Source-linked AI summary
GPU-based Iterative Cone Beam CT Reconstruction Using Tight Frame Regularization
Xun Jia, Bin Dong, Yifei Lou, Steve B. Jiang
TL;DR
Repeated CBCT imaging raises dose concerns, especially when undersampling or lower mAs makes conventional reconstructions highly degraded. The paper develops a GPU-accelerated iterative tight-frame reconstruction method that enforces projection consistency, tight-frame sparsity, and positivity. The method produces high-quality reconstructions across phantom and clinical cases, while TF iteration times are lower than TV iteration times in the reported comparisons.
Problem
Repeated CBCT scans can expose healthy organs to clinically concerning imaging dose, while fewer projections or lower mAs degrade conventional reconstructions from insufficient and noisy data.
Method
The paper develops an iterative CBCT algorithm that imposes projection consistency, tight-frame sparsity, and positivity, accelerated with GPU implementation and multi-grid computation.
Results
The TF algorithm yields high-quality CBCT results across NCAT, Catphan, and head-and-neck patient studies, with quantitative image-quality analyses confirming the reported quality.
Takeaways & Limitations
The reported results support TF-based reconstruction for CBCT with undersampled and noisy projections, including real clinical patient data.
Takeaways & Limitations
Comparisons between TF and TV are based on initial studies of only a few cases and are not conclusive.
Abstract
from arXiv · showhide
X-ray imaging dose from serial cone-beam CT (CBCT) scans raises a clinical concern in most image guided radiation therapy procedures. It is the goal of this paper to develop a fast GPU-based algorithm to reconstruct high quality CBCT images from undersampled and noisy projection data so as to lower the imaging dose. For this purpose, we have developed an iterative tight frame (TF) based CBCT reconstruction algorithm. A condition that a real CBCT image has a sparse representation under a TF basis is imposed in the iteration process as regularization to the solution. To speed up the computation, a multi-grid method is employed. Our GPU implementation has achieved high computational efficiency and a CBCT image of resolution 512\times512\times70 can be reconstructed in ~5 min. We have tested our algorithm on a digital NCAT phantom and a physical Catphan phantom. It is found that our TF-based algorithm is able to reconstrct CBCT in the context of undersampling and low mAs levels. We have also quantitatively analyzed the reconstructed CBCT image quality in terms of modulation-transfer-function and contrast-to-noise ratio under various scanning conditions. The results confirm the high CBCT image quality obtained from our TF algorithm. Moreover, our algorithm has also been validated in a real clinical context using a head-and-neck patient case. Comparisons of the developed TF algorithm and the current state-of-the-art TV algorithm have also been made in various cases studied in terms of reconstructed image quality and computation efficiency.
1. Introduction
Serial CBCT imaging can deliver substantial dose to healthy organs, while reducing projections or mAs degrades conventional reconstructions. This motivates methods for high-quality reconstruction from undersampled and noisy data with improved computational efficiency.
- Clinical motivation: A few cGy per 5 scan can expose healthy organs to clinically concerning CBCT imaging dose during repeated IGRT imaging.The concern is greatest when scans precede each treatment fraction.
- Clinical motivation: Reducing x-ray projections or mAs lowers CBCT imaging dose but leaves conventional FDK reconstructions highly degraded.The degradation arises from insufficient and noisy projection data.
- Related work: Compressed-sensing techniques have demonstrated recovery from incomplete measurements and have been applied to CT and CBCT reconstruction from undersampled data.The cited approaches establish feasibility for optimization-based reconstruction in these settings.
- Related work: Tight-frame methods offer high convergence rates and simple matrix-vector or vector operations that facilitate parallel implementation.These numerical properties support computationally efficient GPU execution.
2. Methods
The method reconstructs CBCT volumes by alternating data-consistency, tight-frame sparsity, and positivity operations. GPU parallelization, convolution-based transforms, CGLS updates, and an analytic projection expression accelerate the iterative computation.
- Model and algorithm: CBCT reconstruction from few projections is underdetermined, so the algorithm imposes three solution conditions iteratively.These conditions address projection consistency, tight-frame sparsity, and physically required positivity.
- Model and algorithm: CGLS updates minimize the projection-consistency objective to generate solutions matching the observed projections.The method uses iterative least-squares updates rather than directly solving the unstable underdetermined system.
- Model and algorithm: The reconstruction assumes sparse tight-frame coefficients and uses a piece-wise linear tight-frame basis with one low-pass and 26 high-pass 3D basis functions.The transform is computed through convolution, with high-frequency coefficients regularized by vector shrinkage.
- Model and algorithm: Voxel-wise vector shrinkage regularizes tight-frame coefficients, while negative attenuation values are truncated to zero.The algorithm alternates update, shrinkage, and correction steps until convergence.
- GPU implementation: GPU threads parallelize voxel-wise shrinkage, positivity correction, tight-frame convolutions, and matrix-vector operations.These operations apply the same computation independently across data elements and therefore suit GPU data parallelism.
- GPU implementation: An analytic projection expression avoids storing the approximately 16 GB sparse projection matrix and supports parallel computation of its transpose operation.The expression is evaluated using the x-ray source, imager, ray geometry, pixel size, and voxel size.
