Dual‐Branch Multimodal Attention Fusion Networks for Electrical Impedance and Microwave Dual‐Mode Tomography of Stroke

ABSTRACT Electrical impedance tomography (EIT) and microwave tomography (MWT), as two emerging noninvasive imaging techniques, have been widely used in stroke diagnosis. However, single‐modality imaging exhibits inherent limitations such as insufficient information and low resolution, which pose challenges in meeting the requirements of clinical diagnosis. To enhance the resolution of stroke imaging, a dual‐branch multimodal attention fusion network (DMAFusion) for electrical impedance and microwave dual‐mode tomography (EI/MDT) is proposed. The network employs a dual‐branch interactive encoder module to train the encoders separately for different modalities while enabling cross‐modal information fusion. With an improved efficient multiscale attention module, the module enhances feature extraction capabilities. An attentional feature fusion strategy is applied to deeply fuse features from different modalities, obtaining more comprehensive and accurate information. Through comparative and robustness experiments, the results demonstrate that DMAFusion outperforms single‐modality imaging methods, achieving higher‐resolution reconstructed images and improved robustness. Additionally, compared to other multimodal imaging networks, the proposed method significantly enhances image quality and reconstruction accuracy, further validating the effectiveness of EI/MDT and the superiority of DMAFusion in multimodal imaging. Therefore, through multimodal information fusion, the network provides a new technical means for clinical application in noninvasive and precise stroke diagnosis.