Diffeomorphic Image Registration with Neural Velocity Field

Kun Han, Shanlin Sun, Hao Tang, Deying Kong, Xiangyi Yan, Xiaohui Xie, Ma, Haoyu, Kong, Deying, Xie, Xiaohui

Abstract

Diffeomorphic image registration, offering smooth transformation and topology preservation, is required in many medical image analysis tasks. Traditional methods impose certain modeling constraints on the space of admissible transformations and use optimization to find the optimal transformation between two images. Specifying the right space of admissible transformations is challenging: the registration quality can be poor if the space is too restrictive, while the optimization can be hard to solve if the space is too general. Recent learning-based methods, utilizing deep neural networks to learn the transformation directly, achieve fast inference, but face challenges in accuracy due to the difficulties in capturing the small local deformations and generalization ability. Here we propose a new optimization-based method named DNVF (Diffeomorphic Image Registration with Neural Velocity Field) which utilizes deep neural network to model the space of admissible transformations. A multilayer perceptron (MLP) with sinusoidal activation function is used to represent the continuous velocity field and assigns a velocity vector to every point in space, providing the flexibility of modeling complex deformations as well as the convenience of optimization. Moreover, we propose a cascaded image registration framework (Cas-DNVF) by combining the benefits of both optimization and learning based methods, where a fully convolutional neural network (FCN) is trained to predict the initial deformation, followed by DNVF for further refinement. Experiments on two large-scale 3D MR brain scan datasets demonstrate that our proposed methods significantly outperform the state-of-the-art registration methods.

Introduction

Medical image registration is a fundamental task in medical image analysis, involving the alignment of two or more images into a common coordinate system. Diffeomorphic registration is particularly desirable as it ensures that the resulting transformations are smooth, invertible, and preserve the topology of anatomical structures. Historically, this field has been divided between traditional optimization-based algorithms and modern deep learning approaches.

Traditional methods like SyN or LDDMM are highly accurate but computationally intensive, often taking minutes or hours per image pair. Conversely, deep learning models like VoxelMorph offer near-instantaneous inference but often struggle with fine-grained local deformations and generalization to unseen data distributions. This paper introduces DNVF (Diffeomorphic Image Registration with Neural Velocity Field) and its cascaded variant, Cas-DNVF, to bridge the gap between these two paradigms, combining the precision of optimization with the representational power of neural fields.

Methodology: Neural Velocity Fields and Cascaded Refinement

The core innovation of DNVF is the representation of the transformation's velocity field as a continuous Neural Field. Instead of a discrete grid of vectors, the authors employ a Multi-Layer Perceptron (MLP) that maps 3D coordinates $(x, y, z)$ to velocity vectors.

The Cas-DNVF framework combines a pretrained FCN for initial alignment with a DNVF module for high-precision refinement.

Sinusoidal Activations and Diffeomorphisms

To overcome the "spectral bias" of standard ReLU-based networks—which tend to learn low-frequency components first—the authors utilize sinusoidal activation functions (SIREN). This allows the MLP to capture high-frequency, fine-grained anatomical details essential for accurate brain registration. To ensure the transformation is diffeomorphic, the stationary velocity field is integrated over time using the scaling and squaring (SS) method, producing a smooth, invertible deformation map.

The Cascaded Framework (Cas-DNVF)

While DNVF is powerful, optimizing an MLP from scratch for every image pair can be slow. The Cas-DNVF framework introduces a two-stage process:

  1. Initial Alignment: A pretrained Fully Convolutional Network (FCN) quickly predicts a coarse initial deformation.
  2. Refinement: DNVF then optimizes a residual velocity field to correct the fine-scale errors left by the FCN. This "residual" approach allows the system to focus its optimization effort on the most challenging local deformations.

Experimental Results

The proposed methods were evaluated on three major 3D MRI brain scan datasets: OASIS, Mindboggle101, and CANDI. These datasets provide a rigorous test of the model's ability to handle diverse anatomical structures and age ranges.

Accuracy (DSC) vs. running time on Mindboggle and OASIS datasets, showing Cas-DNVF (red) achieving superior performance.

Performance and Accuracy

Cas-DNVF demonstrated significant improvements over state-of-the-art methods. On the OASIS dataset, it achieved a Dice Similarity Coefficient (DSC) of 0.815, notably higher than previous learning-based and traditional benchmarks. Furthermore, the method maintains excellent topology preservation, with a near-zero rate of folding (negative Jacobian determinants), ensuring physically plausible results.

The Speed-Accuracy Trade-off

As shown in the performance plots, Cas-DNVF (labeled as C2F-DNVF in the experimental figures) occupies a superior position on the Pareto front. While it is slower than pure learning-based methods (which take milliseconds), it reaches peak accuracy in approximately 48–72 seconds—a fraction of the time required by traditional high-accuracy optimization methods.

Conclusion and Future Directions

DNVF and Cas-DNVF represent a significant step forward in diffeomorphic image registration. By leveraging the continuous nature of neural fields and the efficiency of cascaded refinement, the authors achieve a new state-of-the-art in accuracy while maintaining reasonable computational costs. Future work may explore end-to-end training of the cascaded components and the extension to non-stationary velocity fields to capture even more complex, fluid-like anatomical changes.

Summary updated

Abstract

Diffeomorphic image registration, offering smooth transformation and topology preservation, is required in many medical image analysis tasks.Traditional methods impose certain modeling constraints on the space of admissible transformations and use optimization to find the optimal transformation between two images. Specifying the right space of admissible transformations is challenging: the registration quality can be poor if the space is too restrictive, while the optimization can be hard to solve if the space is too general. Recent learning-based methods, utilizing deep neural networks to learn the transformation directly, achieve fast inference, but face challenges in accuracy due to the difficulties in capturing the small local deformations and generalization ability. Here we propose a new optimization-based method named DNVF (Diffeomorphic Image Registration with Neural Velocity Field) which utilizes deep neural network to model the space of admissible transformations. A multilayer perceptron (MLP) with sinusoidal activation function is used to represent the continuous velocity field and assigns a velocity vector to every point in space, providing the flexibility of modeling complex deformations as well as the convenience of optimization. Moreover, we propose a cascaded image registration framework (Cas-DNVF) by combining the benefits of both optimization and learning based methods, where a fully convolutional neural network (FCN) is trained to predict the initial deformation, followed by DNVF for further refinement. Experiments on two large-scale 3D MR brain scan datasets demonstrate that our proposed methods significantly outperform the state-of-the-art registration methods.
BibTeX
@inproceedings{han2023diffeomorphic,
  author = {Kun Han and Shanlin Sun and Hao Tang and Deying Kong and Xiangyi Yan and Xiaohui Xie and Ma, Haoyu and Kong, Deying and Xie, Xiaohui},
  title = {{Diffeomorphic Image Registration with Neural Velocity Field}},
  booktitle = {IEEE/CVF Winter Conference on Applications of Computer Vision},
  year = {2023},
  doi = {10.1109/wacv56688.2023.00191},
}
Figure