Abstract: Image reconstruction from computed tomography (CT) measurement is a challenging statistical inverse problem since a high-dimensional conditional distribution needs to be estimated. Based on training data obtained from high-quality reconstructions, we aim to learn a conditional density of images from noisy low-dose CT measurements. To tackle this problem, we propose a hybrid conditional normalizing flow, which integrates the physical model by using the filtered back-projection as conditioner. We evaluate our approach on a low-dose CT benchmark and demonstrate superior performance in terms of structural similarity of our flow-based method compared to other deep learning based approaches.
Introduction: The Challenge of Low-Dose CT
Computed Tomography (CT) reconstruction is a fundamental inverse problem in medical imaging. The goal is to recover the internal cross-sectional image of a patient from X-ray attenuation measurements. In clinical practice, reducing radiation exposure is critical, leading to "low-dose" CT protocols. However, low-dose measurements are inherently noisy and incomplete, making the reconstruction problem ill-posed: multiple valid images could potentially correspond to the same noisy data.
Traditional deep learning approaches typically treat this as a regression problem, training models (like U-Nets) to minimize Mean Squared Error (MSE). While effective at reducing noise, MSE-based models tend to produce blurry results because they essentially output the average of all possible solutions, losing fine structural details and textures.
Proposed Method: Hybrid Conditional Normalizing Flows
To address these limitations, we propose a Hybrid Conditional Normalizing Flow. Unlike standard regression models, Normalizing Flows are generative models that learn an invertible mapping between a complex data distribution (medical images) and a simple latent distribution (e.g., a standard Gaussian).
Physics-Informed Conditioning
A key innovation in our approach is the use of Filtered Back-Projection (FBP) as a conditioner. Instead of feeding raw, high-dimensional sinograms directly into the neural network, we leverage the known physics of CT by first performing an FBP. This transforms the measurement data into the image domain, providing a spatially-aligned "hint" that guides the conditional Invertible Neural Network (cINN).
Architecture and Sampling
The cINN architecture utilizes affine coupling layers, ensuring the mapping is bi-directional and allows for exact likelihood evaluation. To handle the high resolution of CT images, we employ a multi-scale design with Haar downsampling. Because the model learns the full posterior distribution, we can generate multiple possible reconstructions for a single measurement. By averaging these samples, we obtain a "Conditioned Mean" reconstruction that is both clean and structurally accurate.
Key Insights: Visualizing Uncertainty
One of the most significant advantages of a flow-based approach is the ability to quantify uncertainty. By calculating the standard deviation across multiple posterior samples, we can generate a "Standard Deviation Map."
As shown in the figure above, this map highlights regions where the reconstruction is less certain, often corresponding to areas with high noise or potential artifacts. This provides clinicians with valuable diagnostic information that standard "single-best-guess" models cannot offer.
Experimental Results
We evaluated our model on the LoDoPaB-CT dataset, which contains over 40,000 human chest CT images with simulated low-dose Poisson noise.
Quantitative Performance
Our cINN achieved state-of-the-art results in terms of Structural Similarity (SSIM), reaching a score of 0.88, significantly outperforming baselines like TV Regularization (0.74) and FBP+U-Net (0.81). In terms of Peak Signal-to-Noise Ratio (PSNR), the model achieved 35.68 dB, remaining highly competitive with models explicitly optimized for MSE (such as Learned Primal-Dual at 36.12 dB).
Qualitative Comparison
Visual inspection confirms that the cINN reconstructions are significantly sharper and cleaner than standard FBP.
Compared to other deep learning methods, our approach preserves better edge definitions and anatomical textures across various patient cases, which are crucial for accurate medical diagnosis.
Conclusion and Future Work
By integrating the physics of medical imaging with the generative power of Conditional Normalizing Flows, we have developed a robust framework for low-dose CT reconstruction. Our method not only provides high-quality images but also offers a principled way to visualize reconstruction uncertainty. Future work will explore more advanced invertible architectures and the use of iterative reconstruction algorithms as conditioners to further enhance performance.
Summary updated