Breast cancer is the most common form of cancer among women, and combating this requires effective diagnosis. Over the years, computer-aided diagnosis systems have been deemed as invaluable tools, however, devising such systems require large datasets that many medical institutions lack access to. Federated learning solves this problem by allowing medical institutions to collaborate building a model without sharing their own data. This study examines the use of federated learning to train models that can predict breast cancer diagnosis from breast ultrasounds. Three benchmark breast ultrasound datasets were used: BUSI, BUS-BRA, and BrEaST. Through a series of experiments, federated learning showed an average increase of 13.5%, 14.1%, and 10.1% in AUROC for each dataset, respectively. The study suggests that federated learning is a promising solution for developing robust models in healthcare, where data privacy concerns limit direct data sharing.