NEURAL NETWORKS FOR ELECTRICAL-IMPEDANCE TOMOGRAPHY IMAGE CHARACTERIZATION

The Southampton electrical impedance tomography (EIT) system used a Sheffield data acquisition unit and a PC based 'Harlequin' transputer card to reconstruct and display images of the distribution of internal conductivity within the thorax. The system produces real-time images relating to both cardiac and pulmonary function. As a first step towards diagnosis using these images neural nets have been applied to the identification of regions of interest in the EIT images for which some activity with time, such as ventricular ejection, is sought. This paper addresses the use of a back-projection network to identify characteristic regions within the images. The network facilitates the production of automated real-time activity plots by defining their effective extent in the images of specific organs. The application is novel within the medical imaging field as the aim is to use neural networks for real-time image analysis.