Abstract
In a Computer Aided Diagnosis (CAD) System for breast cancer detection, two main steps are mass detection from mammograms and then classification of suspicious regions of interest (ROIs) into masses and normal cases. This paper introduces three features extraction methods for mass classification. ROI is divided into s x s parts then each part is treated as input image to the three methods. Two methods employ Discreet Cosine Transform (DCT) in novel ways to encode texture information of mammogram images. The third method (WTDFB) extracts multi-resolution and multi-directional texture information in a novel way by employing a hybrid of Discrete Wavelet Transform (DWT) and Directional filter bank (DFB). Support Vector Machine (SVM) is used to classify suspicious ROIs into masses and normal cases using features extracted by the three methods. For validating the usefulness of the methods, the benchmark database Digital Database for Screening Mammography (DDSM) has been used. The experimental results show that one DCT based method (DCT1) achieves an accuracy 98.03% and sensitivity 98.48% with a small number (only 256) of features. The other DCT based method (DCT2) got an accuracy 98.6% and sensitivity 97.6%, whereas the WTDFB method obtained accuracy 98.04:98.43% and sensitivity 98:98.35%. DCT2 is the best in terms of accuracy among the three proposed methods whereas DCT1 is the best in terms of sensitivity. Also the DCT based methods are superior to the WTDFB method in terms of the dimension of the feature space. This contribution is suitable to be taken as a second opinion for radiologists in classifying suspicious ROIs.