3112
Towards a harmonized protocol for structural MRI imaging of the brain for multisite studies in the Italian IRCCS advanced neuroimaging network1IRCCS Fondazione Stella Maris, Pisa, Italy, 2INFN, Pisa Division, Pisa, Italy, 3Fondazione IRCCS Istituto neurologico “Carlo Besta”, Milan, Italy, 4Radiology Department, IRCCS Fondazione Mondino, Pavia, Italy, 5Brain MRI 3T Research Center, IRCCS Fondazione Mondino, Pavia, Italy, 6Department of Brain and Behavioural Sciences, University of Pavia, Pavia, Italy, 7Queen Square MS Centre, Department of Neuroinflammation, UCL Queen Square Institute of Neurology, Faculty of Brain Sciences, University College London, London, United Kingdom, 8The Italian IRCCS advanced neuroimaging network, Milan, Italy
Synopsis
MRI derived brain structural measurements from multicenter datasets can strongly be affected by factors such as the acquisition protocol, the static magnetic field strength and the scanner manufacturer. A preliminary study was performed to assess the homogeneity of population metrics from 3DT1 scans acquired with already established routine protocols in a dataset of 174 healthy subjects from 18 Italian Research Hospital Centers (IRCCS). The impact of each center acquisition parameters on outcomes was assessed with quality control measurements and FreeSurfer volumetric metrics of cortical and subcortical structures. Future multicenter studies will benefit from harmonizing the acquisition protocols.
INTRODUCTION
In recent years novel MRI biomarkers for the diagnosis and the prognosis of neurodegenerative and neurodevelopmental diseases emerged1–6. These biomarkers include neuroimaging derived measures such as subcortical structure volumes, cortical areas and thicknesses from 3DT1 high resolution scans. However, given the non-quantitative intrinsic nature of structural MRI imaging, the calculation of brain structural metrics from large multisite datasets can strongly be affected by many factors such as the acquisition protocol, the static magnetic field and scanner manufacturer, which can cause increased variability of the measured metrics7–9. In this study, we explored the variability of routine brain scan acquisitions among Italian neuroimaging centers (IRCCS) on healthy volunteers and we assessed the impact that different parameters can have on both image quality and values of derived metrics.METHODS
We pooled 174 3DT1-weighted MRI scans acquired on healthy volunteers (age 295 years) recruited in 18 IRCCS. The acquisition parameters were directly extracted from the DICOM header files. For each scan, the following quality control measurements were computed according to the Preprocessed Connectome Project (PCP) protocol (http://preprocessed-connectomes-project.org):
-Signal-to-Noise Ratio (SNR): The mean intensity within gray matter divided by the standard deviation of the values outside the brain.
-Contrast-to-Noise Ratio (CNR): The mean of the gray matter intensity values minus the mean of the white matter intensity values divided by the standard deviation of the values outside the brain.
-Entropy-Focus-Criterion (EFC): The Shannon entropy of voxel intensities proportional to the maximum possible entropy, indicative of ghosting and head motion-induced blurring.
The scans were also processed through the FreeSurfer recon-all workflow10 (version FS 6.0) and the volumes of the subcortical structures, the area and thickness of cortical structures were extracted. On these metrics a linear SVM classifier was trained in a 5-fold validation scheme to assess how much the derived measures were actually independent from the acquisition site, the static magnetic field strength and the scanner vendor. To test whether an analogous classifier was able to perform a well-known task, an SVM classifier was also trained on the FreeSurfer metrics to distinguish male vs female subjects.
RESULTS
The acquisition parameters of the different centers are the following:
-Static Magnetic Field: 1.5T/3T
-Scanning sequence: Fast Gradient
Vendor 1: TR within the range [7, 11.5] ms TE within the range [2.5, 5.5] ms
Vendor 2: TR within the range [1900, 2400] ms TE within the range [2.75, 3.4] ms
Vendor 3: TR within the range [6.9, 12.6] ms TE within the range [2.9, 12] ms
-FA within the range [8, 15]°
-Receiving Coil [8, 64] channels
An example of quality control measurement variability is reported in figure 1 (SNR).
An SVM classifier was able to classify the FreeSurfer derived metrics as belonging to a scan coming from a specific site with an accuracy of 55.1% (chance level = 5.6% for 18-class classification). A similar classifier was capable to identify from the same features the static magnetic field used with an accuracy of 85.6% (chance level = 50%). The 3 different vendors were also correctly classified with an accuracy of 75.4% (chance level = 33%). Very similar performances were obtained in the well-known task of male vs. female discrimination starting from FS metrics (accuracy = 76.6%). We considered as a benchmark for the overall variability of the data the hippocampal volume, which we found to be 4240 $$$\pm$$$ 340 (8%) mm3 in our cohort of young volunteers.
