A Pattern Recognition-Based Deep Learning Framework for Breast Cancer Classification in Digital Breast Tomosynthesis Using a Hybrid Feature Fusion Approach

Early detection of breast cancer is essential for improving patient prognosis, and digital breast tomosynthesis (DBT) has emerged as a more advanced imaging technique compared to traditional 2D mammography. This study investigates the use of deep learning models for breast cancer classification using DBT images, incorporating both single-slice and multi-slice inputs. The preprocessing pipeline involves normalization, resizing, contrast enhancement via CLAHE, and data augmentation to improve image quality and diversity. Feature extraction includes morphological, texture, and density-based analyses, along with deep features obtained from convolutional neural networks (CNNs). These features are systematically combined into single, double, triple, and quadruple sets to enhance feature representation. To optimize feature selection and minimize redundancy, Exhaustive Feature Selection (EFS) is applied, improving computational efficiency. The classification framework integrates deep learning architectures such as ResNet v2, EfficientNet, and a customized Inception v3[Formula: see text] within a hybrid model. Additionally, ensemble learning with XGBoost is employed, with hyperparameter tuning conducted through Grid Search (GS) for performance optimization. The model’s effectiveness is assessed using accuracy, sensitivity, specificity, and area under the curve (AUC) metrics. Experimental results indicate that the proposed method achieves high diagnostic accuracy, with the integration of EFS, CLAHE, and multi-slice processing significantly enhancing model performance and robustness for clinical breast cancer detection.