Image Quality and Tumor Shape Correction in Sparse Projection Cone-Beam CT Using Conditional Generative Adversarial Networks

PurposeTo reveal the effectiveness of image correction using a conditional generative adversarial network for sparse-view cone-beam computed tomography (CBCT) images, we generated pseudo-CBCT images with data projection angles ranging from 1 degrees to 20 degrees from CT images and evaluated the image quality and tumor shape reproducibility.MethodsCone beam projection was simulated using multi-slice CT images with the ray summation method, and pseudo-CBCT images were reconstructed for 15 patients. These projections were collected at each rotation angle of 1 degrees- 20 degrees to acquire a sparse-view CBCT image. The dataset comprised paired sparse pseudo-CBCT and original CT images. Overall, 14 and one cases were used in the training and validation datasets, respectively. The structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR) were calculated to evaluate image quality, and the Dice index evaluated tumor shape similarity in the corrected CBCT images.ResultsThe sparse-view CBCT image quality improved, resembling the original CT images. SSIM increased from an average of 0.11 to 0.80, and PSNR improved from 9.50 to 21.2 dB. The dice index improved from 0.79 to 0.82. However, data acquisition at larger sparse projection angles resulted in a blurring of reconstructed tumor shapes, indicating limitations in maintaining structural fidelity.ConclusionConditional generative adversarial networks-based correction significantly improves the image quality and tumor shape reproducibility of sparse-view CBCT images. However, higher projection angles result in structural blurring, suggesting limitations in maintaining tumor shape integrity at larger angles.