ABSTRACT

This paper investigates detection of Architectural Distortions (AD) and spiculated masses in mammograms based on their physical characteristics. We have followed a model based approach which separates the abnormal patterns of AD and spiculated masses from normal breast tissue. The model parameters are retrieved from Gabor filters which characterize the texture features and synthetic patterns were generated using pplanes to retrieve specific patterns of abnormalities in mammographic images. In addition, eight discriminative features are extracted from region of interest (ROI) which describes the patterns representing AD and spiculated masses. Support vector machine (SVM) and Multi-layer Perceptrons (MLP) classifiers are used to classify the discriminative features of AD and spiculated masses from normal breast tissue. This study concentrates on classifying AD and spiculated masses from the ones which actually are normal breast parenchyma. Our proposal is based on the texture pattern that represents salient features of AD and spiculated mass. Once the descriptive features are extracted SVM and MLP classifiers are used. We have used receiver operating characteristic curve (ROC) to evaluate the performance and we have compared our method with several other existing methods. Our method outperformed other existing methods by achieving 90% of sensitivity, 86% specificity in distinguishing AD from normal breast tissue and 93% sensitivity and 88% specificity in classifying spiculated mass from normal breast parenchyma. In first stage of this study we consider ROI’s that include AD, spiculated masses and normal breast tissue as input. Our method was tested on 190 ROI’s( 19 AD , 19 spiculated mass and 152 normal breast tissue) from Mini-MIAS database and 150 ROI’s( 23 AD , 30 spiculated mass and 97 normal breast tissue ) collected from DDSM database. In the second stage we have applied SVM classification model on whole images and the performance is analyzed by plotting Free Response Operating Characteristic (FROC) curves. SVM classifiers achieved 96% sensitivity with 9.6 false positives per image in detection of spiculated mass and 97% sensitivity with 6.6 false positives per image while detecting AD in digital mammograms.