Alzheimer’s disease (AD) strains healthcare systems worldwide, generating annual costs exceeding $1 trillion. Early detection through MRI segmentation can improve patient outcomes and reduce long-term expenses, but traditional deep-learning approaches require expensive infrastructures and fail to adapt to multi-scanner variability. We propose a Hybrid Dynamic Dropout(HDD) framework for 3D U-Nets that cuts computational overhead by up to 40% while preserving high segmentation accuracy. Our method fuses three dropout strategies (loss-based, variance-based, and gradient-based) with a confidence-gated module targeting uncertain regions, plus a domain-adaptive regularization term for cross-hospital data. In tests on the ADNI dataset, we reach a Dice score of 0.90 on hippocampal segmentation, using 35% less GPU memory compared to standard models. Beyond technical gains, economic analyses indicate potential annual savings of $2.8 M for a mid-sized hospital network, factoring in reduced hardware investments, streamlined operational costs, and improved diagnostic workflows. This synergy of clinical accuracy, computational efficiency, and economic feasibility positions our framework as a sustainable e-health solution for large-scale AD screening and monitoring.