Multitaper Spatial Frequency Domain Signal Noise Reduction in Magnetic Resonance Imaging

Medical data, such as MRI scan information is not being recorded directly as images, but is firstly stored in k-space, which corresponds to the MRI signal’s spatial frequency spectrum. The high-frequency part of this signal is exposed to the noise of biological and nonbiological nature, possibly complicating the interpretation of the medical scan. To mitigate this, noise reduction is traditionally performed by applying different types of filters and windowing functions to the fully processed MRI scans such as T2-weighted images, or by transforming them back to the frequency domain. However, when starting with these types of data, that is specified not for denoising but for its readability to a medical professional, there is a risk of mis-reconstructing the original signal leading to the loss of image's fine details. Recently, an approach called multitaper analysis which utilizes multiple orthogonal windows (tapers) to process the noisy image was highlighted in the scientific literature on the topic. Here we further investigate its denoising capabilities using as a starting material not the ready-to-read MRI scans, but the data containing both phase and magnitude information to form a complex image from which k-space can be obtained, closely replicating the original MR signal acquisition. We assume that the usage of multiple tapers in k-space can potentially reduce bias and variance, resulting in less noise in the spectral estimate. To determine how the type and count of tapers can affect the denoising results, the comparison with the single periodogram and the non-local means filter is made. The results, evaluated using peak signal-to-noise ratio, show that the multitaper method outperforms the two others, while feature similarity index measure metric gives identical values for multitaper and standard periodogram methods, while both being lower values compared to non-local means filter.