Stochastic optimization with randomized smoothing for image registration

Stochastic optimization with randomized smoothing for image registration

Sun, Wei (Erasmus MC; University of Southern California)
Poot, D.H.J. (TU Delft ImPhys/Quantitative Imaging; Erasmus MC)
Smal, Ihor (Erasmus MC)
Yang, Xuan (Shenzhen University)
Niessen, W.J. (TU Delft ImPhys/Quantitative Imaging; Erasmus MC)
Klein, S. (Erasmus MC)

2017

Image registration is typically formulated as an optimization process, which aims to find the optimal transformation parameters of a given transformation model by minimizing a cost function. Local minima may exist in the optimization landscape, which could hamper the optimization process. To eliminate local minima, smoothing the cost function would be desirable. In this paper, we investigate the use of a randomized smoothing (RS) technique for stochastic gradient descent (SGD) optimization, to effectively smooth the cost function. In this approach, Gaussian noise is added to the transformation parameters prior to computing the cost function gradient in each iteration of the SGD optimizer. The approach is suitable for both rigid and nonrigid registrations. Experiments on synthetic images, cell images, public CT lung data, and public MR brain data demonstrate the effectiveness of the novel RS technique in terms of registration accuracy and robustness.

Image registration
Local minima
Randomized smoothing
Stochastic gradient descent

http://resolver.tudelft.nl/uuid:54341eb8-4282-4d17-aa74-ff41bbe65ff2

https://doi.org/10.1016/j.media.2016.07.003

2018-07-13

1361-8415

Medical Image Analysis, 35, 146-158

Accepted Author Manuscript

Institutional Repository

journal article

© 2017 Wei Sun, D.H.J. Poot, Ihor Smal, Xuan Yang, W.J. Niessen, S. Klein