Twelve neurological diagnostic classes of interest were identified: dementia (Alzheimer's disease, frontotemporal dementia, vascular dementia), epilepsy, demyelinating diseases, motor neuron diseases, spinal cord pathologies, Parkinson's disease and parkinsonism, brain tumors, disorders of consciousness, rare diseases, pediatric diseases, headaches, and cerebrovascular diseases.

For each of these neurological diagnostic classes, neuroradiologists have established guidelines for the acquisition of a minimum standard clinical MRI protocol that includes the sequences required for radiological evaluation, as well as relevant geometry imaging parameters (i.e., slice orientation, phase encoding, the range for the field of view, voxel size, slice thickness). These guidelines were shared with the Italian Association of Neuroradiology for a wider application and diffusion to various centers at the national level. Neuroradiologists and researchers delineated MRI sequences for an advanced qMRI protocol, which includes techniques for both clinical, structural, and functional assessment of the brain according to the state of the art in the neuroimaging field. The working group identified its case study in dementia so that it can initiate the translation of all harmonization efforts into large-scale multimodal data collection for patients and animal models (Figure 3).

The 3T scanner survey showed that within the Network there was a prevalence of Philips vendor (8 scanners), compared to General Electric (GE) (3 scanners) and Siemens (7 scanners) vendors. Moreover, heterogeneity in scanner configurations was found, particularly in the gradient system and the transmitter/receiver head coils. A head coil with 32 or fewer channels was available in 15 sites, while the latest technology with a 64 channels head coil was available in 3 centers. The size of the head coils affected the choice of phantoms for the QCs. QC phantom scans were conducted approximately monthly for 2 to 4 years. Through QC, we assessed the reproducibility over time among scanners, using published reference performance indices even when significant hardware and software upgrades were performed intra-site in response to the progress of technology.

For the evaluation of geometry and contrast parameters in QC, the American College of Radiology (ACR) phantom was identified and purchased in the large and/or small versions (37): the ACR large phantom fits head coils with 32 channels or less, while the ACR small phantom can fit head coils with 64 channels. For both phantoms, the MRI protocol, including T1- and T2-weighted sequences, was set up starting from the ACR recommendations (37). Additionally, specifications for the in-plane resolution, post-processing filters, and receiver bandwidth were introduced because these parameters are known to affect distortions of the images (38). Thus, an SOP for QC with ACR phantom (Figure 3) was established including care of the phantom (e.g., refill), its placement within the head coil, implementation of the MRI protocol, and image acquisition. To implement an automatic QC pipeline, Matlab script available at http://jidisun.wixsite.com/osaqa-project/resources/ was improved (39). This script includes the evaluation of the following quantitative parameters defined in ACR recommendation: geometry distortion, slice thickness accuracy, intensity uniformity, ghosting artifact, and high contrast spatial resolution. A further measure considering the geometric accuracy along the feet-to-head axis was introduced: the ratio between the elliptical areas of the first and the third/fifth slices was still obtained for small/large ACR. Each value derived from QC MRI images was compared with tolerance ranges defined by standard ACR recommendations (37) to define outlier measures.

For the evaluation of the stability of gradient echo-planar imaging (GE-EPI) sequences over time, the FUNSTAR (https://www.goldstandardphantoms.com/products/funstar/) phantom, compatible with all scanner head coils, was chosen. The MRI protocol was set up to use echo-planar sequences with similar parameters affecting gradients' stability and performance. Similar to the ACR phantom, a SOP for the QC with FUNSTAR phantom (Figure 3), including placement within the head coil, implementation of the MRI protocol, and image acquisition, was written. Data analysis was performed through a script implemented in Python (https://github.com/mri-group-opbg/stabilitycalc). This analysis estimated the number of parameters that are of interest for neuroimaging functional studies. It computed Signal to Noise Ratio (SNR), Signal to Fluctuation Noise Ratio (SFNR), Percentage Signal Change (PSC), Signal drift, temporal SNR (tSNR) (16, 40). It performed even-odd analysis producing an output to visualize structured noise and spike detection to identify anomalous volumes. Moreover, an additional analysis was implemented in our QC assessment on the FUNSTAR phantom. The Weiskoff analysis that we proposed covered not only a single plane as in the original paper (40), but innovatively all three orthogonal planes as well as the 3D volume in order to estimate the radius of DeCorrelation across all planes. This approach enables us to monitor the effect of noise across the slices and estimate poor slice selection and signal leakage. Values derived from FUNSTAR data were extracted from a central ROI as well as from multiple peripherical ROIs and they were compared with tolerance ranges defined as two standard deviations from the mean value among all the first nine acquisitions from each site, after excluding outliers.

