Source-linked AI summary
Shape-guided Gaussian Splatting for Sparse-View X-ray 3D Reconstruction
Pranav Poudel, Florence Dell'Aniello Picard, Nairouz Shehata, Frédéric Lavoie, Herve Lombaert
TL;DR
Sparse-view X-ray reconstruction is ill-posed, while existing Gaussian-splatting methods lack anatomical constraints under very few projections. The paper guides Gaussian positions with a statistical shape model and regularizes density using population statistics, improving 5-view femoral reconstruction by 2.83 dB PSNR over a state-of-the-art baseline.
Problem
Sparse-view X-ray reconstruction must recover a density field from few projections, but existing methods lack anatomical priors and degrade under extreme sparsity.
Method
The framework constrains Gaussian positions with a population shape model and uses population density information for initialization and density regularization.
Results
2.83 dB PSNR and 0.022 SSIM improvements are reported at 5 views over competitive state-of-the-art methods.
Takeaways & Limitations
The framework produces anatomically consistent femoral reconstructions under extreme 5-view sparsity.
Abstract
from arXiv · showhide
Sparse-view X-ray 3D reconstruction is essential for reducing radiation exposure, but recovering a density field from a handful of X-ray projections is severely ill-posed. Recently, 3D Gaussian Splatting has achieved state-of-the-art performance in sparse-view reconstruction by representing the volume using explicit, optimized primitives, but it requires dozens of projected views. With fewer views, reconstruction quality degrades severely since the explicit primitives are optimized freely without any anatomical information. Anatomical structures, in contrast, share similar geometry and density across a population. Their variations are bounded within a limited range that statistical shape models can capture. This paper proposes a shape-guided Gaussian splatting framework for sparse-view X-ray 3D reconstructions. Our contribution lies in driving Gaussian positions toward anatomically valid configurations, alongside atlas-based density regularization. Our method ensures anatomically consistent reconstruction and improves PSNR by 2.83 dB over a state-of-the-art Gaussian splatting baseline with as few as 5 views. Code Available: https://github.com/polyshape-lab/ShapeGuidedGaussian
1 Introduction
Sparse-view X-ray volumetric reconstruction is ill-posed: existing methods degrade with few projections and lack anatomical constraints. The paper introduces shape-guided Gaussian splatting that uses population priors to constrain geometry and improve reconstruction under extreme sparsity.
- Motivation: Sparse-view X-ray reconstruction recovers a volumetric density field from few 2D projections, reducing radiation exposure but remaining severely ill-posed.Volumetric methods directly reconstruct X-ray attenuation, whereas surface methods recover boundaries and depend on contours or landmarks.
- Limitations of prior work: Existing NeRF- and Gaussian-splatting methods infer density primarily from projections, without anatomical priors during optimization.R2-Gaussian achieves state-of-the-art sparse-view performance at 25–75 views, but the cited methods lack anatomical constraints.
- Anatomical priors: Population anatomy exhibits bounded geometric variation and internal density regularities that statistical shape models can capture.Organs contain homogeneous regions with sharp boundary changes, motivating population-based geometric and density guidance.
- Proposed framework: The proposed framework guides Gaussian splatting with population shape and density information during initialization and optimization.Initialization seeds positions, orientations, and densities from population information; optimization deforms positions through a shape prior and regularizes density statistically.
- Contributions: The method binds Gaussian primitive positions to a shape model, constraining geometry to anatomically valid configurations.This is the framework’s stated primary contribution for reducing unconstrained primitive motion.
- Contributions: At 5 views, the method shows consistent improvements over the state-of-the-art baseline on femoral reconstruction.The experiments evaluate femoral data under extreme sparsity, including 5-view reconstruction.
2 Method
The method represents the volume with shape-guided and free Gaussian primitives, using a statistical shape model to deform anatomical positions while leaving appearance parameters adaptable. Population density statistics initialize primitives and provide confidence-weighted regularization during optimization.
- 2.1 Object Representations: The volume is partitioned into shape-guided Gaussians Γs and free Gaussians Γz, with only Γs positions driven by the shape model.Free primitives learn position, covariance, and density independently, while shape-guided primitives retain shape-constrained positions.
- 2.2 Shape Model: The shape model learns a low-dimensional geometry space from registered population segmentations using PCA over velocity fields.Shape coefficients generate a velocity field that is integrated into a displacement field for deforming shape Gaussian primitives.
