Transformed Grid Distance Loss for Supervised Image Registration

Xinrui Song, Hanqing Chao, Sheng Xu, Baris Turkbey, Bradford Wood, Pingkun Yan

Keywords: Image registration, Loss function, Transformation parameter, Supervised learning

TL;DR: A novel loss function for rigid image registration that unites rotation and translation.

Abstract: Many deep learning image registration tasks, such as volume-to-volume registration, frame-to-volume registration, and frame-to-volume reconstruction, rely on six transformation parameters or quaternions to supervise the learning-based methods. However, these parameters can be very abstract for neural networks to comprehend. During the optimization process, ill-considered representations of rotation may even trap the objective function at local minima. This paper aims to expose these issues and propose the Transformed Grid Distance loss as a solution. The proposed method not only solves the problem of rotation representation but unites the gap between translation and rotation. We test our methods both with synthetic and clinically relevant medical image datasets. We demonstrate superior performance in comparison with conventional losses while requiring no alteration to the network input, output, or network structure at all.