Application of Artificial Neural Networks

Impedance tomography (EIT) is a modern imaging technique used to visualize the internal distribution of electrical conductivity within an object based on voltage measurements recorded on its surface. As an inverse problem, EIT requires advanced computational methods for effective image reconstruction. In the latest study published in Przegląd Elektrotechniczny, a team of researchers conducted a comprehensive comparison of three selected machine learning methods: Elastic Net, Least Angle Regression (LARS), and Artificial Neural Networks (ANN), analyzing their effectiveness in the context of EIT image reconstruction.

The results clearly indicate that artificial neural networks are the most effective approach among the tested techniques. In the scenario with a single inclusion, the neural network achieved the lowest RMSE value (0.03016), significantly outperforming both Elastic Net and LARS, which maintained RMSE values around 0.078. This means that ANN allows for much more accurate reconstruction of the electrical conductivity distribution compared to traditional linear regression algorithms.

*Upper panel – image reconstruction using an artificial neural network (3750 training cases). Lower panel – original image with a single inclusion.*

An important aspect of the study was the analysis of the impact of the number of training data on the quality of reconstruction. Elastic Net and LARS stabilized their results after approximately 750 samples, while the neural network showed continued improvement in accuracy as the number of training examples increased, reaching a plateau only after about 2250 cases. This observation has significant practical implications – it allows for optimizing the training time of models, avoiding excessive computational load when there are no noticeable benefits.

The research team also examined the impact of the number of inclusions on reconstruction effectiveness. It turned out that, although an increase in the number of inclusions makes the task more difficult for the models, the artificial neural network still maintains a clear advantage. For a single inclusion, the ANN achieved a 58% reduction in error compared to models trained on smaller datasets. For two and three inclusions, the improvements were 34% and 25%, respectively, confirming the great flexibility and adaptive potential of neural networks in more complex cases.

The results of these studies not only confirm the high effectiveness of artificial neural networks in EIT image reconstruction but also point to further directions for development. One of the key steps will be the development of methods for automatically detecting the point at which the plateau is reached during training, which will further optimize learning time. At the same time, deeper neural network architectures will be tested, which may prove more efficient in analyzing complex, multi-inclusion conductivity structures.

The full version of the article is available in Przegląd Elektrotechniczny, issue 4/2024:
link to publication