This work reports baseline results for the CLEF 2008 Medical Automatic Annotation Task (MAAT) by applying a classifier with a fixed parameter set to all tasks 2005 2008. The classifier performs a weighted combination of three distance and similarity measures operating on global image features: Scaled-down representations of the images are compared via metrics that model the typical variability in the image data, mainly translation, local deformation, and radiation dose. In addition, a distance measure based on texture features is used. For classification, a k nearest neighbor classifier is used. In 2008, the baseline classifier yields error scores of 170.34 and 182.77 for k=1 and k=5 when the full code is reported, which corresponds to error rates of 51.3% and 52.8% for 1-NN and 5-NN, respectively. Judging the relative increases of the number of classes and the error rates over the years, MAAT 2008 is estimated to be the most difficult in the four years.
The content of one radiograph is represented by Tamura’s texture measures (TTM) proposed in
[
The single classifiers are combined within a parallel scheme, which performs a weighting of the normalized distances obtained from the single classifiers Ci, and applies the nearest-neighbordecision function C to the resulting distances: dcombined(q, r) = X λi · d′i(q, r),
i d′i(q, r) =
di(q, r) Pr′∈R di(q, r′) (1) (2) where 0 ≤ λi ≤ 1, Pi λi = 1 denotes the weight for the normalized distance di(q, r) obtained from classifier Ci for a sample q and a reference r from the set of reference images, R. Values 0 ≤ si(q, r) ≤ 1 obtained from similarity measures are transformed via di(q, r) = 1 − si(q, r).
The three content descriptors and their distance measures use the following parameters: • TTM: texture histograms from down-scaled image (256 × 256), 384 bins, Jensen-Shannon divergence as a distance measure, • CCF: 32 × 32 icon, 9 × 9 translation window • IDM: X × 32 icon, gradients, 5 × 5 window, 3 × 3 context The weighting coefficients were set empirically during CLEF MAAT 2005: λIDM = 0.42, λCCF = 0.18, and λTTM = 0.4. 3
A rough estimation of the task difficulty can be derived from the baseline error rates: Comparing 2005 and 2006, the number of classes increased by 103%, while the error rate only increased by 63% and 48% for 1-NN and 5-NN, respectively. This suggests that the task in 2006 was easier than in 2005. Since the challenges in 2006 and 2007 use the same class definitions, the obtained error rates are directly comparable and show a slightly reduced task difficulty in 2007. In 2008, the number of classes increased by 70% compared to 2007, while the error rate increased by 157% and 193%, respectively. The 2008 task can therefore be considered to be more difficult than the 2007 task. Applying the same estimation, the 2008 task is also found to be more difficult than the 2005 task, as the number of classes increased by 246% and the error rate increased by 286% and 257%, respectively.