A shallow convolutional neural network-based method for enhanced conductivity distribution reconstruction under limited measurement

Purpose In the brain imaging based on electrical impedance tomography, it is sometimes not able to attach 16 electrodes due to space restriction caused by craniotomy. As a result of this, the number of boundary measurements decreases, and spatial resolution of reconstructed conductivity distribution is reduced. The purpose of this study is to enhance reconstruction quality in cases of limited measurement. Design/methodology/approach A new data expansion method based on the shallow convolutional neural network is proposed. An eight-electrode model is built from which fewer boundary measurements can be obtained. To improve the imaging quality, shallow convolutional neural network is constructed which maps limited voltage data of the 8-electrode model to expanded voltage data of a quasi-16-electrode model. The predicted data is compared with the quasi-16-electrode data. Besides, image reconstruction based on L1 regularization method is conducted. Findings The results show that the predicted data generally coincides with the quasi-16-electrode data. It is found that images reconstructed with the data of eight-electrode model are the poorest. Nevertheless, imaging results when the limited data is expanded by the proposed method show large improvement, and there is a minor difference with the images recovered with the quasi-16-electrode data. Also, the impact of noise is studied, which shows that the proposed method is robust to noise. Originality/value To enhance reconstruction quality in the case of limited measurement, a new data expansion method based on the shallow convolutional neural network is proposed. Both simulation work and phantom experiments have demonstrated that high-quality images of cerebral hemorrhage and cerebral ischemia can be obtained when the limited measurement is expanded by the proposed method.