Purpose: Recently, clinical use of MV cone beam CT (CBCT) for radiotherapy guidance has increased. However, even in systems capable of lower MV energies, image quality suffers from detection efficiency limitations in traditional x-ray detectors. By combining a high DQE multi-layer detector prototype with dose-efficient, iterative Penalized Weighted Least Squares (PWLS) reconstruction, improved MV-CBCT image quality is pursued.
Methods: The proposed imager consists of 4 detector layers, each with a 0.506 mm GOS scintillator and a 0.7 mm photodiode sub-layer. Copper blocking layers are omitted to improve DQE at lower (2.5 MV) energies. The iterations of PWLS are accelerated by using the Ordered Subsets Momentum-Based (OSMOM) technique to enhance convergence rate, and the public-domain Tomographic Iterative GPU-based Reconstruction (TIGRE) library to alleviate forward/back projection time. Refinements to TIGRE were made that allowed it to communicate input/output as MATLAB gpuArray objects. This enabled all iterative computations to be executed within MATLAB on the GPU, eliminating overhead from CPU-to-GPU transfers. The overall technique is evaluated in terms of noise and contrast-to-noise (CNR) in scans of the Catphan 604, and qualitatively in a pelvis phantom.
Results: The proposed technique demonstrated 3 times lower noise as a function of spatial resolution compared to a conventional MV detector architecture and reconstruction algorithm (FDK). Reconstructed noise levels closely approached that of the Varian Edge LINAC’s clinical kV-CBCT Spotlight protocol at matched dose (CTDI=12.5 mGy). This resulted in CNR about half that of kV-CBCT in a soft tissue (LDPE) region of interest. On the GTX 1080 Ti graphics card, reconstruction time for the pelvis was about 30 sec. using a 400x400x170 voxel grid and 200x320x400 projections.
Conclusion: Combining a multi-layer imager with PWLS reconstruction greatly diminishes the image quality gap between MV-CBCT and standard kV-CBCT. Moreover, short reconstruction times are implementable with MATLAB and non-proprietary GPU software.
Funding Support, Disclosures, and Conflict of Interest: This work was partially supported by award number R01CA188446 from the National Institute of Health.