<![CDATA[Electrical Impedance Tomography (EIT) is a vital non-invasive imaging technique for dynamic monitoring, such as lung ventilation. The primary challenge in EIT lies in the inverse problem, which is non-linear, ill-posed, and computationally slow, especially when high accuracy and real-time speed are simultaneously required. Conventional EIT reconstruction algorithms often yield blurred images and are highly susceptible to measurement noise and geometrical uncertainties, such as variations in electrode placement and unknown boundary shapes. This research proposes the Deep Physics-Informed Neural Network (D-PINN), an extended deep learning framework, to achieve accurate and real-time dynamic EIT reconstruction. Unlike purely data-driven methods, our D-PINN integrates the governing Laplace’s Equation directly into the network’s loss function, providing a strong physical constraint to significantly enhance image quality. The innovative focus of this study is addressing the critical gap in model uncertainty robustness. We develop a stochastic D-PINN training scheme that not only solves the conventional inverse problem (predicting conductivity) but also simultaneously accounts for small variations in boundary geometry or electrode positions. Initial simulation results are expected to show that D-PINN consistently:1. Reduces the reconstruction inference time to the millisecond scale, enabling true real-time monitoring. 2. Significantly improves the spatial resolution and image contrast (measured by the Structural Similarity Index / SSIM) compared to standard iterative methods. 3. Maintains high accuracy even when the input measurement data is noisy and the assumed forward geometrical model is intentionally perturbed, which is crucial for real-world instrumentation applications. This work is expected to advance EIT into a more reliable and robust real-time imaging tool.]]>
Copyrights © 2025
