Dynamic cone beam CT reconstruction via spatiotemporal Gaussian neural representation

In image-guided radiotherapy (IGRT), four-dimensional cone-beam computed tomography (4D-CBCT) is critical for assessing tumor motion during a patient's breathing cycle prior to beam delivery. However, generating 4D-CBCT images with sufficient quality requires significantly more projection images than a standard 3D-CBCT scan, leading to extended scanning times and increased imaging dose to the patient.

To introduce a novel spatiotemporal Gaussian neural representation framework to reconstruct high-temporal dynamic CBCT images from 1-minute acquisition, preserving motion dynamics and fine spatial details without relying on prior images or motion models.

Our framework employs a differentiable 4D Gaussian representation initialized from average CBCT images. Gaussian points are characterized by position, covariance, rotation, and density, offering a compact and dynamic model for CBCT scenes. A Gaussian deformation network, incorporating a HexPlane encoder and multi-head decoder, predicts Gaussian deformations to minimize L1 and structural similarity index measure (SSIM) losses between rendered and measured projections. Adaptive Gaussian control refines the representation by pruning underutilized Gaussians and densifying points in high-gradient regions. The method was benchmarked on the AAPM SPARE challenge datasets and further validated with clinical CBCT scans from a Varian TrueBeam system. For the AAPM SPARE challenge datasets, the performance of the proposed method was evaluated using the root-mean-squared-error (RMSE) and the structural similarity index (SSIM) in the four regions of interest: Body, Lung, PTV, and Bone. The geometric accuracy was evaluated by calculating the registration error when aligning the tumor to the ground truth using the Elastix package, focusing on pixels within the planning target volume (PTV). To demonstrate our method's capability in high-temporal motion dynamic modeling using extremely undersampled projections, the clinical half-fan projections from a 1-minute Varian TrueBeam acquisition were sorted into 50 phases with approximately 18 projections per phase, significantly finer than the commonly used 10-phase binning.

Compared to the AAPM SPARE challenge participant methods, our method achieved superior geometric accuracy in terms of PTV alignment error, and comparable RMSE and SSIM when no prior 4DCT or motion model is used for our reconstruction. For PTV alignment, our method achieved translational and rotational errors of 0.54 mm (LR), 0.76 mm (SI), 1.36 mm (AP), 0.55° (rAP), and 0.93° (rSI), and 1.31° (rLR), respectively. For high temporal dynamic CBCT reconstruction, our method successfully reconstructed a 50-phase CBCT from a 1-minute Varian Truebeam half-fan scan, demonstrating effective streak artifact suppression, respiratory motion preservation, and fine detail restoration. Reconstruction on a single NVIDIA RTX A6000 GPU required approximately 30-80 min, depending on the number of Gaussian points used (ranging from 50 to 400K), to reconstruct CBCT from 680 projections acquired with a 30 × 40 cm detector. Our code and reconstruction results can be found at: https://github.com/fuyabo/4DGS_for_4DCBCT/tree/main.

The spatiotemporal Gaussian framework is a novel data-driven dynamic CBCT reconstruction technique that features excellent geometric accuracy in terms of PTV alignment and high-temporal motion modeling, indicating promise for tumor motion assessment and high-temporal respiratory motion modeling based on a 1-minute half-fan scan prior to beam delivery.