- Abstract: Computer-aided detection or decision support systems aim to improve breast cancer screening programs by helping radiologists to evaluate digital mammography (DM) exams. Commonly, such systems proceed in two steps: selection of candidate regions, and subsequent false positive reduction of the candidates as either suspicious lesions or inconspicuous breast tissue. In this study, we present a method based on deep learning for automatic detection of soft tissue lesions in DM using a one-step approach. A database of DM exams (mostly bilateral and two views) was collected from our institutional archive. In total, 7192 DM exams (23405 DM images) were acquired with systems from three different vendors (General Electric, Siemens, Hologic), of which 2883 contained malignant lesions verified with histopathology. The performance of our automated detection system was assessed using the free receiver operating characteristic (FROC) analysis. A maximum sensitivity of 0.97 at 3.56 false positives (FP) per image was achieved. The best model achieved a sensitivity of 0.73, 0.45, 0.31 at 0.1, 0.02 and 0.01 FP per image, respectively. Overall, the results of our evaluation suggests that our soft tissue lesion detection system can replace current two stage detectors.
- Keywords: object detection, digital mammography, breast cancer, women's cancers, regional cnn, mask rcnn, faster rcnn
- Author Affiliation: Radboud University Medical Center