Non-coplanar CBCT image reconstruction using a generative adversarial network for non-coplanar radiotherapy.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Artificial Intelligence in Medicine

Background:

• Non-coplanar radiotherapy offers potential advantages in dose distribution and sparing of organs at risk.
• Limited-angle cone-beam computed tomography (CBCT) is crucial for image-guided radiotherapy but faces reconstruction challenges.
• Accurate 3D patient positioning verification is essential for effective radiotherapy delivery.

Purpose of the Study:

• To develop and evaluate a non-coplanar CBCT image reconstruction method using limited-angle projections.
• To enable intra-treatment 3D position verification for non-coplanar radiotherapy.
• To leverage generative adversarial networks (GANs) for improved CBCT reconstruction.

Main Methods:

• A dual-branch encoder-decoder generative adversarial network (GAN) was designed for non-coplanar CBCT reconstruction.
• The GAN utilized coplanar CBCT and non-coplanar projections as input for image generation.
• A patch-based discriminator and a novel joint loss function were employed to enhance reconstruction accuracy and global consistency.

Main Results:

• The proposed GAN method successfully reconstructed non-coplanar CBCT images from limited-angle projections.
• Reconstruction accuracy was evaluated using root mean square error (RMSE) and registration error (ε).
• Models trained with ±45° couch angles demonstrated superior performance compared to ±90° angles in both patient and phantom data.

Conclusions:

• The developed non-coplanar CBCT reconstruction method shows promise for intra-treatment 3D position verification.
• Accurate image reconstruction is vital for enhancing the precision of non-coplanar radiotherapy delivery.
• Further research may explore optimizing GAN architectures and loss functions for diverse non-coplanar radiotherapy scenarios.