A new approach for the detection of architectural distortions using textural analysis of surrounding tissue.

Aiming at improving the performance of computer-aided detection of architectural distortion (AD) in mammograms, this paper investigates whether textural patterns of AD surrounding tissue (ST) have the potential to detect AD signatures. More specifically, for characterizing the presence of AD; we investigated the application of textural analysis for discriminating between AD surrounding tissue and normal breast parenchyma. We evaluated the underlying hypothesis using a dataset of 2544 regions of interest (ROI) obtained from the Digital Database for Screening Mammography (DDSM). The ROI dataset contained 353 ST regions and 2191 normal parenchyma related regions. The bidimensional empirical mode decomposition (BEMD) algorithm was, first, applied to extract, from each ROI, the 2D intrinsic mode functions (2DIMF) or detail subbands. Then, statistical signatures of IMF layers were computed and used along with the fractal dimension, estimated from the original ROI, for discriminating ST from the normal breast tissue. The statistical analysis of various textural descriptors demonstrated the significant difference between characteristics of AD surrounding tissue and normal breast parenchyma. The highest AD recognition results of Az = 0.869, obtained from the textural analysis of AD surrounding tissue, is very promising and comparable with Az = 0.913 produced from characterizing AD regions.