Enhanced Breast Cancer Detection Using Dual-View Mammography and Deep Learning Models

Background and Aims: Breast cancer is the most prevalent cancer among women, and early detection is essential for improving survival rates. Mammography is widely used for screening, but interpreting mammograms requires specialized expertise, which is often limited. This study aims to enhance breast cancer detection by leveraging deep learning techniques, including convolutional neural networks (CNNs) and object detection algorithms like YOLOv۵. The hypothesis is that combining dual-view mammography (cranio-caudal and mediolateral oblique views) with these techniques will improve detection accuracy and model efficiency for deployment on mobile/edge devices. The dataset used includes ۵,۰۰۰ mammogram images from the VinDr-mammo dataset and ۳,۱۲۳ images from a local dataset of ۱,۰۲۸ patients. Methods: The study employs a dual-model approach: an EfficientNetB۳-based abnormality classifier and a YOLOv۵-based mass detection model. Both models process mammograms from two views, and transfer learning is utilized to improve model performance. Pruning is applied to reduce the size and computational cost of the models, making them suitable for edge deployment. Performance is evaluated using metrics such as accuracy, sensitivity, specificity, F۱-score, and area under the curve (AUC). Results: The dual-view approach achieves a sensitivity of ۰.۸۹ and an F۱-score of ۰.۷۹, significantly outperforming single-view methods. The integration of YOLOv۵ improves sensitivity by detecting suspicious masses more effectively. After pruning and transfer learning, the model size is reduced by ۷۵% with minimal loss in accuracy (۹۷.۲۳%), making it deployable on mobile devices with real-time capabilities. Conclusion: This study demonstrates that dual-view mammography combined with deep learning significantly enhances breast cancer detection. The model provides high sensitivity and is optimized for deployment on mobile/edge devices, making it accessible in resource-constrained settings. Future work will focus on expanding its detection capabilities and testing it in clinical environments.

کلیدواژه ها:

Breast Cancer ، Deep Learning ، Transfer Learning