Deep Learning in Medical Image Registration: A Survey (1903.02026v2)

Abstract: The establishment of image correspondence through robust image registration is critical to many clinical tasks such as image fusion, organ atlas creation, and tumor growth monitoring, and is a very challenging problem. Since the beginning of the recent deep learning renaissance, the medical imaging research community has developed deep learning based approaches and achieved the state-of-the-art in many applications, including image registration. The rapid adoption of deep learning for image registration applications over the past few years necessitates a comprehensive summary and outlook, which is the main scope of this survey. This requires placing a focus on the different research areas as well as highlighting challenges that practitioners face. This survey, therefore, outlines the evolution of deep learning based medical image registration in the context of both research challenges and relevant innovations in the past few years. Further, this survey highlights future research directions to show how this field may be possibly moved forward to the next level.

Deep Learning in Medical Image Registration: A Survey

The paper delineates advancements in the domain of medical image registration through the application of deep learning techniques. As image registration is pivotal in numerous clinical tasks, this survey underscores the transformative impact of deep learning methods that have led to state-of-the-art developments in this field. The authors provide a comprehensive overview of the evolution and innovations in deep learning-based registration, critically evaluating challenges and identifying future research trajectories.

Key Contributions

The paper categorizes deep learning approaches to image registration into distinct methodologies:

1. Deep Iterative Registration: Initially, deep learning was employed to augment traditional iterative methods for both unimodal and multimodal registration tasks. Techniques such as deep similarity-based registration and reinforcement learning have been leveraged to improve the accuracy and robustness of registration, particularly in complex multimodal scenarios.
2. Supervised Transformation Estimation: Methods in this category have been developed to predict transformations in a single step using fully supervised and weakly/dual supervised approaches. While fully supervised methods show promise, generating reliable ground truth is a significant challenge. In contrast, dual/weakly supervised techniques mitigate this by using segmentation-based or dual loss functions to reduce dependency on ground truth data.
3. Unsupervised Transformation Estimation: Such methods have gained considerable attention due to their ability to circumvent the need for annotated data. These approaches often involve similarity metrics and feature-based registration, which offer robust performance in unimodal tasks. Recent advancements in GANs and spatial transformer networks have begun to address multimodal registration challenges.

Results and Implications

The survey highlights strong numerical results evidenced by deep learning techniques surpassing traditional methods across various datasets and applications. One key implication of these methodologies is the potential reduction in computational time and enhancement in registration accuracy, which is critical for real-time clinical applications. Moreover, unsupervised and weakly supervised approaches present substantial benefits in terms of scalability and applicability, especially when large annotated datasets are not available.

Future Directions

The paper points to several future research directions in deep medical image registration:

• Adversarial Frameworks: The potential of GANs to generate realistic deformations and enhance image similarity metrics is anticipated to play an influential role. GANs can also aid in multimodal scenarios by transforming datasets across domains.
• Reinforcement Learning: The promising outcomes of reinforcement learning for registration tasks suggest further exploration, particularly in enhancing deformable registration models.
• Integration of Reconstruction and Registration: Emerging techniques in deep learning for image reconstruction from raw data could evolve to form end-to-end pipelines integrating reconstruction and registration, improving workflow efficiency in clinical settings.

Conclusion

The paper provides a detailed survey that captures the current landscape and future prospects of deep learning in medical image registration. Despite substantial progress, the field continues to face challenges related to data annotation and multimodal similarity quantification. Nevertheless, the innovative strategies and methodologies discussed pave the way for continued advancements, with the potential to significantly influence clinical practices through improved medical image analysis. The paper stands as a critical resource for researchers aiming to contribute to or understand the evolving field of deep learning-based medical image registration.