Large Scale Unsupervised Brain MRI Image Registration Solution for Learn2Reg 2024

Yuxi Zhang, Xiang Chen, Jiazheng Wang, Min Liu, Yaonan Wang, Dongdong Liu, Renjiu Hu, Hang ZhangarXiv 2024 · OpenReview

Abstract

In this paper, we summarize the methods and experimental results we proposed for Task 2 in the Learn2Reg 2024 Challenge. This task focuses on unsupervised registration of anatomical structures in brain MRI images between different patients. The difficulty lies in: (1) the absence of segmentation labels, and (2) the large volume of data. To address these challenges, we built an efficient backbone network and explored several schemes to further enhance registration accuracy. Under the guidance of the NCC loss function and smoothness regularization loss function, we obtained a smooth and reasonable deformation field. According to the leaderboard, our method achieved a Dice coefficient of 77.34%, which is 1.4% higher than TransMorph. Overall, we won second place on the leaderboard for Task 2.

Introduction

Medical image registration—the process of aligning multiple images into a single coordinate system—is a cornerstone of neuroimaging, essential for surgical planning and longitudinal disease tracking. While deep learning has accelerated this process, two major hurdles remain: label scarcity and scalability. Most high-performing models require expensive "ground truth" segmentation labels, and handling large-scale datasets (thousands of 3D volumes) requires computationally efficient architectures.

This work addresses Task 2 (LUMIR) of the Learn2Reg 2024 challenge, which demands unsupervised inter-subject registration on the OpenBHB dataset. This dataset comprises 3,374 unlabeled 3D T1-weighted brain MRI volumes for training, with images standardized to 160 × 224 × 192 pixels.

Methods

The proposed framework utilizes a robust encoder-decoder backbone enhanced by four key architectural innovations designed for large-scale registration:

  1. Co-Attention Mechanism (CA): Integrated into the encoder to improve the simultaneous feature extraction from "moving" and "fixed" images.
  2. Large Kernel Convolution (LK): Expands the receptive field to capture broader contextual information without the overhead of Transformers.
  3. Large Convolution Channels (LC): Increases channel width during deformation field prediction for more detailed displacement maps.
  4. Bilateral Filtering (BF): A post-processing step to ensure the deformation field remains sharp at boundaries while smooth elsewhere.
Overview of the proposed registration architecture featuring Large Kernel (LK), Co-attention (CA), and Large Channel (LC) modules.

The model is trained using an unsupervised loss function combining Normalized Local Correlation (NCC) for intensity robustness and a smoothness regularization term to ensure physically plausible deformations.

Results and Discussion

The proposed method, ours(next-gen-nn), achieved a Dice coefficient of 77.34% on the leaderboard, outperforming the TransMorph baseline by 1.4%. Precision was further evidenced by a reduced 95% Hausdorff Distance (HD95) of 3.3415 mm. Crucially, the Non-Diffeomorphic Volumes (NDV) score was maintained at 0.0001%, indicating that the deformations were almost entirely free of folding or tissue overlap.

Training loss and validation Dice coefficient curves over 200 epochs, showing stable convergence and performance improvement.

Experimental insights suggest that in large-scale scenarios, backbone efficiency is paramount. While complex schemes often show promise in small-scale validation, a "next-gen" convolutional approach proved most effective for the challenge's massive dataset.

Conclusion

The developed unsupervised registration framework secured second place in the Learn2Reg 2024 Task 2. Future work will address dataset imbalances and the "hard case" problem, where specific image pairs currently yield significantly lower Dice scores compared to the average. Focusing on these outliers will be key to further improving the robustness of large-scale medical image registration.

Summary updated

Abstract

In this paper, we summarize the methods and experimental results we proposed for Task 2 in the learn2reg 2024 Challenge. This task focuses on unsupervised registration of anatomical structures in brain MRI images between different patients. The difficulty lies in: (1) without segmentation labels, and (2) a large amount of data. To address these challenges, we built an efficient backbone network and explored several schemes to further enhance registration accuracy. Under the guidance of the NCC loss function and smoothness regularization loss function, we obtained a smooth and reasonable deformation field. According to the leaderboard, our method achieved a Dice coefficient of 77.34%, which is 1.4% higher than the TransMorph. Overall, we won second place on the leaderboard for Task 2.
BibTeX
@misc{zhang2024large,
  author = {Yuxi Zhang and Xiang Chen and Jiazheng Wang and Min Liu and Yaonan Wang and Dongdong Liu and Renjiu Hu and Hang Zhang},
  title = {{Large Scale Unsupervised Brain MRI Image Registration Solution for Learn2Reg 2024}},
  year = {2024},
  eprint = {2409.00917},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2409.00917},
}
Figure