Author: Si-Wa Chan, Yuan-Yu Lee, Zhi-Ying Li, Jia-Wei Liao, Hui-Yu Cathy Tsai 👨🔬
Affiliation: Department of Radiology, Taichung Veterans General Hospital, Institute of Nuclear Engineering and Science, National Tsing Hua University 🌍
Purpose: Dense breast tissue reduces the sensitivity of mammography, posing diagnostic challenges, especially for Asian women with high breast density (up to 50%). Current single-modality techniques often fail in such cases. This study aims to develop a computer-aided diagnosis (CAD) system integrating MRI and digital breast tomosynthesis (DBT) to enhance lesion detection, offering a tailored solution for this high-risk population.
Methods: This study is the first to combine multimodal image registration for MRI, DBT and mammography. A 3D breast model was constructed from MRI and simulated the compression of mammography by finite element method, then generating virtual mammograms. Lesion locations from MRI and DBT were registered onto mammograms with high accuracy. A YOLOv9 object detection model was trained using multimodal annotated data. Transfer learning was employed with pretraining on the large dataset, then, followed by clinical data training and validation.
Results: The results showed that our multimodal registration technique had high accuracy, the average target registration errors are 5.13 mm (MRI) and 2.02 mm (DBT), which surpassed most mammography-based studies. This study also demonstrated the value of clinically challenging cases with uncertain lesion boundaries and dense breasts that cannot be labelled. The multimodal data-trained lesion detection model also showed excellent performance, the average accuracy was 96%, with superior precision, recall, F1 score and mean average precision (mAP) compared to single-modality models.
Conclusion: The developed CAD system successfully integrates MRI, DBT and mammography, providing a reliable multimodal imaging registration method and lesion detection model. This innovation significantly improves lesion detectability in dense breasts, particularly benefiting high-risk populations. The findings offer a foundation for advancing breast cancer diagnosis and suggest promising directions for future clinical applications.