Details
| Originalsprache | Englisch |
|---|---|
| Seitenumfang | 6 |
| Publikationsstatus | Elektronisch veröffentlicht (E-Pub) - 2019 |
Abstract
Zitieren
- Standard
- Harvard
- Apa
- Vancouver
- BibTex
- RIS
2019.
Publikation: Arbeitspapier/Preprint › Preprint
}
TY - UNPB
T1 - Deformable Medical Image Registration Using a Randomly-Initialized CNN as Regularization Prior
AU - Laves, Max-Heinrich
AU - Ihler, Sontje
AU - Ortmaier, Tobias
PY - 2019
Y1 - 2019
N2 - We present deformable unsupervised medical image registration using arandomly-initialized deep convolutional neural network (CNN) as regularizationprior. Conventional registration methods predict a transformation by minimizingdissimilarities between an image pair. The minimization is usually regularizedwith manually engineered priors, which limits the potential of theregistration. By learning transformation priors from a large dataset, CNNs haveachieved great success in deformable registration. However, learned methods arerestricted to domain-specific data and the required amounts of medical data aredifficult to obtain. Our approach uses the idea of deep image priors to combineconvolutional networks with conventional registration methods based on manuallyengineered priors. The proposed method is applied to brain MRI scans. We showthat our approach registers image pairs with state-of-the-art accuracy byproviding dense, pixel-wise correspondence maps. It does not rely on priortraining and is therefore not limited to a specific image domain.
AB - We present deformable unsupervised medical image registration using arandomly-initialized deep convolutional neural network (CNN) as regularizationprior. Conventional registration methods predict a transformation by minimizingdissimilarities between an image pair. The minimization is usually regularizedwith manually engineered priors, which limits the potential of theregistration. By learning transformation priors from a large dataset, CNNs haveachieved great success in deformable registration. However, learned methods arerestricted to domain-specific data and the required amounts of medical data aredifficult to obtain. Our approach uses the idea of deep image priors to combineconvolutional networks with conventional registration methods based on manuallyengineered priors. The proposed method is applied to brain MRI scans. We showthat our approach registers image pairs with state-of-the-art accuracy byproviding dense, pixel-wise correspondence maps. It does not rely on priortraining and is therefore not limited to a specific image domain.
U2 - 10.48550/arXiv.1908.00788
DO - 10.48550/arXiv.1908.00788
M3 - Preprint
BT - Deformable Medical Image Registration Using a Randomly-Initialized CNN as Regularization Prior
ER -