Characterization of Architectural Distortion in Mammograms

Abstract

We present a technique to characterize architectural distortion in mammograms based upon oriented texture analysis. The local texture orientation is computed for all pixels in a region of interest (ROI), thus obtaining the corresponding orientation field. The orientation fields are then analyzed using phase portraits. Six features are extracted from the phase portraits, and a quadratic discriminant classifier is used to classify the ROIs as a site of architectural distortion or other breast parenchymal pattern. The methods were tested with 37 ROIs: 15 with architectural distortion, eight with spiculated malignant tumors, two with malignant calcifications, and 12 with normal parenchymal patterns. The results obtained indicated a sensitivity of 80%, a specificity of 80%, and area under the receiver operating characteristics curve of 0.86.