Deep Learning in Medical Image Registration: A Review (1912.12318v1)

Abstract: This paper presents a review of deep learning (DL) based medical image registration methods. We summarized the latest developments and applications of DL-based registration methods in the medical field. These methods were classified into seven categories according to their methods, functions and popularity. A detailed review of each category was presented, highlighting important contributions and identifying specific challenges. A short assessment was presented following the detailed review of each category to summarize its achievements and future potentials. We provided a comprehensive comparison among DL-based methods for lung and brain deformable registration using benchmark datasets. Lastly, we analyzed the statistics of all the cited works from various aspects, revealing the popularity and future trend of development in medical image registration using deep learning.

• The paper provides a comprehensive categorization of seven DL-based methods, detailing key contributions and limitations.
• The paper demonstrates that techniques like supervised and unsupervised transformation prediction and GAN approaches offer promising real-time registration capabilities.
• The paper identifies future research directions focused on combined frameworks and enhanced validation to improve clinical application outcomes.

Deep Learning in Medical Image Registration: A Review

This review paper explores the advancements in deep learning (DL) methods applied to medical image registration, detailing the categorization of various approaches, highlighting critical contributions, and presenting challenges associated with these techniques. The review categorizes DL-based registration methods into seven distinct categories and provides an assessment of each, alongside comparative analyses using benchmark datasets.

Overview

Medical image registration aims to align images to ensure coherence across multiple imaging sessions or modalities. The complex nature of anatomical structures and variations between patients introduces several challenges in registration tasks. This paper systematically categorizes DL-based methods, offering clarity on the state of these techniques and potential future directions.

Categories of DL-Based Methods

1. Deep Similarity-Based Methods: These methods involve learning similarity metrics through deep networks, offering an alternative to traditional metrics like SSD or MI. However, the paper notes computational intensity remains a hurdle, along with difficulties in acquiring well-aligned image pairs for training.
2. Reinforcement Learning in Registration: RL methods decompose the problem into sequences of classifications. Challenges persist in handling high-dimensional transformation models, and there is decreased interest due to limited representational capabilities compared to direct prediction methods.
3. Supervised Transformation Prediction: Direct transformation prediction via supervised learning shows significant progress, yet generating realistic ground truth transformations remains challenging. Future growth is expected as accuracy improves to match traditional methods.
4. Unsupervised Transformation Prediction: Overcoming the lack of ground truth data, unsupervised methods have emerged as promising alternatives. The integration of spatial transformer networks (STNs) provides differentiability for unsupervised loss calculations, enhancing correspondence maps.
5. GAN in Registration: GANs offer innovative regularization and domain translation capabilities. These techniques play a dual role, ensuring plausible transformations and reducing modality discrepancies, although accuracy evaluations remain ongoing.
6. Registration Validation Using DL: Advances have been made in registration accuracy assessment through DL, predominantly treating error prediction as a supervised regression task. Future exploration may focus on generalized validation frameworks.
7. Other Learning-Based Methods: Including architectures beyond CNNs, this category spans various DL models applied to unique registration challenges, showcasing exploratory aspects of the field.

Implications and Future Directions

The review highlights the rapid evolution and growing sophistication of DL methods in medical image registration. Despite DL methods not consistently surpassing traditional methods, their speed and potential for real-time applications are notable. Practical challenges persist, including limited ground truth data and defining optimal regularization strategies. Future research is likely to intensify around combined supervised-unsupervised frameworks, further GAN exploration, and enhanced model generality.

Conclusion

The application of DL to medical image registration is an emerging field with significant potential. The paper provides a comprehensive exploration of current methods and sets the stage for ongoing research to bridge gaps between theoretical advances and practical medical implementations. As the research community continues to innovate, it is expected that DL methods will progressively become integral to medical image registration workflows.