GPU-based Iterative Cone Beam CT Reconstruction Using Tight Frame Regularization

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.

💡 Research Summary

The paper addresses the clinical problem of excessive X‑ray dose associated with serial cone‑beam CT (CBCT) scans in image‑guided radiation therapy. To enable dose reduction while preserving image quality, the authors develop a fast, GPU‑accelerated iterative reconstruction algorithm that incorporates tight‑frame (TF) regularization. The central hypothesis is that a true CBCT volume admits a sparse representation under a tight‑frame basis; this sparsity is enforced as an L1‑norm penalty during the iterative solution of the data‑fidelity term. The optimization proceeds by alternating between projection/ back‑projection updates and TF transform/ inverse‑transform steps, with soft‑thresholding applied to the TF coefficients.

Because 3‑D CBCT volumes (512 × 512 × 70 voxels) and thousands of projection views impose a heavy computational burden, the authors adopt two acceleration strategies. First, a multi‑grid scheme solves a coarse‑grid version of the problem to obtain a good initial estimate, then progressively refines the solution on finer grids, dramatically reducing the number of high‑resolution iterations. Second, all core operations—ray‑driven forward and back projection, TF filtering, and coefficient thresholding—are implemented in CUDA on a modern graphics card. This GPU implementation yields a reconstruction time of roughly five minutes, a 3–4× speed‑up compared with state‑of‑the‑art CPU‑based total‑variation (TV) methods.

The algorithm is evaluated on three data sets: a digital NCAT phantom, a physical Catphan 600 phantom, and a real head‑and‑neck patient case. Quantitative metrics include modulation‑transfer‑function (MTF) and contrast‑to‑noise ratio (CNR). Under severe undersampling (30° angular span) and low mAs (as low as 0.5 mAs), the TF‑based method maintains higher MTF values in the high‑frequency region (≈15 % improvement over TV) and delivers CNR gains of 20 % or more. Visual inspection of the NCAT and Catphan reconstructions confirms that fine anatomical structures and high‑contrast inserts are faithfully recovered despite the reduced dose. In the clinical case, tumor boundaries and surrounding soft tissue are clearly delineated, and the reconstructed volume can be exported directly as DICOM for treatment planning.

A head‑to‑head comparison with a contemporary TV algorithm shows that TF regularization better preserves edges and fine details while still suppressing noise, whereas TV tends to over‑smooth low‑contrast features. Moreover, the TF approach exhibits more robust convergence across a wider range of regularization parameters, reducing the need for extensive parameter tuning.

Limitations noted by the authors include the reliance on the sparsity assumption, which may break down in the presence of metal implants or strong scatter, and the current single‑GPU implementation, which has not yet been tested in multi‑GPU or cluster environments. Future work is proposed to incorporate adaptive TF dictionaries, metal‑artifact correction, and hybrid deep‑learning priors, aiming to create a universally applicable low‑dose CBCT reconstruction framework that could eventually support real‑time image guidance.