Kolmogorov-Arnold Vision Transformer for Image Reconstruction in Lung Electrical Impedance Tomography

Electrical impedance tomography is a non-invasive and non-ionizing imaging technique, which can provide real-time monitoring of the internal structures and function of the human body, and has been particularly popular in lung monitoring. However, the associated inverse problem is ill-posed, leading to suboptimal image quality with low spatial resolution, which hinders its practical use in the clinical settings. To achieve reliable image reconstruction, this work proposes a novel deep learning approach, applied to lung monitoring. The proposed model is a hybrid of the vision transformer and the recently introduced Kolmogorov Arnold Network (KAN). The fully connected layers in the transformer are replaced with KAN layers, which enhances its ability to learn the complex relationship between the voltage measurements and the conductivity distribution within the lungs. In comparison with the use of convolutional models and Vision Transformer, the proposed method achieves outstanding performance with a mean squared error of 0.0045, structural similarity index of 0.96, relative error of 0.11, and correlation coefficient of 0.98.