💻 computer science

Feed-Forward Hierarchical Gaussian Diffusion for Extreme CT Reconstruction

The paper proposes HiGDiff, a feed-forward hierarchical Gaussian diffusion framework that decomposes CT reconstruction into global structure and local detail stages to achieve state-of-the-art performance in recovering images from severely constrained projections across various degradation settings.

Original authors: Yuezhe Yang, Li Cheng

Published 2026-08-18

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuezhe Yang, Li Cheng

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Medical imaging relies on a delicate balance between seeing clearly and keeping patients safe. Computed tomography, or CT, creates detailed three-dimensional maps of the body by measuring how X-rays pass through tissues. To build a complete picture, a scanner usually rotates around the patient, taking hundreds of measurements from every angle. However, in many real-world situations, this ideal process is disrupted. A patient might move, making it impossible to capture every angle. A doctor might need to scan a child quickly, limiting the time the machine can run. Or, to reduce radiation exposure, the machine might be set to use fewer X-ray photons. When any of these constraints happen, the resulting images often suffer from streaks, blurring, or a grainy texture that hides the very details doctors need to see. For decades, scientists have tried to fix these broken images using mathematics and computer algorithms, but the most difficult cases—where multiple problems happen at once—have remained stubbornly hard to solve.

A team of researchers has now introduced a new approach called HiGDiﬀ, designed specifically to tackle these severely constrained imaging scenarios. Instead of trying to fix the entire image at once, their method breaks the problem down into two distinct layers: the broad, overall shape of the anatomy and the fine, sharp details of the tissue boundaries. Imagine trying to restore a faded photograph; you would first sketch the general outlines of the face and then fill in the specific features like eyes and lips. The researchers apply this same logic to the complex mathematics of CT reconstruction. They use a system that first learns to recover the large-scale structure of the body, such as the overall shape of the lungs or the spine, and then uses that recovered structure as a guide to fill in the missing fine details, like the edges of blood vessels or small lesions.

The core of this new method involves a clever way of organizing the data. Traditional computer models often treat the entire 3D volume of the body as a uniform grid of tiny cubes, trying to guess the value of every single cube simultaneously. This can be inefficient and prone to errors when data is scarce. The new system, instead, uses a collection of flexible, floating shapes that can move and change size to fit the important parts of the image. These shapes are first placed in the most critical areas, guided by the physical evidence from the X-ray measurements. The system then runs a two-step process. In the first step, it focuses entirely on getting the big picture right, smoothing out the major structures and removing the large-scale distortions caused by missing angles. Once this solid foundation is built, the second step kicks in. It looks at the difference between what the scanner actually measured and what the first step produced, using that leftover information to add back the sharp edges and subtle textures that were lost.

To test if this approach truly works, the researchers applied it to three different sets of medical data, covering everything from chest scans to abdominal views. They challenged the system with the worst-case scenarios: images taken with very few angles, images taken from only a partial circle, and images taken with very low radiation, as well as combinations of all three. In these difficult tests, the new method consistently outperformed existing techniques. On a standard dataset of low-dose chest scans, it improved the clarity of the reconstructed images by a significant margin, boosting a standard measure of image quality by nearly six points compared to the best previous methods. It also produced images that looked more structurally similar to the original, healthy scans, preserving the continuity of tissues that other methods often blurred or broke apart.

The success of this work suggests that separating the task of seeing the "big picture" from the task of seeing the "fine print" is a powerful strategy for medical imaging. By letting the computer first establish a reliable global framework and then refining the local details, the system avoids the confusion that arises when it tries to do everything at once. This is particularly important when the data is incomplete or noisy, as the system can rely on the strong structural clues it found in the first step to guide the recovery of the finer details in the second. While the researchers note that the most extreme cases still leave some subtle details difficult to recover, the method represents a significant step forward in making CT scans clearer and more reliable, even when the scan itself is compromised by movement, time limits, or the need to minimize radiation. The ability to reconstruct high-quality images from such limited information could eventually allow for faster scans, lower radiation doses, and better diagnoses for patients who cannot undergo a standard, full rotation scan.

Technical Summary: Feed-Forward Hierarchical Gaussian Diffusion for Extreme CT Reconstruction

Problem Statement
Reconstructing three-dimensional computed tomography (CT) volumes from severely constrained projections is a highly ill-posed inverse problem. The authors address scenarios where acquisition is degraded by sparse-view sampling (reducing angular density), limited-angle coverage (restricting angular span), and low-dose acquisition (increasing Poisson noise). These constraints often occur individually, in pairs, or jointly. Joint degradation couples directional information loss with signal-dependent noise, obscuring both long-range anatomical continuity and subtle local tissue transitions. Existing learned methods often target a single dominant degradation or utilize dense voxel parameterizations that uniformly distribute reconstruction capacity, failing to efficiently allocate resources to informative regions under extreme constraints. Furthermore, current diffusion and Gaussian approaches typically recover global structure and local detail within a shared representation, lacking an explicit hierarchy to handle the distinct recovery demands of long-range geometry versus residual boundaries.

