Deep learning for dentomaxillofacial cone-beam computed tomography enhancement: A systematic review and meta-analysis

Background Cone-beam computed tomography (CBCT) is widely used in dentomaxillofacial imaging and radiotherapy but is limited by noise and low contrast resolution. Deep learning (DL) has emerged as a promising tool for CBCT enhancement, improving image quality and diagnostic accuracy. This systematic review and meta -analysis evaluated DL-based techniques for dentomaxillofacial CBCT enhancement and their impact on objective image quality metrics. Methods A systematic search across six databases identified studies applying DL for CBCT enhancement. Performance metrics included mean absolute error (MAE), peak signal-to-noise ratio (PSNR), structural similarity index (SSIM), and root mean squared error (RMSE). Risk of bias was assessed using the modified QUADAS-2 tool. Meta-analyses were performed using random-effects models, with effect sizes expressed as standardized mean differences (SMD) for MAE and RMSE and mean differences (MD) for PSNR and SSIM. Heterogeneity was evaluated using Cochran's Q and I 2 statistics. Results Thirty-seven studies met inclusion criteria, covering CBCT-to-CT synthesis (27), metal artifact reduction (4), noise reduction (3), motion artifact reduction (1), and super-resolution (2). DL significantly reduced MAE (SMD: −4.24, 95% CI: −5.68 to −2.80, p < 0.0001) and RMSE (SMD: −1.21, 95% CI: −1.55 to −0.88, p < 0.0001), while improving PSNR (MD: 3.60, 95% CI: 3.00 to 4.20, p < 0.0001) and SSIM (MD: 0.072, 95% CI: 0.055 to 0.089, p < 0.0001). High heterogeneity (I 2 > 99%) limits direct comparisons. Conclusion DL improves CBCT image quality, but methodological variability and limited clinical validation necessitate further research.