Abstract
Low-dose computed tomography (LDCT) is essential for reducing radiation risks but often suffers from significant noise and artifacts. While supervised deep learning methods have shown promise, they frequently struggle with generalization and require extensive paired datasets. This paper presents EASEL (dEep grAdient priorS of genErative modeL), an unsupervised framework that integrates score-based generative models into an iterative reconstruction pipeline. By learning the gradient of the data distribution (the "score") from normal-dose images, EASEL provides a robust prior that, when combined with a physical data-fidelity term via Separable Quadratic Surrogates (SQS), achieves state-of-the-art reconstruction quality. Experimental results on the AAPM Grand Challenge dataset demonstrate that EASEL outperforms existing supervised and iterative methods in both noise suppression and fine detail preservation, while exhibiting superior robustness to varying noise levels.
Introduction
Computed Tomography (CT) is a cornerstone of modern medical diagnostics. However, the ionizing radiation associated with standard-dose CT scans poses potential health risks, including carcinogenesis. Reducing the X-ray dose (LDCT) is a primary clinical goal, but it leads to degraded image quality characterized by severe noise and streaking artifacts.
Traditional solutions rely on Iterative Reconstruction (IR) with hand-crafted priors, such as Total Variation (TV). While mathematically grounded, these often produce "blocky" artifacts and lose subtle textures. Conversely, supervised deep learning (DL) models learn direct mappings from low-dose to normal-dose images. Despite their high performance, these models are often "black boxes" that require large paired datasets and lack robustness when the test data's noise characteristics deviate from the training set.
To address these limitations, the authors propose EASEL, which leverages score-based generative models. As shown in the taxonomy above, these models fall under explicit generative models that learn the gradient of the log-probability density ($\nabla_x \log p(x)$), providing a powerful unsupervised prior for iterative solvers.
Methodology: The EASEL Framework
The EASEL framework treats CT reconstruction as a posterior sampling problem, combining a learned image prior with a physical data-fidelity term.
Score-based Generative Prior
Instead of estimating the probability density $p(x)$ directly, EASEL learns the score function—the vector field pointing toward regions of higher data density. This is achieved through Denoising Score Matching (DSM), where a network is trained to estimate the score of a noise-perturbed version of normal-dose images. This approach is unsupervised, requiring only high-quality images for training.
Annealed Langevin Dynamics and Data Consistency
Reconstruction is performed using Annealed Langevin Dynamics. Starting from a noisy initial estimate, the algorithm iteratively updates the image by following the learned score function. To ensure the result remains faithful to the actual X-ray measurements, a data-fidelity term is integrated using the Separable Quadratic Surrogates (SQS) algorithm.
As illustrated in the pipeline diagram, the process alternates between the generative prior (moving the image toward the manifold of clean CT scans) and the SQS step (ensuring consistency with the raw sinogram data $y$). A momentum update is further applied to accelerate convergence and stabilize the trajectory.
Efficient Architecture: Channel-Copy Guided RefineNet
To handle the high dimensionality of CT data efficiently, the authors employ a modified RefineNet. A "channel-copy" strategy is used at the input to expand the single-channel CT image into multiple feature channels without the computational cost of high-dimensional initial convolutions.
Experimental Results
The framework was evaluated on the AAPM Low Dose CT Grand Challenge dataset and anthropomorphic CIRS Phantom data.
Quantitative Performance
EASEL achieved a Peak Signal-to-Noise Ratio (PSNR) of 41.58 dB on the AAPM dataset, surpassing leading supervised models like DP-ResNet (41.03 dB) and RED-CNN (39.37 dB). Notably, EASEL maintained high performance even when the noise level in the test data differed from the training assumptions, whereas supervised models experienced significant degradation.
Qualitative Analysis
Visual comparison revealed that EASEL effectively suppresses noise while preserving sharp edges and fine anatomical structures. Unlike TV-based methods that often "over-smooth" images, EASEL maintains the natural texture of the tissue, which is critical for accurate clinical diagnosis.
Conclusion
EASEL represents a significant step forward in LDCT reconstruction by bridging the gap between unsupervised generative modeling and physics-based iterative optimization. By utilizing deep gradient priors, it offers a robust, high-fidelity solution that generalizes better than traditional supervised networks. Future work will focus on improving convergence speed and extending the framework to 3D volumetric reconstruction.
Summary updated