Hybrid CNN-RELM with Improved Butterfly Optimization Algorithm and Deep Adaptive Spatial Feature Fusion for Automated Breast Cancer Detection and Classification from Mammographic Images

Abstract

Breast cancer is the most prevalent malignancy in women worldwide, with mortality rates strongly inversely correlated to the stage at which the disease is detected. Computer-Aided Diagnosis (CAD) systems powered by deep learning have shown substantial promise for automating mammographic breast cancer screening; however, existing methods suffer from redundant feature representations, premature convergence in optimization, limited multi-scale spatial reasoning, and poor generalization under class imbalance. To address these interrelated limitations, this paper proposes a novel integrated framework designated CNN-RELM-IBOA-DASFF, which synergistically combines Convolutional Neural Network–Regularized Extreme Learning Machine (CNN-RELM) with a Hyperbolic Secant (HS) activation function, an Improved Butterfly Optimization Algorithm (IBOA) incorporating six chaotic mapping strategies for robust feature selection, and Deep Adaptive Spatial Feature Fusion (DASFF) for multi-scale representational refinement. The proposed pipeline begins with data augmentation and Contrast Limited Adaptive Histogram Equalization (CLAHE) for image quality enhancement, followed by precise tumor region delineation via an Optimized Region Growing (ORG) algorithm guided by Dragonfly Optimization. Dual-stream transfer learning with fine-tuned MobileNetV2 and NasNet Mobile architectures provides complementary deep feature vectors, which undergo IBOA-driven selection before multi-scale DASFF consolidation and final CNN-RELM-HS classification. Exhaustive experiments on the MIAS mammography dataset demonstrate that CNN-RELM-IBOA-DASFF achieves 99.98% accuracy, 99.82% precision, 99.68% sensitivity, 99.50% specificity, and a 99.74% F1-score—establishing new state-of-the-art benchmarks over seven competitive methods including EACO-ResNet101, IMPA-ResNet50, CSVM, and IFSGA-DNN. A systematic ablation study confirms the statistically significant and independent contribution of each architectural component to the aggregate performance gain. The proposed framework additionally achieves a 68.3% feature dimensionality reduction via IBOA with a total inference time of approximately 4.1 seconds per mammographic image, confirming deployment feasibility for high-throughput clinical screening programs.