Abstract
Image registration plays an important role in medical image analysis. Conventional optimization based methods provide an accurate estimation due to the iterative process at the cost of expensive computation. Deep learning methods such as learn-to-map are much faster but either iterative or coarse-to-fine approach is required to improve accuracy for handling large motions. In this work, we proposed to learn a registration optimizer via a multi-scale neural ODE model. The inference consists of iterative gradient updates similar to a conventional gradient descent optimizer but in a much faster way, because the neural ODE learns from the training data to adapt the gradient efficiently at each iteration. Furthermore, we proposed to learn a modal-independent similarity metric to address image appearance variations across different image contrasts. We performed evaluations through extensive experiments in the context of multi-contrast 3D MR images from both public and private data sources and demonstrate the superior performance of our proposed methods.
Introduction
Medical image registration—the alignment of multiple images into a common coordinate system—is essential for clinical tasks such as longitudinal disease monitoring and surgical planning. Historically, this field has been divided between two paradigms. Conventional iterative methods, such as ANTs, offer high accuracy but are computationally intensive, often requiring minutes or hours per volume. Conversely, recent deep learning "learn-to-map" approaches, like VoxelMorph, provide near-instantaneous inference but frequently struggle with large anatomical deformations and multi-modal data where image contrasts differ significantly.
This paper introduces MS-ODENet, a framework designed to bridge this gap. By treating registration as a continuous optimization process governed by Neural Ordinary Differential Equations (ODEs), the model learns an efficient "registration optimizer" that combines the accuracy of iterative methods with the speed of deep learning.
Methodology
The MS-ODENet framework incorporates two primary innovations to enhance registration performance:
Multi-scale Neural ODEs
Instead of using a standard feed-forward network to predict a static displacement field, MS-ODENet models the registration as a continuous trajectory. The network learns to predict the derivative of the transformation parameters over time. To handle large anatomical motions, this ODE is solved across a multi-scale image pyramid. This coarse-to-fine strategy allows the model to resolve large-scale misalignments at lower resolutions before refining fine-grained details at higher resolutions.
Furthermore, by using the "adjoint method" to train the Neural ODE, the model avoids the need to store every intermediate step in memory. This makes it significantly more memory-efficient than "cascaded" deep learning models that simply stack multiple networks.
Modality-Independent Similarity Metric
Registering images across different modalities (e.g., T1-weighted to T2-weighted MRI) is challenging due to varying image appearances. The authors address this by using a Generative Adversarial Network (GAN) to disentangle images into modality-specific "style" and modality-independent "content" features. By calculating similarity based solely on these content features, the registration process becomes robust to contrast variations, effectively treating multi-modal registration as a mono-modal task.
Experiments and Results
The model was evaluated on the public BraTS 2020 dataset (494 subjects with four MRI contrasts) and a private volunteer dataset. The training was conducted in an unsupervised manner, utilizing a self-supervised loss where the model learns to reverse random perturbations applied to the data.
Performance and Speed
MS-ODENet demonstrated significant improvements over both classical and deep learning baselines:
- Accuracy: In rigid registration tasks on the BraTS dataset, MS-ODENet achieved a Dice score of 90.6%, substantially outperforming the industry-standard ANTs, which scored 63.0%.
- Efficiency: Inference for rigid registration was completed in just 0.55 seconds, compared to over 17 seconds for ANTs. Even for complex hybrid (rigid + deformable) tasks, MS-ODENet remained under 1 second, while ANTs exceeded 70 seconds.
Visual Comparison
Qualitative results highlight the model's ability to accurately align complex brain structures across different modalities, maintaining anatomical integrity where other methods might fail or produce artifacts.
Conclusion
MS-ODENet represents a significant step forward in medical image registration by framing the problem as a learned continuous optimization. The integration of Neural ODEs with a multi-scale architecture and modality-independent features allows for high-accuracy, high-speed registration across varying MRI contrasts. Future work will explore the extension of this framework to 3D-2D registration tasks, which are critical for real-time image-guided interventions.
Summary updated