Hybrid Learning-Based Cell Aggregate Imaging With Miniature Electrical Impedance Tomography
Zhou Chen, Yunjie Yang, Pierre Bagnaninchi
Abstract
Real-time, nondestructive, and label-free imaging of 3-D cell culture process using miniature Electrical Impedance Tomography (mEIT) is an emerging topic in tissue engineering. Image reconstruction of mEIT for cell culture is challenging due to weak sensing signals and increased sensitivity to sensor imperfection. Conventional regularization-based image reconstruction methods cannot always achieve satisfactory performance in terms of image quality and computational efficiency for this particular setup. Recent advances in deep learning have pointed out a promising alternative. However, with a single neural network, it is still difficult to reconstruct multiple objects with varying conductivity levels; these cases are widespread in the application of cell imaging. Aiming at this challenge, in this article, we propose a deep learning and group sparsity (GS) regularization-based hybrid algorithm for cell imaging with mEIT. A deep neural network is proposed to estimate the structural information in form of binary masks given the limited amount of data sets. Then, the structural information is encoded in the GS regularization to obtain the final estimation of conductivity. The proposed approach is validated by both simulation and experimental data on MCF-7 human breast cancer cell aggregates, which demonstrates its superior performance and generalization ability compared with a number of existing algorithms.