Domain agnostic 2D-3D deformable registration Application to fluoroscopic guidance without contrast agent

We present a method for estimating, in real time, a 3D displacement field from a single fluoroscopic image. Our approach uses a fully convolutional network architecture to solve the associated inverse problem. Supervised learning is performed on synthetic data, using Digitally Reconstructed Radiographs as input and displacement fields as output. We use randomized Gaussian kernels to produce a synthetic training dataset with displacement fields that are smooth and diffeomorphic. In contrast to other 2D-3D registration methods, our novel data generation approach does not rely on a statistical motion model. This enables our model to accurately predict deformations unrelated to breathing or other predetermined motion patterns. As an example of clinical application, we show that our model is able to predict deformations related to percutaneous needle insertions accurately, potentially removing the need for contrast agent injection.