3. Experimental Results
Experiments on NCAT, Catphan, and patient data evaluated TF reconstruction under sparse-view and low-dose conditions. TF and TV generally produced high-quality images, while TF offered sharper structures and stronger spatial resolution in several comparisons, with performance depending on noise and regularization.
- NCAT phantom and Catphan phantom: 40 projections produced severe streak artifacts with FDK, whereas TF and TV reconstructed high-quality NCAT and Catphan images under extreme undersampling.The comparisons used a digital NCAT phantom and a physical Catphan phantom; the Catphan scan used 1.0 mAs/projection.
- NCAT phantom and Catphan phantom: TF and TV image quality was quite similar, but TF captured finer anatomical structures and sharper edges while TV produced smoother, more blurred images.TF also exhibited small residual streaks, whereas TV's explicit gradient penalty increased smoothness.
- Quantitative analysis: At 0.1 mAs/projection, the dotted Catphan structure was almost unresolved in both TF and TV reconstructions.The results show degradation in spatial resolution as projection noise increases at lower mAs levels.
- Quantitative analysis: The optimal TF regularization parameter depends on noise level, scanning parameters, reconstruction resolution, and the scanned object, and its precise relationship remains future work.The reported cases used parameter values selected as optimal for their settings.
4. Conclusion and Discussions
The paper presents a fast TF-based iterative CBCT reconstruction algorithm that achieves high-quality results from undersampled and noisy data, with GPU and multigrid acceleration. Comparisons with TV show similar image quality overall, distinct edge and artifact behavior, and different computational efficiency, while broader clinical comparison remains future work.
- Algorithm and validation: The TF-based algorithm reconstructs high-quality CBCT images from undersampled and noisy projection data using iterative TF regularization.The method assumes real CBCT images have sparse TF representations and applies three iterative conditions.
- Algorithm and validation: The algorithm was tested on digital NCAT, physical Catphan, and head-and-neck patient data, with image quality assessed using MTF under varied scanning conditions.The clinical case produced promising results, while systematic assessment across larger patient sets remains planned.
- TF versus TV: TF and TV produced quite similar image quality across the studied cases, but TF preserved sharper edges while often leaving residual artifacts during limited iterations.With more iterations, both methods can remove their characteristic artifacts satisfactorily; TV tends toward greater smoothness and edge blurring.
- TF versus TV: TV controls noise better than TF at low mAs, whereas TF maintains image contrast through unchanged low-frequency components.The comparison describes a trade-off between TF contrast preservation and TV noise suppression under different dose conditions.
- Computational efficiency: TF required 1.1, 4.1, and 15.2 sec per iteration across three multiscale levels, compared with about 1.8, 5.1, and 17.7 sec for TV.The difference is attributed mainly to deterministic TF-coefficient thresholding versus iterative ROF-model optimization in TV.
- TF versus TV: The TF–TV connection is related to high-pass filtering and smoothness, with a mathematical relationship established only under certain conditions.The paper presents the connection as interesting but does not make it the basis for a conclusive comparison.
1. CGLS algorithm
The CGLS algorithm solves the least-square reconstruction subproblem iteratively using conjugate gradients. In this CBCT method, it enforces data fidelity without requiring convergence at every outer TF iteration, preserving computational efficiency.
- CGLS algorithm: CGLS solves the least-square problem iteratively using a conjugate gradient method.The algorithm initializes an iterate and repeatedly updates it before outputting the solution.
- Role in reconstruction: Because few-projection CBCT reconstruction is underdetermined, CGLS is used to generate a solution from an initial guess while minimizing the residual norm.The paper uses CGLS to ensure the data-fidelity condition during each outer TF iteration.
- Role in reconstruction: Only a few CGLS steps are performed within each outer iteration because full convergence is unnecessary for enforcing data fidelity and would considerably slow computation.The paper states that this limited inner iteration is sufficient in practice.
2. Derivation of Eq. (3)
The derivation establishes the adjoint relationship between the x-ray projection and backprojection operators through an inner-product condition. It then uses functional variation and a delta-function reformulation to obtain the projection expression, with a discrete implementation caveat affecting CT-number accuracy.
- Adjoint operator: The backprojection operator is defined as the adjoint of the x-ray projection operator through an inner-product condition over image and projection domains.The condition is stated for smooth functions in the CBCT image and x-ray projection domains.
- Derivation: Functional variation of the operator identity and interchange of integration and variation produce the next equation in the derivation.This step connects the continuous inner-product formulation to the explicit operator expression.
- Derivation: A delta-function reformulation followed by substitution yields the explicit projection-equation form used for the CBCT operator.The resulting expression includes geometric quantities defined relative to the source and imager.
- Implementation: The implementation sums over projection angles to account for all x-ray projection images.The operator expression uses source-to-imager geometry and distances along the ray.
- Implementation caveat: Using vector rather than integral inner products introduces a pixel-size-to-voxel-size factor, and numerical error in the adjoint condition is below 1%.The resulting discretization may cause CT-number or Hounsfield Unit inaccuracy, although absolute CT-number accuracy is not crucial for patient setup.