Reconstruction for MRI by autoencoder and periodogram method

In MRI, noise can arise from a patient’s movement during scanning or from the imperfections of the equipment’s sensors. The application of neural networks significantly improves the analysis of noise-impacted medical images. However, these networks typically require a large training dataset. To mitigate this, in this study the effectiveness of an autoencoder-based neural network combined with periodogram reconstruction was evaluated. The periodogram as a non-parametric method was employed in conjunction with the autoencoder for preliminary filtering and noise reduction, when spectral characteristics of the image are unknown. The findings indicate that a hybrid method can be effective in noise-impacted image reconstruction in low data availability situations, as confirmed by SSIM and PSNR values.