GPU accelerated voxel-driven forward projection for iterative reconstruction of cone-beam CT

Area of Science:

• Medical Imaging
• Computational Science

Background:

• Cone-beam computed tomography (CBCT) is crucial in clinical settings.
• Iterative reconstruction algorithms offer superior image quality over FDK methods.
• Computational speed, particularly forward projection, limits iterative CBCT applications.

Purpose of the Study:

• To analyze the cone-beam forward projection problem using a voxel-driven model.
• To propose and detail a GPU-based acceleration method for CBCT forward projection.

Main Methods:

• Developed a GPU-based acceleration strategy for voxel-driven CBCT forward projection.
• Detailed the method rationale and implementation workflow.
• Conducted computational simulations for validation and performance evaluation.

Main Results:

• The proposed GPU method effectively addresses inter-thread interference.
• Achieved an acceleration ratio exceeding 100x compared to single-threaded CPU implementation.
• Demonstrated significant computational speed-up for CBCT forward projection.

Conclusions:

• The voxel-driven forward projection calculation for CBCT is highly parallelized by the proposed method.
• This GPU-based approach is a critical module for developing advanced iterative reconstruction and correction methods in CBCT imaging.