We present a SwinTransformer-based model for breast cancer detection in full-field digital mammograms (FFDM) that iteratively zooms in on the most suspicious region of each image using Grad-CAM attention. Unlike our previous two-stage convolutional model–which used separate networks for coarse localization and fine-grained classification–our new approach trains a single model to evaluate crops at multiple resolutions through a sequence of focused attention steps. The model was trained on a large dataset of Siemens mammograms with confirmed diagnosis as the only learning signal. The model was evaluated on test sets from multiple vendors, including Siemens, Philips, Hologic/Lorad, and GE. On Siemens images, our model achieved an AUC of 0.978 for screen-detected cancers, improving by 1.3% points over the previous ResNet-based model–a substantial gain given the high performance baseline. The SwinTransformer also generalized markedly better to non-Siemens vendors, particularly GE with a huge improvement from 76.7% to 92.3% on screen-detected cancers. A ResNet trained using the same zooming algorithm failed to achieve similar gains, suggesting that transformer architectures are better suited for handling both spatial resolution changes and domain shifts. Our findings demonstrate that attention-guided zooming combined with transformer backbones offers a scalable and generalizable approach.