PaperlayerView exact revision on arXiv

Source-linked AI summary

Gaussian Material Fields for Volumetric Multi-Energy CT Decomposition

Jian Lin, Jiancheng Fang, Hongming Shan, Shaoyu Wang, Yang Chen, Qiegen Liu

arXiv:2610.09492v1cs.CV

TL;DR

Gaussian material fields address how to share spatial organization across material fields while preserving independent composition and material-specific refinement needs. The method jointly fits shared Gaussian geometry and material coefficients from multi-energy projections; across 15 simulated-projection cases, it leads all three reported metrics in every case.

  • Problem

    Shared spatial organization must preserve independent material composition, and a region represented adequately for one material may still need capacity for another.

    Source: §I · pp. 1–2

  • Method

    Gaussian material fields combine shared anisotropic Gaussian geometry and independent nonnegative material coefficients, jointly fitted from projections with a differentiable spectral model and material-aware density control.

    Source: §I · p. 2

  • Results

    Across 15 simulated-projection cases, the proposed method leads all three reported metrics in every case.

    Source: B. Comparative Results · p. 6

  • Takeaways & Limitations

    The tested representation unifies spatial organization, material composition, and local adaptation, with experiments reporting improved material reconstruction and localized detail recovery.

    Source: §V · p. 9

Abstract

from arXiv · show

Volumetric material decomposition in multi-energy computed tomography requires a representation that organizes multiple three-dimensional material fields in a common spatial domain while retaining differences in composition and local structure. We observe that spatial primitives can be shared across materials without tying their coefficients, but their local capacity must respond to material-specific reconstruction needs. We introduce Gaussian material fields, which represent multiple material distributions with shared anisotropic 3D Gaussian primitives and independent nonnegative material coefficients. The shared geometry defines a continuous spatial basis, while the coefficients determine each primitive's contribution to the individual material fields. To reconstruct this representation from multi-energy projections, a differentiable spectral forward model combines Gaussian material path integrals with a calibrated basis matrix, enabling joint optimization of spatial geometry and material composition. Material-aware adaptive density control retains material-specific refinement evidence before aggregation and adjusts local representation capacity to accommodate both spatially extensive components and sparse details. Experiments use synthesized multi-energy projections generated from pseudo-reference material maps constructed by conventional methods from publicly available CT data. Across 15 cases, our approach improves average PSNR by 4.03 dB and SSIM by 4.96% over the strongest baseline, while reducing NRMSE by 33.45%. Material-wise comparisons and component ablations support improved recovery of localized structures, while runtime and memory measurements show favorable computational scaling. These results establish Gaussian material fields as an explicit, adaptive representation for volumetric multi-material reconstruction.

I. INTRODUCTION

Volumetric material decomposition must organize multiple 3D fields in a common spatial domain without erasing material-specific composition or local structure. Gaussian material fields pair shared spatial primitives with independent coefficients, joint projection-domain fitting, and material-aware capacity adjustment.

I. INTRODUCTION

  • A shared basis can still leave one material under-resolved, so refinement evidence must preserve material-specific demand before aggregation.Localized components may require refinement even where other materials provide little evidence.Source: §I · pp. 1–2
  • Gaussian material fields combine shared anisotropic geometry with independent nonnegative material coefficients, jointly fitted from multi-energy projections.The calibrated spectral forward model projects material path integrals and updates geometry and composition together.Source: §I · p. 2
  • Evaluation uses synthesized multi-energy data derived from public CT data and conventionally constructed pseudo-reference material maps.The paper reports overall comparisons, material-wise analyses, ablations, and configuration studies.Source: §I · p. 2

A. Spectral CT Material Decomposition and Reconstruction

  • Spectral CT methods estimate material coefficients from reconstructed energy images, material path integrals before reconstruction, or material volumes directly through joint reconstruction; neural methods learn image-domain classification, concentration estimation, or projection-consistent dual-energy decomposition.Source: A. Spectral CT Material Decomposition… · p. 2

B. Gaussian-Based CT Reconstruction

  • Gaussian methods have been applied to X-ray reconstruction, while SpectralCTGaussians fits spectral coefficients on shared Gaussians and derives material labels by clustering [16].The proposed method instead reconstructs material coefficient fields with a calibrated log-domain model and material-aware density control.Source: B. Gaussian-Based CT Reconstruction · p. 2

III. METHOD

The method represents multiple material fields with shared Gaussian geometry and independent coefficients, then fits them directly against multi-energy projections. Material-specific refinement evidence guides local changes to the shared representation.

