Sinusoidal Representation of the Electrical Impedance Tomography Inverse Problem

This work explores the use of Sinusoidal Multi-Layer Perceptrons (SMLPs) for image reconstruction in Electrical Impedance Tomography (EIT). Leveraging periodic activation functions and a principled weight initialization, SMLPs demonstrate a strong capacity to recover fine spatial details and maintain robustness to measurement noise. A hierarchical hyperparameter optimization strategy is proposed to efficiently identify performant architectures with reduced computational cost. Compared to a baseline convolutional neural network, the optimized SMLP achieved lower reconstruction errors across multiple datasets, while using fewer than one-third of the trainable parameters. These results highlight the potential of sinusoidal activations for solving high-dimensional inverse problems. Future work includes validation on experimental data and integration with hybrid or physics-informed models.