Marzena Wylezinska-Arridge1, Mark J White1,2, Indran Davagnanam1, M Jorge Cardoso3, Sjoerd B Vos3,4, Sebastien Ourselin3, Olga Ciccarelli5, Tarek Yousry1, and John Thornton1,2
1Neuroradiological Academic Unit, UCL Institute of Neurology, University College London, London, United Kingdom, 2Lysholm Department of Neuroradiology, National Hospital for Neurology and Neurosurgery, London, United Kingdom, 3Translation Imaging Group, Centre for Medical Imaging Computing, University College London, London, United Kingdom, 4MRI Unit, Epilepsy Society, Chalfont, St Peters, United Kingdom, 5Institute of Neurology, University College London, London, United Kingdom
Synopsis
The huge
number of hospital MRI examinations routinely obtained for clinical purposes
offers a potentially valuable “big data” resource for largescale experimental
neurology. However, acquisition-scheme variation may compromise the research
value of clinical imaging data. A first step towards reducing variation by
prospective protocol harmonization is to systematically capture sequence-use
statistics. Using an in-house tool developed to automate capture of long-term,
MRI sequence deployment statistics in routine practice within our
neuroradiological service, we identified “core“, most used sequences and the
deployment frequency of their respective variants, to enable efficient,
targeted protocol harmonization.PURPOSE
The huge
number of hospital MRI examinations routinely obtained for clinical purposes is
a potentially valuable “big data” resource alongside established genetic and
phenotypic data streams for largescale experimental neurology. However, acquisition
variation may compromise the research value of clinical imaging data: radiology
units addressing a complex range of diagnostic questions commonly employ many
indication-specific protocols differing in detail between operators, scanners,
and institutions.
A first step
towards reducing this variation is to systematically capture the sequence
variants and their frequency of use. To this end, we developed an automated
method to capture long-term MRI sequence-variant usage statistics in routine
practice across multiple scanners at our neuroradiology center, aiming to quantify
protocol variation and identify a ‘core’ suite of the most-used sequences to facilitate
acquisition protocol harmonization.
METHODS
Using an in-house developed software tool [1] we parsed the system scanning
logs of four clinical MRI scanners (Siemens Magnetom: 1.5T Avanto, 1.5T Espree, 3T Skyra, and 3T
Trio). Individual scanner records consisted of ~12K measurements over a ~ 4 months period
(Sep 2014 – Jan 2015), and for one system (1.5T Avanto) we analysed data from a
longer period (June2013- Oct 2015) ~60K measurements. Heuristic tests
identified individual pulse-sequence execution occurrences, recording user-defined
MRI sequence names to provide a machine-generated statistical record of pulse-sequence
variant use. The user-defined sequence name strings include the core contrast
(e.g. ‘t2_tse’) with additional characters denoting sequence variants, employed
for specific indications with, generally, small variations in acquisition
parameters. Head and neck only acquisitions were included, identified by cross-referencing
scanner protocol tree paths and RF coil identification, and localiser sequences
were excluded. Final data reduction was performed in Microsoft Excel (V.14.07).The most frequently used sequences were
ascribed to our group of core-contrast clinical sequences. For each of these we
determined the frequency of execution of respective sequence-variants.
RESULTS
Across 4
scanners, a total of ~5060 patient examinations were performed during the ~4
months period.
Among the
most frequently deployed were variants of the following: 2D multi-slice pulse
sequences: T2 weighted (T2w) turbo-spin echo (TSE), T2w FLAIR, T1w FLASH, T1w
spin-echo, and diffusion weighted DWI; and 3D volume sequences T1w MPRAGE and
3D SWI -FLASH.
Figures 1-2
show histograms of sequence-variant execution frequency for representative 1.5T
and 3T scanners. Across our 4 scanners,
the most frequently deployed basic sequences were common, but there were differences
in their relative execution frequencies.
Figures 3-4
show the number of sequence-variants for each of the core-contrast sequences
for 1.5T and 3T scanners respectively. In general there were more sequence-variants
on our 1.5T than 3T systems, with T2w TSE and T1w SE being deployed with the
largest number of variations. For 2D acquisitions, scan orientation choices increased
the number of variants (e.g. on the Avanto 1.5T, for the T1_SE core contrast there were
16 axial, 25 coronal and 6 sagittal variants,
and for T2_TSE
- 36 axial, 11 coronal and 4 sagittal). Among the 3D sequences, there
were more MPRAGE variants than SWI FLASH for both field strengths.
DISCUSSION AND CONCLUSIONS
A
statistical approach to scanner activity analysis revealed the most commonly
used core “menu” of sequences used in clinical brain examination. For all scanners,
groups of the same most frequently used sequences were observed, differences
between their relative frequencies presumably reflecting scanner-dependent
differences in the clinical workflow. However, sequence variations may also
occur for historical or arbitrary reasons: this data collection will permit
inspection of the sequence variants to determine if they truly reflect indication-specific
requirements, or can be standardized to reduce data variation. A bespoke system
analytics tool enables the capture of MRI acquisition statistics not practical
by manual record keeping, and a means to derive ongoing service quality control
data. We hypothesize that harmonization of clinical MRI protocols across
sites/hospitals will increase the value of clinical MRI data as a large-scale
research resource, with decreased acquisition variation helping to improve the
statistical power of future large-scale patient-data studies.
Acknowledgements
We are
grateful to UCL NIHR Biomedical Research Centre for funding.References
[1]
M.J.White, T. Yousry, J.S. Thornton. MRI activity analytics: IPEM Conference
Book of Abstracts, 2014, Vol 2 : 28.