A hybrid Capsule Network–based computer-aided diagnosis framework for breast cancer detection in mammographic images
Farnaz Hoseini, Masume Kheyri · Intelligence-Based Medicine · 2026
Breast cancer early diagnosis is critical for reducing mortality rates. This study proposes a fully automated hybrid Computer-Aided Diagnosis (CAD) framework integrating unsupervised saliency-based Region-of-Interest (ROI) extraction with a statistical feature-driven Capsule Network (CapsNet) classifier. Unlike methods relying on manual annotations, which risk data leakage, our approach utilizes a gradient and Laplacian-of-Gaussian saliency mechanism to localize ROIs automatically without ground-truth priors. From each ROI, a compact 26-dimensional feature vector describing texture and intensity is computed and classified via CapsNet to capture hierarchical feature dependencies. The framework was extensively evaluated using a rigorous 5-fold cross-validation with patient-level splitting across four heterogeneous datasets: MIAS, DDSM, CBIS-DDSM, and INbreast. Experimental results demonstrated robust generalization, achieving an average accuracy of 97.8% (AUC = 0.99) for Normal/Abnormal and 86.0% (AUC = 0.91) for Benign/Malignant classification. Comparative analyses confirm that this hybrid approach not only outperforms conventional classifiers (e.g., SVM, Naive Bayes) but also exhibits superior stability and data efficiency compared to end-to-end Convolutional Neural Networks (CNNs) trained directly on raw images. These findings suggest that combining domain-informed statistical features with structured capsule learning offers a reliable, interpretable, and dataset-agnostic solution for clinical decision support.