Reducing the effect of false positives in classification of detected clustered microcalcifications
Yongyi Yang, Robert M. Nishikawa, Maria V. Sainz de · 2018
In classification of clustered microcalcifications (MCs), a typical approach is to extract a set of quantitative features from the MC objects for subsequent analysis by a machine learning algorithm. The presence of false positives (FPs) among the identified MCs will inevitably cause an adverse impact on the classification accuracy. In this work, we present a new feature extraction approach to accommodate the presence of potential FPs among detected MCs in a cluster. We introduce a quality factor to enhance the contribution of true MCs in the extracted features, and devise a set of cluster features to reduce the impact of FPs on the spatial distribution of the cluster. In the experiments, we demonstrated the performance of the proposed approach on a set of 186 FFDM images. The results show that it could lead to improved classification performance consistently over a wide range of detection sensitivity and FP levels.