Reliable 3D EIT reconstruction for masonry condition monitoring using conditional diffusion and cascaded FEM-space refinement

Three-dimensional electrical impedance tomography (EIT) reconstruction of moisture-related conductivity in masonry requires preservation of both global structure and local finite-element variations. This study proposes a conditional denoising diffusion probabilistic model with cascaded FEM-space residual refiners (cDDPM-SFR). A 448-component voltage vector conditions a 3D U-Net on a 16 × 16 × 16 voxel grid, followed by three multilayer perceptron (MLP) stages producing 3050 outputs. Training used 49,000 synthetic cases and evaluation 1,000. Compared with MLP, mean squared error (MSE) decreased from 0.119 to 0.048, peak signal-to-noise ratio (PSNR) rose from 21.692 to 26.961 dB, structural similarity index (SSIM) from 0.978 to 0.992, and Pearson correlation coefficient from 0.981 to 0.992 (Holm-adjusted p < 0.001). Four field acquisitions showed correlations of 0.196–0.935 and mean median absolute deviation of 0.689. Independent dielectric profiles and the EIT surface proxy identified the same moisture-related zone near 360 cm, supporting physical plausibility.