Deep Cascade Classifiers to Detect Clusters of Microcalcifications

Recent advances in Computer-Aided Detection (CADe) for the automatic detection of clustered microcalcifications on mammograms show that cascade classifiers can compete with high-end commercial systems. In this paper, we introduce a deep cascade detector where the learning algorithm of each binary pixel classifier has been redesigned in the early stopping mechanism conventionally used to avoid overfitting to the training data. In this way, we strongly increase the number of features considered in each stage of the cascade (hence the term “deep”), yet we still benefit from the cascade framework by obtaining a very fast processing of mammograms (less than one second per image). We evaluated the proposed approach on a database of full-field digital mammograms; the experiments revealed a statistically significant improvement of deep cascade with respect to the traditional cascade framework. We also obtained statistically significantly higher performance than one of the most widespread commercial CADe systems, the Hologic R2CAD ImageChecker. Specifically, at the same number of false positives per image of R2CAD (0.21), the deep cascade detected 96 % of true lesions against the 90 % of R2CAD, whereas at the same lesion sensitivity of R2CAD (90 %), we obtained 0.05 false positives per image for the deep cascade against the 0.21 of R2CAD.