Spatiotemporal context feedback bidirectional attention network for breast cancer segmentation based on DCE-MRI

Breast cancer is a highly heterogeneous, both between patients (inter-tumor) and within individual tumors (intra-tumor), leads to indistinct boundaries and varying sizes, shapes, appearances and densities of tumors. Current 3D structural imaging-based methods face challenges to segment breast cancer. The key ingredient to the problem is how to exploit the temporal correlations of 4D functional imaging that depict the heterogeneity of vascular permeability in cancer for accurate tumor segmentation. In this paper, we propose a unique spatiotemporal context feedback bidirectional attention network, which segments breast cancer by modeling dynamic contrast-enhanced dependency to exploit pharmacokinetics feature representations. Specifically, we design a temporal context feedback encoder to learn pharmacokinetics feature representations, which embeds bidirectional temporal attention for bidirectionally propagating contextual semantics across time sequences. Additionally, learned representations are fed into a temporal context feedback decoder to obtain a voxel-level classification of breast tumors. Experimental results demonstrated that the proposed method outperforms recent tumor segmentation methods. Furthermore, our approach achieves competitive results on a small training data and avoids the over-fitting phenomenon due to the model-driven skill to capture dynamic contrast-enhanced temporal correlations.