Independent Component Analysis for Microcalcifications Detection in Digital Mammograms
Jun Jia Zheng, Emma E. Regentova · 2008
2 Abstract—In this paper, a method is presented that employs independent component analysis (ICA) for detecting the microcalcifications (MC) in digital mammograms. ICA is used to estimate unknown, statistically independent sources which form the regions of interest (ROIs) in digital mammograms according to the linear mixing model. Two different ICA architectures are applied on the ROIs, i.e. the first one treats the ROIs as random variables and the pixels as observations, and in the second one, the ROIs are observations and the pixels are variables. The transformation coefficients of ICA are used as classification features which are fed into a three-layer back-propagation neural network. The neural network classifier decides on whether a region contains MCs or not. The classification results certify that the first ICA architecture provides higher classification accuracy compared to that yielded by the second one. The system performance is evaluated using the free-response receiver operating characteristic (FROC) curves. The results show that the method developed based on the first ICA architecture is able to detect approximately 90% of the true clusters with an average of one false cluster per image.