Source-linked AI summary
Large Deformation Diffeomorphic Image Registration with Laplacian Pyramid Networks
Tony C. W. Mok, Albert C. S. Chung
TL;DR
Existing deep learning registration methods are often limited to small deformations and may not preserve bijective, topology-preserving transformations. LapIRN uses Laplacian-pyramid, coarse-to-fine optimization within diffeomorphic maps, achieving state-of-the-art accuracy with 0.3-second running time while maintaining diffeomorphic properties.
Problem
Deep learning image registration methods are often limited to small deformations, while bijective mapping and topology preservation may not be guaranteed.
Method
LapIRN combines a Laplacian pyramid and pyramid similarity metric with stationary-velocity-field optimization within the space of diffeomorphic maps.
Results
State-of-the-art registration accuracy was achieved with 0.3 sec running time while maintaining desirable diffeomorphic properties.
Takeaways & Limitations
The formulation can be transferred to various applications with minimum effort and shows potential for time-sensitive medical studies.
Abstract
from arXiv · showhide
Deep learning-based methods have recently demonstrated promising results in deformable image registration for a wide range of medical image analysis tasks. However, existing deep learning-based methods are usually limited to small deformation settings, and desirable properties of the transformation including bijective mapping and topology preservation are often being ignored by these approaches. In this paper, we propose a deep Laplacian Pyramid Image Registration Network, which can solve the image registration optimization problem in a coarse-to-fine fashion within the space of diffeomorphic maps. Extensive quantitative and qualitative evaluations on two MR brain scan datasets show that our method outperforms the existing methods by a significant margin while maintaining desirable diffeomorphic properties and promising registration speed.
1 Introduction
The paper targets large-deformation registration, where conventional methods can be slow and deep learning methods may struggle with optimization and diffeomorphic guarantees. LapIRN addresses these challenges through coarse-to-fine Laplacian-pyramid optimization, a pyramid similarity metric, and a diffeomorphic formulation.
- Challenge: Conventional registration methods can provide accurate, diffeomorphic transformations but become time-consuming for highly misaligned, high-resolution 3D images.Their runtime depends on the degree of misalignment.
- Challenge: Deep learning registration methods offer speed and quality but may struggle with unsupervised large-deformation registration.At finest resolution, similarity gradients can be rough and the optimization has many transformation degrees of freedom.
- Approach: The proposed pyramid similarity metric captures both large and small misalignments and helps avoid local minima during optimization.It is designed for use within the pyramid network.
- Approach: LapIRN uses a diffeomorphic setting that guarantees invertibility and topology preservation of the computed transformations.These properties address limitations of displacement-field approaches.
2 Methods
LapIRN performs 3D deformable registration through a three-level Laplacian pyramid, optimizing alignment from coarse to fine while using a diffeomorphic deformation model. Its similarity pyramid combines multi-resolution image similarity, and its velocity-field formulation targets smooth, invertible, topology-preserving transformations.
- Deep Laplacian Pyramid Image Registration Networks: LapIRN uses three identical CNN-based registration networks across pyramid levels to mimic conventional multi-resolution registration.The architecture combines feature encoders, residual blocks, and decoders, with lower-level feature embeddings passed to finer levels.
- Deep Laplacian Pyramid Image Registration Networks: Each finer pyramid level upsamples the previous deformation and velocity fields, warps the moving image, and predicts a refined deformation.Skip connections transfer lower-level feature embeddings to improve receptive field and nonlinear correspondence learning.
- Coarse-to-fine Training: LapIRN uses coarse-to-fine training with a stable warm start because balancing multiple resolution losses makes end-to-end training difficult.The training scheme is designed to stabilize optimization across pyramid levels.
- Similarity Pyramid: The similarity pyramid combines local normalized cross-correlation across resolutions, assigning lower weights to coarser levels to reduce local-minimum problems.Coarser-resolution similarity is smoother and less sensitive to noise, while the implementation sets the window size to w = 1 + 2i.
- Diffeomorphic Deformation: LapIRN parameterizes deformation with a stationary velocity field and integrates it over unit time under the Log-Euclidean framework to obtain a diffeomorphic map.The reported LapIRNdisp variant instead parameterizes deformation with displacement fields.
3 Experiments
The experiments evaluate LapIRN on OASIS and LPBA40 brain-atlas registration using accuracy, deformation quality, topology change, and runtime measures. LapIRN achieves the strongest reported registration performance while maintaining smooth deformation fields, very few folding voxels, and fast inference.
- Datasets: The evaluation uses 425 OASIS and 40 LPBA40 T1-weighted brain MR scans for brain-atlas registration.OASIS includes subjects aged 18 to 96, including 100 subjects with very mild to moderate Alzheimer’s disease.
- Evaluation Metrics: The study measures Dice accuracy, non-positive Jacobian percentage, Jacobian standard deviation, topology change, and registration time.These metrics jointly assess anatomical overlap, folding, smoothness, topology, and computational speed.
- Baseline Methods: LapIRN is compared with Demons, SyN, VM, and DIF-VM using official implementations or carefully tuned parameters, with DLIR methods trained from scratch.Demons and SyN represent conventional approaches, while VM and DIF-VM represent deep learning-based methods.
- Quantitative Results: 0.808 Dice is achieved by LapIRNdisp on OASIS, exceeding Demons, SyN, DIF-VM, and VM by 13%, 18%, 17%, and 11% of Dice score, respectively.LapIRNdisp does not guarantee smooth and locally invertible solutions, as reflected by high Jacobian variation and folding percentages.
- Quantitative Results: 0.319 Jacobian standard deviation and < 0.01% folding voxels are achieved by LapIRN, alongside 0.33 sec inference time per MR-scan pair.The paper reports LapIRN as having the best registration performance across the baseline methods while producing plausible, smooth deformation fields.
4 Conclusion
LapIRN combines a similarity pyramid for large misalignments with diffeomorphic stationary-vector-field formulation, achieving accurate, fast, and topology-preserving registration.
- The similarity pyramid mimics conventional multi-resolution registration to capture large misalignments between input scans.
- LapIRN formulates deformation using stationary vector fields under the Log-Euclidean framework to guarantee desirable diffeomorphic properties.
- 0.3 sec running time accompanies state-of-the-art registration accuracy in the reported experiments.
- The formulation can be transferred to various applications with minimum effort, supporting time-sensitive medical studies.