Dottorato in MORFOGENESI E INGEGNERIA TISSUTALE

Titolo della tesi: Comparison of denoising techniques in ultra-high field fMRI data and their effect on different brain tissues

BACKGROUND. One of the main challenges in functional Magnetic Resonance Imaging (fMRI) is reducing data variability caused by thermal and physiological noise, in order to isolate the signal resulting from neuronal activity. Ultra-high magnetic field fMRI (7T and above) allows for high spatial resolution, making it feasible to localize the neuronal activity at the level of brain laminae. However, higher resolution implicates higher levels of thermal noise, and high magnetic fields lead to greater physiological noise (due to cardiac and respiratory-induced brain motility) because of the stronger susceptibility effects. Nevertheless, there is a lack of consensus on the best denoising strategy for ultra-high field fMRI data. GENERAL OBJECTIVES. This thesis explores the applicability and compares the efficacy of different denoising methods on 7T fMRI data of human brain, focusing on gray matter (GM), white matter (WM), cerebrospinal fluid (CSF) and the GM layers in the hand-knob region of the primary motor cortex. RESULTS. The results of this thesis indicate that the thermal denoising algorithm NORDIC had the strongest effect among all the steps in the evaluated preprocessing pipelines, and it also enhanced the effectiveness of the subsequent physiological denoising algorithms (performed either with RETROICOR or aCompCor). Pipeline including the aCompCor algorithm showed a greater impact on quality metrics (tSNR, DVARS, laminar CoV, laminar FA) compared to those using RETROICOR. The reproducibility of noise spatial distribution was similar across physiological denoising algorithms. Regarding the noise maps extracted with RETROICOR, cardiac noise had higher spatial reproducibility than respiratory noise. In the pipeline including aCompCor, GM was more affected by noise extracted from the cerebrospinal fluid (roiCSF noise) than by noise extracted from white matter (roiWM noise), with roiCSF noise showing higher reproducibility. Finally, only aCompCor was able to reduce the spurious differences in functional connectivity (FC) between upper and lower laminae of the primary motor cortex (M1) with the premotor cortex, while preserving the FC differences between the upper and lower laminae of M1 with the somatosensory cortex. CONCLUSIONS. The findings of this thesis confirmed that thermal noise in submillimeter 7T fMRI data can be greatly reduced using the NORDIC algorithm. The aCompCor algorithm was found to be more efficient than RETROICOR in reducing physiological noise. CSF was identified as the brain tissue most impacted by physiological noise, with high reproducibility, making CSF the recommended choice for the optimization of denoising algorithms based on signals extracted from a region of interest. Finally, the connectivity analysis suggested that a laminar functional connectivity pattern could emerge after accounting for physiological noise, which is often disregarded in laminar fMRI.