Att-MoE: Attention-based Mixture of Experts for nuclear and cytoplasmic segmentation

Att-MoE: Attention-based Mixture of Experts for nuclear and cytoplasmic segmentation

Liu, Jinhua and Desrosiers, Christian and Zhou, Yuanfeng

Neurocomputing 2020

Abstract : Cell segmentation is a critical step in histology images analysis. Recently, Convolutional Neural Network (CNN) has shown outstanding performance for various segmentation problems, however, the segmentation of histology images remains challenging due to the tight arrangement of cells and their weak boundaries. This paper proposes a novel architecture called Attention-based Mixture of Experts (Att-MoE) for nuclear and cytoplasmic segmentation in fluorescent histology images, which integrates multiple Expert networks using a single Gating network. Expert networks complement each other to accomplish sub-tasks under the direction of the Gating network, which enforces the adaptive use of multiple networks to complete the segmentation task. The Att-MoE also introduces attention gates and residual blocks in the Expert networks to improve segmentation accuracy. The attention gate is used to emphasize useful features and suppress irrelevant features for segmentation in a self-adaptive manner. On the other hand, residual blocks are employed to enhance gradient flow in training and improve the stability and segmentation accuracy of the network. Experiments on fluorescent histology images of mouse liver show that Att-MoE is superior to recent segmentation methods and has the potential for cancer diagnosis based on histology images.