FaCT-GS: Fast and Scalable CT Reconstruction with Gaussian Splatting

Pawel Tomasz Pieta; Rasmus Juul Pedersen; Sina Borgi; Jakob Sauer Jørgensen; Jens Wenzel Andreasen; Vedrana Andersen Dahl

cs.CVApr 2, 2026

FaCT-GS: Fast and Scalable CT Reconstruction with Gaussian Splatting

Authors: Pawel Tomasz Pieta, Rasmus Juul Pedersen, Sina Borgi, Jakob Sauer Jørgensen, Jens Wenzel Andreasen, Vedrana Andersen Dahl

Organizations: Technical University of Denmark, Kgs. Lyngby, Denmark

Abstract

Gaussian Splatting (GS) has emerged as a dominating technique for image rendering and has quickly been adapted for the X-ray Computed Tomography (CT) reconstruction task. However, despite its growing popularity, the benefits of GS are typically not substantial enough to motivate a transition from well-established reconstruction algorithms. This paper addresses the most significant remaining limitations of the GS-based approach by introducing FaCT-GS, a framework for fast and flexible CT reconstruction. Enabled by an in-depth optimization of the voxelization and rasterization pipelines, our new method is significantly faster than its predecessors and scales well with projection and output volume size. Furthermore, the improved voxelization enables rapid fitting of Gaussians to pre-existing volumes, which can serve as a prior for warm-starting the reconstruction, or simply as an alternative, compressed representation. FaCT-GS is over 4X faster than the State of the Art GS CT reconstruction on standard 512x512 projections, and over 13X faster on 2k projections. Implementation and data available through: https://github.com/PaPieta/fact-gs.

Figures & tables

Figure 1: Reconstruction results of select baselines and FaCT-GS given a fixed time for a certain problem size. FaCT-GS provides superior CT reconstruction capabilities and scales exceptionally well with increasing projection size.

Figure 2: FaCT-GS optimization pipeline. We propose two initialization schemes: one, based on the gradient of the FDK reconstruction from collected projections, and one based on the Gaussians fitted directly to a prior volume. We accelerate the core components of the pipeline: rasterization, voxelization, and compute-intensive loss functions.

Panel of Figure 3 — Figure 3: Effective acceleration of the rasterization and voxelization components enabled through CUDA kernel optimization. Outlined relationships follow a power law. Plotted in log scale, baselines sourced from $R^{2}$ -Gaussian [ 77 ] .

	75-view			50-view			25-view
	PSNR	SSIM	Time	PSNR	SSIM	Time	PSNR	SSIM	Time
FDK	26.76	0.388	-	24.42	0.296	-	20.69	0.181	-
SIRT	33.96	0.844	2m13s	32.54	0.815	1m37s	30.11	0.764	1m7s
FISTA	37.07	0.934	55s	35.85	0.923	46s	33.51	0.898	39s
IntraTomo	29.67	0.829	29m24s	29.24	0.821	19m40s	28.71	0.808	9m51s
NAF	35.76	0.903	33m21s	34.41	0.891	22m14s	31.61	0.849	11m8s

Table 1: Quantitative results on the

512^{2}

-projection dataset from [ 77 ] . Colors signify best , second-best , and third-best numbers. FDK reconstruction was instantaneous. Qualitative comparison of the reconstructions can be found in Supplementary Material .

Panel of Figure 4 — Figure 4: Impact of increasing volume and projection size on the performance of chosen methods. For larger data sizes, FaCT-GS performs exceptionally well, providing faster and better reconstruction than both FISTA and R 2 -Gaussian at a fraction of the execution time. R 2 -Gaussian iteration speed scales better with data size, but FISTA reaches convergence quicker.

Figure 5: Comparison of the output of reconstruction initialization strategies on four examples with a similar prior volume available. Gradient-based sampling better outlines the high-frequency features than intensity sampling. FaCT-GS rapid volume fitting demonstrates a similar impact as using the prior directly – a standard approach in algebraic methods.

	Reverted pipeline components					Reverted rasterizer components
Ablation scenario	None	Fused TV	Fused SSIM	Accel. Vox.	Accel. Rast.	AA BBox	Per-Gauss. backward	Others
[I]	161 s	+10 s	+34 s	+54 s	+359 s	+16 s	+339 s	+4 s
[II]	217 s	+6 s	+187 s	+37 s	+3568 s	+155 s	+3556 s	+57 s

Table 3: Impact of the most significant pipeline accelerations on execution time under two scenarios: [I] – the 75-view instance of the main experiment ( Table 1 ) on an RTX 4070 GPU, and [II] – the 50-view 2k Coral experiment ( Fig. 1 ) on an LS40 GPU.

Figure 7: Example scan pair used in the hyperresolution warm-start study, and a 3D volume visualization. Two scans of the same patient form a pair, where the high-resolution scan serves as the reconstruction target and the low-resolution scan provides the volumetric prior.

