A web-accessible framework for the automated storage and texture analysis of biomedical images
Michael Barnathan, Jingjing Zhang, Vasileios Megalooikonomou · 2008
We present a framework for automated image texture analysis that utilizes vector quantization (VQ), an image compression technique, to perform common data mining operations, such as classification, clustering, and similarity searches, on 2D and 3D image datasets. We additionally demonstrate the effectiveness of this framework in a medical imaging context through MIDMS (Medical Image Data Mining System), a web-based system written in Perl and Matlab. MIDMS is capable of automating submission, normalization, compression, and real-time querying of general user-submitted medical image data. Our framework processes submissions by generating a locally optimal codebook for submitted datasets using the Generalized Lloyd Algorithm (GLA). After generating the codebook, our framework uses the codeword usage frequency of each image as the image's feature vector. Users of the framework may then perform highly accurate classification and similarity search experiments on these vectors using either the histogram model (HM) or summed Euclidean distance (SED) metrics, as confirmed by previous experiments utilizing these techniques. Our framework has the potential to assist researchers and clinicians in the sharing, mining, and analysis of large quantities of medical imaging data.