CM2-Net: A Hybrid CNN-Mamba2 Net for 3D Electromagnetic Tomography image reconstruction

Three-dimensional electromagnetic tomography (3D EMT) image reconstruction is characterized by larger data volumes, more complex structures, and stronger spatial correlations. These characteristics often result in insufficient reconstruction accuracy when using traditional imaging algorithms. To address this challenge, we propose a deep learning-based reconstruction algorithm, Hybrid CNN-Mamba2-Net (CM2-Net), specifically designed for 3D-EMT image reconstruction. The architecture of CM2-Net consists of three main components: an initial reconstruction module, a hybrid encoder, and a fusion decoder. The CNN pathway is capable of capturing fine-grained local features, while the Mamba-2 pathway, equipped with a linear attention mechanism, excels at learning and representing the three-dimensional spatial structure of the input data. Features from the CNN and Mamba2 pathways are fused at three different scales, summed point-wise, and restored to the original spatial resolution through the decoder. This design enables more accurate feature extraction and representation. The proposed model outperforms the baseline across five quantitative evaluation metrics, achieving a Dice score of 0.718, accuracy of 0.943, precision of 0.657, recall of 0.794, and an IoU of 0.561, demonstrating strong performance in the 3D-EMT reconstruction task.