Pretraining Deformable Image Registration Networks with Random Images

TL;DR

This paper introduces a pretraining method for image registration networks using random images, which improves accuracy, reduces data requirements, and speeds up training for medical image registration tasks.

Contribution

It proposes using random image registration as a pretraining task, a novel approach that enhances model performance and efficiency in medical image registration.

Findings

Pretraining with random images improves registration accuracy.

Reduces the amount of domain-specific data needed.

Speeds up convergence during downstream training.

Abstract

Recent advances in deep learning-based medical image registration have shown that training deep neural networks~(DNNs) does not necessarily require medical images. Previous work showed that DNNs trained on randomly generated images with carefully designed noise and contrast properties can still generalize well to unseen medical data. Building on this insight, we propose using registration between random images as a proxy task for pretraining a foundation model for image registration. Empirical results show that our pretraining strategy improves registration accuracy, reduces the amount of domain-specific data needed to achieve competitive performance, and accelerates convergence during downstream training, thereby enhancing computational efficiency.