Efficient de-noising of high-resolution fMRI using local and sub-band information

High-resolution fMRI, useful for accurate brain mapping, suffers from low functional sensitivity at a reasonable acquisition time. Conventional smoothing techniques although reduce the noise and boost the sensitivity, but degrade the spatial resolution of fMRI.

We propose a novel spatial de-noising technique to increase sensitivity while preserving the boundaries of active regions in the high-resolution fMRI. A modified version of PCA that utilizes adjacent voxels information (LPCA) is first suggested for de-noising. This technique is then further empowered by its application to wavelet sub-bands (WLPCA).

Proposed techniques were assessed on both simulated and experimental data. Identifiablity index was calculated for evaluation of the denoising on the simulated data. Maximum and mean z-scores along with LAE and SSIM were reported on experimental data for two presented techniques as well as Guassian smoothing. WLPCA outperformed other techniques in Identifiablity index, for simulation, and in preserving maximum z-score, for experimental study.

The presented technique was developed to simultaneously suppress the noise and preserve the boundaries of active areas against leakage. For first aim, its achievable mean z-score was compared to conventional Gaussian. For second aim, its maximum z-score was compared to that of no-smoothing. While Gaussian and no-smoothing can work fine with only one measure, WLPCA was able to improve both measures concurrently.

The local PCA based methods, and in particular WLPCA, is an effective noise reduction step that preserves the spatial resolution by preventing activity leakage of high-resolution fMRI data.