SOPs for QC are available on request at the following link (https://zenodo.org/record/6320896).

After a period of trial and error, an advanced harmonized qMRI protocol in a clinical setting was then finalized. The final protocol now includes the following sequences: 3D T1 weighted images (3D T1w), 3D T2 weighted Fluid Attenuated Inversion Recovery (3D T2-FLAIR) images, quantitative susceptibility mapping imaging (QSM), 2-shell diffusion-weighted imaging (DWI), and resting-state functional magnetic resonance imaging (rs-fMRI). 3D T1w and 3D T2-FLAIR images, sagittally oriented, were acquired with the same geometry parameters (e.g., voxel = 1 x 1 x 1 mm³, number of slices = 175–180, depending on scanner). For the QSM data, a multi-echo gradient echo sequence (voxel size = 1 × 1 × 1 mm³) was defined. Given that different sequence implementations were available across vendors, standardizing the echo-train and echo time (TE) values, without using the scanners in research mode, was quite difficult. So, across all sites, we aimed to achieve a uniform average TE, according to literature standards at the time of implementation (41). For the DWI data, a 2-shell standard single-shot echo-planar imaging sequence (EPI) sequence (voxel size = 2.5 × 2.5 × 2.5 mm³, two shells with 30/32-depending on the scanner- isotropically distributed diffusion-weighted directions, diffusion weightings of 1,000 and 2,000 s/mm², 4–7 non-diffusion weighted b = 0 s/mm² images equally distributed among diffusion-weighted images) was implemented. In addition, 3 non-diffusion weighted images with the reversed phase-encoding acquisition were acquired for distortion correction. For the rs-fMRI data, a GE-EPI sequence (voxel size = 3 × 3 × 3 mm³, repetition time = 2,400 ms, echo time = 30 ms) was set. GE-EPI inverted blip acquisition was acquired as well. No advanced acceleration parameters were used in any of the protocols, as hardly any of the IRCCS scanners were equipped with options such as simultaneous excitation or compressed sensing as part of their software versions.

The total acquisition time for the qMRI protocol was approximately 40 min.

Thereafter, three test sites were identified: site CB for Philips, site DG for Siemens, site MA for GE. For each sequence of interest, the technical working group defined fundamental geometry and contrast parameters so that each scanner in the Network could implement them. The sequences were implemented and optimized in these 3 test sites. The guidelines on the correct implementation and acquisition of the harmonized qMRI protocol were summarized in an SOP document.

To verify the repeatability and reproducibility of qMRI measures extracted from each sequence, the protocol was acquired in 4 “traveling brains” at each test site. Not only the presence of motion artifacts in both structural and functional sequences but also excessive distortions along the antero-posterior phase encoding direction in DWI and rs-fMRI sequences were evaluated, together with contrast to noise ratio between gray and white matter, signal to noise ratio in different brain tissues and the coefficient of joint variation for the assessment of intensity non-uniformity. After this initial quality check, specific analysis pipelines were implemented based on Freesurfer (https://surfer.nmr.mgh.harvard.edu/) for cortical and subcortical thickness/volume assessment, on STI suite (https://people.eecs.berkeley.edu/~chunlei.liu/software.html) for QSM, on MRTrix/FSL (https://www.mrtrix.org/, https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FSL) for DWI, and CONN (https://web.conn-toolbox.org/) for rs-fMRI. Several qMRI measures were obtained: cortical and/or subcortical volumes and thickness in regions of interest (ROIs) and corpus callosum shape for 3D T1w; quantitative susceptibility values in subcortical ROIs for QSM; fractional anisotropy, mean diffusivity, mean kurtosis in white matter and gray matter masks for DWI; tSNR in white matter, gray matter, and cerebrospinal fluid masks for rs-fMRI. Similarity within each qMRI measure extracted from the “traveling brains” data was assessed by evaluating intra- and inter-scanner coefficients of variation. In particular, the extracted qMRI measures were considered robust if the coefficients of variation were <10% and no outliers were identified. Therefore, a tolerance range for each qMRI measure was defined as two standard deviations from the mean value calculated across all “traveling brains” data.

After the protocol optimization step in the three test sites, the remaining sites implemented the qMRI protocol following the SOP and/or through direct import of the file/s generated by the test scanners (e.g., examcard, DICOM files). To verify the protocol implementation at each site, the technical working group examined both the quality of the images acquired on a control subject-as outlined for the “traveling brain” -as well as the parameters set through checking automatically specific DICOM tags. Once the protocol was approved, each site proceeded with the acquisition of data using the qMRI protocol in 5 healthy participants (mean age: 29.7 ± 5.0; 32 male/ 45 female). To assess the reliability and robustness of the harmonized protocol across sites, the previously developed analysis pipelines were applied and the qMRI measurements obtained from each subject and each imaging modality were compared with the tolerance range derived from “traveling brains.”

