Feature-based image registration in structured light endoscopy
Proceedings of the Fourth Conference on Medical Imaging with Deep Learning, PMLR 143:369-383, 2021.
Images offer a two-dimensional (2D) representation of a three-dimensional (3D) environment. However, in many biomedical tasks, a 3D view is crucial for diagnosis. Projecting structured light, such as a regular laser grid, onto the surface of interest allows to reconstruct its 3D structure. For reconstruction, it is crucial to correctly identify and assign each laser ray to its respective position in the laser grid. Current methods for this task use semi-automatic, yet highly manual annotations. Hence, a fully automatic, reliable method is desired. Here, we show that this assignment can be approached as an image registration. We first separate the laser rays from the background using semantic segmentation. We found that registration of the extracted laser rays directly to the fixed laser grid image fails, when we use state-of-the-art intensity-based image registration techniques, such as ANTs. Using our feature-based custom loss and a deep neural network, we are able to use a U-Net-like architecture to compute deformation fields to successfully register the laser rays onto the fixed image accompanied with a custom post-processing sorting step. Using synthetic data, we show that the network is in general able to learn affine and non-linear transformations. Our method is also robust to missing or occluded rays. Using an ex vivo dataset, we achieved an registration accuracy of 91%. In summary, we provide a new platform to perform feature-based registration and showcase this on a biomedical dataset. In future, we will evaluate different architectural designs and more complex datasets.