The mask R-CNN algorithm was implemented to search entire images to identify suspicious lesions for further evaluation of malignancy probability. One training (N=98, Siemens 1.5T, non-fat-sat) and two independent testing (N=241, Siemens 3T, non-fat-sat; and N=91, GE 3T, fat-sat) datasets were used. The pre-contrast and subtraction image, and the subtraction image of the contralateral breast, were used as three inputs. The training set had a total of 1353 positive slices (containing lesion), 8055 negative slices without lesion. The 10-fold cross-validation showed accuracy=0.80 and mean DSC= 0.82. The accuracy was 0.73 and 0.62 for two testing datasets, lower for fat-sat images.