ResCA-UNet: Residual coordinate attention enhanced UNet for nonlinear CT geometric calibration

In CT imaging, mechanical vibrations can cause unstable projection geometry, leading to geometric artifacts. This issue is critical in CT systems sensitive to perturbations, like C-arm CT systems. Offline calibration methods are limited by the precision of marker-based phantoms and the accuracy of marker positioning algorithms, and they do not allow for real-time calibration. Online calibration, however, is more suitable for random system perturbations, as it does not require special markers and can be performed in real-time. Among various methods, 3D–2D online calibration provides high accuracy but is computationally expensive. The 2D–2D online calibration model reduces computation time, though its accuracy still needs improvement. To address these challenges, this study proposes a nonlinear CT geometric calibration workflow using neural networks for efficient and accurate calibration. Specifically, the study introduces ResCA-UNet, an enhanced UNet network with a coordinate attention mechanism that extracts global information both horizontally and vertically, improves projection registration by focusing on key features, and enables precise calibration through a nonlinear 2D deformation model. Qualitative analysis shows that our method generates visually clearer images and provides more accurate projection parameter estimates compared to other 2D–2D calibration methods. Quantitative results demonstrate that our method outperforms other 2D–2D calibration methods in RMSE, SSIM, and NCC, while operating 200 times faster than the 3D-2D method and 5 times faster than iterative methods.