Structurally-Informed 3D Gaussian Splatting for Limited-Angle CBCT.

Area of Science:

• Medical Imaging
• Computational Imaging
• Image Reconstruction

Background:

• Limited-angle cone-beam computed tomography (LA-CBCT) offers faster imaging and lower radiation doses but suffers from artifacts due to incomplete data.
• Current 3D Gaussian Splatting (3D-GS) methods for tomographic reconstruction are sensitive to initialization, impacting their effectiveness.

Purpose of the Study:

• To introduce SPARK (Structurally-Informed Projection-Accelerated Reconstruction), a two-stage framework designed to improve LA-CBCT reconstruction quality and speed.
• To leverage structurally informed, generative initialization to enhance 3D-GS performance in low-data scenarios.

Main Methods:

• SPARK employs a two-stage approach: first, a geometry-conditioned network generates 3D Gaussian parameters from sparse projections using anatomical priors.
• Second, physics-based 3D-GS optimization refines the generated scene for high-fidelity reconstruction.

Main Results:

• SPARK achieved superior Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity Index Measure (SSIM) in severely limited-angle scenarios compared to existing methods.
• The framework demonstrated significant improvements in both image quality and convergence speed on public datasets.
• SPARK reconstructions serve as enhanced inputs for downstream networks, further improving image fidelity.

Conclusions:

• SPARK presents a promising prior-informed 3D-GS framework for simulated LA-CBCT reconstruction with limited angular coverage.
• The method effectively bridges data-driven anatomical priors with physics-based refinement for improved tomographic imaging.