Deep Neural Networks for Estimating the Bladder Boundary Using Electrical Impedance Tomography

Abstract

Electrical Impedance Tomography (EIT) is a noninvasive imaging technology that aims to reconstruct the cross sectional image of internal conductivity distribution of electrically conducting objects such as human body or any domain. In the working principle of EIT, an array of electrodes are attached on the surface of the human body or object domain to inject an alternating current (AC) and measure the induced voltages on the electrodes, and the corresponding electrical conductivity information is estimated according to Ohm's law. Reconstruction of the internal resistivity distribution using EIT is an ill-posed problem and highly non-linear. EIT image has a lower spatial resolution due to ill-posedness and makes it difficult to reconstruct the boundaries of regions inside the object. If the internal resistivity distribution values of the object can be known a priori, then the inverse problem in EIT becomes estimating the boundary, size, position of the regions inside the object. The approach to estimate region boundaries rather than internal resistivity distributions is known as boundary estimation in EIT. In the clinical applications, boundary estimation of the organ can provide additional clinical information to diagnosis human health. As a result, here we show interest in the boundary estimation of the organ rather than the inner conductivity distribution of the human body. However, the performance of the conventional inverse algorithms for estimation of organ boundaries using EIT is often sub-optimal. To overcome this problem, a deep neural network algorithm is proposed to estimate urinary bladder boundary inside the pelvic domain using EIT. Two deep neural network models were considered and trained the model with pairs of the boundary voltage measurements of pelvic area as input data and the corresponding Fourier coefficients of internal bladder boundary as output data. First 5-layer DNN model is intended to estimate the boundary of a bladder shaped target inside the pelvic domain and trained with pairs of the boundary voltage measurements of pelvic area as input data and the corresponding Fourier coefficients of internal bladder boundary as output data. Whereas the second 5-layer DNN model is designed to estimate boundary of bladder shaped targets surrounded by three other neighboring tissue shaped targets placed inside the pelvic domain. The estimated results of simulations and phantom experiments show that DNN algorithm has significantly better estimation performance as compared to the traditional algorithms.