R²-Gaussian is a novel 3D Gaussian splatting (3DGS) based framework for sparse-view tomographic reconstruction. The paper identifies an integration bias in standard 3DGS that hampers accurate density retrieval in volumetric reconstruction. To address this, the authors propose R²-Gaussian, which introduces tailored Gaussian kernels, extends rasterization to X-ray imaging, and develops a CUDA-based differentiable voxelizer. The method achieves high-quality results in 4 minutes, which is 12× faster than NeRF-based methods and comparable to traditional algorithms. Experiments on synthetic and real-world datasets show that R²-Gaussian outperforms state-of-the-art methods in both accuracy and efficiency. The method rectifies the integration bias by adjusting the density scaling factor during projection rendering, enabling accurate 3D density retrieval. The framework also includes a differentiable voxelizer for efficient volume retrieval and a training pipeline that uses photometric losses and 3D total variation regularization. The results demonstrate that R²-Gaussian achieves superior performance in both reconstruction quality and training speed, making it a promising solution for real-world applications in medical imaging and industrial inspection.R²-Gaussian is a novel 3D Gaussian splatting (3DGS) based framework for sparse-view tomographic reconstruction. The paper identifies an integration bias in standard 3DGS that hampers accurate density retrieval in volumetric reconstruction. To address this, the authors propose R²-Gaussian, which introduces tailored Gaussian kernels, extends rasterization to X-ray imaging, and develops a CUDA-based differentiable voxelizer. The method achieves high-quality results in 4 minutes, which is 12× faster than NeRF-based methods and comparable to traditional algorithms. Experiments on synthetic and real-world datasets show that R²-Gaussian outperforms state-of-the-art methods in both accuracy and efficiency. The method rectifies the integration bias by adjusting the density scaling factor during projection rendering, enabling accurate 3D density retrieval. The framework also includes a differentiable voxelizer for efficient volume retrieval and a training pipeline that uses photometric losses and 3D total variation regularization. The results demonstrate that R²-Gaussian achieves superior performance in both reconstruction quality and training speed, making it a promising solution for real-world applications in medical imaging and industrial inspection.