Multi-contrast MRI Imputation Using Mamba with U-Net Shaped Architecture

Deep learning techniques have significantly advanced the generation of missing imaging modalities from available acquisitions. However, effectively capturing the complex, nonlinear tissue transformations that arise from varying spatial interactions remains a key challenge. Conventional convolutional neural networks and attention-based architectures have either a lack of global context or high complexity. In this work, we introduce a novel approach for multi-contrast MRI imputation that combines the strengths of a U-Net shaped backbone with state space processing (Mamba). Our method uses Mamba blocks in the network to obtain extended spatial dependencies while maintaining minimal model complexity. Experimental results show that our model consistently outperforms existing approaches, demonstrating its effectiveness in faithfully reconstructing target modalities with superior structural and contextual fidelity.