WP3 of the RIN-Neuroimaging Network defined the first harmonized MRI study-both at the setup and analysis pipeline levels-carried out on a mouse model. The WP3 produced a SOP for advanced harmonized qMRI protocol in the preclinical setting. The survey showed that the 7T preclinical scanners were only from Bruker. Thus, all the QC (SNR, ghosting, intensity stability) were conducted using Bruker standard phantoms. The only exception was for the QC of diffusion measurements for which a homemade phantom filled with n-Tridecane was used to test the stability and reproducibility of diffusion maps (35). The harmonized preclinical protocol includes the following sequences: 3D T2-weighted images (3D T2w), DWI, Flow-sensitive Alternating Inversion Recovey Arterial Spin Labeling (FAIR-ASL), and Glutamate Chemical Exchange Saturation Transfer (Glu-CEST).

3D T2w provide a high resolution 0.1 x 0.1 x 0.1 mm³ isotropic anatomical scan. In preclinical imaging, it is imperative to pay attention to the placement of the surface head coil concerning the animal's head to avoid subsequent inaccurate brain extraction and segmentation. The diffusion protocol was set up as a single shell EPI sequence (voxel size = 0.115 × 0.115 × 0.66 mm³, with 30 isotropically distributed diffusion-weighted directions, diffusion weightings of 900 s/mm², 5 non-diffusion weighted b = 0 s/mm² images). Very precise positioning of the slice package is requested to prevent any misalignment along the head-tail axis that might introduce strong partial volume effects. FAIR-ASL is characterized by two T1 inversion recovery maps under both global and selective inversion regimes (voxel size = 0.156 × 0.208 × 1 mm³, T1 map with 15 inversion times). Measures of endogenous Glutamate levels were performed using a Glu-CEST acquisition with voxel size = 0.2 × 0.2 × 1 mm³, with the Z-spectrum generated with 22 off-resonance saturation pulses with lengths = 1 s and intensity = 5 mT.

The harmonized protocol and analysis pipelines allow obtaining the following qMRI measures: the whole brain volumes for 3D T2w, the white matter integrity through the fractional anisotropy, radial diffusion value, and axial diffusion values for DWI, the cerebral blood flow for FAIR-ASL, and the endogenous glutamate levels for Glu-CEST.

According to the outcome of discussions within WP1 regarding the first pathology to tackle, it was decided to study longitudinally (from 4 to 19 months) the transgenic (TG) mouse model APP/PS1. This model is a well-established model of Alzheimer's disease (42, 43).

RIN–Neuroimaging Network identified three different web-based platforms developed under previous European and National calls, such as NeuGRID (44), ARIANNAFORIRCCS (45), and Medical Informatic Platform (MIP) (46) to provide complementary tools.

This project echoes the FAIR guiding principles of open science developed by the contemporary neuroscience community researchers for enhancing the reusability of research data (36, 47).

NeuGRID (https://neugrid2.eu) is a High-Performance Computing e-infrastructure aiming to collect a large amount of image data paired with computationally intensive data analyses. NeuGRID has been identified by the Re3data initiative as an official research data repository compatible with the FAIR principle. NeuGRID provides the following services to the RIN–Neuroimaging Network: (i) persistent storage and monitoring of harmonized acquisitions of phantoms and murine models with the settings of each scanner checked over time; (ii) imaging analysis using publicly available analysis software and artificial intelligence tools (46). Once data transfer from IRCCS hospitals to the central platform is complete, a QC procedure is automatically triggered for the recently uploaded data (i.e., ACR and FUNSTAR phantoms). QC results are exported to a spreadsheet and PNG snapshots, which are archived along with DICOM images for long-term performance reporting and automatically sent to the user as a final report.

MIP (https://ebrains.eu/service/medical-informatics-platform/), which is cross-linked with NeuGRID, is part of a large-scale European initiative, the Human Brain Project and aims at the integration of multiparametric neuroscientific data using advanced machine learning and deep learning algorithms for the investigation of the human brain in pathologies in the future development of this Network (46).

ARIANNAFORIRCCS platform (https://ariannaforirccs.pi.infn.it/), stemmed from the Arianna project (45) is a web-based research environment for the evaluation of advanced harmonized qMRI protocols on healthy participants. ARIANNAFORIRCCS provides the following services: (i) upload of demographics and brain imaging data; (ii) imaging analysis using public available analysis software.

Each of these platforms responds to a specific need of the RIN–Neuroimaging Network and the FAIR principles (36, 47). According to them, data are characterized by the following standards: