Enhanced Image Reconstruction in Electrical Impedance Tomography using Radial Basis Function Neural Networks

Enhanced Image Reconstruction in Electrical Impedance Tomography using Radial Basis Function Neural Networks

Článek WoS

This paper presents a novel cascade algorithm for image reconstruction in electrical impedance tomography (EIT) using radial basis function neural networks. The first subnetwork applies a density-based algorithm and k-nearest neighbors (KNN) to determine the center and width of the radial basis function neural networks, with the aim of preventing ill-conditioned connection weights between the hidden and output layers. The second subnetwork is a generalized regression neural network dedicated to functional approximation. The combined subnetworks result in a reduced mean square error and achieve an accuracy of 89.54 % without noise and an accuracy between 82.90 % and 89.53 % with noise levels ranging from 30 to 60 dB. In comparison, the original radial basis function neural networks (RBFNN) method achieves an accuracy of 85.44 % without noise and between 80.90 % and 85.31 % under similar noise conditions. The total variation (TV) method achieves 83.13 % without noise, with noise-influenced accuracy ranging from 34.28 % to 45.15 %. The Gauss-Newton method achieves 82.35 % accuracy without noise, with accuracy ranging from 33.21 % to 46.15 % in the presence of noise. The proposed method proves to be resilient to various types of noise, including white Gaussian noise, impulsive noise, and contact noise, and consistently delivers superior performance. It also outperforms the other methods in noise-free conditions. The reliability of the method in noisy environments supports its potential application in the development of new modular systems for electrical impedance tomography.

This paper presents a novel cascade algorithm for image reconstruction in electrical impedance tomography (EIT) using radial basis function neural networks. The first subnetwork applies a density-based algorithm and k-nearest neighbors (KNN) to determine the center and width of the radial basis function neural networks, with the aim of preventing ill-conditioned connection weights between the hidden and output layers. The second subnetwork is a generalized regression neural network dedicated to functional approximation. The combined subnetworks result in a reduced mean square error and achieve an accuracy of 89.54 % without noise and an accuracy between 82.90 % and 89.53 % with noise levels ranging from 30 to 60 dB. In comparison, the original radial basis function neural networks (RBFNN) method achieves an accuracy of 85.44 % without noise and between 80.90 % and 85.31 % under similar noise conditions. The total variation (TV) method achieves 83.13 % without noise, with noise-influenced accuracy ranging from 34.28 % to 45.15 %. The Gauss-Newton method achieves 82.35 % accuracy without noise, with accuracy ranging from 33.21 % to 46.15 % in the presence of noise. The proposed method proves to be resilient to various types of noise, including white Gaussian noise, impulsive noise, and contact noise, and consistently delivers superior performance. It also outperforms the other methods in noise-free conditions. The reliability of the method in noisy environments supports its potential application in the development of new modular systems for electrical impedance tomography.

Artificial neural network, density-based algorithm, electrical impedance tomography, k-nearest neighbors, radial basis function neural networks, EIDORS framework

Artificial neural network, density-based algorithm, electrical impedance tomography, k-nearest neighbors, radial basis function neural networks, EIDORS framework

MIKULKA, J.; KOUAKOUO NOMVUSSI, S.

2025

16.12.2024

Sciendo

1335-8871

Measurement Science Review

24

6

Slovenská republika

200

210

11

https://sciendo.com/article/10.2478/msr-2024-0027

http://hdl.handle.net/11012/251601

Enhanced-Image-Reconstruction-in-Electrical-Impedance-Tomography-using-Radial-Basis-Function-Neural-Networks