Abstract
This paper presents the development of a novel miniature electrical impedance tomography (EIT) system made out of glass, along with the training, validation, and testing of an accompanying open-source machine learning image reconstruction model. Our 1-dimensional convolutional neural network (1D-CNN) models were uniquely benchmarked, both qualitatively and quantitatively, using synthetic and experimental data, against well-established image reconstruction methods: the one-step Gauss-Newton method and the total variation reconstruction method. Image reconstruction results obtained using our 1D-CNN show significant advantages over these traditional methods, achieving an up to 5-fold reduction in mean square error on synthetic data. These results were replicated for two common excitation/measurement modes and extended to objects with varying conductivity and quantities. The superior EIT reconstruction capabilities of our 1D-CNN were further validated experimentally across a similarly broad range of parameters, achieving an average positional accuracy of 147 μm and an average dimensional resolution of 70 μm. To demonstrate potential applications in in vitro monitoring, we used our platform to observe zebrafish development through three distinct phases, from embryo to larvae, showcasing our platform's compatibility with biological imaging.