X-Ray to DRR Images Translation for Efficient Multiple Objects Similarity Measures in Deformable Model 3D/2D...

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• 3D/2D registration accuracy relies on image correspondences between digitally reconstructed radiographs (DRR) and X-ray images.
• Differences between DRR and real X-rays, like superimposed tissues, hinder accurate registration.
• Existing methods struggle with registration in complex scenes with multiple objects.

Purpose of the Study:

• To enhance the robustness and accuracy of intensity-based 3D/2D registration.
• To overcome challenges posed by differences between simulated DRR and actual X-ray images.
• To improve the registration of 3D vertebra models to biplanar spinal radiographs.

Main Methods:

• Implemented a Generative Adversarial Network (GAN)-based cross-modality image-to-image translation to convert X-ray images into DRR images (XRAY-to-DRR).
• Utilized standard similarity measures on translated images, enabling efficient registration with simple DRR projections.
• Applied the method to 3D/2D fine registration of deformable vertebra models to biplanar spine radiographs.

Main Results:

• XRAY-to-DRR translation significantly enhances 3D/2D registration results.
• The method increases the capture range of the registration process.
• Registration performance shows decreased dependence on the choice of similarity measure, making it mono-modal.

Conclusions:

• GAN-based XRAY-to-DRR translation is an effective prior step for improving 3D/2D registration.
• This approach standardizes image domains, facilitating more accurate matching of 3D models to X-ray data.
• The technique offers a more robust and versatile solution for medical image registration tasks.