Cross-Modal Self-Attention Distillation for Prostate Cancer Segmentation
Guokai Zhang, Xiaoang Shen, Ye Luo, Jihao Luo, Zeju Wang, Weigang Wang, Binghui Zhao, Jianwei Lu · 2020
The automatic segmentation of prostate cancer (PCa) from the multi-modal magnetic resonance imaging (MRI) is of prime importance for the initial staging and prognosis of patients. Nevertheless, the challenge of how to utilize multi-modal image features more efficiently still needs resolving, especially in the segmentation scenario. In this paper, we propose a crossmodal self-attention distillation network that can fully exploit the encoded information of the intermediate layers from different modalities, and the learned attention maps of different modalities are then transferred among modalities to provide significant spatial information with more details incorporated. Furthermore, we propose a novel spatial correlated feature fusion module that is able to learn more complementary correlation and nonlinear information from different modality images. To evaluate the effectiveness of the proposed approach, we conduct extensive experiments on the PCa MRI dataset, and the experiment results demonstrate that our proposed approach could achieve state-of-the-art performance.