Effective Color Features for Content Based Image Retrieval in Dermatology

We are concerned with the extraction of effective color features for a content-based image retrieval (CBIR) application in dermatology. Effectiveness is measured by he rate of correct retrieval of images from four color classes of skin lesions. We employ and compare two different methods: Limited Rank Matrix Learning Vector Quantization (LiRaM LVQ) and a Large Margin Nearest Neighbor (LMNN) approach. Both methods use supervised training data and provide a discriminant linear transformation of the original features to a lower-dimensional space. The extracted color features are used to retrieve images from a database by a k-nearest neighbor search. We perform a comparison of retrieval rates achieved with extracted and original features for eight different, standard color spaces. We achieved significant improvement in every examined color space. The best results were obtained with features extracted from original features in the color spaces YCrCb, CIE-Lab, CIE-Lch, CIELuv and RGB. The increase of the mean correct retrieval rate lies between 10% and 27% in the range of k = 1 to k = 25 retrieved images, and the correct retrieval rate lies between 84% for k = 1 and 70% for k = 50. We present explicit RGB and CIE-Lab color feature combinations of healthy and lesion skin which lead to this improvement. LiRaM LVQ and LMNN give comparable results for large values of the method parameter kappa of LMNN (\kappa > 25) and LiRaM LVQ outperforms LMNN for smaller values of kappa. We conclude that feature extraction by LiRaM LVQ leads to considerable improvement in retrieval by color of dermatologic images. Institute for Mathematics and Computing Science University of Groningen P.O. Box 407, 9700 AK Groningen The Netherlands