WCSE 2020 Summer
ISBN: 978-981-14-4787-7 DOI: 10.18178/wcse.2020.06.030

AM-AN: Adversarial Network Based on Attention Mechanism for Medical Image Segmentation

Lingjing Qin, Wenxin Hu, Jun Zheng

Abstract—Image segmentation is crucial to a series of medical applications. One of the current problems is that imbalanced pixel classes have a negative effect on the results of medical image segmentation. However, the method of employing weighted loss cannot address this problem well. Therefore, we are motivated to propose a novel method based on Generative Adversarial Networks (GANs) and attention mechanisms to train our segmentation model with effective loss functions. Firstly, the proposed model consists of the segmentor network and the critic network, which are trained by adversarial learning. Moreover, the attention module we used in each residual block takes not only channel attention but also spatial attention into consideration and makes the segmentor perform better. The overhead of computation and parameters costed by our attention module is negligible. In addition, our work extends other methods by means of using pixel-wise loss functions, which include multi-scale loss and binary cross entropy loss. Finally, our work presents the results of comparing three different methods on medical image segmentation and our method yields a higher performance than theirs on MICCAI PROMISE12, DSB2018 and the lung nodule datasets

Index Terms— Medical image segmentation, Attention mechanism, Adversarial network, Pixel-wise Loss

Lingjing Qin, Wenxin Hu, Jun Zheng
School of Data Science & Engineering, East China Normal University, CHINA

[Download]


Cite: Lingjing Qin, Wenxin Hu, Jun Zheng , " AM-AN: Adversarial Network Based on Attention Mechanism for Medical Image Segmentation " Proceedings of 2020 the 10th International Workshop on Computer Science and Engineering (WCSE 2020), pp. 185-194, Shanghai, China, 19-21 June, 2020.