Multi-View Mammographic Breast Cancer Detection Using Deep Learning: A Comprehensive Review of Fusion, Cross-View Consistency, and Uncertainty Methods

Abstract

Breast cancer is currently among the primary causes of cancer-related mortality in women worldwide, and mammographic screening enables earlier detection of the disease, which plays a central role in enhancing survival outcomes. Standard screening protocols acquire craniocaudal (CC) and mediolateral oblique (MLO) projections, but the majority of deep learning systems process these views independently and without regard to the complementary diagnostic value that multi-view examination provides. This is a systematic review of deep learning methods in multi-view mammographic breast cancer detection, which discusses four interconnected themes: single-view CNN architectures and their inherent limitations, multi-view fusion strategies ranging from simple concatenation to cross-view attention mechanisms, cross-view consistency enforcement through dedicated loss formulations, and Bayesian uncertainty quantification methods including Monte Carlo dropout, deep ensembles and evidential learning. Surveyed benchmark datasets and evaluation protocols include RSNA, DDSM, MIAS, and VinDr-Mammo. Other emerging directions, such as transformer-based multi-view models, vision-language pretraining, and explainability by gradient-weighted visualisations, are also addressed. The main open challenges, especially dense tissue detection, calibration quality, and generalisation across imaging centres, are identified, and future research directions are outlined.