- 2.3 Parameterization and Initialization: Shape Gaussian positions are obtained by deforming template reference centers with the displacement field generated by shape coefficients.The deformation binds primitive centers to anatomically plausible configurations rather than allowing independent position updates.
- 2.3 Parameterization and Initialization: Scaling, rotation, and density remain independently optimized, allowing primitive appearance and orientation to adapt to the projections.This separates anatomical positional constraints from appearance-related parameter adaptation.
- 2.3 Parameterization and Initialization: Initialization uses template-derived orientations, population-mean density, boundary-weighted sampling, and anisotropic surface primitives.Density values are set from the atlas mean at reference locations and adjusted for local overlap so overlapping Gaussians reproduce the mean density.
- 2.4 Optimization Strategy: Optimization combines photometric and total-variation losses with shape and density priors.The shape prior penalizes deviations in standardized shape-coefficient space, while the density prior pulls reconstructed volume toward population statistics.
- 2.4 Optimization Strategy: The density prior uses confidence weighting that is stronger where population density is consistent and weaker where it varies substantially.The prior is computed from shape Gaussian primitives queried on randomly sampled sub-volume grids.
3 Experimental Setup
Experiments use six held-out femur subjects reconstructed from synthetic datasets derived from 758 CT scans, with qualitative comparisons across coronal and sagittal axes.
- Dataset and evaluation: Six held-out subjects were used for independent per-subject reconstruction evaluations.The shape model was built from 100 subjects, while six were reserved for reconstruction.
- Dataset and evaluation: The dataset comprised 758 NMDID CT scans resampled to 1 mm isotropic spacing and aligned using femoral segmentations.TotalSegmentator generated segmentations, and ANTs performed rigid alignment to a common reference.
- Qualitative comparison: Qualitative comparisons were shown across coronal and sagittal axes against R2-Gaussian and ground truth.The comparison focused on cortical boundaries and internal structure.
- Optimization settings: The proposed method used empirically selected loss weights, learning rates, exponential decay schedules, and a 50-iteration shape-coefficient warm-up.The warm-up was used to stabilize appearance during optimization.
4 Results
Across six subjects, anatomical guidance consistently improved sparse-view reconstruction, with the largest gains at five views and sharper qualitative structure than R2-Gaussian.
- Quantitative results: +2.83 dB average 3D PSNR and +0.022 3D SSIM were achieved at 5 views over the baseline.The average PSNR increased from 34.55 to 37.38, while SSIM increased from 0.937 to 0.959.
- Quantitative results: +1.46 dB average 3D PSNR and +0.012 3D SSIM were achieved at 10 views over the baseline.The gains narrowed as additional projections supplied more constraints.
- Qualitative results: At 5 views, the proposed method recovered sharper cortical boundaries and finer internal structure than R2-Gaussian.The reconstructions more closely followed the ground-truth contour across coronal and sagittal slices.
- Ablation study: +4.33 dB resulted from adding the shape prior to density initialization alone in the 5-view ablation.Independent primitives could match the training-view projection while producing an inaccurate volume.
- Ablation study: Collective deformation from the shape prior resolved mottled, inconsistent interiors, while the density prior provided a smaller further refinement.The density prior refined reconstructed density after the shape prior improved the structure.
5 Conclusion
The paper concludes that population-guided Gaussian splatting constrains sparse-view reconstructions to anatomically valid configurations and improves 5-view quality over competitive methods.
- Conclusion: The framework constrains Gaussian primitive locations with a shape model and uses population statistics for regularization.Population statistics also initialize Gaussian parameters with an anatomically valid starting point.
- Conclusion: At 5 views, the method improves PSNR by 2.83 dB and SSIM by 0.022 relative to competitive state-of-the-art methods.These results are reported under extreme sparsity.
- Conclusion: The method is theoretically not limited to the tested femur and may extend to other anatomical structures.This is presented as a potential extension rather than an evaluated result.
- Acknowledgments: The project used NMDID CT imaging data and received support from the NSERC Alliance Advantage grant in partnership with Eiffel Medtech Inc.Computational resources were provided by the Digital Research Alliance of Canada.
- Disclosure of interests: The work received research funding from Eiffel Medtech Inc., whose founder is an author; the other authors declared no competing interests.This disclosure identifies a potential institutional conflict of interest.