Image-and-Spatial Transformer Networks for Structure-Guided Image Registration
This work addresses the challenge of precise alignment in medical image registration, particularly for Structures-of-Interest, offering a novel framework that improves accuracy in a domain-specific context.
The paper tackles the problem of accurate image registration for aligning anatomical structures in medical imaging by introducing Image-and-Spatial Transformer Networks (ISTNs), which leverage Structures-of-Interest information to learn optimized representations and enable iterative refinement, resulting in highly accurate registration even with limited training data.
Image registration with deep neural networks has become an active field of research and exciting avenue for a long standing problem in medical imaging. The goal is to learn a complex function that maps the appearance of input image pairs to parameters of a spatial transformation in order to align corresponding anatomical structures. We argue and show that the current direct, non-iterative approaches are sub-optimal, in particular if we seek accurate alignment of Structures-of-Interest (SoI). Information about SoI is often available at training time, for example, in form of segmentations or landmarks. We introduce a novel, generic framework, Image-and-Spatial Transformer Networks (ISTNs), to leverage SoI information allowing us to learn new image representations that are optimised for the downstream registration task. Thanks to these representations we can employ a test-specific, iterative refinement over the transformation parameters which yields highly accurate registration even with very limited training data. Performance is demonstrated on pairwise 3D brain registration and illustrative synthetic data.