Abstract
By modeling a precise geometry and avoiding data rebinning or interpolation, our physics-based reconstruction achieves a higher spatial resolution and fewer image artifacts with smaller magnitude than the clinical standard hybrid IR.
Introduction: The Challenge of Modern CT Reconstruction
Modern computed tomography (CT) scanners have evolved significantly, employing advanced hardware such as Dual Sources (two X-ray tubes) and Flying Focal Spots (FFS). In FFS systems, the electron beam wobbles to create multiple interleaved projection sets, which is essential for high temporal resolution imaging of moving organs like the heart or for patients who cannot remain still.
However, the software used to reconstruct images from these raw X-ray measurements has lagged behind. Most clinical systems still rely on Filtered Back-Projection (FBP) or "hybrid" iterative methods. These traditional approaches require "rebinning"—a process of interpolating raw data into a simplified parallel-beam geometry to reduce computational complexity. At high helical pitches, this interpolation leads to significant blurring and "windmill" artifacts. While statistical Iterative Reconstruction (IR) has been successful for single-source scanners, a model that fully accounts for the complex, interleaved physics of a DS-FFS scanner was previously missing.
The JENG Algorithm: A Physics-Based Approach
To address these limitations, the researchers developed JENG (Joint Estimation for Native Geometry). JENG is the first physics-based statistical iterative reconstruction algorithm designed specifically for DS-FFS CT scanners. It treats reconstruction as a massive joint optimization problem rather than a sequence of approximations.
Native Geometry Modeling
Unlike FBP, JENG utilizes a "discrete-to-discrete" mapping. It models the exact path of every X-ray from the deflected focal spots through the voxels to the detector units. By avoiding pre-processing interpolation or rebinning, JENG preserves the high-frequency information often lost in traditional pipelines. This approach allows the system to break the traditional Nyquist limits associated with FBP-based methods.
Statistical Noise and Consensus Equilibrium
JENG incorporates a noise physics model to account for data acquisition noise and a prior image model (Generalized Markov Random Field) to maintain realistic texture while reducing graininess. Because reconstructing these images involves massive system matrices, JENG employs the Consensus Equilibrium (CE) numerical method. CE allows the problem to be partitioned across multiple compute nodes and is provably guaranteed to converge to the exact global minimum, ensuring reconstruction accuracy even with non-uniform data.
Experimental Evaluation and Results
The algorithm was rigorously tested using both phantom and clinical data, including thoracic and abdominal scans from a Siemens Somatom Force scanner. JENG demonstrated superior performance across several key metrics:
- Spatial Resolution: In phantom studies, JENG’s Modulation Transfer Function (MTF) remained above the 10% visibility threshold up to 1.0 mm⁻¹, significantly outperforming the clinical standard (ADMIRE), which dropped below the threshold at 0.7 mm⁻¹.
- Artifact Reduction: JENG effectively eliminated "windmill" and streaking artifacts common in high-pitch helical scans. In clinical thoracic cases, it decimated aliasing streaks near the periphery of the image.
- Clinical Utility: Qualitative assessments showed better detectability of small bone structures and clearer visualization of lung nodules and liver cysts, providing a more uniform soft-tissue texture.
Limitations and Future Work
While JENG marks a significant advancement, some areas for improvement remain. The current implementation does not yet correct for beam hardening, which can cause mild artifacts near dense objects. Additionally, the algorithm assumes a point-like focal spot; modeling the actual physical shape of the focal spot could further enhance resolution for ultra-high-resolution applications. Finally, the computational cost of pure iterative methods remains higher than FBP, though the CE framework provides a scalable path forward for clinical implementation.
Summary updated