Multiscale Rolling Attention Network With Enhanced Local and Global Features for Medical Image Segmentation

Medical image segmentation plays a key role in disease diagnosis, but its accuracy is often constrained by the morphological variability and scale variability of lesions. Although existing methods alleviate this problem by fusing local and global features, they suffer from the defects of low feature fusion efficiency and insufficient multiscale modeling. To this end, we propose the LLA network, the core of which is to enhance the model's ability to extract detailed features from images by learning global contextual information in multiple directions of the whole image through a parallel dual orthogonal rolling multilayer perceptron (DOR-MLP), as well as by enhancing the extraction of detailed features from images through the local perceptual ability of the windowed attention module. We introduce multiscale field blocks (MSF) in skip connections containing four parallel convolutional branches of different sizes to extract more comprehensive and richer feature information at different scales. The encoder and decoder utilize double-layer convolution and residual concatenation for efficient feature extraction. Experiments on BUSI, PH2, and DDTI datasets show that the IoU reaches 73.32%, 90.96%, and 70.89%, respectively, and our method effectively captures local and global information and achieves better segmentation results compared to other state-of-the-art methods.