Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

uniGradICON: A Foundation Model for Medical Image Registration

Lin Tian, Hastings Greer, Roland Kwitt, Francois-Xavier Vialard, Raul San Jose Estepar, Sylvain Bouix, Richard Rushmore, Marc Niethammer

TL;DR

UniGradICON unifies the speed and accuracy benefits of learning-based registration algorithms with the generic applicability of conventional non-deep-learning approaches.

Abstract

Conventional medical image registration approaches directly optimize over the parameters of a transformation model. These approaches have been highly successful and are used generically for registrations of different anatomical regions. Recent deep registration networks are incredibly fast and accurate but are only trained for specific tasks. Hence, they are no longer generic registration approaches. We therefore propose uniGradICON, a first step toward a foundation model for registration providing 1) great performance \emph{across} multiple datasets which is not feasible for current learning-based registration methods, 2) zero-shot capabilities for new registration tasks suitable for different acquisitions, anatomical regions, and modalities compared to the training dataset, and 3) a strong initialization for finetuning on out-of-distribution registration tasks. UniGradICON unifies the speed and accuracy benefits of learning-based registration algorithms with the generic applicability of conventional non-deep-learning approaches. We extensively trained and evaluated uniGradICON on twelve different public datasets. Our code and the uniGradICON model are available at https://github.com/uncbiag/uniGradICON.

uniGradICON: A Foundation Model for Medical Image Registration

TL;DR

UniGradICON unifies the speed and accuracy benefits of learning-based registration algorithms with the generic applicability of conventional non-deep-learning approaches.

Abstract

Conventional medical image registration approaches directly optimize over the parameters of a transformation model. These approaches have been highly successful and are used generically for registrations of different anatomical regions. Recent deep registration networks are incredibly fast and accurate but are only trained for specific tasks. Hence, they are no longer generic registration approaches. We therefore propose uniGradICON, a first step toward a foundation model for registration providing 1) great performance \emph{across} multiple datasets which is not feasible for current learning-based registration methods, 2) zero-shot capabilities for new registration tasks suitable for different acquisitions, anatomical regions, and modalities compared to the training dataset, and 3) a strong initialization for finetuning on out-of-distribution registration tasks. UniGradICON unifies the speed and accuracy benefits of learning-based registration algorithms with the generic applicability of conventional non-deep-learning approaches. We extensively trained and evaluated uniGradICON on twelve different public datasets. Our code and the uniGradICON model are available at https://github.com/uncbiag/uniGradICON.
Paper Structure (14 sections, 1 equation, 2 figures, 7 tables)

This paper contains 14 sections, 1 equation, 2 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Material & method
  3. Dataset curation and pre-processing
  4. Registration network
  5. Training protocol and experimental setting
  6. Model availability and development plan
  7. Experiments
  8. Performance on in-distribution tasks
  9. Performance on out-of-distribution tasks
  10. Performance of finetuning on out-of-distribution dataset
  11. Conclusion, limitations, and future work
  12. Acknowledgements
  13. Dataset and implementation details
  14. Some example registration results

Figures (2)

  • Figure 1: Example uniGradICON registrations. Prediction only w/o IO.
  • Figure 2: Visualization of uniGradICON registration results for zero-shot inference. We display images as they are presented to our uniGradICON foundation model, i.e., not necessarily based on the typical anatomical convention. Note that while for a task-specific network it is important to use consistent image orientations it is much less clear that this is the case for a universal registration network, as such a network is ideally expected to be able to handle images of any orientation. In future work this could be further studied, for example, by exploring specific orientation-changing data augmentation strategies.