NAS Powered Deep Image Prior for Electrical Impedance Tomography
Haoyuan Xia, Qianxue Shan, Junwu Wang, Dong Liu
Abstract
In this paper, we introduce a novel approach that combines neural architecture search (NAS) with the deep image prior (DIP) framework for electrical impedance tomography (EIT) reconstruction. Deep neural networks have proven effective as DIPs in various image reconstruction tasks, but the appropriate prior is task-dependent. Manually designing network architectures for EIT reconstruction is challenging. Our method automates this process by using NAS to identify optimal neural network configurations tailored for EIT reconstruction. This approach eliminates the need for rare labeled data, which is a significant advantage in EIT applications. Extensive validation using both simulated and experimental data showcases the effectiveness of our NAS-powered DIP approach. Comparative evaluations against traditional methods and state-of-the-art techniques consistently demonstrate superior reconstruction results and robustness against noise. Our approach opens up exciting possibilities for advancing EIT reconstruction methods, with potential applications in medical imaging and industrial testing.