Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks

Chengjia Wang; Gillian Macnaught; Giorgos Papanastasiou; Tom MacGillivray; David Newby

doi:10.1007/978-3-030-00536-8_6

Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks

Chengjia Wang^*, Gillian Macnaught, Giorgos Papanastasiou, Tom MacGillivray, David Newby

^*Corresponding author for this work

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

24 Citations (Scopus)

Abstract

Recently, the cycle-consistent generative adversarial networks (CycleGAN) has been widely used for synthesis of multi-domain medical images. The domain-specific nonlinear deformations captured by CycleGAN make the synthesized images difficult to be used for some applications, for example, generating pseudo-CT for PET-MR attenuation correction. This paper presents a deformation-invariant CycleGAN (DicycleGAN) method using deformable convolutional layers and new cycle-consistency losses. Its robustness dealing with data that suffer from domain-specific nonlinear deformations has been evaluated through comparison experiments performed on a multi-sequence brain MR dataset and a multi-modality abdominal dataset. Our method has displayed its ability to generate synthesized data that is aligned with the source while maintaining a proper quality of signal compared to CycleGAN-generated data. The proposed model also obtained comparable performance with CycleGAN when data from the source and target domains are alignable through simple affine transformations.

Original language	English
Title of host publication	Simulation and Synthesis in Medical Imaging. SASHIMI 2018
Editors	Orcun Goksel, Ipek Oguz, Ali Gooya, Ninon Burgos
Publisher	Springer
Pages	52-60
Number of pages	9
ISBN (Electronic)	9783030005368
ISBN (Print)	9783030005351
DOIs	https://doi.org/10.1007/978-3-030-00536-8_6
Publication status	Published - 12 Sept 2018
Event	3rd International Workshop on Simulation and Synthesis in Medical Imaging 2018 - Granada, Spain Duration: 16 Sept 2018 → 16 Sept 2018

Publication series

Name	Lecture Notes in Computer Science
Volume	11037
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	3rd International Workshop on Simulation and Synthesis in Medical Imaging 2018
Abbreviated title	SASHIMI 2018
Country/Territory	Spain
City	Granada
Period	16/09/18 → 16/09/18
Other	Held in Conjunction with 21st International Conference on Medical Image Computing and Computer-Assisted Intervention, MICCAI 2018.

Keywords

Deep learning
GAN
Synthesis
Unsupervised learning

ASJC Scopus subject areas

Theoretical Computer Science
General Computer Science

Access to Document

10.1007/978-3-030-00536-8_6

Cite this

Wang, C., Macnaught, G., Papanastasiou, G., MacGillivray, T., & Newby, D. (2018). Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. In O. Goksel, I. Oguz, A. Gooya, & N. Burgos (Eds.), Simulation and Synthesis in Medical Imaging. SASHIMI 2018 (pp. 52-60). (Lecture Notes in Computer Science; Vol. 11037). Springer. https://doi.org/10.1007/978-3-030-00536-8_6

Wang, Chengjia ; Macnaught, Gillian ; Papanastasiou, Giorgos et al. / Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. Simulation and Synthesis in Medical Imaging. SASHIMI 2018. editor / Orcun Goksel ; Ipek Oguz ; Ali Gooya ; Ninon Burgos. Springer, 2018. pp. 52-60 (Lecture Notes in Computer Science).

@inproceedings{07a2b5a2b77d40908a10df8deb06e611,

title = "Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks",

abstract = "Recently, the cycle-consistent generative adversarial networks (CycleGAN) has been widely used for synthesis of multi-domain medical images. The domain-specific nonlinear deformations captured by CycleGAN make the synthesized images difficult to be used for some applications, for example, generating pseudo-CT for PET-MR attenuation correction. This paper presents a deformation-invariant CycleGAN (DicycleGAN) method using deformable convolutional layers and new cycle-consistency losses. Its robustness dealing with data that suffer from domain-specific nonlinear deformations has been evaluated through comparison experiments performed on a multi-sequence brain MR dataset and a multi-modality abdominal dataset. Our method has displayed its ability to generate synthesized data that is aligned with the source while maintaining a proper quality of signal compared to CycleGAN-generated data. The proposed model also obtained comparable performance with CycleGAN when data from the source and target domains are alignable through simple affine transformations.",

keywords = "Deep learning, GAN, Synthesis, Unsupervised learning",

author = "Chengjia Wang and Gillian Macnaught and Giorgos Papanastasiou and Tom MacGillivray and David Newby",

note = "Funding Information: C. Wang—This work was supported by British Heart Foundation. Publisher Copyright: {\textcopyright} 2018, Springer Nature Switzerland AG.; 3rd International Workshop on Simulation and Synthesis in Medical Imaging 2018, SASHIMI 2018 ; Conference date: 16-09-2018 Through 16-09-2018",

