Abstract

The rapid development of deep learning, a family of machine learningtechniques, has spurred much interest in its application to medical imagingproblems. Here, we develop a deep learning algorithm that can accurately detectbreast cancer on screening mammograms using an "end-to-end" training approachthat efficiently leverages training datasets with either complete clinicalannotation or only the cancer status (label) of the whole image. In thisapproach, lesion annotations are required only in the initial training stage,and subsequent stages require only image-level labels, eliminating the relianceon rarely available lesion annotations. Our all convolutional network methodfor classifying screening mammograms attained excellent performance incomparison with previous methods. On an independent test set of digitized filmmammograms from Digital Database for Screening Mammography (DDSM), the bestsingle model achieved a per-image AUC of 0.88, and four-model averagingimproved the AUC to 0.91 (sensitivity: 86.1%, specificity: 80.1%). On avalidation set of full-field digital mammography (FFDM) images from theINbreast database, the best single model achieved a per-image AUC of 0.95, andfour-model averaging improved the AUC to 0.98 (sensitivity: 86.7%, specificity:96.1%). We also demonstrate that a whole image classifier trained using ourend-to-end approach on the DDSM digitized film mammograms can be transferred toINbreast FFDM images using only a subset of the INbreast data for fine-tuningand without further reliance on the availability of lesion annotations. Thesefindings show that automatic deep learning methods can be readily trained toattain high accuracy on heterogeneous mammography platforms, and holdtremendous promise for improving clinical tools to reduce false positive andfalse negative screening mammography results.