Resting-state networks (RSNs) and functional connectivity has been increased exploited to investigate synchronous activity and identify abnormalities in different pathologies, including epilepsy¹. Most of the studies have been based on blood-oxygenation-level-dependent (BOLD) contrast in combination with spatial independent component analysis (ICA) to separate signals into maps of covarying voxels^2,3.Promising results have been recently achieved in RSN analysis using Arterial Spin Labeling (ASL)^4,5, which demonstrated its ability to provide direct and quantitative measures of brain physiology⁶. However, its wide application has been hampered by the relatively low sensitivity and robustness to noise sources, which have to be properly filtered when mapping resting-state fluctuations. In this study, we aim to: 1) assess the ability of ASL combined with two preprocessing pipelines in detecting RSNs derived from ICA compared to literature networks; 2) quantify spatial RSNs in healthy controls (HC) and epilepsy patients; 3) evaluate the temporal properties of the spatial independent components (IC) for deriving measures of brain connectivity and identifying significant pathological alterations.

Experimental protocol:Seven healthy volunteers (32±6years) and seven patients (31±8years) with drug-resistant temporal epilepsy were scanned on a 3T PET/MR scanner (Biograph mMR) using 2D-EPI pulsed PICORE Q2TIPS ASL⁷(3.6x3.6x5mm3;19 slices;TI1/TIs/TI2:800/1200/1800ms;TR/TE:2860/17ms). Four resting ASL runs of 100 volumes each were acquired 5min apart from one another.

HC data analysis: data preprocessing on HC was carried out with FSL tools, in particular two preprocessing pipelines were applied for each subject before spatial component identification:1)Basic pipeline³: each ASL run underwent head motion correction, non-brain removal, spatial smoothing with a 5-mm Gaussian kernel and high-pass filtering (100s);2)Advanced pipeline: the basic pipeline is followed by single-session spatial ICA using MELODIC⁸ separately applied to each dataset. The main artefactual components were visually identified and removed by linear regression, obtaining cleaned data. For each HC, the four preprocessed ASL runs (basic and advanced pipelines) were transformed in the 2-mm MNI space and underwent network decomposition by means of group-ICA with multi-session temporal concatenation. ICs were converted to z-statistic maps and threshold at z=3³. ICs representing the main RSNs were identified according to the spatial distribution proposed in SOA and compared to the RSN template from BOLD with 10 ICs³ by quantifying the degree of spatial similarity/overlap (Dice’s Coefficient [DSC]) and spatial cross-correlation for each network (r-value)⁹.

Patients’ data analysis: for each patient, only the advanced pipeline has been applied and cleaned datasets underwent the same group-ICA procedure as for HC.

Functional connectivity analysis: the mean time course of each RSN was extracted in all subjects and Pearson’s correlation between all networks’ pairs was calculated deriving matrices of brain connectivity. Connectivity differences between controls and patients were tested by Wilcoxon rank-sum test (p<0.05).

For HC, ICA decompositions on basic and advanced preprocessed data revealed seven common RSNs: medial visual (VIS), default mode network (DMN), cerebellum (CER), sensorimotor (SM), auditory (AUD), executive (EXE) and joint right-left frontoparietal (FP). In terms of DSC, the results consistently showed across subjects a good level of overlap with reference networks for VIS, DMN, CER, AUD and FP which further increased with the application of the advanced pipeline. The same networks presented an overall r-value>0.25, the minimum value for classifying a good component from BOLD data³ (Fig.1).DMN maps are reported for a representative subject in Fig.2, where the good spatial similarities with the BOLD template can be appreciated by using the advanced pipeline (DSC=0.55;r-value=0.63). Since subjects’ results consistently showed the improvement introduced by the advanced pipeline, this was used for patient’ data, revealing the same seven RSNs were present in this group except for CER which was not detected in two patients (Fig.3).The mean matrices of correlations computed between the seven RSN time-courses of both HC and patients are reported in Fig.4. The statistical analysis revealed significant changes in connectivity between DMN and CER, with a decreased pattern in patients vs HC (z-value=1.936,p=0.04), and between VIS and FP, with an increased correlation between these areas in patients (z-value=-2.025,p=0.038).In this study, we first demonstrated ASL is able to detect not only DMN but all the different networks, with good level of overlap and correlation with the corresponding reference BOLD ICs. Moreover, ICA on cleaned data allows detecting more precisely RSNs, increasing their similarities. Second, the spatial patterns are consistent with commonly reported networks, both for HC and patients. Finally, functional connectivity analyses derived from ICA results allow to map pathological changes, in particular the decreased patterns involving DMN are in line with previous studies on temporal epilepsy and BOLD, showing reduced connectivity between DMN and different areas as cerebellum, temporal and frontal lobes¹⁰.