Independent component analysis in angiography images
Ewa Snitkowska, Wëodzimierz Kasprzak · 2006
An important source for information about digital image content is the texture of image regions. This paper presents a feature extraction approach that is based on independent component analysis (ICA). In ICA a transformation of measured vectored time series is discovered via blind signal processing that gives statistically independent source signals. In our approach every textured region is considered as a mixture of (initially unknown) statistically independent source regions, scanned to 1-D time series. After these sources called independent components are extracted by ICA, optimally w.r.t. given mage type, the mixing coefficients of particular region constitute its feature vector. The quality of such features is experimentally verified and compared to other common feature schemas. The comparison procedure explores the Fisher information criterion and classification results for feature evaluation. Our application field is the analysis of angiography images. It is difficult for medical doctors properly to classify such images, hence an automatic tool could provide support in this matter. We demonstrate the usefulness of ICA-based features for automati evaluation of angiography images.