B. Shared-Geometry Gaussian Material Field

  • Each material field is a differently weighted combination of shared kernels: moving or reshaping a primitive changes the basis for all materials, while its coefficient changes one material's contribution.The kernel centers, scales, and orientations encode shared spatial organization.Source: B. Shared-Geometry Gaussian Material… · p. 3

C. Projection-Domain Joint Reconstruction

  • Material path integrals reduce to weighted sums of Gaussian ray responses; the infinite-line response approximates finite segments only when omitted tails are negligible.The rasterized model is exact for its defined approximation, not for the original finite-ray integral.Source: C. Projection-Domain Joint… · p. 4
  • The projection model connects material fields to measurements through Gaussian ray responses and calibrated spectral mixing, allowing geometry and composition to affect the same predicted projections.The calibrated log-domain model is not an exact polychromatic photon-count model, and energy-dependent volumes are derived rather than independent reconstruction variables.Source: C. Projection-Domain Joint… · pp. 3–4, B. Shared-Geometry Gaussian Material… · p. 3, Figure 3 · p. 4
  • Material-specific projection sensitivities are retained before aggregation and combined with Gaussian scale to guide cloning and splitting; pruning tests amplitudes across channels.This capacity adjustment addresses regions that are resolved in one material but not another.Source: Figure 4 · p. 5
  • The method initializes from energy-wise FDK followed by voxelwise NNLS, then jointly optimizes geometry and composition with Adam, interleaving parameter updates with density control.The projection objective uses an ℓ1 loss without volumetric ground truth or an auxiliary volume/support loss.Source: C. Projection-Domain Joint… · pp. 4–5

D. Material-Aware Adaptive Density Control

Material-aware density control preserves refinement evidence from individual material branches before adapting shared capacity. It uses branch responses and primitive scale to guide refinement, while restricting edits to prescribed rules and budgets.

D. Material-Aware Adaptive Density Control

  • Density control adjusts local capacity beyond parameter updates, using aggregated material-branch evidence and primitive scale to guide adaptation.Parameter updates fit the current spatial basis but cannot change how many primitives represent a local structure.Source: D. Material-Aware Adaptive Density… · p. 5
  • Branch responses are retained before aggregation because summing can cancel material-specific signals and averaging norms can dilute a response confined to one material.The maximum is taken before temporal accumulation, with zero evidence for invisible branches.Source: D. Material-Aware Adaptive Density… · p. 5
  • Primitives are cloned when refinement evidence meets τg and scale is at most τs, split when evidence meets τg and scale exceeds τs, and otherwise kept.Cloning inserts identical kernels with activated coefficients ai/2; splitting offsets centers, reduces scales, and divides coefficients across children while preserving their sum.Source: D. Material-Aware Adaptive Density… · pp. 5–6
  • Refinement is limited to a prescribed iteration window and Gaussian budget, and residual-guided edits neither exactly minimize primitive count nor guarantee a loss decrease.Low-content pruning requires maxm ai,m < τa, and centers outside the valid reconstruction box are also removed.Source: D. Material-Aware Adaptive Density… · p. 6

A. Datasets and Experimental Settings

The evaluation uses 15 simulated multienergy cone-beam cases, including known synthetic maps and pseudo-reference maps derived from public CT data. At 50 views, the proposed method leads all three reported metrics in every case, with additional comparisons examining structure recovery, sparse views, and runtime.

A. Datasets and Experimental Settings

  • The 15 cases comprise nine AAPM phantoms with known synthetic maps and six cases with pseudo-reference maps derived from public CT data.The maps generate the multi-energy projections and serve as evaluation references.Source: A. Datasets and Experimental Settings · p. 6
  • Each case uses 50 uniformly spaced full-rotation training views and 100 disjoint test views; PSNR, NRMSE, and SSIM are averaged across material channels.AAPM training projections include prescribed Poisson and Gaussian noise; their test projections and the other six cases’ projections are clean.Source: A. Datasets and Experimental Settings · p. 6

