Domain shifts between mammography datasets driven by variations in imaging equipment, clinical protocols, and patient populations present a major challenge to the generalizability of artificial intelligence (AI)-based breast cancer diagnostic tools. Recent unsupervised domain adaptation (UDA) frameworks have introduced contrastive learning to mitigate these discrepancies. This chapter investigates domain shifts across four publicly available mammography datasets from different geographic and clinical settings: VinDr-Mammo (Vietnam), INbreast (Portugal), CDD-CESM (Egypt), and KAU-BCMD (Saudi Arabia). We employ a ConvNeXt V2-based masked autoencoder to extract robust and representative features from the mammography images. For UDA, we adapt the CoTMix framework, originally designed for time-series data. Unlike conventional UDA approaches based on adversarial training or statistical alignment, CoTMix introduces a novel contrastive mix-up strategy. It generates intermediate augmented representations from both source and target domains and aligns them by maximizing feature similarity with their original views. This enables effective alignment of diverse domain distributions in a shared latent space. Experimental evaluation shows that the proposed framework improves generalization across domains, achieving higher precision, F1 scores, and AUC compared to baseline adaptation methods. These findings demonstrate the potential of combining masked autoencoders with contrastive mix-up techniques for reliable AI-based diagnostics in heterogeneous clinical environments.