MAFUNet: Mamba with adaptive fusion UNet for medical image segmentation

In medical image segmentation tasks, accurately capturing lesion contours and understanding complex lesion information is crucial, which relies on efficient collaborative modeling of local details and global contours. However, methods based on convolutional neural networks (CNNs) and transformers are limited by local receptive fields and high computational complexity, respectively, making it difficult for existing approaches to achieve a balance between the two. Recently, state-space models represented by Mamba have gained attention due to their significant advantages in capturing long-range dependencies and computational efficiency. Based on the above advantages of Mamba, we propose Mamba with Adaptive Fusion UNet (MAFUNet). First, we design a hierarchy-aware Mamba (HAM) module. HAM progressively transmits local and global information across different channel branches through Mamba and balances feature contributions through a dynamic gating mechanism, improving the accuracy of lesion region recognition. The multi-scale adaptive fusion (MAF) module combines HAM, convolution block, and cascaded attention mechanisms to achieve efficient fusion of lesion features at different scales, thereby enhancing the model’s robustness and precision. To address the feature alignment issue, we propose adaptive channel attention (ACA) and adaptive spatial attention (ASA) modules, where the former achieves channel enhancement through dual-scale pooling and the latter strengthens spatial representation using a dual-path convolution strategy. Extensive experiments on the BUSI, CVC-ClinicDB, and ISIC-2018 three public datasets show that MAFUNet achieves excellent performance in medical image segmentation tasks.