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

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.

💡 Research Summary

This paper addresses the pressing clinical need to reduce imaging dose in image‑guided radiation therapy (IGRT) while preserving the volumetric image quality required for accurate patient alignment. Conventional cone‑beam CT (CBCT) protocols typically acquire 300–400 projections at relatively high tube current‑exposure (mAs), resulting in a substantial cumulative dose when scans are repeated over a treatment course. The authors propose a fast, GPU‑accelerated iterative reconstruction algorithm that can produce high‑quality CBCT images from severely undersampled and noisy projection data, thereby enabling dose reductions on the order of 30–40×.

The reconstruction problem is formulated as the minimization of an energy functional consisting of a data‑fidelity term (the squared ℓ₂ norm of the difference between measured projections and forward‑projected estimates) and a total variation (TV) regularization term that promotes piecewise‑smooth images while preserving edges. Mathematically, the objective is
 E(f) = ‖A f – p‖₂² + λ TV(f),
where A denotes the forward projection operator, p the measured sinogram, f the unknown 3‑D volume, and λ a regularization weight. The TV term mitigates the amplified noise inherent in low‑dose, few‑projection data, preventing the stair‑casing artifacts that would otherwise dominate a simple least‑squares solution.

To solve this non‑smooth convex problem efficiently, the authors adopt a forward‑backward splitting (FBS) scheme. In each iteration, a gradient descent step on the smooth data‑fidelity term is followed by a proximal mapping of the TV term. The gradient step requires forward projection and back‑projection, operations that are highly parallelizable on modern graphics processing units (GPUs). The proximal TV step is implemented using a fast gradient‑projection (FGP) algorithm (or equivalently the Chambolle‑Pock method), which also maps naturally onto a 2‑D thread block layout. By carefully arranging memory accesses (coalesced reads/writes, shared‑memory buffering) the authors achieve near‑peak throughput on an NVIDIA Tesla C1060 (≈1 TFLOP).

A multigrid strategy further accelerates convergence. The reconstruction starts on a coarse voxel grid (e.g., 64³), where the FBS iterations rapidly reduce the bulk error. The coarse solution is then interpolated to a finer grid (e.g., 256³) and refined with additional FBS cycles. This hierarchical approach reduces the total number of expensive high‑resolution back‑projections by a factor of 5–10, making the algorithm feasible for clinical volumes without exceeding GPU memory limits.

Experimental validation comprises three parts: (1) a digital NCAT phantom, (2) physical Catphan and Rando head‑and‑neck phantoms scanned at 0.1 mAs per projection, and (3) a real patient case involving a head‑and‑neck tumor. In all scenarios the authors used only 40 equally spaced projections (≈1° angular spacing), a dramatic reduction from the standard 360‑projection protocol. Quantitative metrics—root‑mean‑square error (RMSE), structural similarity index (SSIM), and contrast‑to‑noise ratio (CNR)—show that the TV‑regularized reconstructions achieve RMSE ≈ 2 HU, SSIM > 0.92, and CNR improvements of 3–4× relative to filtered back‑projection (FBP) under the same low‑dose conditions. Importantly, alignment errors measured after importing the reconstructed volumes into a treatment planning system remain below 0.3 mm translation and 0.2° rotation, well within clinical tolerances (≤1 mm, ≤1°).

Performance measurements reveal a total reconstruction time of approximately 130 seconds on the Tesla C1060, which the authors estimate to be about 100× faster than comparable CPU‑based TV‑regularized iterative methods that typically require several hours for a full 3‑D CBCT volume. This speed brings the method into the realm of practical clinical use, where a reconstruction can be completed while the patient remains on the treatment couch.

The discussion acknowledges that TV regularization, while powerful, can oversmooth fine anatomical details if the regularization weight λ is set too high. Consequently, adaptive or data‑driven λ selection (e.g., via L‑curve analysis or Bayesian approaches) is identified as a future research direction. Moreover, the current implementation assumes a circular, evenly spaced trajectory; extending the framework to non‑circular or patient‑specific trajectories, as well as to dual‑energy or spectral CBCT, would broaden its applicability.

In conclusion, the paper delivers a compelling solution to the low‑dose CBCT problem: by integrating a mathematically sound TV‑regularized model with a GPU‑optimized forward‑backward splitting algorithm and a multigrid acceleration scheme, the authors achieve clinically acceptable image quality with as few as 40 projections and as low as 0.1 mAs per projection. This translates to an estimated 36‑fold reduction in imaging dose compared with standard protocols, while maintaining reconstruction times compatible with routine IGRT workflows. The work paves the way for routine low‑dose CBCT in modern radiotherapy, potentially improving patient safety without compromising treatment accuracy.