Breast cancer is the most common type of cancer among women in Taiwan. Except regular breast self-exams, mammography is regarded as the most effective tool for early detection. Clinically, clustered microcalcifications (MCCs) on mammograms are considered to be a strong indicator of malignancy. However, their tiny size and low contrast to surrounding tissues always make radiologists hard to interpret. In this study, we develop a computer-aided detection system (CADe) for MCCs, which is able to help radiologists reducing missing rates as well as arising diagnostic efficiencies.
Via image pre-processing procedures, some candidates of MCCs were found. Afterwards, the textural features were extracted from the GLCM (gray-level co-occurrence matrix) and the SVM (Support Vector Machine) classifier was well trained to detect MCCs. Sequential forward selection, sequential backward selection and F-score were used to find discriminative features. We took 111 full-field digital mammography raw images obtained from CMUH, among them 57 images were used as training, and the rest 54 images were used for test. The performance of the proposed scheme was evaluated by means of the receiver operating characteristic (ROC) curves and area under curve (Az).
According to the experimental results, the texture features extracted from Laws’ LS method have the best accuracy in MCCs detection, which Az is equal to 0.9896. The composition of texture features extracted from LS and LE images has the performance of Az = 0.9897. The proposed method has 90% of sensitivity and 0.74 FPs/image in detection of MCCs.
