Artigo Internacional: On the Implicit Representation of the Electrical Impedance Tomography Inverse Problem
Durante o período da Bolsa de Iniciação Científica, o orientador Marcos de Sales Guerra Tsuzuki propôs que os alunos orientados por ele participassem do desenvolvimento de um artigo científico que seria publicado no congresso internacional IEEE EMBS 2025, sobre a aplicação de uma técnica de Deep Learning (Implicit Layers) no cenário de Tomografia por Impedância Elétrica.
Artigo científico feito para IEEE EMBS 2025 pelos autores: Jungeui Choi, Lucas Song, Henrique S. Martins, Rodrigo H. Imai, Mateus S. Parducci, Luigi F. Romani, Marcos A. B. Campos, Guilherme C. Duran, Thiago C. Martins e Marcos S. G. Tsuzuki.
Abstract-Electrical Impedance Tomography (EIT) is a non-invasive imaging method that reconstructs internal conductivity maps by applying low-frequency electrical currents through surface electrodes and measuring the resulting voltages. While EIT’s inverse problem is inherently ill-posed, requiring regularization techniques to stabilize solutions, recent advances in neural networks offer promising alternatives for improving reconstruction accuracy. This study investigates the application of Deep Implicit Layers to EIT reconstruction, combining fixed-point iteration and Newton’s method to solve implicit constraints. The forward problem is solved using the finite element method (FEM), and numerical phantoms are employed for training. To ensure a more realistic representation of measurement conditions, contact conductivity was modeled as a random variable sampled from a Gaussian distribution, improving the model’s robustness and ability to handle uncertainties. The implicit neural network effectively distinguishes closely positioned objects; however, contrast limitations persist at the boundaries of the phantoms. The results highlight the potential of implicit models in EIT and suggest future improvements through advanced optimization techniques and expansion to 3D thorax reconstruction.