Dual-domain attention Mamba-based 3D U-net architecture for multiple sclerosis lesion segmentation

Accurate segmentation of small lesions of Multiple Sclerosis (MS) remains challenging in clinical diagnosis because of their subtle size, low contrast, and heterogeneous appearance in MRI. While deep learning models, including U-Net and other fused models, have advanced the field, they often struggle to balance global context modeling, local detail preservation, and computational efficiency. In this paper, we propose a novel hybrid 3D architecture on the basis of the traditional U-Net for precise small MS lesion segmentation. In our proposed architecture, the encoder integrates a Dual-Domain Attention (DDA) block that extracts frequency domain features by Fourier transformation and enhances spatial features via a channel attention mechanism, improving the representation of subtle lesion boundaries. In the bottleneck, we replace the standard 3D convolutional layers in the bottleneck of U-Net layers with the Vision Mamba (ViM) block, which aims at efficiently modeling long-range dependencies with linear complexity. Within the decoding path, the Convolutional Block Attention Module (CBAM) further refines spatial details. Besides, inspired by ResUNet, residual units are integrated into both the encoder path and decoder path to better address the issue of vanishing and exploding gradients and therefore improve the performance of the model. Extensive experiments on the MSLesSeg 2025 prove that our architecture realizes outstanding performance in segmenting small MS lesions with an average DSC of 0.6529, reaching state-of-the-art performance while ensuring low time and space complexity at the same time. The results verify that our framework significantly enhances performance.