RCLD-UNet: An Asymmetric U-Shaped Encoder-Decoder Based on Recurrent Criss-Cross Attention and Linear Deformable Convolution for Medical Image Segmentation

Abstract

With the continuous advancement of medical technology, medical image segmentation plays a pivotal role in various aspects of disease diagnosis, treatment planning, and therapeutic evaluation. In recent years, methods based on U-shaped Convolutional Neural Networks (CNNs), U-shaped Transformer architectures, and U-shaped Mamba models have achieved significant progress in this domain, greatly promoting research and application development. Although CNNs excel at local feature extraction, they are inherently limited in modeling global dependencies due to the locality of convolution operations. Transformers address this limitation through self-attention mechanisms that capture long-range dependencies effectively, but their quadratic computational complexity with respect to sequence length severely impacts efficiency. While the Mamba architecture demonstrates promising performance in long-sequence modeling, its generalization to other tasks remains limited. To address these challenges, this paper proposes RCLD-UNet, an efficient asymmetric U-shaped encoder-decoder network designed for 2D medical image segmentation. Built upon the U-Net backbone, RCLD-UNet incorporates the Recurrent Criss-Cross Attention (RCCA) module in the encoder, which leverages sparse connections and recurrent operations to efficiently capture global contextual information, thereby overcoming the limitations of standard convolutions in modeling global dependencies. In the decoder, a Linear Deformable Convolution (LDConv) module is introduced to enhance feature extraction through irregular convolutions, while ensuring linear parameter growth to balance performance and computational cost. Extensive experimental results on three public datasets indicate that RCLD-Unet achieves superior performance compared to existing methods. Specifically, the proposed model achieves average Dice Similarity Coefficients (DSC) of 84.14 %, 92.02 %, and 80.84 % on the Synapse, ACDC, and MoNuSeg datasets, respectively.