Introduction
X-ray Computed Tomography (CT) is an essential diagnostic tool, yet the ionizing radiation it employs poses significant health risks. To mitigate these risks, clinicians often utilize Low-Dose CT (LDCT). However, reducing radiation dose introduces substantial noise and artifacts, potentially obscuring critical clinical features. While Diffusion Probabilistic Models (DDPMs) have set a high bar for image restoration, their iterative nature—requiring dozens or hundreds of steps—makes them computationally prohibitive for real-time clinical use.
This paper introduces High Noise Improved Consistency Training (HN-iCT), a framework designed for high-quality, single-step denoising. By refining the training process of Consistency Models (CMs), the authors achieve state-of-the-art results that match the fidelity of multi-step diffusion models at a fraction of the computational cost.
Methodology
The core of the HN-iCT framework lies in three primary innovations designed to make consistency models robust enough for medical imaging:
1. Beta Noise Distribution
Standard consistency training often struggles with the high-noise regimes typical in medical imaging. The authors replace the conventional log-normal noise distribution with a Beta distribution. This allows for precise control over the noise levels encountered during training via two tunable parameters, $\alpha$ and $\beta$. Crucially, the researchers found that ensuring a small portion (4-5%) of training samples are at very high noise levels is essential for the model to learn robust structural recovery from extreme degradation.
2. Sinusoidal Training Curriculum
Consistency models typically use a "curriculum" that gradually increases the complexity of noise levels. Unlike previous methods that used abrupt step-wise increases, this work proposes a sinusoidal curriculum. This smooth progression maintains "trajectory point traceability," facilitating a more stable learning path from highly degraded inputs back to clean reconstructions.
3. Weighted Attention Gates (WAG)
For image-conditioned tasks like LDCT denoising, the authors introduced the HN-iCT-CN architecture. This model uses the noisy LDCT image as a condition to guide the generation process. The architecture incorporates Weighted Attention Gates to effectively extract features from the source image while preventing input noise from dominating the final output.
Experimental Results
The proposed methods were evaluated on both general benchmarks (CIFAR10, CelebA) and the Mayo Clinic AAPM Low-Dose CT Grand Challenge dataset.
General Benchmarks
On the CIFAR10 dataset, the HN-iCT-Small model achieved a Fréchet Inception Distance (FID) of 10.50 in a single step (NFE=1), a significant improvement over the standard iCT baseline of 13.50. Ablation studies confirmed that the combination of the sinusoidal curriculum and Beta noise distribution was key to this performance leap.
Medical Denoising Performance
In the LDCT denoising task, the HN-iCT-CN model demonstrated exceptional efficiency and quality. As shown in the table below, the model achieves competitive PSNR and SSIM scores in just one step, rivaling models that require significantly more computation.
The HN-iCT-CN Medium model achieved a PSNR of 44.07 and an SSIM of 0.96 with NFE=1. This performance is on par with or superior to multi-step models like EDM (NFE=79) and DDPM (NFE=50), highlighting the efficiency of the proposed consistency training approach for clinical applications.
Conclusion
High Noise Improved Consistency Training (HN-iCT) represents a significant step forward in making generative denoising practical for clinical environments. By optimizing the noise distribution and training curriculum, the authors enable high-fidelity, single-step image restoration. While training remains computationally intensive and hyperparameter tuning is required for different modalities, the dramatic reduction in inference time offers a viable path toward real-time, high-quality LDCT imaging.
Summary updated