$K$-NeAS: Scalable Multi-Material CT Reconstruction Using Neural SDFs

2026-07-15 · Source: Computer Vision and Pattern Recognition · Field: Science & Research — Artificial Intelligence & Machine Learning, Computer Vision & Pattern Recognition, Health & Medical Research · Depth: Expert, quick

Summary

$K$-NeAS introduces a unified and scalable architecture for automated, multi-material surface reconstruction in Computed Tomography (CT). This method addresses limitations of previous implicit scene representations (ISRs) by replacing independent material networks with a shared latent backbone and incorporating a fully differentiable $K$-material sequential soft selector, enabling modeling of arbitrary overlapping tissues. It automates attenuation bounding using a Gaussian Mixture Model (GMM) and employs a scheduled auxiliary floater loss to mitigate geometric hallucinations under extreme sparsity. Evaluated on four clinical Cone-Beam CT (CBCT) datasets, $K$-NeAS achieves superior 3D volumetric fidelity, demonstrating \$33.28\text{ dB}$ 3D PSNR at $K=3$ materials on Abdomen regions, a $+1.88\text{ dB}$ improvement over the \$31.40\text{ dB}$ single-material NeAS baseline. The model also shows enhanced robustness, outperforming the baseline by up to \$1.17\text{ dB}$ under 5- and 10-view constraints.

Key takeaway

For AI Scientists or Computer Vision Engineers developing CT reconstruction methods who need to improve fidelity and material segmentation from sparse data, $K$-NeAS offers a robust approach to automate multi-material surface reconstruction, eliminating manual tuning and rigid constraints. You should consider integrating its shared latent backbone and GMM-based attenuation bounding to achieve superior volumetric fidelity and robustness in clinical CBCT applications, especially when dealing with limited projection views.

Key insights

$K$-NeAS enables scalable, automated multi-material CT reconstruction from sparse views using neural signed distance functions.

Principles

• Shared latent backbones improve multi-material modeling.
• Automated attenuation bounding enhances reconstruction fidelity.
• Auxiliary floater loss mitigates geometric hallucinations.

Method

The method uses a shared latent backbone, a $K$-material sequential soft selector, a Gaussian Mixture Model (GMM) for automated attenuation bounding, and a scheduled auxiliary floater loss to prevent geometric hallucinations.

In practice

• Scale CT reconstruction to arbitrary material counts.
• Improve fidelity in complex multi-tissue regions.
• Enhance robustness under extreme sparse-sampling.

Topics

• Computed Tomography
• Neural Signed Distance Functions
• Multi-Material Reconstruction
• Sparse-View CT
• Implicit Scene Representations
• Gaussian Mixture Model

Best for: AI Scientist, Research Scientist, Computer Vision Engineer

Related on AIssential

• Neu-PiG: Neural Preconditioned Grids for Fast Dynamic Surface Reconstruction on Long Sequences — Neu-PiG uses a preconditioned latent-grid encoding for fast, drift-free dynamic 3D surface reconstruction.
• Enabling self-supervised learned primal dual with Noise2Inverse — Self-supervised learning can enable high-quality CT reconstruction without ground-truth data.
• Extracting Neural Materials from Multi-view Images — NeuMatEx uses an LMRM and uncertainty-guided inverse rendering to extract complex neural materials from multi-view images.
• PSCT-Net: Geometry-Aware Pediatric Skull CT Reconstruction via Differentiable Back-Projection and Attention-Guided Refinement — PSCT-Net reconstructs 3D pediatric skull CT from sparse X-rays using geometry-aware differentiable back-projection and attention-guided refinement, reducing radiation risk.
• UnDA: Unpaired Domain Alignment for Cross-Modal Knowledge Transfer in Medical Imaging — UnDA enables robust cross-modal knowledge transfer in medical imaging using unpaired data by mitigating noise and aligning features.
• Enabling self-supervised learned primal dual with Noise2Inverse — N2I-LPD enables self-supervised CT reconstruction by extending Noise2Inverse to Learned Primal-Dual, using noise independence for training without ground truth.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Computer Vision and Pattern Recognition.