Abstract:
|
Image registration is a core problem in Imaging and Data Sciences. Although many models and computational algorithms have been developed in recent years, it remains a big challenge to achieve both an accurate solution and fast speed for real time applications. In this talk, after reviewing a few classes of models, we discuss the deep learning framework. To offer wider applicability, we consider the scenario where ground-truth deformation fields are not available for training. We propose that the deformation fields are self-trained by a variational model compromised by an image similarity metric and a regularization term. The latter builds in a constraint on the determinant of the transformation in order to obtain a diffeomorphic solution.
The proposed algorithm is first trained and tested on synthetic and real mono-modal images. The results show how it deals with large deformation registration problems and leads to a real time solution with no folding. It is then generalized to multi-modal images. To improve the robustness, we combine the deep learning algorithm with a pre-processing approach. The initial given pair of images, which are non-linearly correlated, are first processed and optimized to serve the purpose of intensity or edge correction. This pre-processing step is based on the reproducing Kernel Hilbert space theory and yields intermediate new images which are more strongly correlated than the original ones and which will be used for training the model. Initial experiments and comparisons with learning and non-learning models demonstrate that this approach can deliver good performances and simultaneously generate an accurate diffeomorphic transformation.
|