classifier is internally composed of two blocks: an initial module in charge of selecting those images in the training set whose vectors are at a distance lower than a given threshold from the vector associated to the image to classify, and then a second module that actually generates the IRMA code, depending on the codes and similarity of nearby images.
Finally, we submitted four runs to be evaluated, described in Table 1. For all of them, the returned IRMA code is generated from the combination of the first N images in the training set that are most similar to the image to classify. The combination consists of a simple “addition” of strings characters in which, if both characters are different, the result is the wildcard “*” representing the ambiguity (or “hesitation” to choose). This algorithm actually could be considered as a variation of the classical k-Nearest Neighbour algorithm [8] with a specific definition of the generating the output class.
Additionally, two runs use relevance feedback (RF) with the first two images in the training set that are at a lowest distance. Vectors of those images are added and averaged to build a new vector that is used for querying the system again.
Results are shown in Table 2. The “Error score” column contains the experiment score as computed by the task
organizers [
Based on the analysis performed over each axis, the first conclusion to be drawn is that the first weak point of our experiments is the prediction of the Direction (D) and Biological (B) axis. Some extra effort must be invested on determining which image features could be most useful to predict those axis.
In addition, it can be observed that although the number of incorrect predictions is relatively low, this does not correspond to a high number of correct predictions (which in fact is also relatively low), as it would be expected applying a binary “correct” vs. “incorrect” logic. This is due to the fact that, as the cost of making an incorrect decision is higher than the cost of not actually making a decision, the design criteria of the system is biased for “hesitation”, i.e., the system is very cautious and assigns a wildcard “*” if there is any kind of ambiguity. This explanation also confirmed by the result of the run that takes 3 codes for generating the final IRMA code: when the number of codes increases, so ambiguity does, thus the number of complete correct predictions decreases and also the error score.
Finally, in all runs, the calculation of the distance among vectors assigns the same weight to every dimension of the vectors, regardless of the nature of the feature to which this component belongs and/or the number of components belonging to that feature. This was actually our mistake when carrying out the experiments and the feature matrix should have been divided into the different feature sub-matrixes that employ different distances for calculating similarity and are combined to each other using different weight strategies. For sure this will be taken into account for future participations.
This work has been partially supported by the Spanish R+D National Plan, by means of the project BRAVO (Multilingual and Multimodal Answers Advanced Search – Information Retrieval), TIN2007-67407-C03-03 and by Madrid R+D Regional Plan, by means of the project MAVIR (Enhancing the Access and the Visibility of Networked Multilingual Information for the Community of Madrid), S-0505/TIC/000267. line http://www-i6.informatik.rwth-aachen.de/