B. Comparative Results

  • At 50 views, the proposed method leads PSNR, SSIM, and NRMSE in all 15 cases.Table II reports material-averaged full-volume metrics, with the final column equally averaging the cases.Source: B. Comparative Results · p. 6, Table II · p. 7
  • The proposed method’s mean PSNR margin over the strongest competitor is 2.38 dB across nine AAPM phantoms, and the two animal cases also favor it.The largest stated gaps are 41.39 versus 32.98 dB for the powder phantom and 37.77 versus 29.86 dB for walnut.Source: B. Comparative Results · p. 6
  • The comparisons report more continuous calcification arcs, cortical boundaries, powder features, ore-Pb rings, MUSIC boundaries, and walnut rims with fewer spurious responses or less mottling.Fig. 5 covers calcification and cortical boundaries; Figs. 6–7 show material slices including powder features, the ore-Pb ring, MUSIC boundaries, and the walnut rim.Source: B. Comparative Results · p. 6, Figure 5 · p. 7, Figure 6 · p. 7
  • Across ore and lizard-head cases, the proposed method leads all three metrics at six tested view-count settings; selected 25-view slices retain the ore-Pb ring and cortical edge more continuously than two baselines.The view-count evaluation uses 25, 50, and 75 views with one seed per method and setting.Source: B. Comparative Results · p. 6, Figure 8 · p. 8, Figure 9 · p. 8
  • Different scenes and bases, along with a different grid in one case, prevent attributing case differences to channel count alone.Source: B. Comparative Results · p. 8
  • On the 50-view ore case, the proposed method reaches 35.01 dB mean PSNR versus 33.34 dB for R2+NNLS, with median primary times of 354.514 s and 183.081 s, respectively.Both methods were repeated three times; the remaining four methods have one observation each.Source: B. Comparative Results · p. 8

C. Ablation Studies

The ablations compare the complete method with variants that alter Gaussian updates, gradient aggregation, or the loss domain. The conclusion reports improved material reconstruction and localized-detail recovery, with computational benefits in the tested settings.

C. Ablation Studies

  • The ablation uses four controlled variants across three cases, with three paired seeds per case and 15,000 iterations per setting.The cases are lizard head, powder phantom, and MUSIC 3D; settings share a 50,000-point initialization within each case and seed.Source: C. Ablation Studies · p. 8, PSNR (dB) · p. 9

PSNR (dB)

  • Fig. 12 displays paired variant-minus-Full PSNR and SSIM differences across three cases and seeds, with blue negative and ochre positive.The upper and lower maps encode PSNR and SSIM differences, respectively.Source: Figure 12 · p. 9, PSNR (dB) · p. 9
  • Material-domain ℓ1 has the highest 3-seed mean PSNR and SSIM and lowest mean NRMSE in each case; energy-domain ℓ1 performs worse on all three mean metrics.Mean aggregation improves all three mean metrics over the complete setting, while holding the Gaussian set fixed reduces mean PSNR in every case; final representation sizes differ, so aggregation effects are not isolated.Source: PSNR (dB) · p. 9

V. CONCLUSION

  • The conclusion reports improved material reconstruction and localized-detail recovery, alongside computational benefits in the tested settings.The evaluation uses synthesized projections from public-data-derived pseudo-reference material maps.Source: §V · p. 9

APPENDIX

The appendix supplies derivations and implementation details for the method, including its ray-response model, conditioning analysis, refinement evidence, and computational accounting.

Gaussian ray response

  • The appendix derives finite Gaussian ray responses and covariance-dependent rectification, while noting that perspective-ray use retains a local approximation.The rectification factor arises by integrating the locally transformed unit-peak Gaussian over the third coordinate.Source: APPENDIX · p. 10

Conditioning

  • The conditioning analysis does not establish global uniqueness with trainable geometry; cloned kernels can create duplicate columns in the spatial basis.Its quadratic model diagnoses conditioning only and is not the Hessian of the actual ℓ1 loss or a replacement training objective.Source: APPENDIX · p. 10

Material-specific refinement evidence

  • Material-specific refinement branches are retained before aggregation, so averaging across materials does not suppress localized refinement evidence.The maximum branch norm is related to the shared projected-center derivative; this rule prevents averaging-induced dilution but does not guarantee accurate recovery or optimal allocation.Source: APPENDIX · p. 11

Primitive edits

  • A clone with identical kernels and halved activated coefficients preserves the field and prediction at insertion under unchanged linear rendering; splitting generally changes them.Splitting conserves coefficient sums but not, in general, spatial integrals, and the loss change has no guaranteed sign.Source: APPENDIX · p. 11

Pruning

  • Pruning error is not bounded by small peak amplitude alone, because the perturbation bound is not scale-independent.The appendix also notes that a center outside the reconstruction domain need not have zero Gaussian contribution inside it.Source: APPENDIX · p. 11

Parameterization and initialization

  • Initialization uses measured projections, while finite parameter initialization requires a positive floor for zero NNLS coefficients and an interior convention for bounded scales.Finite Softplus logits cannot represent zero coefficients, and sigmoid endpoints require numerical handling.Source: APPENDIX · p. 11

Computational accounting

  • Continuous outputs are point samples, and separate material rasterization has view-evaluation cost O(MKv) + O(EMHW) and trainable storage O(N(10 + M)).The accounting also excludes additional costs such as sorting/culling and optimizer states; no fused multi-channel acceleration or material-count-independent runtime is assumed.Source: APPENDIX · p. 12

Related papers

Loading 2610.09492v1…