Combining convolutional neural networks for multi-context microcalcification detection in mammograms

Breast cancer is the most frequent cancer among women, and also causes the greatest number of cancer-related deaths. One effective way to reduce breast-cancer related deaths is to use mammography as a screening strategy. In this framework, cluster of microcalcifications can be an important indicator of breast cancer. To help radiologists in their diagnostic operations, Computer Aided Detection systems have been proposed, which are based Deep Learning methodologies. Such solutions showed remarkable performance, but further improvements can be gained if the design of the detector takes advantage of specific knowledge on the problem. We present an approach for the automated detection of microcalcifications in Full Field Digital Mammograms which involves an ensemble of CNN. The rationale is to employ one CNN trained on ROIs strictly containing the lesions to be detected together with other CNNS trained on ROIs centered on the same lesions, but progressively larger. In this way, shallower networks become specialized in learning local image features, whereas deeper ones are well suited to learn patterns of the contextual background tissues. Once trained, the detectors are combined together to obtain a final ensemble that can effectively detect lesions with a substantial reduction of false positives. Experiments made on a publicly available dataset showed that our approach obtained significantly better performance with respect to the best single detector in the ensemble, so demonstrating its effectiveness.