Genetic Programming for Feature Construction in Breast Ultrasound Tumor Classification

We present a feature construction approach based on Genetic Programming (GP) for classifying breast tumors into benign and malignant cases on ultrasonography. GP is an evolutionary algorithm that codifies potential solutions in syntax trees, representing mathematical expressions that combine the inputs (i.e., morphological and texture features) using arithmetic operators and mathematical functions to construct new features that increase the descriptive power of the original ones. The experimental results show that the GP-based method achieves an accuracy of 87% with four constructed features. It is compared against two feature selection techniques, attaining 87% and 88% accuracy with 86 and 13 selected characteristics, respectively, and two deep networks proposed for breast ultrasound (BUS) classification, reaching 82% and 84% accuracy each. Thus, using a reduced feature set, the proposed method is an adequate alternative for distinguishing between benign and malignant breast tumors on BUS images.