Super-resolution CBCT on a new generation flat panel imager of a C-arm gantry linear accelerator

Kilovoltage cone-beam computed tomography (kV CBCT) is vital for image-guided radiotherapy (IGRT). The new RTI4343iL panel on the Varian TrueBeam LINAC offers higher resolution but requires binning to achieve practical frame rates, leading to projection resolution loss. Existing super-resolution (SR) techniques have been applied to enhance CBCT image quality but primarily operate in the image domain, struggling to restore resolution loss in the projection domain.

This study aimed to evaluate the feasibility of a deep learning (DL) SR model, based on a conditional Generative Adversarial Networks (cGANs) architecture, for enhancing the spatial resolution of CBCT acquired with the new RTI4343iL panel in the projection domain. We hypothesize that projection-domain deblurring will primarily depend on the detector and minimally on patient anatomy, enhancing primary signal resolution without significantly altering scatter distribution. The study quantitatively assessed the impact of SR-enhanced projections on the quality of reconstructed CBCT images.

A DLSR model was developed to enhance CBCT resolution in the projection domain.  For data acquisition, a Varian TrueBeam system equipped with the RTI4343iL panel was used, which features a native high-resolution image size of 2848 × 2144 pixels, but operates in 2 × 8 binning mode (1424 × 268 pixels) during CBCT scans to mitigate data readout speed limitations. Following thorax CBCT protocols, 576 pairs of CBCT projections were acquired at two resolutions using Rando, Longman, and Steeve phantoms. Of these, 460 pairs were allocated for model training, while 116 were reserved for validation. Model testing involved 144 Dynamic Thorax projections and CBCT reconstructions utilizing Catphan 604 phantoms. The DL SR model was built on a cGANs framework with a U-Net generator. Image enhancement was quantitatively evaluated with metrics including peak signal-to-noise ratio (PSNR), mean square error (MSE), structural similarity index measure (SSIM), feature similarity index measure (FSIM), and mean absolute percentage error (MAPE).

The DL SR model effectively enhanced image resolution, producing SR projections with greater detail and improved structural clarity. Quantitative analysis showed that the SR-enhanced projections outperformed upscaled low-resolution (LR) projections with higher PSNR (44.4 vs. 43.7, p < 0.001), lower MSE (187,083.7 vs. 205,364.4, p < 0.001), and improved MAPE (7.6% vs. 13.5%, p < 0.001). While SSIM and FSIM values were similar for both methods, the SR-enhanced projections demonstrated a slight advantage, achieving an FSIM of 0.998. Reconstructed CBCT images from SR-enhanced projections exhibited improved spatial resolution as well, increasing from 0.6 lp/mm to 0.9 lp/mm on the Catphan 604 phantom image. Enhanced structural detail and sharper intensity profiles in SR CBCT images further validated the model's potential to restore resolution lost during the acquisition process.

This study underscores the efficacy of a projection-domain DL SR method for CBCT enhancement. The developed model presents a promising avenue for attaining high-resolution CBCT, potentially benefiting for many clinical applications.