	Cold start		Warm start		$\tau_{\mathrm{iter}}$ [%]
Index	PSNR	SSIM	PSNR	SSIM	PSNR	SSIM
1	34.74	0.924	35.03	0.927	54.0	58.0
2	34.94	0.916	35.19	0.920	63.0	56.0
3	34.10	0.908	34.36	0.912	59.5	47.5
4	34.99	0.921	35.18	0.922	68.0	77.0
5	33.80	0.902	33.96	0.904	69.0	67.5

Table 4: Quantitative impact of volume prior warm start on the reconstruction quality. Warm-started reconstructions achieve slightly higher scores and reach comparable performance in significantly fewer iterations.

Panel of Figure 8 — Figure 8: Gaussian number ablation study for varying type and resolution of reconstructed data. Complex, noisy, or large volumes can require significantly more Gaussians to reach maximum possible reconstruction quality. Increasing Gaussians beyond a certain point tends to yield diminishing returns.

Panel of Figure 9 — Figure 9: Example volumes from the R 2 -Gaussian dataset that required the least and most Gaussians to reach the 99% reconstruction quality threshold. Noise, structural complexity, and varying density are the primary factors leading to higher Gaussian count requirements (assuming the same volume size).

Figure 10: Qualitative comparison of reconstruction quality on select volumes from the 25-view setting of the R 2 -Gaussian [ 78 ] dataset.

Figure 11: Qualitative comparison of reconstruction quality on select volumes from the 50-view setting of the R 2 -Gaussian [ 78 ] dataset.

Figure 12: Qualitative comparison of reconstruction quality on select volumes from the 75-view setting of the R 2 -Gaussian [ 78 ] dataset.

Computed Tomography

Computed tomography uses X-ray measurements from multiple angles to reconstruct cross-sectional and volumetric images for diagnosis and scientific imaging.

See all 272 papers

CTBakeGS: training-free baking of photorealistic CT volume renderings into 3D Gaussian splats
Oct 27, 2026Haruki Fukuda, Constantin Lux, Wolf Schweitzer +23D Gaussian Splatting 3D Medical Imaging
Department of Forensic Medicine, Gunma University Graduate School of Medicine, Maebashi, Japan · Institute of Forensic Medicine, Universität Zürich, Zürich, Switzerland · Biometrics/Forensic Imaging, Zurich Forensic Science Institute, Zurich, Switzerland
Render-FM: Feedforward Model for Real-time Photorealistic Volumetric Rendering
May 22, 2025Zhongpai Gao, Benjamin Planche, Meng Zheng +4Neural Rendering 3D Medical Imaging
United Imaging Intelligence, Boston, MA, USA
TR-GS: High-Fidelity Sparse-View CT Volumetric Rendering via t-Distribution Gaussian Splatting and Ray-Confidence Modeling
Aug 17, 2026Zedong Xiao, Yiren Wang, Zhou Liu +23D Medical Imaging Sparse-View CT Reconstruction
Shenzhen University Shenzhen, China · Guangdong Laboratory of Artificial Intelligence and Digital Economy (SZ) Shenzhen, China

Oct 27, 2026J/K move · Enter open · S save

CTBakeGS: training-free baking of photorealistic CT volume renderings into 3D Gaussian splats

Haruki Fukuda, Constantin Lux, Wolf Schweitzer, Tadashi Hosoya+1

Department of Forensic Medicine, Gunma University Graduate School of Medicine, Maebashi, Japan · Institute of Forensic Medicine, Universität Zürich, Zürich, Switzerland · Biometrics/Forensic Imaging, Zurich Forensic Science Institute, Zurich, Switzerland

3D Gaussian Splatting 3D Medical Imaging Computed Tomography

Open paper AskPDF arXiv

Photorealistic volume rendering of computed tomography (CT) may convey anatomy intuitively, but it needs the CT data, dedicated software and a powerful graphics processor. Three-dimensional Gaussian splatting (3DGS) displays photorealistic scenes in real time on less powerful devices, but existing medical approaches train each scene for minutes to hours. We aimed to convert a CT volume rendering directly into a 3DGS model without training. Splats were placed on an adaptive octree of millimetre-sized cells. Their colour and opacity were computed on the graphics processor with the renderer's lighting, cell opacity was converted analytically into splat opacity, and faint splats were retained stochastically above the 1/255 opacity floor of standard 3DGS rasterisers. Six public CT scans (1.5-mm voxels) were converted with three transfer functions and compared with converged references by peak signal-to-noise ratio (PSNR), structural similarity and a learned perceptual metric. Two-million-splat models reached a PSNR of 32.6 +/- 0.9 dB for bone, 30.1 +/- 1.0 dB for contrast-enhanced vessels and 32.3 +/- 1.5 dB for skin and bone. A conversion pass (usually two or three per model) took 4-30 s on a current laptop and 9-60 s on a desktop with a 4-GB graphics card. On the laptop, the models rendered at 68 frames per second at 800 x 800 pixels. Training-free conversion produced real-time 3DGS models of photorealistic CT renderings in about a minute or less per conversion pass on ordinary computers, with fidelity that we consider sufficient for interactive viewing.

Abstract

Computed Tomography

Explore similar work

CTBakeGS: training-free baking of photorealistic CT volume renderings into 3D Gaussian splats

Render-FM: Feedforward Model for Real-time Photorealistic Volumetric Rendering

TR-GS: High-Fidelity Sparse-View CT Volumetric Rendering via t-Distribution Gaussian Splatting and Ray-Confidence Modeling