# TransMorph: Transformer for unsupervised medical image registration

**Authors:** Junyu Chen, Yong Du, Yufan He, William P. Segars, Ye Li, Eirc C. Frey  
**arXiv:** [2111.10480](https://arxiv.org/abs/2111.10480)  
**Date:** 2021-11-19  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2111.10480v1/

## Key points

- Convolutional neural networks (ConvNets) have been widely used in medical image analysis but have limitations in capturing long-range spatial relations between voxels.
- Vision Transformers have emerged as promising alternatives to ConvNets, showing state-of-the-art performances in various medical imaging applications.
- TransMorph is a hybrid model combining Transformer and ConvNet architectures designed for volumetric medical image registration tasks.
- TransMorph offers three variants: two diffeomorphic variants for topology-preserving deformations and a Bayesian variant for registration uncertainty estimation.
- Extensive testing against existing methods and Transformer architectures using volumetric medical images showed significant performance improvement with TransMorph.
- The research demonstrates the effectiveness of Transformers in enhancing accuracy and efficiency in medical image registration processes.

## Summary

, , , , 

In the rapidly evolving field of medical image analysis, convolutional neural networks (ConvNets) have been at the forefront for the past decade. However, despite their widespread use, ConvNets have shown limitations in capturing long-range spatial relations between voxels within an image. Recent advancements in vision Transformers have emerged as promising alternatives to ConvNets, showcasing state-of-the-art performances in various medical imaging applications. One such innovative approach is presented in the paper titled "TransMorph: Transformer for unsupervised medical image registration," authored by Junyu Chen, Yong Du, Yufan He, William P. Segars, Ye Li, and Eirc C. Frey. The research introduces TransMorph, a hybrid model that combines Transformer and ConvNet architectures specifically designed for volumetric medical image registration tasks. The proposed TransMorph model offers three distinct variants: two diffeomorphic variants that ensure topology-preserving deformations and a Bayesian variant that provides a well-calibrated estimate of registration uncertainty. To validate the efficacy of TransMorph and its variants, extensive testing was conducted against a range of existing registration methods and Transformer architectures using volumetric medical images from inter-patient brain MRI registration and phantom-to-CT registration applications. The results obtained from both qualitative and quantitative analyses demonstrate a significant performance improvement over baseline methods. This highlights the effectiveness of Transformers in enhancing accuracy and efficiency in medical image registration processes. With 32 pages and 28 figures detailing their findings, the authors provide valuable insights into the potential of TransMorph as a cutting-edge solution for unsupervised medical image registration. By leveraging the unique capabilities of Transformers to capture intricate spatial relationships within images, this research opens up new possibilities for advancing precision medicine through enhanced imaging technologies.