We developed two deep learning methods for breast MRI evaluation, first using Mask R-CNN for detection of suspicious areas, and then using ResNet50 for estimating the malignancy probability. These two networks were combined to test its diagnostic validity in two datasets. In Dataset-1, sensitivity=96.1% and specificity=78.1%. In Dataset-2, sensitivity=81.1% and specificity= 80.6%. We further characterized all false positives (FPs), and found other than confirmed benign lesions, FPs may come from vessels and asymmetric parenchymal enhancements, which can be further eliminated by other algorithms. The results suggest the potential of combined deep learning networks as a fully-automatic breast MRI CAD.
This abstract and the presentation materials are available to members only; a login is required.