A hybrid deep learning architecture for breast cancer classification

Accurate classification of breast lesions in mammography and ultrasound images remains a clinically critical yet challenging task, requiring both high diagnostic precision and interpretability. While machine learning (ML) and deep learning (DL) approaches have shown strong performance in computer-aided diagnosis, most existing methods rely on end-to-end convolutional architectures that lack transparency and fail to explicitly capture geometric and morphological characteristics of lesion structures. In this study, we propose a hybrid feature extraction framework that integrates topological data analysis (TDA) with classical shape and texture descriptors for breast cancer classification. Persistent homology is applied via sublevel-set filtration on contrast-enhanced grayscale images to generate persistence diagrams, from which persistence values of β ₀ β 0 and β ₁ β 1 features are extracted. These descriptors capture structural properties while reducing redundancy in topological representations. The resulting topological features are combined with Zernike moments and Gray-Level Co-occurrence Matrix (GLCM)-based Haralick features to form a comprehensive feature pool. An optimization-based feature selection process identifies the most discriminative features, which are then used to train three lightweight architectures: a residual Multi-Layer Perceptron (Topo-Opt-MLP), a Transformer-based model capturing global feature dependencies (Topo-Opt-Trf), and a Feature Tokenizer Transformer (Topo-Opt-FT-Trf) that models fine-grained interactions between features. Experiments on the BUSI and CBIS-DDSM datasets demonstrate competitive performance, achieving up to $$95.4%$$ 95.4 % sensitivity and $$93.8%$$ 93.8 % accuracy. The proposed framework provides an interpretable and data-efficient alternative to conventional DL models, supporting reliable clinical decision-making.