Methodology: HiGDiﬀ
The authors propose HiGDiﬀ, a feed-forward hierarchical Gaussian diffusion framework designed to decompose reconstruction both spatially and hierarchically (from structure to detail). The pipeline consists of three main stages:

Foreground-Aware Gaussian Initialization:
- Physics-Conditioned Anchors: A physics feature encoder processes analytic initialization (FBP) and reliability-weighted backprojection residuals to generate a feature pyramid. From this, a landmark-free anchor network predicts $K$ anatomical anchors (coordinates, scales, orientations, confidence, and embeddings) to form a coarse anatomical scaffold.
- Foreground Capacity Field: A continuous scalar field predicts the utility of allocating Gaussian primitives to specific regions, modulating initialization and denoising features to emphasize boundaries and residual-sensitive areas.
- Dual Banks: The framework initializes two fixed-cardinality banks of explicit 3D Gaussians: a Structure Bank ( $N_s$ primitives) for global attenuation geometry and a Detail Bank ( $N_d$ primitives) for signed residual boundaries. The total budget $N$ is split based on a structure fraction $\pi_s$ .
Structure-to-Detail Hierarchical Diffusion:
- Structure Diffusion: A diffusion stage ( $D_s$ ) first recovers the low-to-mid-frequency attenuation geometry. It uses the physics features and anatomical anchors to denoise the structure bank.
- Conditioned Detail Diffusion: The completed structure bank generates tokens ( $R_s$ ) and a rendered field ( $\hat{\mu}_s$ ). These, combined with the unexplained measurement residual ( $B_s$ ), condition a second diffusion stage ( $D_d$ ). This stage recovers the signed residual details. Crucially, detail primitives attend to the nearest structure tokens via local attention, ensuring anatomically compatible corrections.
- Sequential Dependency: The process is strictly feed-forward; the structure bank conditions the detail bank, establishing a functional hierarchy where global geometry guides local refinement.
Gradient-Isolated Attenuation Refinement:
- The two completed Gaussian banks are rendered into an additive attenuation field ( $\hat{\mu}_G$ ).
- A refinement network ( $H$ ) operates on a dense voxel grid, taking the rendered field, the residual backprojection, and the foreground field as input. It predicts a bounded residual correction ( $r_\mu$ ) to correct grid-scale errors at sharp interfaces.
- Stop-Gradient: The refinement network is trained with a stop-gradient on the Gaussian output, preserving the learned primitive parameters and preventing the refinement stage from overriding the hierarchical structure.

Key Contributions

Unified Task Formulation: The work formulates a unified CT reconstruction task covering isolated, paired, and joint degradation settings (sparse-view, limited-angle, and low-dose) under severely constrained acquisition.
Sequential Structure-to-Detail Diffusion: The authors introduce a novel framework over explicit Gaussian banks where a foreground-aware initialization allocates primitives to informative regions, and a completed structure bank conditions a subsequent detail diffusion stage. This is coupled with a gradient-isolated refinement module that preserves the Gaussian reconstruction as the backbone.
State-of-the-Art Performance: The method achieves superior performance across three distinct benchmark datasets (LDCT-PD, LoDoPaB-CT, and TCIA Pancreas), particularly in joint degradation settings.

Experimental Results
Experiments were conducted on three datasets with various acquisition constraints:

LDCT-PD (3D Thorax): HiGDiﬀ achieved the best macro-average PSNR (32.655 dB) and SSIM (0.8661), improving over the strongest metric-specific baselines by 5.81 dB and 0.113 SSIM, respectively. It maintained robustness even under the most demanding joint settings (e.g., 1/12 view ratio, 180° angle, and low dose).
LoDoPaB-CT (2D Thorax): The method achieved the highest mean SSIM (0.9241) and was effectively tied with the best diffusion baseline (DOLCE) in macro-average PSNR. It ranked first in 10/16 PSNR and 14/16 SSIM comparisons at the setting level, showing superior robustness under compound degradation.
TCIA Pancreas (3D Abdomen): Demonstrating cross-dataset generalization, HiGDiﬀ achieved a macro-average PSNR of 35.298 dB and SSIM of 0.9261, outperforming the strongest comparators by 8.571 dB and 0.1464 SSIM. This indicates the learned representation transfers effectively from thoracic to abdominal anatomy.

Significance and Claims
The paper claims that HiGDiﬀ addresses the gap in existing methods by explicitly decomposing ambiguity spatially (via foreground-aware allocation) and hierarchically (via structure-to-detail diffusion). The authors argue that separating the recovery of global attenuation geometry from residual tissue transitions allows the model to handle the coupled nature of joint degradations more effectively than single-representation approaches.

The significance lies in the ability to preserve anatomical continuity and subtle details under extreme constraints where traditional analytical or iterative methods fail. The framework is presented as an explicit and adaptable representation for severely constrained CT reconstruction. The authors note that while the method shows significant improvements, performance decreases in the most severe joint settings (e.g., 27.172 dB at 1/3 view, 120° angle, low dose), where information lost to the angular null space and photon noise remains fundamentally ambiguous. They emphasize that current evaluations rely on simulated or protocol-derived projection corruption, and clinical conclusions require further validation with raw photon-count data and task-based assessments.

Technical Summary: Feed-Forward Hierarchical Gaussian Diffusion for Extreme CT Reconstruction

More like this