A deep neural network for estimating the bladder boundary using electrical impedance tomography

Accurate bladder size estimation is an important clinical parameter that assists physicians, enabling them to provide better treatment for patients who are suffering from urinary incontinence. Electrical impedance tomography (EIT) is a non-invasive medical imaging method that estimates organ boundaries assuming that the electrical conductivity values of the background, bladder, and adjacent tissues inside the pelvic domain are known a priori. However, the performance of a traditional EIT inverse algorithm such as the modified Newton-Raphson (mNR) for shape estimation exhibits severe convergence problems as it heavily depends on the initial guess and often fails to estimate complex boundaries that require greater numbers of Fourier coefficients to approximate the boundary shape. Therefore, in this study a deep neural network (DNN) is introduced to estimate the urinary bladder boundary inside the pelvic domain.We designed a five-layer DNN which was trained with a dataset of 15 subjects that had different pelvic boundaries, bladder shapes, and conductivity. The boundary voltage measurements of the pelvic domain are defined as input and the corresponding Fourier coefficients that describe the bladder boundary as output data. To evaluate the DNN, we tested with three different sizes of urinary bladder.Numerical simulations and phantom experiments were performed to validate the performance of the proposed DNN model. The proposed DNN algorithm is compared with the radial basis function (RBF) and mNR method for bladder shape estimation. The results show that the DNN has a low root mean square error for estimated boundary coefficients and better estimation of bladder size when compared to the mNR and RBF.We apply the first DNN algorithm to estimate the complex boundaries such as the urinary bladder using EIT. Our work provides a novel efficient EIT inverse solver to estimate the bladder boundary and size accurately. The proposed DNN algorithm has advantages in that it is simple to implement, and has better accuracy and fast estimation.