Real-Time Deep Pose Estimation With Geodesic Loss for Image-to-Template Rigid Registration

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• 3-D rigid registration of medical images is crucial for inter-subject and subject-to-template analysis.
• Current methods face limitations in capture range and performance speed.
• Accurate 3-D pose estimation is essential for aligning medical scans.

Purpose of the Study:

• To develop and evaluate deep learning-based methods for accelerated 3-D rigid registration.
• To improve the capture range and real-time performance of medical image registration.
• To enable accurate 3-D pose estimation of arbitrarily oriented anatomy in a canonical space.

Main Methods:

• Proposed regression convolutional neural networks (CNNs) to predict 3-D rotations and translations.
• Utilized angle-axis representation for pose estimation.
• Compared mean square error and geodesic loss for training CNNs in slice-to-volume and volume-to-volume registration scenarios.
• Applied methods to fetal brain MRI scans for registration to a standard atlas space.

Main Results:

• Deep learning methods with geodesic loss achieved wide-capture range, real-time (<100ms) 3-D pose estimation.
• Trained CNNs demonstrated generalization capabilities on expanded age ranges and different MR image contrasts (T1-weighted newborn brains from T2-weighted fetal models).
• Conditional generative adversarial networks facilitated domain adaptation for cross-modality and cross-contrast image registration.

Conclusions:

• Deep learning regression CNNs offer a robust and efficient solution for 3-D medical image registration.
• Geodesic loss minimization enhances pose estimation accuracy and capture range.
• The proposed methods show potential for broad application in future automatic imaging and image processing systems, even across different imaging modalities and contrasts.