Residual Convolutional Neural Network-Based Stroke Classification With Electrical Impedance Tomography
Yanyan Shi, Zhiwei Tian, Meng Wang, Yuehui Wu, Bin Yang, Feng Fu
Abstract
Electrical impedance tomography (EIT) is an emerging technique for medical imaging. According to reconstructed images, the human disease can be identified. Note that the reconstruction of EIT is an ill-posed inverse problem. To solve this problem, various sensitivity-based methods have been proposed. Nevertheless, a sensitivity matrix is affected by a number of factors. Therefore, classification reflected by images is sometimes not very reliable. In this work, a novel method based on a residual neural network is proposed for stroke classification. It is realized by directly processing measured difference data from EIT. After training of the residual neural network, the proposed method is quantitatively evaluated when noise is exerted, shape deviation occurs, and conductivity varies in three different layers. Also, impacts of contact impedance and small-sized inclusion are studied. Furthermore, phantom experiments are conducted. Compared with fully connected neural networks and shallow convolution neural networks, the results demonstrate that the proposed method has a better performance in classification for various cases. This article offers an effective alternative for stroke classification in EIT.