FiRework: Field Refinement Framework for Efficient Enhancement of Deformable Registration

Haiqiao Wang,Dong Ni,Yi Wang

2024-10-13

Abstract:Deformable image registration remains a fundamental task in clinical practice, yet solving registration problems involving complex deformations remains challenging. Current deep learning-based registration methods employ continuous deformation to model large deformations, which often suffer from accumulated registration errors and interpolation inaccuracies. Moreover, achieving satisfactory results with these frameworks typically requires a large number of cascade stages, demanding substantial computational resources. Therefore, we propose a novel approach, the field refinement framework (FiRework), tailored for unsupervised deformable registration, aiming to address these challenges. In FiRework, we redesign the continuous deformation framework to mitigate the aforementioned errors. Notably, our FiRework requires only one level of recursion during training and supports continuous inference, offering improved efficacy compared to continuous deformation frameworks. We conducted experiments on two brain MRI datasets, enhancing two existing deformable registration networks with FiRework. The experimental results demonstrate the superior performance of our proposed framework in deformable registration. The code is publicly available at <a class="link-external link-https" href="https://github.com/ZAX130/FiRework" rel="external noopener nofollow">this https URL</a>.

Computer Vision and Pattern Recognition

What problem does this paper attempt to address?

The paper aims to address the complex deformation issues in deformable image registration. Specifically, existing deep learning-based registration methods, while capable of handling larger deformations, often suffer from cumulative errors and inaccurate interpolation, leading to suboptimal results. Additionally, these methods typically require multiple cascading stages, which not only increase the demand for computational resources but also reduce training efficiency.

To tackle these problems, the authors propose a novel method—Field Refinement Framework (FiRework), specifically designed for unsupervised deformable registration. FiRework reduces the aforementioned errors by redesigning the continuous deformation framework and requires only one recursive level during training, supporting continuous inference, thereby improving the efficiency and accuracy of registration.

**Main Contributions**:
1. **Innovative Framework**: A new deformable registration framework is proposed, which avoids interpolation and cumulative errors through continuous deformation field refinement.
2. **Single-Level Recursion**: Only one recursive level is needed during training, while multiple iterations can be performed during inference to optimize the deformation field.
3. **Performance Improvement**: By combining two existing registration networks (VoxelMorph and TransMorph) with FiRework, experimental results show that this framework significantly enhances registration performance.

**Experimental Validation**:
The authors conducted experiments on two brain MRI datasets (LPBA and Mindboggle), and the results show that FiRework outperforms various existing advanced methods in terms of registration accuracy, surface distance, and deformation field quality. Notably, FiRework-VM and FiRework-TM excel in metrics such as Dice coefficient and mean symmetric surface distance.

In summary, by introducing the FiRework framework, this paper effectively addresses the cumulative error and inaccurate interpolation issues present in existing deformable registration methods, providing a more efficient and accurate image registration solution for clinical practice.