Вы находитесь на странице: 1из 2

Submitter information:

Erik Röcher, M.Sc. research associate, eroecher@ukaachen.de

Machine Learning Approaches for Real-Time Quality Assurance of MR Images


Roecher E1, Moesch L², Thiele F3, Meyer-Baese L4, Eisner P1, Zweerings J1, Sarkheil P1,
Mathiak K1,5
1
Department of Psychiatry, Psychotherapy and Psychosomatics, RWTH Aachen University, Germany
²RWTH Aachen University, Germany
3
Philips IntelliSpace Discovery, Aachen, Germany
4
Georgia Institute of Technology, GA, USA
5
Jülich Aachen Research Alliance (JARA), Translational Brain Medicine, Germany

Introduction
In the context of the ever-increasing importance of structural high-resolution magnetic resonance
imaging (MRI), it is widely acknowledged that quality assurance (QA) is a key aspect for
reliable medical diagnosis and analysis of imaging data [1] [2]. QA should be reproducible,
objective and accessible. However, recent studies indicate a large within- and inter-rater
variability for MRI QA even for trained experts [3] [4]. Furthermore, visual inspection may not
be feasible for very large datasets. Automated approaches for quality control of structural MR
images exist [3]. However, those approaches are mostly limited to noise and inhomogeneity.
Additional image degradations, e.g. motion artifacts and anatomical changes cannot be detected.
Systematic QA has even higher significance in the field of real-time functional imaging
(rtfMRI). rtfMRI-based neurofeedback studies rely on the provision of a feedback signal with
sufficient quality. While online artefact correction is available, a standardized evaluation of the
quality parameters is still lacking (e.g. motion parameters).
In recent years, machine-learning approaches based on deep artificial neural networks (DNN)
demonstrated their potential in advanced pattern recognition and classification tasks in the
medical domain [5]. A DNN trained on specific task is capable to perform this task rapidly,
making DNNs perfect candidates for real-time online image analysis. Consequently, we
investigated the applicability of deep neural networks for automated QA of MR images.

Materials and Methods


We trained DNNs in an initial QA project on structural MR images. The dataset used for training
and validation of our networks originates from the 1000 Brains Study [6], a project investigating
structural and functional variability in the human brain during ageing. The dataset comprises 1184
T1-weighted brain images with isotropic resolution of one millimeter (scanning parameters:
repetition time TR = 2.25 s, echo time TE = 3.03 ms, inversion time TI = 900ms, field of view
FoV = 256x256 mm2, flip angle = 9°). All images were acquired using the same 3-Tesla whole-
body MR scanner. Two trained experts examined all images under supervision of the QA team of
the Psychiatric Imaging Network Germany (PING) and rated severance of several image artefacts
on a three level scale.
For training and validation, we used an equal number of images from each level to prevent biased
learning. Network architecture is a three times repeated set of two convolutional layers followed
by maxpooling and two final fully connected layers followed by a softmax classification layer.

Results and Conclusions

1
Network accuracy is measured by the ratio of correctly classified image volumes to the total
number of volumes in a dataset. Our trained network yielded a mean accuracy of 80.14±1.52 in
five-fold cross-validation for random head movement classification and 95.42±1.37 for anatomical
left-right flips due to coding errors.
In summary, real-time QA of brain MR images using neural networks is feasible for anatomical
images and can be used to detect images with relevant artefacts and unexpected anatomical
anomalies. Computational time for quality parameters of a single MR image is below one second
on standard personal computer, confirming that DNNs can be utilized in real-time applications.
Application of DNN in real-time QA to rtfMRI and other imaging modalities will be further
investigated.

References
[1] Power JD, Barnes KA, Snyder AZ, Schlaggar BL, Petersen SE, Spurious but systematic
correlations in functional connectivity MRI networks arise from subject motion,
Neuroimage. 2012, 59, p. 2142–2154
[2] Rosen, A., Roalf, D. R., Ruparel, K., Blake, J., Seelaus, K., Villa, L. P., Satterthwaite, T.
D, Quantitative assessment of structural image quality, NeuroImage, 2018, 169, p. 407–
418
[3] Gaser C, Dahnke R, CAT-A Computational Anatomy Toolbox for the Analysis of
Structural MRI Data, HBM, 2016, p. 336–348
[4] Dale A.M., Fischl B., Sereno M.I., Cortical surface-based analysis. I. Segmentation and
surface reconstruction, Neuroimage 9, 1999, p.179–194
[5] Jose Bernal, Kaisar Kushibar, Daniel S. Asfaw, Sergi Valverde, Arnau Oliver, Robert
Martí, Xavier Lladó, Deep convolutional neural networks for brain image analysis on
magnetic resonance imaging: a review, CoRR, 2017
[6] Caspers et al., Studying variability in human brain aging in a population-based German
cohort-rationale and design of 1000BRAINS, Frontiers in Aging Neuroscience, 6, 2014,
p. 149

Supported by BMBF (APIC, PING; 01EE1405A-C) and DFG (IRTG 2150).

Вам также может понравиться