SCOUT: Fast Spectral CT Imaging in Ultra LOw-data Regimes via PseUdo-label GeneraTion

Noise and artifacts during computed tomography (CT) scans are a fundamental challenge affecting disease diagnosis. However, current methods either involve excessively long reconstruction times or rely on data-driven models for optimization, failing to adequately consider the valuable information inherent in the data itself, especially medical 3D data. This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes. By leveraging spatial nonlocal similarity and the conjugate properties of the projection domain to generate pseudo-3D data for self-supervised training, high-fidelity results can be achieved in a very short time. Extensive experiments demonstrate that this method not only mitigates detector-induced ring artifacts but also exhibits unprecedented capabilities in detail recovery. This method provides a new paradigm for research using unlabeled raw projection data. Code is available at https://github.com/yqx7150/SCOUT.

💡 Research Summary

The paper introduces SCOUT, a novel zero‑shot self‑supervised framework for reconstructing high‑quality spectral CT images from ultra‑low‑dose raw projection data without any external training data or pre‑training. The authors first identify two major shortcomings of existing approaches: reliance on large labeled datasets and long reconstruction times, and insufficient exploitation of the intrinsic physical properties of CT data. SCOUT addresses these by leveraging (1) spatial non‑local similarity within the 3D projection volume and (2) the conjugate symmetry inherent in the projection geometry.

For any voxel in the low‑dose projection volume, SCOUT searches globally for voxels that are highly similar in appearance, aggregating them into a low‑rank voxel bank. This bank captures the repetitive anatomical structures that appear throughout the volume (bones, soft tissue, lesions, etc.). Simultaneously, the conjugate theorem—stating that a point p(s, θ) on the detector has a symmetric counterpart p(‑s, θ + π)—is used to generate a physically constrained conjugate voxel bank. The combination of statistical similarity and geometric constraints yields pseudo‑labels that faithfully represent the underlying signal while preserving noise as a high‑rank component.

These pseudo‑labels train a 3‑D neural network in a self‑supervised manner; the network learns to map noisy raw projections to their denoised counterparts using only the data itself. Because the signal resides in a low‑rank subspace and the noise is high‑rank, the network naturally separates them without explicit regularization or external priors.

Experiments cover four scenarios: (1) mouse photon‑counting CT (PCCT) data, where SCOUT dramatically reduces Poisson noise and ring artifacts while preserving fine vascular details; (2) dual‑energy walnut scans, demonstrating superior PSNR, SSIM, and RMSE across both energy channels; (3) simulated single‑energy CT datasets (May‑2016, May‑2020, LIDC‑IDRI, CT‑spine1K) with mixed Gaussian‑Poisson noise, where SCOUT improves PSNR by ~3 dB and SSIM by ~7 % over state‑of‑the‑art self‑supervised methods; and (4) speed benchmarks, showing total training and inference time of only 3.5 minutes for a volume of 300–600 slices. This is over 30× faster than Neighbor2Neighbor, >800× faster than Prompt‑SID, and roughly two orders of magnitude faster than Noise2Sim or Noise2detail.

The authors also discuss the method’s universality. Although designed for PCCT, the reliance on projection‑domain physics and non‑local similarity makes SCOUT applicable to conventional single‑energy CT and potentially to other tomographic modalities with known geometry.

Limitations include scalability to volumes with thousands of slices, dependence on accurate knowledge of the scanning geometry for conjugate symmetry, and sensitivity to hyper‑parameters governing similarity thresholds and voxel bank size. Future work may focus on adaptive parameter selection, extension to non‑standard trajectories, and integration with hardware‑level acceleration.

In summary, SCOUT presents a compelling paradigm shift: by generating pseudo‑labels directly from the raw data using physical constraints and statistical redundancy, it achieves ultra‑fast, high‑fidelity spectral CT reconstruction in data‑starved settings, opening new avenues for clinical deployment of photon‑counting CT and other advanced imaging systems.