Weak Bounding Box Supervision for Image Registration Networks

Image registration is a fundamental task in medical image analysis. Many deep learning based methods use multi-label image segmentations during training to reach the performance of conventional algorithms. But the creation of detailed annotations is very time-consuming and expert knowledge is essential. To avoid this, we propose a weakly supervised learning scheme for deformable image registration that uses bounding boxes during training. By calculating the loss function based on these bounding box labels, we are able to perform an image registration with large deformations without using densely labeled annotations. The performance of the registration of inter-patient 3D Abdominal CT images can be enhanced by approximately 10% only with little annotation effort in comparison to unsupervised learning methods. Taken into account this annotation effort, the performance also exceeds the performance of the label supervised training.