METrans: Multi‐encoder transformer for ischemic stroke segmentation
Jing Wang, Shuyu Wang, Wei Liang
Abstract
Abstract Ischemic stroke is the most common brain disease. Segmentation of the stroke lesion from a medical scan is vital to plan the surgical procedure. The purpose of this work is to develop an efficient network based on self‐attention and spatial‐channel attention mechanisms. In this letter, a novel multi‐encoder transformer (METrans) is proposed, which overcomes the inability of U‐Nets to model long‐range contextual interactions. Different from traditional segmentation methods, four encoders with different scales are explored to extract the multi‐scale features. Then, the feature maps are fed into a transformer for global feature modelling. The proposed methodology is tested on ATLAS, ISLES 2015, and ISLES 2018 datasets. The extensive experimental results suggest that METrans achieves consistent improvements over the state‐of‐the‐art for segmentation tasks.