TY - GEN
T1 - Inhomogeneity Correction in Magnetic Resonance Images Using Deep Image Priors
AU - Han, Shuo
AU - Prince, Jerry L.
AU - Carass, Aaron
N1 - Funding Information:
The authors would thank colleagues from the Image Analysis and Communications Laboratory at the Johns Hopkins University for their support and help. Data were provided in part by OASIS-3: Principal Investigators: T. Benzinger, D. Marcus, J. Morris; NIH P50AG00561, P30NS09857781, P01AG026276, P01AG003991, R01AG043434, UL1TR000448, R01EB009352. AV-45 doses were provided by Avid Radiopharmaceuticals, a wholly owned subsidiary of Eli Lilly.
Publisher Copyright:
© 2020, Springer Nature Switzerland AG.
PY - 2020
Y1 - 2020
N2 - Intensity inhomogeneity in magnetic resonance (MR) images can decrease the performance of image processing, such as segmentation and registration. In this work, we propose an unsupervised learning approach to correct the inhomogeneity of an MR image based on deep image priors (DIPs). In DIPs, the structure of the convolutional neural networks was previously shown to capture the prior probability of an image, which has been demonstrated in several applications such as image denoising, segmentation, and super resolution. To obtain an inhomogeneity-free MR image, the problem was formulated in a Bayesian inference framework. The priors of the image and inhomogeneity field were captured by two DIPs and their likelihood was modeled based on the observed image. The approximated expectation of the posterior was calculated to get the corrected image using a stochastic gradient Langevin dynamics algorithm. Since we modeled the noise distribution, the proposed method is simultaneously capable of denoising to some extent. We compared our method with N4, a popular inhomogeneity correction method, in a simulated data set and a couple of real data sets, statistically showing that it has comparable or even superior performance than N4 when the inhomogeneity is severe or noise is high.
AB - Intensity inhomogeneity in magnetic resonance (MR) images can decrease the performance of image processing, such as segmentation and registration. In this work, we propose an unsupervised learning approach to correct the inhomogeneity of an MR image based on deep image priors (DIPs). In DIPs, the structure of the convolutional neural networks was previously shown to capture the prior probability of an image, which has been demonstrated in several applications such as image denoising, segmentation, and super resolution. To obtain an inhomogeneity-free MR image, the problem was formulated in a Bayesian inference framework. The priors of the image and inhomogeneity field were captured by two DIPs and their likelihood was modeled based on the observed image. The approximated expectation of the posterior was calculated to get the corrected image using a stochastic gradient Langevin dynamics algorithm. Since we modeled the noise distribution, the proposed method is simultaneously capable of denoising to some extent. We compared our method with N4, a popular inhomogeneity correction method, in a simulated data set and a couple of real data sets, statistically showing that it has comparable or even superior performance than N4 when the inhomogeneity is severe or noise is high.
UR - http://www.scopus.com/inward/record.url?scp=85092734326&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85092734326&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-59861-7_41
DO - 10.1007/978-3-030-59861-7_41
M3 - Conference contribution
AN - SCOPUS:85092734326
SN - 9783030598600
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 404
EP - 413
BT - Machine Learning in Medical Imaging - 11th International Workshop, MLMI 2020, Held in Conjunction with MICCAI 2020, Proceedings
A2 - Liu, Mingxia
A2 - Lian, Chunfeng
A2 - Yan, Pingkun
A2 - Cao, Xiaohuan
PB - Springer Science and Business Media Deutschland GmbH
T2 - 11th International Workshop on Machine Learning in Medical Imaging, MLMI 2020, held in conjunction with the 23rd International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2020
Y2 - 4 October 2020 through 4 October 2020
ER -