DISCUSSION
The collected scans allowed us to perform a survey on the routine protocols used to acquire 3DT1-weighted MRI brain scans in 18 IRCCS. Quality controls measurements confirmed that acquisition protocol, the static magnetic field strength and vendor have a strong impact on image quality and image characterization. Even high-level measures, such as FreeSurfer derived metrics, are strongly dependent on site, magnetic field and vendor. Nevertheless, a significant clinically-meaningful task performed with a standard classifier was successful in distinguishing with a good accuracy male and female subjects; This result is encouraging as it suggests that by treating different acquisition features as covariate can help mitigating variability in a big-data multisite approach. Hippocampal volume is measured with an 8% standard deviationCONCLUSION
This study provides an informative analysis of routine not-harmonized 3DT1 scans from the Italian IRCCS network in terms of structural metrics variability. MRI brain imaging harmonization of the acquisition protocols could improve results, in terms of metrics variability at population level. Future studies that aggregate data coming from the centers spread across the country will benefit from an increased statistical power derived from the harmonization of the protocols.Acknowledgements
The Italian IRCCS advanced neuroimaging network is constituted by the following centers: IRCCS Istituto Auxologico Italiano (Milan); IRCCS Ospedale pediatrico Bambino Gesù (Rome); Fondazione IRCCS Istituto neurologico “Carlo Besta” (Milan); IRCCS Centro Neurolesi “Bonino Pulejo” (Messina); Centro IRCCS “Santa Maria nascente” - Don Gnocchi (Milan); IRCCS Centro San Giovanni di Dio – Fatebenefratelli (Brescia); IRCCS Ospedale pediatrico “Giannina Gaslini” (Genoa); IRCCS Istituto Clinico Humanitas (Milan); Istituto di Ricerche Farmacologiche “Mario Negri” IRCCS (Milan); Istituti Clinici Scientifici Maugeri, IRCCS (Pavia); IRCCS Eugenio Medea (Bosisio Parini); Fondazione IRCCS Istituto Neurologico “Casimiro Mondino” (Pavia); IRCCS NEUROMED – Istituto Neurologico Mediterraneo (Pozzilli); IRCCS Associazione Oasi Maria SS Onlus – Troina (Enna); Fondazione IRCCS Ca’ Granda Ospedale Maggiore Policlinico (Milan); IRCCS Fondazione Ospedale San Camillo (Venice); IRCCS Ospedale San Raffaele (Milan); IRCCS Fondazione Santa Lucia (Rome); IRCCS Istituto di Scienze Neurologiche (Bologna); IRCCS SDN Istituto di ricerca diagnostica e nucleare (Naples); IRCCS Fondazione Stella Maris (Pisa)References
1. Jack, C. R. Alliance for Aging Research AD Biomarkers Work Group: structural MRI. Neurobiol. Aging 32, S48–S57 (2011).
2. Frisoni, G. B., Fox, N. C., Jack, C. R., Scheltens, P. & Thompson, P. M. The clinical use of structural MRI in Alzheimer disease. Nat. Rev. Neurol. 6, 67–77 (2010).
3. Dickerson, B. C. et al. Alzheimer-signature MRI biomarker predicts AD dementia in cognitively normal adults. Neurology 76, 1395–402 (2011).
4. Ecker, C., Bookheimer, S. Y. & Murphy, D. G. M. Neuroimaging in autism spectrum disorder: brain structure and function across the lifespan. Lancet Neurol. 14, 1121–1134 (2015).
5. Thayyil, S. et al. Cerebral Magnetic Resonance Biomarkers in Neonatal Encephalopathy: A Meta-analysis. Pediatrics 125, e382–e395 (2010).
6. Woodward, L. J., Anderson, P. J., Austin, N. C., Howard, K. & Inder, T. E. Neonatal MRI to Predict Neurodevelopmental Outcomes in Preterm Infants. N. Engl. J. Med. 355, 685–694 (2006).
7. Martinez-Murcia, F. J. et al. On the brain structure heterogeneity of autism: Parsing out acquisition site effects with significance-weighted principal component analysis. Hum. Brain Mapp. 38, 1208–1223 (2017).
8. Jovicich, J. et al. Reliability in multi-site structural MRI studies: Effects of gradient non-linearity correction on phantom and human data. Neuroimage 30, 436–443 (2006).
9. Suckling, J. et al. The neuro/PsyGRID calibration experiment. Hum. Brain Mapp. 33, 373–386 (2012).
10. Fischl, B. et al. Whole Brain Segmentation: Neurotechnique Automated Labeling of NeuroanatomicalStructures in the Human Brain. Neuron 33, 341–355 (2002).
Figures
