Samir Brahim Belhaouari

Abstract. This paper introduces a novel feature extraction method for breast cancer diagnosis in digital mammograms using multiresolution representations. The proposed method transforms mammogram images into coefficient vectors via wavelet and curvelet transforms, constructing a matrix where each row represents an image and each column a coefficient. Feature significance is ranked using a statistical t-test, optimizing the feature set to maximize classification accuracy. Support Vector Machine (SVM) classifiers distinguish between normal and abnormal tissues, as well as benign and malignant tumors. The feature capability to differentiate classes is given by \(CT = \frac{m_a - m_b}{\sqrt{\left(\frac{s_a^2}{n_a}\right) + \left(\frac{s_b^2}{n_b}\right)}}\). Validation using the MIAS dataset achieved classification accuracy rates of 94.79% with 1238 features and 100% with 150 features for wavelets, and 95.67% with 5663 features and 100% with 333 features for curvelets, demonstrating the efficacy of the method in improving diagnostic accuracy.

Illustration of the proposed experimental design.