AUTOMATED BREAST CANCER DETECTION USING HYBRID CNN AND VISION TRANSFORMER NETWORKS ON MAMMOGRAPHIC IMAGES

Abstract

Breast cancer is a highly prevalent malignant tumor worldwide, arising from the uncontrolled proliferation of abnormal cells in breast tissue. It primarily affects women, though cases also occur in men. Currently, mammography is the primary technology for early breast cancer screening. Detected imaging abnormalities are classified as benign or malignant. However, when radiologists manually interpret these images, the complexity of breast tissue patterns complicates interpretation, often leading to inconsistent readings and missed diagnoses. To address this challenge, this study proposes a hybrid deep learning framework that integrates convolutional neural networks (CNNs) and Vision Transformers (ViTs). CNNs excel at extracting hierarchical local image features, while ViTs capture long-range dependencies and global context. That said, ViTs have relatively high computational and data requirements, so this hybrid model is designed to improve robustness and diagnostic reliability. This study trains its model on the Kaggle breast mammography dataset enhanced with Contrast Limited Adaptive Histogram Equalization (CLAHE). Four pre-trained models, namely DenseNet, Inception, SE-ResNet, and XceptionNet, are selected for performance comparison. The proposed model achieves a validation accuracy of 90.1%, with balanced precision and recall of 89.4% and 90.8%, respectively, demonstrating robust generalization. Although XceptionNet attained perfect accuracy, such performance may suggest potential overfitting. Overall, the hybrid approach effectively balances local and global feature learning, offering a reliable solution for automated breast cancer classification.