Development of an Unpaired Deep Neural Network for Synthesizing X-ray Fluoroscopic Images from Digitally Reconstructed Tomography in Image Guided Radiotherapy.

OBJECTIVE: We developed and evaluated a deep neural network (DNN) to generate flat-panel detector (FPD) images from digitally reconstructed radiography (DRR) in lung cancer treatment, aiming to streamline image-guided radiotherapy workflows. METHODS: A modified CycleGAN was trained on over 400 paired DRR-2D-FPD images from lung tumor patients and tested on 100 2D-FPD images. Mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index measure (SSIM), and Kernel Inception Distance (KID) assessed the synthetic images' similarity to ground truth. Generation time was also recorded. RESULTS: Despite some positional mismatches, synthetic 2D-FPD images closely resembled ground-truth images. The DNN outperformed both DRR inputs and a U-Net-based method in MAE, PSNR, SSIM, and KID. Image generation averaged milliseconds per image. Qualitatively, the DNN replicated realistic noise patterns, potentially reducing the need for manual noise adjustments. CONCLUSIONS: This DNN efficiently converts DRR to realistic 2D-FPD images for thoracic cases, potentially enhancing patient setup verification and overall clinical workflow. Future work should validate this approach across various imaging systems and address marker visualization challenges for broader clinical adoption. ADVANCES IN KNOWLEDGE: This novel CycleGAN-based method rapidly produces realistic 2D-FPD images from DRR, generating noise patterns that closely resemble those of real images and outperforming traditional methods, and may thereby help streamline clinical workflows.