Robust retrieval from compressed medical image archives

Paper addresses the computational aspects of extracting important features directly from compressed images for the purpose of aiding biomedical image retrieval based on content. The proposed method for treatment of compressed medical archives follows the JPEG compression standard and exploits algorithm based on spacial analysis of the image cosine spectrum coefficients amplitude and location. The experiments on modality-specific archive of osteoarticular images show robustness of the method based on measured spectral spatial statistics. The features, which were based on the cosine spectrum coefficients' values, could satisfy different types of queries' modalities (MRI, US, etc), which emphasized texture and edge properties. In particular, it has been shown that there is wealth of information in the AC coefficients of the DCT transform, which can be utilized to support fast content-based image retrieval. The computational cost of proposed signature generation algorithm is low. Influence of conventional and the state-of-the-art compression techniques based on cosine and wavelet integral transforms on the performance of content-based medical image retrieval has been also studied. We found no significant differences in retrieval efficiencies for non-compressed and JPEG2000-compressed images even at the lowest bit rate tested.