Electrical Impedance Tomography for Stroke Diagnosis Using Deep Neural Networks
Electrical impedance tomography (EIT) is a portable, non-ionizing imaging modality with potential for bedside stroke monitoring. However, its nonlinear, ill-posed inverse problem is sensitive to measurement noise, modeling errors, and electrode faults. This research develops a physics-based computational framework for robust conductivity reconstruction, electrode-fault localization, and anatomically informed imaging. The forward problem is modeled using the complete electrode model and finite element method, with Jacobian-based sensitivity analysis. Conductivity reconstruction combines hybrid L1/L2 data fidelity and regularization using a constrained Levenberg-Marquardt iteratively reweighted least-squares algorithm. Numerical experiments with a 32-electrode system show that robust data fidelity improves reconstruction under single- and two-electrode faults, while the optimal prior mixture depends on the evaluation metric. The lowest mean reconstruction errors are 0.0278 S/m and 0.0410 S/m for single- and two-electrode faults, respectively. A multi-task convolutional neural network is also evaluated for electrode-fault detection and localization, achieving 97.41% exact-set localization accuracy under nominal conditions, although its performance degrades substantially under mild fault conditions. Methods based on anatomical contour processing of a two-dimensional head geometry, Bayesian statistical priors and learning-assisted reconstruction are formulated for further investigation. The findings demonstrate the potential of integrating physical modeling, robust optimization, and learning-based fault identification to support future developments in anatomically informed EIT reconstruction for stroke assessment.