Electrical Impedance Tomography Shape Classification Using Deep Neural Networks

Electrical impedance tomography (EIT) is often formulated as an inverse problem of reconstructing internal conductivity distribution from boundary-voltage measurements. In contrast to conventional image reconstruction approaches, this work addresses a supervised task classification of the shape of a single inclusion directly from experimentally acquired EIT voltage data. The dataset consists of measurements obtained from a circular saline-filled phantom instrumented with 16 boundary electrodes and containing one polyethylene terephthalate glycol (PETG) inclusion with circular, square, or triangular geometry. A protocol-ordered voltage feature was used as input to a fully connected deep neural network classifier. The proposed model achieved approximately 97% test accuracy under the adopted experimental split, while a random forest baseline produced comparable performance. These findings show that measured EIT voltage measurements signals contain discriminative information for object-shape classification. However, they also indicate that the dataset is strongly separable and that future work may include additional phantom configurations, and comparisons with reconstruction-based and convolutional approaches.