Taming Neural Fields with a Frequency-Lipschitz Curriculum for Sparse-View CT Reconstruction

Songtao Hu, Jingrui Wu, Hong Ding, Fanhao Zhu, Yunhao Liu, Suyang XiarXiv 2026 · arXiv

TL;DR

The researchers introduce FLIC (Frequency-Lipschitz INR Curriculum), a framework that reconstructs high-quality CT images from limited X-ray views by explicitly scheduling the network's learning capacity. By coupling frequency exposure with Lipschitz constraints, FLIC achieves state-of-the-art unsupervised reconstruction quality in a single training stage, eliminating the need for slow, iterative external denoising.

What Problem Does This Paper Address?

Computed Tomography (CT) is vital for diagnosis, but high radiation doses are a major safety concern. Sparse-view CT (SVCT) reduces radiation by taking fewer X-ray projections. However, this creates a "severely under-determined" inverse problem: there isn't enough data to perfectly reconstruct the image.

Standard algorithms often fill the gaps with "streak artifacts" or hallucinated structures. While Implicit Neural Representations (INRs)—networks that map coordinates to tissue density—have shown promise, they are hard to control. If a network is too expressive too early, it "memorizes" the noise and artifacts. Previous solutions relied on slow, external denoising steps (like BM3D) that can take nearly an hour per image. The authors wanted a way to bake regularization directly into the network's training schedule.

Methods

The core of the paper is the Frequency-Lipschitz INR Curriculum (FLIC). Instead of penalizing the output, the authors control when the network is allowed to become complex.

  1. Dual-Path Architecture: The model uses two inputs: spatial coordinates ($p$) and a "self-prior" image ($z$) derived from a quick, low-quality Filtered Back Projection (FBP).
  2. Joint Curriculum: The authors couple two controls:
  3. Frequency Control: They use Random Fourier Features (RFF) with a mask that gradually "opens up" to higher frequencies as training progresses.
  4. Lipschitz Control: They apply spectral normalization to the MLP layers. The "Lipschitz budget" (a measure of how much the output can change relative to the input) starts small (smooth) and grows linearly (allowing for sharp details).
  5. Scan-Adaptive Calibration: Rather than manual tuning, the terminal Lipschitz budget is automatically calculated based on the gradient statistics of the initial FBP image. This allows the model to adapt to different anatomical complexities (e.g., soft tissue vs. sharp bone interfaces).
  6. Prior-Guided Adaptive Sampling: Instead of sampling X-ray integration points uniformly, FLIC uses the FBP prior to concentrate samples near high-gradient areas like anatomical boundaries, improving the fidelity of the forward model.
Figure 2. Overview of the proposed framework. (a) Overall single-stage framework. (b) Dual-path Lipschitz architecture. (c) Prior-guided adaptive sampling.

Datasets

The authors evaluated FLIC on four diverse public CT datasets to ensure robustness across different body parts:

  • AAPM: Abdominal scans.
  • COVID-19: Chest scans.
  • CQ500: Head scans.
  • ToothFairy: Dental/mandible scans.

They tested these under noise-free, normal-dose, and low-dose (high noise) settings with 60 and 90 projection views.

Key Insights and Ideas

  • The "Loophole" Effect: The authors point out that controlling frequency bandwidth alone isn't enough. A high-capacity MLP can still map low-frequency inputs to high-frequency artifacts. By coupling frequency and Lipschitz constraints, they close this "loophole," forcing the network to follow a strict coarse-to-fine trajectory.
  • Spectral Fidelity: As shown in Figure 5, the joint curriculum allows the network to match the ground-truth frequency spectrum much more closely than an unconstrained baseline, which tends to inflate mid-to-high frequency energy (artifacts).
  • Efficiency: Because FLIC is a single-stage optimization without external iterative denoising, it is significantly faster. It takes roughly 7 minutes per image, compared to 50 minutes for previous state-of-the-art methods like Spener.
Comparison of frequency spectra showing how the joint curriculum suppresses spurious high-frequency energy.

Key Results

FLIC consistently outperformed both analytical and other unsupervised INR methods across all datasets:

  • Quantitative Gains: On the AAPM dataset (60 views), FLIC achieved a PSNR of 33.52 dB, surpassing the previous best unsupervised method (Spener) by 1.39 dB. On the ToothFairy dataset, the gain was even more dramatic: +3.29 dB.
  • Noise Robustness: In low-dose settings, FLIC's performance dropped by only 0.42 dB, while competitors saw drops of up to 1.37 dB. This is attributed to the explicit Lipschitz constraint on the "prior path," which prevents the network from amplifying noise in the FBP input.
  • Qualitative Quality: As seen in Figure 4, FLIC produces the cleanest reconstructions, effectively suppressing the "streaks" common in sparse-view CT while preserving sharp edges in bone and dental structures.
Figure 3. Qualitative comparison of reconstructed CT images on four datasets under the 90-view noise-free setting. From left to right: FBP, FBPConvNet, RegFormer, CoIL, NeRP, SCOPE, Spener, Ours, and Ground Truth. Blue boxes highlight regions of interest, where the proposed method better suppresses streak artifacts while preserving fine anatomical structures.

Limitations and Future Work

While highly effective, the current work focuses on 2D slice reconstruction. The authors note that extending this to 3D cone-beam CT is a natural next step. Additionally, while the total Lipschitz budget is adaptive, the distribution of that budget across different layers of the MLP currently follows a fixed exponential strategy; the authors suggest that learned layer-wise budget allocation could further optimize performance.

Summary updated

Abstract

Sparse-view computed tomography (SVCT) reduces radiation dose by acquiring fewer projection views, but the resulting under-determined inverse problem requires careful regularization to prevent reconstruction algorithms from fitting streak artifacts and hallucinated structures. Existing implicit neural representation (INR) methods rely on implicit spectral bias or iterative external denoising, yet they provide limited control over when high-frequency capacity is exposed during optimization. To address this limitation, we propose FLIC (Frequency-Lipschitz INR Curriculum), a capacity-controlled self-prior INR framework that makes the network's representational capacity explicit and schedulable. The core of our method is a joint frequency-Lipschitz curriculum that couples random Fourier feature encoding with spectrally normalized MLP decoding under non-uniform Lipschitz budgets, guiding reconstruction along a coarse-to-fine trajectory from stable global anatomy to fine details. The terminal capacity is automatically calibrated from FBP gradient statistics, enabling scan-adaptive regularization without per-scan manual tuning of the terminal Lipschitz budget. We further introduce a prior-guided adaptive ray sampling strategy that concentrates integration samples near high-gradient anatomical boundaries, improving forward-model fidelity. The entire framework operates in a single training stage without iterative denoising, enabling efficient per-image reconstruction. Extensive experiments on four public CT datasets under noise-free and low-dose settings show that our method achieves state-of-the-art unsupervised reconstruction quality with consistent performance across diverse anatomical structures.
BibTeX
@misc{hu2026taming,
  author = {Songtao Hu and Jingrui Wu and Hong Ding and Fanhao Zhu and Yunhao Liu and Suyang Xi},
  title = {{Taming Neural Fields with a Frequency-Lipschitz Curriculum for Sparse-View CT Reconstruction}},
  year = {2026},
  eprint = {2611.15652},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.15652},
}
Figure