Deep Learning-enabled Fully Automated 3D DCE-MRI Segmentation for Breast Cancer Lesion

Motivation: Automating and optimising the segmentation process for breast tumours aims to reduce the workload of radiologists and improve the consistency and reliability of segmentation. Goal(s): To develop a 3D DCE-MRI-based deep network architecture that would achieve fully automated segmentation of breast cancer lesions. Approach: A dataset consists of 622 (298+128+196) anonymized DCE-MRI scans was collected. We proposed the attention and transformer based method for 3D breast tumor segmentation, incorporating Residual and UNet modules to enhance feature relevance and capture multi-scale context. Results: Our models have achieved significant results in automated breast cancer lesion identification and segmentation. Impact: By constructing and training an efficient deep learning model to achieve high-precision segmentation of breast cancer lesions, it provides a powerful auxiliary tool for clinical diagnosis, treatment planning and prognosis analysis.