Image-to-Graph Convolutional Network for 2D/3D Deformable Model Registration of Low-Contrast Organs

Area of Science:

• Medical imaging
• Artificial intelligence in medicine
• Computational anatomy

Background:

• Accurate organ shape reconstruction is crucial for image-guided radiotherapy and surgical guidance.
• Current methods face challenges with low-contrast 2D projection images and complex organ deformations.

Purpose of the Study:

• To develop an image-to-graph convolutional network for deformable registration of 3D organ meshes from single 2D projection images.
• To enable simultaneous training of 2D-to-3D transformations for accurate organ shape prediction.

Main Methods:

• An image-to-graph convolutional network was proposed for 2D/3D deformable registration.
• The framework simultaneously learns transformations from 2D images to displacement maps and from mesh features to 3D displacements.
• The method was validated on multiple abdominal organs (liver, stomach, duodenum, kidney) for pancreatic cancer treatment scenarios.

Main Results:

• The proposed network achieved accurate deformable registration of 3D organ meshes from low-contrast 2D projection images.
• Shape prediction considering inter-organ relationships effectively predicted respiratory motion and deformation.
• Results demonstrated clinically acceptable accuracy for digitally reconstructed radiographs.

Conclusions:

• The developed image-to-graph convolutional network offers a robust solution for 3D organ shape reconstruction from 2D images.
• This technology has significant potential for advancing image-guided radiotherapy and surgical interventions.
• Accurate prediction of organ motion and deformation can enhance treatment planning and delivery.