Automated identification and categorization of breast lesions in mammographic images using an AMFA-based SVM framework

Abstract Breast cancer is a major cause of cancer-related mortality among women, emphasizing the need for accurate and efficient computer-aided diagnosis. Existing deep learning approaches for mammographic analysis predominantly rely on single feature representations or conventional fusion strategies, limiting their ability to exploit complementary geometric information across multiple feature manifolds. This study proposes an Adaptive Multi-Manifold Feature Aggregation–Support Vector Machine (AMFA-SVM) framework for automated breast lesion classification. The framework was evaluated on the publicly available INbreast full-field digital mammography dataset containing expert-annotated benign and malignant lesions. Mammograms were enhanced using Contrast Limited Adaptive Histogram Equalization (CLAHE), Gabor filtering, and Gaussian filtering, followed by deep feature extraction using EfficientNet-B0. Complementary feature representations were generated using Principal Component Analysis (PCA), Uniform Manifold Approximation and Projection (UMAP), and t-distributed Stochastic Neighbor Embedding (t-SNE). The proposed AMFA module adaptively weights these representations according to their discriminative contribution and integrates them for optimized SVM classification. AMFA-SVM achieved 93.2% accuracy, 81.8% precision, 98.0% recall, 88.9% F1-score, and 93.9% AUC, demonstrating high sensitivity for malignant lesion detection. Comparative evaluation further showed a favorable balance between diagnostic performance, computational efficiency, and feature discrimination. These findings demonstrate that adaptive aggregation of complementary manifold representations can enhance mammographic lesion classification and provide an efficient framework for computer-aided breast cancer diagnosis.