GPU-based Fast Low Dose Cone Beam CT Reconstruction via Total Variation

Xun Jia, Yifei Lou, R. Li, John H. Lewis, C Men, Xuejun Gu, W. Y. Song, S JiangInternational Journal of Radiation Oncology*Biology*Physics 2010

Abstract

Cone-beam CT (CBCT) has been widely used in image guided radiation therapy (IGRT) to acquire updated volumetric anatomical information before treatment fractions for accurate patient alignment purpose. However, the excessive x-ray imaging dose from serial CBCT scans raises a clinical concern in most IGRT procedures. The excessive imaging dose can be effectively reduced by reducing the number of x-ray projections and/or lowering mAs levels in a CBCT scan. The goal of this work is 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. The CBCT is reconstructed by minimizing an energy functional consisting of a data fidelity term and a total variation regularization term. We developed a GPU-friendly version of the forward-backward splitting algorithm to solve this model. A multi-grid technique is also employed. We test our CBCT reconstruction algorithm on a digital NCAT phantom and a head-and-neck patient case. The performance under low mAs is also validated using a physical Catphan phantom and a head-and-neck Rando phantom. It is found that 40 x-ray projections are sufficient to reconstruct CBCT images with satisfactory quality for IGRT patient alignment purpose. Phantom experiments indicated that CBCT images can be successfully reconstructed with our algorithm under as low as 0.1 mAs/projection level. Comparing with currently widely used full-fan head-and-neck scanning protocol of about 360 projections with 0.4 mAs/projection, it is estimated that an overall 36 times dose reduction has been achieved with our algorithm. Moreover, the reconstruction time is about 130 sec on an NVIDIA Tesla C1060 GPU card, which is estimated ~100 times faster than similar iterative reconstruction approaches.

Introduction

In Image-Guided Radiation Therapy (IGRT), clinicians rely on Cone-Beam CT (CBCT) scans to ensure precise patient alignment before each treatment fraction. However, because these scans are performed daily, the cumulative radiation dose to healthy tissue has become a significant clinical concern.

Reducing this dose typically involves two strategies: decreasing the number of x-ray projections (undersampling) or lowering the tube current-time product (mAs) for each projection. Unfortunately, standard reconstruction algorithms, such as the Feldkamp-Davis-Kress (FDK) method, produce severe streak artifacts and high noise levels when data is sparse or low-power. While iterative methods like Total Variation (TV) regularization can produce high-quality images from such data, their computational intensity has historically limited their clinical use, often requiring hours for a single reconstruction.

Methodology

To bridge the gap between image quality and clinical speed, this research introduces a GPU-accelerated reconstruction framework based on TV regularization.

System Geometry and Model

The reconstruction process is modeled based on the geometric relationship between the x-ray source, the object volume, and the detector plane. This geometry is essential for defining the forward projection (mapping the 3D volume to 2D projections) and the back-projection (mapping 2D data back into 3D space).

Geometry of the Cone-Beam CT (CBCT) system illustrating the relationship between the source (S), the object volume f(x,y,z), and the detector plane (P).

Optimization Strategy

The algorithm minimizes an energy functional that balances data fidelity (how well the reconstruction matches the projections) with a TV regularization term. The TV term assumes that anatomical images are piecewise constant, which helps suppress noise while preserving sharp edges at tissue boundaries.

Key technical innovations include:

  1. GPU-Friendly Forward-Backward Splitting: The researchers derived an analytical closed-form gradient for the model. This allows the GPU to process voxels independently and in parallel, avoiding the memory "locks" and conflicts that typically slow down iterative back-projection.
  2. Multi-grid Technique: The algorithm employs a 4-level hierarchy, starting with a coarse grid to find a rough solution quickly and then upsampling to finer resolutions. This significantly accelerates convergence toward the final high-resolution image.

Experimental Validation

The algorithm was validated using a variety of datasets, including the digital NCAT phantom for ground-truth comparison, the Catphan 600 for contrast and resolution metrics, and the Rando head phantom for realistic human geometry. Clinical data from a head-and-neck cancer patient was also used to verify real-world performance.

Key Results

  • Significant Dose Reduction: The study demonstrated that just 40 projections are sufficient for accurate patient alignment. By combining this 9-fold reduction in projections with a 4-fold reduction in mAs (down to 0.1 mAs/projection), the researchers achieved a 36-fold total dose reduction compared to standard protocols.
  • High-Speed Performance: Using an NVIDIA Tesla C1060 GPU, the reconstruction was completed in approximately 130 seconds. This represents a 100-fold speedup over traditional CPU-based iterative methods, making the process practical for real-time clinical workflows.
  • Image Fidelity: Despite the ultra-low dose, the correlation coefficient between the reconstructed images and ground truth remained above 0.99, with a notable absence of the streaking artifacts common in FDK reconstructions.

Conclusion

This work demonstrates that ultra-low-dose CBCT is clinically viable through the use of optimized iterative reconstruction. By solving the computational bottleneck with a GPU-friendly mathematical derivation and multi-grid acceleration, high-quality 3D imaging can now be performed in roughly two minutes, providing a safer and more efficient path for Image-Guided Radiation Therapy.

Summary updated

Abstract

Cone-beam CT (CBCT) has been widely used in image guided radiation therapy (IGRT) to acquire updated volumetric anatomical information before treatment fractions for accurate patient alignment purpose. However, the excessive x-ray imaging dose from serial CBCT scans raises a clinical concern in most IGRT procedures. The excessive imaging dose can be effectively reduced by reducing the number of x-ray projections and/or lowering mAs levels in a CBCT scan. The goal of this work is 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. The CBCT is reconstructed by minimizing an energy functional consisting of a data fidelity term and a total variation regularization term. We developed a GPU-friendly version of the forward-backward splitting algorithm to solve this model. A multi-grid technique is also employed. We test our CBCT reconstruction algorithm on a digital NCAT phantom and a head-and-neck patient case. The performance under low mAs is also validated using a physical Catphan phantom and a head-and-neck Rando phantom. It is found that 40 x-ray projections are sufficient to reconstruct CBCT images with satisfactory quality for IGRT patient alignment purpose. Phantom experiments indicated that CBCT images can be successfully reconstructed with our algorithm under as low as 0.1 mAs/projection level. Comparing with currently widely used full-fan head-and-neck scanning protocol of about 360 projections with 0.4 mAs/projection, it is estimated that an overall 36 times dose reduction has been achieved with our algorithm. Moreover, the reconstruction time is about 130 sec on an NVIDIA Tesla C1060 GPU card, which is estimated ~100 times faster than similar iterative reconstruction approaches.
BibTeX
@article{jia2010gpu,
  author = {Xun Jia and Yifei Lou and R. Li and John H. Lewis and C Men and Xuejun Gu and W. Y. Song and S Jiang},
  title = {{GPU-based Fast Low Dose Cone Beam CT Reconstruction via Total Variation}},
  journal = {International Journal of Radiation Oncology*Biology*Physics},
  year = {2010},
}
Figure