Advanced Image Reconstruction for Electrical Impedance Tomography via Ensemble Learning with Consideration of Measurement Data Quality and Prior-Knowledge-Guided Denoising

Electrical Impedance Tomography (EIT) is a non-invasive imaging method with significant potential for lung state assessment, yet constrained by insufficient image quality. Image reconstruction is a non-linear, ill-posed inverse problem, highly sensitive to measurement perturbations. Existing methods fail to address the dual challenge of conductivity accuracy and structural boundary preservation simultaneously, while post-processing approaches remain computationally intensive and lack prior-knowledge integration, causing persistent residual artifacts. This thesis introduces a holistic framework tackling multiple reconstruction stages. A Gaussian Process regression-based pre-processing achieves 99.97% Mean Squared Error reduction, improving Signal-to-Noise Ratio from 1 dB to 36 dB. An ensemble learning strategy combines a 1D-Residual-CNN-GRU optimized for conductivity accuracy with an enhanced U-Net for structural preservation, integrated via Ridge regression stacking, yielding 3.7% Image Correlation Coefficient (ICC) improvement and 60.8% Relative Image Error (RIE) reduction over state-of-the-art methods. A prior-knowledge-guided post-processing applies targeted denoising, achieving 2.9% ICC improvement and 16.7% RIE reduction. Extended to lung state assessment, the ensemble module achieves 2.9% ICC improvement and 79.3% RIE reduction. Experimental validation on water tank setups and custom PCB thoracic phantoms confirms robustness under real measurement conditions.