MSA-VNet: Multi-scale Attention-based V-Net for DCE-MRI Lesion Segmentation
Chang Yan Xu, Zi Jiang Sang, Ye Qin Shao · 2022 International Conference on Image Processing, Computer Vision and Machine Learning (ICICML) · 2022
Dynamic Contrast Enhancement MRI (DCE-MRI) has become an essential tool for detecting breast cancer in recent years. However, the shape and size of lesions vary widely, and the boundary of lesions is blurry. This work proposes a multi-scale attention-based V-Net (MSA-VNet) for DCE-MRI lesion segmentation, to address these issues. MSA-VNet, based on V-Net, initially employs the 3D multi-receptive-field feature extraction module, which includes multi-branch residual structure, atrous convolutions, and instance normalization layers. Second, to replace the long-range skip connection structure in V-Net, an attention-based long-range skip module is proposed. Finally, the Focal Toversky loss function is introduced in MSA-VNet to enable the model to focus on tiny lesions. The experiments on the breast cancer DCE-MRI dataset show that, the proposed method outperforms the state-of-the-art methods.