<p>Breast cancer is currently ranked as one of the leading causes of death in the world amongst women, and early and precise diagnosis is vital to improve patient outcomes. Although machine learning methods have demonstrated great potential in automated diagnosis, currently existing methods have serious limitations in the generalization of heterogeneous datasets and clinical settings. To overcome these drawbacks, this paper presents a new deep learning network, AFEN (Adaptive Feature Ensemble Network), which incorporates dual-pathway extraction, adaptive attention-based fusion, and hierarchical ensemble learning. The suggested architecture represents the local texture patterns, as well as the global morphological structures, by the parallel processing pathways, weights feature significance dynamically using learnable attention mechanisms, and integrates specialized sub-networks using confidence-calibrated ensemble learning. Comprehensive assessment using three breast cancer datasets such as Wisconsin Breast Cancer Dataset (WBCD), MIAS Mammographic Database, and Coimbra Breast Cancer Dataset, has shown accuracy of 98.68%, 97.21%, and 88.70%, respectively, with an average gain of 2.25% over the best-performing gradient boosting baseline (XGBoost) and 3.09% over traditional neural networks. Transfer learning experiments have demonstrated cross-domain generalization with 4–7% performance gains over baseline systems, and systematic ablation studies validate the contribution of each architectural component. These results suggest that AFEN is a promising framework for feature-based breast cancer classification, though validation on larger multi-centre cohorts remains an important next step before clinical deployment.</p>