Pairwise domain adaptation module for CNN-based 2-D/3-D registration.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate 2-D/3-D registration is crucial for image-guided therapy, enabling precise alignment of preoperative 3-D data with intraoperative 2-D X-ray images.
• Deep learning, particularly Convolutional Neural Networks (CNNs), has advanced 2-D/3-D registration accuracy and efficiency.
• Training deep learning models often requires large annotated clinical datasets, which are difficult to obtain, leading to reliance on synthetic data and a subsequent performance gap on real data.

Purpose of the Study:

• To develop a flexible and generalizable method for adapting deep learning models trained on synthetic data to perform accurately on clinical 2-D/3-D medical image registration.
• To address the challenge of limited annotated clinical data by proposing a pairwise domain adaptation (PDA) module.

Main Methods:

• Proposed a pairwise domain adaptation (PDA) module designed to learn domain-invariant features.
• The PDA module can be integrated into existing deep learning frameworks and applied to pre-trained CNN models.
• Employed a strategy requiring only a small amount of paired real and synthetic data for adaptation.

Main Results:

• Demonstrated significant improvements in generalizability and flexibility for 2-D/3-D medical image registration.
• Quantitative evaluations on two clinical applications using different deep network frameworks confirmed the module's effectiveness.
• Successfully adapted models trained on synthetic data to achieve better performance on clinical data.

Conclusions:

• The proposed pairwise domain adaptation module effectively bridges the gap between synthetic and clinical data for deep learning-based 2-D/3-D registration.
• The module offers a flexible and generalizable solution applicable to various deep learning frameworks and medical imaging scenarios.
• This approach facilitates the use of deep learning in image-guided therapy where clinical data is scarce.