Efficient motion compensation for fetal MRI, Bernhard Kainz
Sprache des Titels:
In this talk we will review a simple but very effective and parallelisable SVR model similar to x_i = W_i y + n_i for 1 <= i < N, where x_i denotes a low resolution (LR) image stack of total N stacks, and y being a high resolution target image. W_i=DBT_i combines motion compensation, sub-sampling and degradation effects, where $D$ is a sub-sampling matrix, B is a blurring matrix, and T_i is a transformation matrix of observation i. n_i adds noise. Thus, the LR image can be considered as a down-sampled, motion corrupted, blurred, and noisy version of a high-resolution output space. The motion compensation problem can be divided into two main parts: (1) motion correction (estimating W_i) and (2) super-resolution (estimating y). Intensity-based image registration is used for estimating W_i, MRI point-spread function-informed super-resolution for obtaining a uniformly spaced motion-free high-resolution image. Expectation-maximization supports the iterative optimisation of (1) and (2) with an outlier rejection and noise mitigation model.