Sinusoidal Representation of the Electrical Impedance Tomography Inverse Problem

Area of Science:

• Medical Imaging
• Computational Electromagnetics
• Machine Learning

Background:

• Electrical Impedance Tomography (EIT) is a non-invasive imaging technique.
• EIT faces challenges in image reconstruction accuracy and noise sensitivity.
• Deep learning models show promise for improving EIT reconstruction.

Purpose of the Study:

• To investigate the efficacy of Sinusoidal Multi-Layer Perceptrons (SMLPs) for EIT image reconstruction.
• To develop an efficient hyperparameter optimization strategy for SMLPs in EIT.
• To compare SMLP performance against baseline convolutional neural networks (CNNs).

Main Methods:

• Utilized SMLPs with periodic activation functions and specific weight initialization.
• Implemented a hierarchical hyperparameter optimization approach.
• Evaluated reconstructions on multiple datasets against a CNN baseline.

Main Results:

• Optimized SMLPs achieved lower reconstruction errors compared to the CNN baseline.
• SMLPs demonstrated superior recovery of fine spatial details and noise robustness.
• The optimized SMLP model required significantly fewer trainable parameters.

Conclusions:

• Sinusoidal activations show significant potential for solving complex inverse problems like EIT.
• SMLPs offer a computationally efficient and high-performing alternative for EIT image reconstruction.
• Further research should include experimental data validation and hybrid model integration.