A Sustainable Computational Framework for Breast Cancer Screening: Optimizing High-Dimensional Feature Spaces via MVP-PCA for Resource-Constrained Environments

The rapid integration of Electronic Health Records (EHR) demands the efficient processing of high-resolution medical images. However, deep learning architectures applied to mammography classification often produce massive, high-dimensional feature spaces susceptible to the curse of dimensionality and anatomical noise. Furthermore, conventional dimensionality reduction approaches tend to cause over-reduction, which destroys crucial microcalcification textures. To address these challenges, this study proposes a CPU-efficient hybrid dimensionality reduction framework integrating Mean Vector Projection (MVP) and Principal Component Analysis (PCA) on features extracted by SqueezeNet. The MVP layer acts as a crucial pre-conditioner to stabilize intra-class variance before PCA decomposition. Experimental results demonstrate that the proposed MVP-PCA framework successfully linearizes the feature space and achieves an extreme compression rate of 99.11%, reducing 264,702 features to 2,355 essential components. The peak accuracy reaches 97.58% with a minimal False Negative rate (2 cases), providing a sustainable diagnostic solution for healthcare facilities with limited technological resources.