Introduction
Computed Tomography (CT) reconstruction is a fundamental inverse problem in medical imaging, where the goal is to recover cross-sectional images from raw X-ray measurements. Traditionally, this has been achieved through Filtered Back-Projection (FBP) or Model-Based Iterative Reconstruction (MBIR). While deep learning has recently introduced powerful generative priors, many current diffusion-based methods suffer from two significant limitations: linearization bias and geometric rigidity.
Linearization bias arises when the complex, inherently nonlinear physics of X-ray transmission is simplified into a linear model for mathematical convenience. This simplification often fails in low-dose settings where noise and "photon starvation" are prevalent. Geometric rigidity refers to the fact that many supervised deep learning models are tied to the specific scanner geometry (e.g., number of views, dose) used during training, requiring costly retraining for new configurations.
This paper introduces a "plug-and-play" reconstruction framework that addresses these issues by combining unsupervised diffusion models with a nonlinear physical measurement model. By utilizing Diffusion Posterior Sampling (DPS), the authors achieve high-fidelity reconstructions that are both physically accurate and flexible across different scanner geometries.
Methods
The proposed framework leverages score-based diffusion models to provide a strong anatomical prior. The process is divided into two main stages: training an unsupervised prior and performing nonlinear posterior sampling.
Unsupervised Prior Training
A score-based diffusion model, typically a U-Net architecture, is trained on a dataset of high-quality, "clean" CT images. This model learns the distribution of human anatomy without requiring paired raw projection data, making it independent of any specific scanner geometry.
Nonlinear Posterior Sampling
During the reconstruction phase, the learned prior is combined with a nonlinear likelihood model using Bayes' rule. The X-ray measurements are modeled using a Poisson distribution and an exponential forward model: $y \sim \text{Poisson}\{I_0 \exp(-Ax)\}$. This approach ensures that the reconstruction process respects the actual physics of X-ray attenuation.
As shown in the figure above, the forward SDE process transforms CT data into noise, while the reverse SDE process reconstructs the image by incorporating a likelihood update. This update term, $\nabla_{x_t} \log p(y|x_t)$, guides the diffusion process to ensure the final image is consistent with the measured raw data $y$.
Implementation and Acceleration
To handle the computational complexity of the nonlinear gradient, the authors implement the DPS Nonlinear algorithm. To further accelerate the process, they introduce an Ordered Subsets (OS) variant, which utilizes only a fraction of the projection data at each iteration to reduce back-projection overhead.
Experimental Evaluation
The method was evaluated using the publicly available CT Lymph Nodes dataset, consisting of 156 subjects. Images were processed into $512 \times 512$ voxels and split into training, validation, and test sets. The authors simulated several challenging clinical scenarios:
- Low-dose CT: $10^3$ photons per pixel to test noise robustness.
- Sparse-view CT: Only 30 projections over 360 degrees to test geometric flexibility.
- Variable Fluence: A range from $10^2$ to $10^5$ photons to investigate potential "hallucinations" in generative models.
Results and Discussion
Performance and Robustness
In low-dose scenarios, the proposed DPS Nonlinear method achieved a PSNR of 28.32 dB, significantly outperforming the linearized diffusion version (27.65 dB) and the state-of-the-art supervised method, DOLCE (27.94 dB).
The framework's robustness was most evident in the sparse-view (30-view) geometry. While the supervised DOLCE model's performance dropped to 24.18 dB because it was not trained for such sparse data, the DPS Nonlinear method maintained a high PSNR of 31.15 dB. This demonstrates that the unsupervised prior generalizes effectively to unseen geometries.
Bias and Hallucinations
By adhering to a nonlinear physical model, the authors reduced reconstruction bias by approximately 4.7% compared to linearized methods. This is critical for diagnostic accuracy, where subtle density variations matter. Furthermore, while generative models can sometimes "hallucinate" features when data is extremely scarce, the nonlinear physical constraint acts as a guardrail, keeping the model more "honest" than purely supervised black-box approaches.
Efficiency
The Ordered Subsets (OS) variant provided a 10.24% speedup in reconstruction time with negligible loss in image quality, though the overall process remains computationally intensive compared to traditional FBP.
Conclusion and Future Work
This work demonstrates that integrating nonlinear physical models into diffusion-based reconstruction significantly improves fidelity and flexibility in CT imaging. While the computational cost (approx. 333 seconds per slice) and variability in fine lung textures remain challenges for real-time clinical use, the framework provides a robust foundation for future developments. Potential future directions include extending the method to 3D cone-beam reconstruction and applying it to other nonlinear problems such as beam-hardening correction and spectral CT.
Summary updated