Inadequate Data Augmentation With Convolutional Neural Network in Brain Electrical Impedance Tomography

Electrical impedance tomography (EIT) is a potential technique for imaging intracranial hemorrhage. In general, 16 electrodes are used in the brain EIT. In some situations, it may be not possible to attach so many electrodes. This reduces the number of measurements and causes poor reconstruction quality. In addition, the skull has poor conductivity which obstructs the injection of current and degrades the quality of the reconstructed image. Thus, inadequate boundary measurement caused by a reduced number of electrodes is a great challenge in brain EIT. To address this problem, a novel approach based on a convolutional neural network is proposed for inadequate data augmentation when only eight electrodes can be equipped. The proposed method enables the establishment of a nonlinear relationship between the inadequate voltage data and the desired augmented data. The voltage data augmented for the eight-electrode brain EIT are compared with the measured voltage of the brain EIT equipped with 16 electrodes. It is found that the two groups of data are very similar. The effectiveness of the proposed approach is further confirmed by imaging hemorrhage in the brain tissue layer. Images reconstructed for various cases are compared with the results, which use directly measured voltage. The qualitative evaluation as well as quantitative analysis demonstrates that the proposed approach greatly enhances the reconstruction quality in the case of inadequate measurement. In summary, the voltage data directly measured from the eight-electrode brain EIT can be well-augmented with the proposed method.