A review of deep learning-based three-dimensional medical image registration methods

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiotherapy

Background:

• Medical image registration is crucial for accurate dose delivery in image-guided radiotherapy (IGRT).
• Deep learning (DL) shows significant promise for advancing 3D medical image registration techniques.
• Recent years have seen a surge in DL applications for 3D medical image registration, yielding promising results.

Purpose of the Study:

• To review and summarize the progress of DL-based 3D medical image registration over the last five years.
• To identify current challenges and explore potential future research directions in the field.
• To statistically analyze collected studies based on region of interest, image modality, supervision, and evaluation metrics.

Main Methods:

• A comprehensive review of DL-based 3D medical image registration studies published in the past five years.
• Classification of studies into three categories: deep iterative, supervised, and unsupervised registration.
• Statistical analysis of study characteristics including region of interest, image modality, supervision method, and evaluation metrics.

Main Results:

• Studies were categorized and reviewed, highlighting unique contributions, advantages, challenges, and trends within each.
• Deep learning approaches have demonstrated significant potential in improving 3D medical image registration accuracy.
• Analysis revealed key areas for statistical examination, including ROI, modality, supervision, and metrics.

Conclusions:

• Deep learning-based 3D medical image registration is a rapidly advancing field with substantial potential.
• Continued research is needed to address common challenges and explore novel avenues for further improvement.
• Future work should focus on overcoming existing limitations and leveraging DL for enhanced medical imaging applications.