다중 스케일 주의기반 네트워크을 통한 의료영상 분할

Abstract

Even though deep learning (DL) based methods have been achieving superior performance in medical image segmentation, such methods still have some downsides. First, the use of skip-connections in encoder-decoder architecture like U-Net allows transferring redundant and superfluous low-level features information at multiple scales. Second, prior methods cannot capture long-range dependencies and hence fail to reconstruct the feature maps adeptly. To subdue these problems, we propose an architecture that adaptively captures global correlations from different scales and utilizes the attention mechanism. This approach integrates the local-features at different scales and underlines the essential features by suppressing noises and unwanted information. We evaluate the proposed architecture in the context of medical image segmentation on two different datasets: Kvasir-SEG and nuclei segmentation. Experimental results show that the proposed model yields better accuracy and outperforms previous methods.