TY - GEN
T1 - Deep harmonization of inconsistent mr data for consistent volume segmentation
AU - Dewey, Blake E.
AU - Zhao, Can
AU - Carass, Aaron
AU - Oh, Jiwon
AU - Calabresi, Peter A.
AU - van Zijl, Peter C.M.
AU - Prince, Jerry L.
N1 - Funding Information:
Acknowledgements. This research was supported by NIH grants R01NS082347 and P41EB015909, as well as a grant from the National Multiple Sclerosis Society (RG-1601-07180).
Publisher Copyright:
© 2018, Springer Nature Switzerland AG.
PY - 2018
Y1 - 2018
N2 - Magnetic resonance image analysis is often hampered by inconsistent data due to upgrades or changes to the scanner platform or modification of scanning protocols. These changes can manifest in three main sources of image inconsistency: contrast, resolution, and noise. Modern analysis techniques that use supervised machine learning can be especially susceptible to these inconsistencies, as existing training data may not be valid after an upgrade or protocol change. In previous work, differences in contrast and resolution have been addressed in isolation. We propose a novel method of image intensity harmonization that addresses each of the three sources of inconsistency. We formulate our method around a multi-planar, multi-contrast U-net, where all of the available contrasts are used as input channels in a single modified U-Net to produce all of the output contrasts simultaneously. The multi-contrast nature of the deep network allows for harmonization of contrast as information can be shared between contrasts. In addition, coherent, biological features are highlighted and matched to the target, while noise, which differs between matched inputs and outputs, is not reinforced. This process also normalizes small differences in resolution due to the influence of the high resolution channels. To combat larger differences in resolution, which would not be recovered by the neural network alone, we use self super-resolution (SSR) on all images with thick (>2 mm) slices before harmonization. To generate consistent images, the target images are also processed in a similar manner so that all resulting images have consistent qualities. Our harmonization process eliminates significant volume bias of multiple brain compartments and lesion estimation. In addition, absolute volume difference and Dice similarity of segmentation volumes were significantly improved (p < 0.005). SSR alone did not affect the statistical significance of the difference, even though the absolute volume difference was reduced.
AB - Magnetic resonance image analysis is often hampered by inconsistent data due to upgrades or changes to the scanner platform or modification of scanning protocols. These changes can manifest in three main sources of image inconsistency: contrast, resolution, and noise. Modern analysis techniques that use supervised machine learning can be especially susceptible to these inconsistencies, as existing training data may not be valid after an upgrade or protocol change. In previous work, differences in contrast and resolution have been addressed in isolation. We propose a novel method of image intensity harmonization that addresses each of the three sources of inconsistency. We formulate our method around a multi-planar, multi-contrast U-net, where all of the available contrasts are used as input channels in a single modified U-Net to produce all of the output contrasts simultaneously. The multi-contrast nature of the deep network allows for harmonization of contrast as information can be shared between contrasts. In addition, coherent, biological features are highlighted and matched to the target, while noise, which differs between matched inputs and outputs, is not reinforced. This process also normalizes small differences in resolution due to the influence of the high resolution channels. To combat larger differences in resolution, which would not be recovered by the neural network alone, we use self super-resolution (SSR) on all images with thick (>2 mm) slices before harmonization. To generate consistent images, the target images are also processed in a similar manner so that all resulting images have consistent qualities. Our harmonization process eliminates significant volume bias of multiple brain compartments and lesion estimation. In addition, absolute volume difference and Dice similarity of segmentation volumes were significantly improved (p < 0.005). SSR alone did not affect the statistical significance of the difference, even though the absolute volume difference was reduced.
KW - Deep learning
KW - MR harmonization
KW - Multiple sclerosis
KW - Super-resolution
UR - http://www.scopus.com/inward/record.url?scp=85053915669&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85053915669&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-00536-8_3
DO - 10.1007/978-3-030-00536-8_3
M3 - Conference contribution
AN - SCOPUS:85053915669
SN - 9783030005351
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 20
EP - 30
BT - Simulation and Synthesis in Medical Imaging - Third International Workshop, SASHIMI 2018, Held in Conjunction with MICCAI 2018, Proceedings
A2 - Goksel, Orcun
A2 - Oguz, Ipek
A2 - Gooya, Ali
A2 - Burgos, Ninon
PB - Springer Verlag
T2 - 3rd International Workshop on Simulation and Synthesis in Medical Imaging, SASHIMI 2018 Held in Conjunction with 21st International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2018
Y2 - 16 September 2018 through 16 September 2018
ER -