DeepEIT: Deep Image Prior Enabled Electrical Impedance Tomography.

Area of Science:

• Medical Imaging
• Computational Imaging
• Electrical Impedance Tomography

Background:

• Neural networks (NNs) are prevalent in tomographic imaging but necessitate extensive training data, a limitation in clinical settings.
• Current methods for electrical impedance tomography (EIT) reconstruction face challenges due to data scarcity.

Purpose of the Study:

• To demonstrate a data-free image reconstruction method for EIT using neural networks.
• To introduce a novel unsupervised approach for EIT by integrating deep image prior (DIP) with neural network reconstruction.

Main Methods:

• The proposed method utilizes deep image prior (DIP) to regularize EIT reconstruction problems.
• Image reconstruction is achieved by optimizing conductivity distribution through NN back-propagation and a finite element solver, without requiring prior training data.

Main Results:

• Quantitative evaluations using simulated and experimental data confirm the method's effectiveness.
• The unsupervised DIP-based EIT reconstruction outperforms current state-of-the-art alternatives.

Conclusions:

• The developed neural network-based EIT reconstruction method offers an effective unsupervised solution for scenarios with limited training data.
• This approach advances medical imaging by overcoming data dependency in image reconstruction.