Related Experiment Video
Updated: Aug 5, 2026
10:25
Brain Infarct Segmentation and Registration on MRI or CT for Lesion-symptom Mapping
Published on: September 25, 2019
Hyperelastic Regularization for Near-Diffeomorphic Transformer-Based Brain MRI Registration
Shiyi Xu1, Mohan Xu2, Erjin Zhou3
1Capital Medical University Second Clinical School, Capital Medical University, Beijing 100050, China.
Journal of Imaging
|July 27, 2026
Summary
Transformer-based brain MRI registration can create problematic foldings. HypEReg, a novel hyperelastic regularizer, effectively suppresses these foldings in displacement fields while maintaining high registration accuracy, improving reliability for morphometry studies.
Area of Science:
- Medical Imaging
- Computational Anatomy
- Machine Learning
Background:
- Transformer-based deformable registration offers high accuracy for brain MRI.
- However, predicted displacement fields may contain local foldings (non-positive Jacobian determinant).
- These foldings violate diffeomorphism assumptions crucial for downstream analyses like tensor-based morphometry and atlas-based segmentation.
Purpose of the Study:
- To introduce HypEReg, a non-linear hyperelastic regularizer for Transformer-based brain MRI registration.
- To suppress local foldings in predicted displacement fields without compromising registration accuracy.
- To enhance the reliability of deformation fields for morphometric analyses.
Main Methods:
- Developed HypEReg, a loss-side module integrating a volume-distortion penalty and an anti-folding hinge acting on the Jacobian determinant.
- Integrated HypEReg into a TransMorph backbone, creating HypEReg-TransMorph.
- Evaluated performance on the IXI and OASIS datasets, assessing registration accuracy (Dice score) and folding suppression (det(Jϕ)≤0 voxel ratio).
Main Results:
- HypEReg-TransMorph significantly reduced the ratio of voxels with non-positive Jacobian determinants (folding) from 1.502×10-2 to 1.5×10-5 on the IXI benchmark, maintaining grouped Dice of 0.7537.
- In zero-shot transfer on OASIS, HypEReg-TransMorph achieved Dice 0.7756 with a folding ratio of 7.6×10-5, outperforming plain TransMorph (Dice 0.7691; ratio 9.6×10-3).
- Downstream multi-atlas label fusion and longitudinal analyses confirmed improved deformation plausibility and accuracy.
Conclusions:
- HypEReg provides effective, determinant-level, non-linear hyperelastic regularization for Transformer-based brain MRI registration.
- It substantially suppresses folding while preserving alignment accuracy with no added inference cost.
- HypEReg is a practical, drop-in solution for improving the reliability of deformation fields in morphometry-oriented registration.
