Unsupervised Cross-Modality Domain Adaptation Network for X-Ray to CT Registration

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Accurate 2D/3D registration is crucial for radiotherapy and image-guided surgeries.
• Convolutional Neural Networks (CNNs) show promise for improving 2D/3D registration accuracy and efficiency.
• Training CNNs requires large datasets, which are challenging to acquire for medical imaging, especially with accurate poses.

Purpose of the Study:

• To develop a method for adapting 2D/3D registration models trained on synthetic data to real X-ray images without requiring labeled target domain data.
• To improve the performance of 2D/3D registration in medical applications by bridging the domain gap between synthetic and real images.

Main Methods:

• Proposed an unsupervised cross-modality domain adaptation network (UCMDAN).
• Utilized adversarial learning to adapt a model trained on synthetic data (source domain) to X-ray images (target domain).
• Employed synergistic alignment in both pixel and feature spaces, including image appearance transformation and domain-invariant feature learning.

Main Results:

• The UCMDAN effectively narrowed the domain gap between synthetic and X-ray image data.
• Achieved superior performance compared to existing state-of-the-art domain adaptation methods.
• Demonstrated effectiveness on CT and CBCT datasets.

Conclusions:

• The proposed UCMDAN is a viable approach for unsupervised domain adaptation in 2D/3D medical image registration.
• This method can significantly improve registration accuracy when labeled clinical data is scarce.
• UCMDAN offers a promising solution for enhancing image-guided interventions.