TY - GEN
T1 - ISMORE
T2 - 4th International Workshop on Simulation and Synthesis in Medical Imaging, SASHIMI 2019, held in conjunction with the 22nd International Conference on Medical Image Computing and Computer Assisted Intervention, MICCAI 2019
AU - Zhao, Can
AU - Son, Seoyoung
AU - Kim, Yongsoo
AU - Prince, Jerry L.
N1 - Publisher Copyright:
© Springer Nature Switzerland AG 2019.
PY - 2019
Y1 - 2019
N2 - In 3D medical imaging, images with isotropic high resolution (HR) are almost always preferred. In practice, however, many acquired images, including magnetic resonance imaging (MRI) and fluorescence microscopy, have HR in the in-plane directions and low resolution (LR) in the through-plane direction. The blurriness and aliasing artifacts that result cannot be solved by simple interpolation. Instead, many researchers have proposed super-resolution algorithms including state-of-art convolutional neural network (CNN)-based methods that require matched training data that have paired LR/HR examples. Since these data are often unavailable in practice, self super-resolution algorithms that do not need external training data have also been proposed. These self super-resolution methods assume that the in-plane slices are HR, and can therefore be used as HR training data. By degrading them into LR images, 2D CNNs can be trained and then used to restore the images in the through-plane. However, there are two issues with these approaches. The first one is that the assumption of HR in-plane slices is actually not solid since these thick in-plane slices are averaged true HR thin slices. Training on thick slices is equivalent to training on averaged true HR images, which is suboptimal. The second one relates to the 2D CNNs used on 3D volume, which cannot guarantee slice consistency. Regarding both issues as well as the generalizability of algorithm, we made four contributions. We show in this paper that one of the existing 2D CNN-based self super-resolution methods, SMORE, can be further improved by iteratively applying it using 2D or 3D networks, yielding 2D and 3D iSMORE. This iterative framework improves training data from thick slices to thinner slices after each iteration, thus improves super-resolution accuracy after each iteration, and solves the first issue. The second contribution is that it uses a 3D network to preserve slice consistency. The third contribution is the use of an edge-based loss function and noise reduction to enhance the performance. Finally, we perform iSMORE on both MRI and two-photon fluorescence microscopy, which demonstrates its generalizability.
AB - In 3D medical imaging, images with isotropic high resolution (HR) are almost always preferred. In practice, however, many acquired images, including magnetic resonance imaging (MRI) and fluorescence microscopy, have HR in the in-plane directions and low resolution (LR) in the through-plane direction. The blurriness and aliasing artifacts that result cannot be solved by simple interpolation. Instead, many researchers have proposed super-resolution algorithms including state-of-art convolutional neural network (CNN)-based methods that require matched training data that have paired LR/HR examples. Since these data are often unavailable in practice, self super-resolution algorithms that do not need external training data have also been proposed. These self super-resolution methods assume that the in-plane slices are HR, and can therefore be used as HR training data. By degrading them into LR images, 2D CNNs can be trained and then used to restore the images in the through-plane. However, there are two issues with these approaches. The first one is that the assumption of HR in-plane slices is actually not solid since these thick in-plane slices are averaged true HR thin slices. Training on thick slices is equivalent to training on averaged true HR images, which is suboptimal. The second one relates to the 2D CNNs used on 3D volume, which cannot guarantee slice consistency. Regarding both issues as well as the generalizability of algorithm, we made four contributions. We show in this paper that one of the existing 2D CNN-based self super-resolution methods, SMORE, can be further improved by iteratively applying it using 2D or 3D networks, yielding 2D and 3D iSMORE. This iterative framework improves training data from thick slices to thinner slices after each iteration, thus improves super-resolution accuracy after each iteration, and solves the first issue. The second contribution is that it uses a 3D network to preserve slice consistency. The third contribution is the use of an edge-based loss function and noise reduction to enhance the performance. Finally, we perform iSMORE on both MRI and two-photon fluorescence microscopy, which demonstrates its generalizability.
UR - http://www.scopus.com/inward/record.url?scp=85075647573&partnerID=8YFLogxK
UR - http://www.scopus.com/inward/citedby.url?scp=85075647573&partnerID=8YFLogxK
U2 - 10.1007/978-3-030-32778-1_14
DO - 10.1007/978-3-030-32778-1_14
M3 - Conference contribution
AN - SCOPUS:85075647573
SN - 9783030327774
T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
SP - 130
EP - 139
BT - Simulation and Synthesis in Medical Imaging - 4th International Workshop, SASHIMI 2019, Held in Conjunction with MICCAI 2019, Proceedings
A2 - Burgos, Ninon
A2 - Gooya, Ali
A2 - Svoboda, David
PB - Springer
Y2 - 13 October 2019 through 13 October 2019
ER -