Date: October, 16th, Tuesday
Time: 11:45 am
Place: Little 339 (the Atrium)
Speaker: Qingguo Zeng


Drinks and Pizza will be provided after the talk

Title:  DEFORMABLE IMAGE REGISTRATION IN RADIATION THERAPY

Abstract:

 In this talk, a fast, robust and accurate deformable image
registration technique is provided. The proposed method
considers the difference between the deformed template and
the reference at each voxel as independent normal random variables
with zero mean and a variance to be optimized. Then
the optimal deformation field and variance are obtained by
minimizing an energy function that consists of the negative
log-likelihood function and a regularization term for the
deformation field.
Since the proposed model accommodates the variance in the
alignment,it is more general and robust to noise than the widely
used Sum of Square Distance (SSD) model, which has limitation in
registering non-linearly related images. Moreover,
the weight between the smoothness of the deformation field and
goodness of the alignment in the proposed model
is adjusted automatically by the updated variance during
iterations.
This feature improves the accuracy of the alignment and speed of
the convergence.
To enhance algorithm efficiency we used the Additive Operator
Splitting(AOS), a fast semi-implicit finite difference scheme, and
multi-resolution method in solving the minimization problem.
The experimental results on synthetic images,
and volumetric CT and MRI lung data indicate
the effectiveness of the proposed scheme.
กก