Advancements in Electrical Impedance Tomography: Addressing electrode displacement with artificial neural networks

Electrode displacement is a common source of error in Electrical Impedance Tomography (EIT), particularly in long-term or dynamic measurements where stable electrode contact is difficult to maintain. This study proposes a comprehensive machine learning framework to detect, classify, and correct electrode displacements prior to image reconstruction. The approach combines tree-based classifiers—such as XGBoost and LightGBM—and Convolutional Neural Networks (CNNs) to identify both the presence and location of displaced electrodes. These models were evaluated across a series of classification tasks with increasing complexity, demonstrating high accuracy even in scenarios involving multiple simultaneous displacements with different magnitudes. For the rectification of distorted voltage measurements, several Denoising Autoencoder (DAE) configurations were investigated. Electrode-specific DAEs trained on all displacement magnitudes achieved an average Mean Squared Error (MSE) reduction of 84.5%, while shift-based DAEs offered computational efficiency for coarse corrections. A hybrid strategy combining fast shift-based and high-accuracy electrode-specific models proved effective and scalable. The pipeline was validated using both synthetic datasets and real EIT measurements, confirming its robustness and generalization capabilities. These results support the use of learning-based correction schemes to improve the reliability and consistency of EIT in practical applications affected by electrode movement.