Multi-level Feature Attention Network for medical image segmentation

Network architectures deriving from the Unet framework and its convolutional neural network variants have garnered significant attention for their impressive feats in computer vision. However, the shallow-level details and deep-level semantic information are underutilized in these methods, leading to the model's inability to adequately localize target regions. In this paper, we put forward a Multi-level Feature Attention Network, a novel method that cross-connects encoder and decoder features and focuses on multi-scale semantic features. Firstly, we extend UperNet using a hierarchical Swin Transformer with shifted windows, giving the network global modeling capabilities. Secondly, we introduce a Cross-connection Multi-level Attention module that connects encoder and decoder to refine the decoder's output features and supplement detailed information. Finally, we employ a Pyramid Collaborative Attention (PCA) module to mine the encoder's deepest semantic features across multiple scales. Our method establishes state-of-the-art performance on the ACDC, ISIC2017 and BUSI datasets, showcasing its exceptional capability in segmenting medical images.