The reconstruction of the internal conductivity distribution of a medium using electrical impedance tomography (EIT) is a highly nonlinear and ill-posed inverse problem that is susceptible to both measurement noise and electrode movement due to motion artifacts. This problem is further intensified when EIT is applied to wearable tactile sensors as it must now conform to scenarios where it is meant to stretch and bend. The classical approach of solving the inverse problem is to use a set of linear equations, however this approach tends to produce solutions which are far from optimal.

Deep learning approaches have shown promising results as they perform well in modeling and generalization complicated nonlinear functions based finite datasets. This thesis proposes a modified Deep Convolutional Neural Network (DCNN) to solve the inverse problem of EIT when exposed to both measurement noise and electrode movement. The modified DCNN is trained using a synthetic dataset consisting of 48,000 samples that were randomly generated by solving the forward problem of EIT using the open-source toolbox Electrical Impedance and Diffuse Optical Reconstruction (EIDORS) for MATLAB. The synthetic dataset consists of a random number of input stimuli with varying shapes, sizes, and conductivity levels. A randomly generated 2D transformation matrix was applied to the model to simulate electrode movement. The transformation matrix includes a strain and shear component in both the X and Y axis ranging from [-0.5,0.5] and [-0.25,0.25] respectively.

The noise immunity and electrode movement compensation capabilities of the network were validated against both the discrete Laplacian approach and an electrode movement compensating augmented forward matrix approach. It was found that the network greatly outperforms the other two approaches when evaluated against a newly generated testing dataset. Specifically, the proposed method yields an image correlation coefficient double that of the augmented forward matrix approach and six times greater than the Laplacian approach. The network also demonstrates a high noise immunity as it can precisely reconstruct the conductivity distribution even when a signal-to-noise-ratio of 40dB is applied to the measurements. The network also demonstrates strong electrode movement compensation as it again precisely reconstructs the conductivity distribution despite extreme electrode movement.