A Cloud-Edge Collaborative Learning-Based Electrical Impedance Tomography Method

Deep learning-based Electrical Impedance Tomography (EIT) technology encounters significant challenges including insufficient image clarity and boundary blurring. To address these issues, we propose a Multi-Scale Residual Attention Model (MSARM) that integrates multi-scale feature fusion with a parameter-free attention mechanism and trains the network using a hybrid L1-L2 loss function to improve the accuracy of image reconstruction. Experiments conducted on a simulated dataset demonstrate that the proposed method achieves a 1.87% improvement in the Correlation Coefficient (CC) metric and a significant 17.42% reduction in Relative Error (RE) compared to the prevailing multi-scale U-Net model. Furthermore, phantom experiments have validated the effectiveness and generalization capability of the proposed method. However, deep learning-based EIT face practical deployment challenges such as high latency, data loss, and privacy breaches. In response, we introduce a novel cloud-edge collaborative EIT system architecture, comprising two-level cloud servers, edge computing nodes, and terminal devices. Experimental results indicate that, compared to the traditional cloud-only service architecture, this architecture reduces the data transmission time by approximately 50% and maintains data integrity during network fluctuations. The proposed EIT system architecture not only improves the real-time and quality of image reconstruction but also provides a viable solution for EIT clinical application and remote monitoring.