year = "2018",

month = sep,

day = "12",

doi = "10.1007/978-3-030-00536-8_6",

language = "English",

isbn = "9783030005351",

series = "Lecture Notes in Computer Science",

publisher = "Springer",

pages = "52--60",

editor = "Orcun Goksel and Ipek Oguz and Ali Gooya and Ninon Burgos",

booktitle = "Simulation and Synthesis in Medical Imaging. SASHIMI 2018",

}

Wang, C, Macnaught, G, Papanastasiou, G, MacGillivray, T & Newby, D 2018, Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. in O Goksel, I Oguz, A Gooya & N Burgos (eds), Simulation and Synthesis in Medical Imaging. SASHIMI 2018. Lecture Notes in Computer Science, vol. 11037, Springer, pp. 52-60, 3rd International Workshop on Simulation and Synthesis in Medical Imaging 2018, Granada, Spain, 16/09/18. https://doi.org/10.1007/978-3-030-00536-8_6

Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. / Wang, Chengjia; Macnaught, Gillian; Papanastasiou, Giorgos et al.
Simulation and Synthesis in Medical Imaging. SASHIMI 2018. ed. / Orcun Goksel; Ipek Oguz; Ali Gooya; Ninon Burgos. Springer, 2018. p. 52-60 (Lecture Notes in Computer Science; Vol. 11037).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution

TY - GEN

T1 - Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks

AU - Wang, Chengjia

AU - Macnaught, Gillian

AU - Papanastasiou, Giorgos

AU - MacGillivray, Tom

AU - Newby, David

PY - 2018/9/12

Y1 - 2018/9/12

N2 - Recently, the cycle-consistent generative adversarial networks (CycleGAN) has been widely used for synthesis of multi-domain medical images. The domain-specific nonlinear deformations captured by CycleGAN make the synthesized images difficult to be used for some applications, for example, generating pseudo-CT for PET-MR attenuation correction. This paper presents a deformation-invariant CycleGAN (DicycleGAN) method using deformable convolutional layers and new cycle-consistency losses. Its robustness dealing with data that suffer from domain-specific nonlinear deformations has been evaluated through comparison experiments performed on a multi-sequence brain MR dataset and a multi-modality abdominal dataset. Our method has displayed its ability to generate synthesized data that is aligned with the source while maintaining a proper quality of signal compared to CycleGAN-generated data. The proposed model also obtained comparable performance with CycleGAN when data from the source and target domains are alignable through simple affine transformations.

AB - Recently, the cycle-consistent generative adversarial networks (CycleGAN) has been widely used for synthesis of multi-domain medical images. The domain-specific nonlinear deformations captured by CycleGAN make the synthesized images difficult to be used for some applications, for example, generating pseudo-CT for PET-MR attenuation correction. This paper presents a deformation-invariant CycleGAN (DicycleGAN) method using deformable convolutional layers and new cycle-consistency losses. Its robustness dealing with data that suffer from domain-specific nonlinear deformations has been evaluated through comparison experiments performed on a multi-sequence brain MR dataset and a multi-modality abdominal dataset. Our method has displayed its ability to generate synthesized data that is aligned with the source while maintaining a proper quality of signal compared to CycleGAN-generated data. The proposed model also obtained comparable performance with CycleGAN when data from the source and target domains are alignable through simple affine transformations.

KW - Deep learning

KW - GAN

KW - Synthesis

KW - Unsupervised learning

UR - http://www.scopus.com/inward/record.url?scp=85053888519&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-00536-8_6

DO - 10.1007/978-3-030-00536-8_6

M3 - Conference contribution

AN - SCOPUS:85053888519

SN - 9783030005351

T3 - Lecture Notes in Computer Science

SP - 52

EP - 60

BT - Simulation and Synthesis in Medical Imaging. SASHIMI 2018

A2 - Goksel, Orcun

A2 - Oguz, Ipek

A2 - Gooya, Ali

A2 - Burgos, Ninon

PB - Springer

T2 - 3rd International Workshop on Simulation and Synthesis in Medical Imaging 2018

Y2 - 16 September 2018 through 16 September 2018

ER -

Wang C, Macnaught G, Papanastasiou G, MacGillivray T, Newby D. Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks. In Goksel O, Oguz I, Gooya A, Burgos N, editors, Simulation and Synthesis in Medical Imaging. SASHIMI 2018. Springer. 2018. p. 52-60. (Lecture Notes in Computer Science). doi: 10.1007/978-3-030-00536-8_6

Unsupervised learning for cross-domain medical image synthesis using deformation invariant cycle consistency networks

Abstract

Publication series

Conference

Keywords

ASJC Scopus subject areas

Access to Document

Other files and links

Fingerprint

Cite this