Dynamic cone beam CT reconstruction via spatiotemporal Gaussian neural representation

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Computational Imaging

Background:

• Four-dimensional cone-beam computed tomography (4D-CBCT) is crucial for assessing tumor motion in image-guided radiotherapy (IGRT).
• Current 4D-CBCT requires extensive projection data, leading to prolonged scan times and increased patient radiation dose.

Purpose of the Study:

• To present a novel spatiotemporal Gaussian neural representation for reconstructing high-temporal dynamic CBCT images.
• To achieve this from a 1-minute acquisition, preserving motion dynamics and spatial details without prior images or motion models.

Main Methods:

• A differentiable 4D Gaussian representation was employed, initialized from average CBCT images.
• A Gaussian deformation network predicted deformations to minimize rendering and projection losses (L1, SSIM).
• Adaptive Gaussian control refined the representation, and the method was validated on AAPM SPARE datasets and clinical scans.

Main Results:

• The framework demonstrated superior geometric accuracy (PTV alignment error) compared to AAPM SPARE participants.
• It achieved comparable RMSE and SSIM without relying on prior 4DCT or motion models.
• Successfully reconstructed a 50-phase CBCT from a 1-minute scan, showing artifact suppression, motion preservation, and detail restoration.

Conclusions:

• The spatiotemporal Gaussian framework offers a novel, data-driven dynamic CBCT reconstruction technique.
• It exhibits excellent geometric accuracy and high-temporal motion modeling capabilities.
• This shows promise for improved tumor motion assessment and respiratory motion modeling using brief scans.