Medical image classification typically relies on CNNs or Vision Transformers, but these approaches face limitations in computational efficiency and long-range dependency modeling. We introduce the first application of Mamba—a novel state space model—to breast ultrasound classification. Our approach treats image patches as sequential tokens, leveraging Mamba’s linear-time complexity and selective memory mechanisms. On the BUSI dataset with 2,400 augmented samples across normal, benign, and malignant categories, our model achieves 91.1% accuracy, 91.3% precision, 90.6% recall, and 0.945 AUC while maintaining significantly lower computational overhead than Transformer-based alternatives. The results demonstrate Mamba’s potential as an efficient architecture for medical imaging tasks requiring both accuracy and deployability in resource-constrained clinical environments.