Abstract
Deep-learning methods have shown promising performance for low-dose computed tomography (LDCT) reconstruction. However, supervised methods face the problem of lacking labeled data in clinical scenarios, and the CNN-based unsupervised denoising methods would cause excessive smoothing in the reconstructed image. Recently, the normalizing flows (NFs) based methods have shown advantages in producing detail-rich images and avoiding over-smoothing, however, there are still issues: (1) Although the alternating optimization in the data and latent space can well utilize the regularization and generation capabilities of NFs, the current two-way transformation strategy of noisy images and latent variables would cause detail loss and secondary artifacts; and (2) Training NFs on high-resolution CT images is hard due to huge computation. Though using conditional normalizing flows (CNFs) to learn conditional probability can reduce the computational burden, current methods require labeled data for conditionalization, and the unsupervised CNFs-based LDCT reconstruction remains a problem. To tackle these problems, we propose a novel CNFs-based unsupervised LDCT iterative reconstruction algorithm. It employs strict one-way transformation when performing alternating optimization in the dual spaces, thus effectively avoiding the problems of detail loss and secondary artifacts. By proposing a novel unsupervised conditionalization strategy, we train CNFs on high-resolution CT images, thus achieving fast and high-quality unsupervised reconstruction. Experiments on different datasets suggest that the performance of the proposed algorithm could surpass some state-of-the-art unsupervised and even supervised methods.
Introduction
Computed Tomography (CT) is a cornerstone of modern medical diagnosis, but the ionizing radiation from standard doses poses health risks. While "low-dose" CT (LDCT) is safer, it introduces significant noise and artifacts that can obscure clinical details.
Current deep learning solutions for LDCT reconstruction face two primary hurdles. First, most high-performing models are supervised, requiring paired low-dose and normal-dose scans which are difficult to obtain in clinical practice. Second, unsupervised CNN-based methods often over-smooth images, losing diagnostic texture. While generative models like Normalizing Flows (NFs) preserve details better, they are computationally expensive and standard "two-way" approaches—mapping noisy images back to latent space—often introduce distribution bias and secondary artifacts.
Methods: One-Way Conditional Normalizing Flows
The proposed OW-CNFs algorithm addresses these challenges through three key innovations:
- One-Way Transformation: Unlike traditional bijective NFs that map $z \leftrightarrow x$, OW-CNFs restrict the optimization to a strict one-way path (latent $z \rightarrow$ image $x$) during reconstruction. This prevents noise from the data space from "poisoning" the latent variables, effectively avoiding the secondary artifacts common in two-way strategies.
- Unsupervised Conditionalization: To handle high-resolution 512x512 images efficiently, the model uses Conditional Normalizing Flows. Instead of requiring paired labels, the authors generate a structural "condition" $c$ directly from the noisy image using a simple denoiser and wavelet filter. This provides a structural skeleton that guides the flow without needing ground-truth pairs.
- Iterative Dual-Space Optimization: The reconstruction alternates between the data space (ensuring fidelity to raw X-ray projections via OS-SART) and the latent space (leveraging the CNF prior to ensure realistic image generation).
Key Results and Performance
The model was evaluated on the Random Rectangle Models (RRM) synthetic dataset and the clinical LIDC-IDRI dataset.
- Superior Accuracy: On the RRM dataset, OW-CNFs achieved a PSNR of 39.61 dB, significantly outperforming other unsupervised methods like EASEL (33.21 dB) and even the supervised RED-CNN (38.03 dB).
- Clinical Utility: On real-world LIDC-IDRI data, the method maintained high structural similarity (SSIM of 0.819), effectively removing noise while preserving high-contrast details critical for identifying lung nodules.
- Efficiency: OW-CNFs are significantly faster than popular diffusion-based models. The algorithm requires only 55 iterations (~49 seconds) to reach convergence, compared to the 1,000+ steps (~700 seconds) often required by diffusion alternatives.
Conclusion and Future Work
OW-CNFs demonstrate that unsupervised generative priors can surpass supervised residual learning in medical imaging. By treating the NF as a one-way generator and utilizing an unsupervised conditioning strategy, the researchers achieved high-fidelity reconstruction without the need for paired clinical data.
Future work will focus on reducing hyperparameter sensitivity (currently requiring manual tuning of $\lambda, \sigma, r_1, r_2$) and exploring automated feature-learning networks to generate the initial structural conditions, potentially making the system more robust for diverse clinical environments.
Summary updated