Cardiac Imaging with Electrical Impedance Tomography (EIT) using Multilayer Perceptron Network

Abstract

This research explores the enhancement of Electrical Impedance Tomography (EIT) for cardiac imaging using Multilayer Perceptron (MLP) networks, focusing on supervised and semi-supervised learning approaches. Using synthetic thoracic datasets simulating dynamic cardiac and respiratory conditions, the study demonstrates that supervised learning achieves lower mean squared error (MSE) values (minimum 4.76) and more stable predictions compared to semi-supervised learning (minimum MSE 5.08). However, semi-supervised learning excels in edge accuracy and noise reduction, particularly in regions with sharp conductivity gradients, making it viable for scenarios with limited labeled data. Dropout regularization at 0.3 provided optimal balance, enhancing model generalization and robustness. While supervised learning outperformed semi-supervised methods in overall accuracy, the latter showed potential for cost-effective and scalable applications in EIT-based cardiac imaging. These findings suggest that integrating advanced machine learning with EIT can improve diagnostic accuracy and enable efficient use of sparse labeled data, paving the way for future optimizations and clinical applications.

Keywords

• electrical impedance tomography
• multilayer perceptron
• semi-supervised learning
• cardiac imaging
• machine learning