Efficient low-dose CT image enhancement using MobileMamba-UNet with wavelet-enhanced long-range modeling

Deep learning has become a dominant paradigm for low-dose computed tomography (LDCT) image reconstruction. Nevertheless, existing approaches still struggle to simultaneously achieve accurate structural detail preservation and computational efficiency, particularly when handling long-range contextual dependencies.

To design a lightweight yet effective LDCT reconstruction framework that captures both global contextual information and fine-grained local details while maintaining low memory consumption and fast inference speed.

We propose MobileMamba-UNet, a hybrid neural network that integrates a MobileMamba backbone with a multi-scale U-Net architecture. The model incorporates a Wavelet Transform Enhanced Mamba mechanism to emphasize high frequency and diagnostically relevant structures, together with a multi-receptive field feature interaction module that jointly models local textures and long-range dependencies. All components are constructed with linear computational complexity to ensure efficiency in large-scale LDCT reconstruction tasks.

Extensive experiments conducted on the Mayo-2016 and Mayo-2020 LDCT datasets demonstrate that MobileMamba-UNet consistently outperforms existing CNN- and Transformer-based methods. The proposed approach achieves superior image quality while significantly reducing memory usage and inference latency.

MobileMamba-UNet represents a promising approach for LDCT image reconstruction, balancing reconstruction performance with computational efficiency and practical applicability.