LAUSR

Computer-aided detection (CADe) and diagnosis (CADx) system of mammographic microcalcification clusters (MCCs) is built for helping human observers to find suspicious areas of MCC and providing risk predictions of malignancy as a reference, since it is challenging and time consuming for radiologists to manually identify some subtle microcalcifications (MCs) and perform precise interpretation in mammograms. However, the performance of traditional CADe and CADx systems is not good enough, thus it is difficult to combine them into a whole system that integrates detection and diagnosis together. The purpose of this study is to develop a fully automatic computer-aided MCC detection and diagnosis system based on deep learning method. In order to detect subtle MCs accurately, a MC candidate detection system is used to obtain a great number of potential MC candidates, then a deep convolution neural network (DCNN) is trained specially to discriminate true MCs from detected MC candidates. Different from previous literatures committing to finding and selecting effective features, the proposed method replaces manual feature extraction step by using DCNN. To accelerate the training procedure of the DCNN, a neutrosophic boosting (NB) strategy is applied in the training stage. Then a density-based regional clustering method is imposed on those true MCs to form MCCs. Finally, another DCNN is employed to differentiate benign from malignant MCC lesions. For cluster-based MCC detection evaluation, a sensitivity of 90% is achieved at 0.14 false positives (FPs) per image. For case-based MCC classification evaluation, the area under the receiver operating characteristic curves (AUCs) on validation and testing sets are 0.945, 0.933 for proposed system, respectively. Our obtained results demonstrated the effectiveness of the proposed method for automated detection and classification of MCCs.