Purpose: The purpose of this study is to investigate the state-of-art end-to-end deep learning techniques for mass detection in mammograms.
Methods: Two state-of-art deep convolutional neural networks (DCNNs), named Faster Region-based Convolutional Network (Faster R-CNN) and Single Shot MultiBox Detector (SSD), were implemented for mass detection in mammograms. In order to train and test two DCNNs, two public mammogram databases including Digital Database for Screening Mammography (DDSM) and INBreast were used in this study. In the training phase, the two DCNNs were pre-trained firstly using a large-scale natural image database for object detection and classification. The DCNNs were trained subsequently using 2,392 mammograms, which consists of 1,118 benign and 1,274 malignant masses obtained from the DDSM. In the test phase, two DCNNs were tested individually by using 105 mammograms from INBreast, including 34 benign and 71 malignant masses. The performances of two DCNNs in mass detection were evaluated based on a free receiver operating characteristic (FROC) curve.
Results: The experimental results show that the performance of SSD is superior to that of Faster R-CNN. Specifically, when the true-positive rate of mass detection is 0.8, the false-positives per image of SSD and Faster R-CNN are 0.4 and 2.8, respectively.
Conclusion: This study investigated two state-of-art deep learning techniques, which can be used for developing an end-to-end mammographic CAD system, for mass detection in mammograms. The experimental results demonstrated (1) that transfer learning is an effective training method when large-scale annotated mammogram databases were not available; and (2) that the SSD was more accurate than Faster R-CNN in mass detection.