• FA-UNext: A Feedback Attention-based MLP Network for Medical Image Segmentation

    Qianyu Li, Bingcai Chen, Jiaxing Tian, RuoLan Liu

    Abstract

    Medical imaging technology and deep learning research have greatly contributed to the advancement of medical image analysis, particularly in medical image segmentation. Although traditional methods tend to train models on large datasets, the process based on single-step prediction prevents them from effectively utilizing information from different training epochs. To address this issue, we propose a novel medical image segmentation network called FA-UNext. It is the first network in medical image segmentation to introduce the feedback attention mechanism into the MLP-based network and combines the advantages of convolutional and MLP structures. The network utilizes the prediction mask from the previous epoch as a feedback attention mechanism to fuse with the feature map. This helps to propagate the information flow in successive training epochs, allowing the network to make full use of information from different learning periods. The attention mechanism helps to focus on important regions while suppressing irrelevant features. Our results show that FA-UNext outperforms current state-of-the-art generalized models on the ISIC2018 skin lesion dataset and the Breast UltraSound Images dataset. Moreover, it has a lower number of parameters and faster inference speed.