µPIU-Net: A Domain-Specific Sinogram Infilling U-Net for Micro-CBCT and the Limitations of Generalized Models

Abstract

Abstract Sparse-view cone-beam computed tomography (CBCT) reduces radiation dose but produces streaking artefacts when reconstructed with conventional algorithms. Deep learning can help, yet it remains unclear whether general-purpose inpainting models pre-trained on natural images can transfer effectively to CT sinogram infilling. We introduce µPIU-Net, a lightweight U-Net for domain-specific micro-CBCT projection infilling, and compare it against five state-of-the-art inpainting models (LaMa, MAT, DeepFill v2, RePaint, Stable Diffusion 2). Evaluation uses both conventional metrics (SSIM, PSNR) and physical image quality assessment (MTF, NPS, NEQ). In the reconstruction domain, µPIU-Net doubles the SSIM of the undersampled baseline while general-purpose models fall below it—despite achieving reasonable sinogram-domain fidelity, demonstrating that sinogram-domain metrics do not reliably predict reconstruction quality. Physical image quality assessment reveals that µPIU-Net maintains noise texture (NPS) while improving contrast transfer (MTF), yielding signal detectability (NEQ) that exceeds the ground truth; these metrics must be interpreted together, as individual measures can be misleading. Our results highlight the importance of domain-specific training for specialised imaging tasks and the need for multi-metric evaluation in medical imaging applications.