Enhanced Multi-Scale Feature Cross-Fusion Network for Impedance–Optical Dual-Modal Imaging

Abstract

The intrinsic issue of low spatial resolution of electrical impedance tomography (EIT) is a long-standing challenge that hinders the capability of performing quantitative analysis based on EIT image. Our recent work demonstrates an impedance–optical dual-modal imaging framework and a deep learning model named multi-scale feature cross-fusion network (MSFCF-Net) to realize information fusion and high-quality EIT image reconstruction. However, this framework’s performance is limited by the accuracy of the mask image obtained from an auxiliary imaging modality. This article further proposes a two-stage deep neural network, which is the enhanced version of MSFCF-Net (named En-MSFCF-Net), to automatically improve the mask image and conduct information fusion and image reconstruction. Compared with MSFCF-Net, En-MSFCF-Net demonstrates the superior ability to correct the inaccurate mask image, leading to a more accurate conductivity estimation. Furthermore, En-MSFCF-Net also maintains the best shape preservation and conductivity prediction accuracy among given learning-based and model-based algorithms. Both the qualitative and quantitative results indicate that En-MSFCF-Net could make dual-modal imaging more robust in real-world situations.