Using real data to train GREIT improves image quality

Image reconstruction in electrical impedance tomography is sensitive to errors in the (forward) model of the measurement system. We propose a new approach, based on the GREIT algorithm, where the reconstruction matrix is trained on real rather than simulated data, obvia t- ing the need for an accurate numerical forward model. We observe a substantial improvement in image quality, pa r- ticularly for changes close to the boundary.