Attention-UNet synthesizes cone-beam projections from CBCT in head and neck cancer patients

Attention-UNet synthesized cone-beam projections from CBCT with PSNR 29.71, SSIM 0.97, VQM 0.20 in 50 head-and-neck cancer patients (retrospective); scatter correction improved PSNR by 21.3%. Method may conserve storage for CB projections.

• Retrospective study of 50 head and neck cancer patients; three DL networks (Attention-UNet, Residual AutoEncoder, Pix2Pix) were compared.
• Attention-UNet achieved the best performance (PSNR 29.71, SSIM 0.97, VQM 0.20); scatter correction further improved PSNR by 21.3%.
• Single-institution, no external validation, and modest sample size limit generalizability.

RadPigeon summaries are original and for information only. They are not clinical advice.