A BI-RADS-based Approach for Tumor Classification in Breast Ultrasound Using Binary Decomposition Strategies

This article presents a BI-RADS-based method for classifying breast tumors on ultrasound. The multiclass classification problem into BI-RADS categories 2 to 5 is addressed by two binary decomposition schemes: one-vs.-all (OVA) and one-vs.-one (OVO). Furthermore, the base classifiers of these strategies are trained with local features that are selected using a search method based on the genetic algorithm. The experimental results show that the proposed approach outperforms the classification performance of traditional OVO and OVA approaches that use a global feature space shared by all base classifiers. Therefore, using local features allows the specialization of the base classifiers to distinguish between two specific classes.