Μpiu-Net: A Domain-Specific Sinogram Infilling U-Net for Micro-CBCT and the Limitations of Generalized Models
Abstract
Purpose
Sparse-view cone-beam computed tomography (CBCT) reduces radiation dose but produces streaking artifacts when reconstructed with conventional algorithms. Deep learning sinogram infilling can address this, but evaluation typically relies on conventional metrics that may not capture clinically relevant image characteristics. We present a metric analysis comparing domain-specific and domain-transferred models, examining how standard metrics can be misleading and why domain-specific training remains essential.
Methods
We developed µPIU-Net, a lightweight U-Net trained on micro-CBCT projection data. Given two adjacent X-ray projections, the network predicts the intermediate projection at the midpoint angle. We evaluated µPIU-Net alongside five general-purpose inpainting models (LaMa, MAT, DeepFill v2, RePaint, Stable Diffusion 2) pre-trained on natural images. All models were assessed on held-out mCTP 610 phantom data using conventional metrics (SSIM, PSNR) and physical image quality metrics (MTF, NPS, NEQ) in both sinogram and reconstruction domains.
Results
In the reconstruction domain, µPIU-Net achieved SSIM of 0.697, doubling the undersampled baseline (0.348). Domain-transferred models achieved SSIM of 0.081-0.107, performing worse than the baseline despite achieving reasonable sinogram-domain fidelity (SSIM 0.52-0.82), demonstrating that sinogram-domain metrics do not reliably predict reconstruction quality. Physical image quality assessment revealed that individual metrics can be misleading: RePaint's smoothing reduced NPS below ground truth, appearing beneficial, but degraded MTF resulted in worse NEQ than baseline.
Conclusion
Domain-specific training is essential for CT sinogram infilling---domain-transferred models achieve high sinogram fidelity yet produce poor reconstructions. Physical image quality metrics (MTF, NPS, NEQ) must be interpreted together to capture clinically relevant characteristics. The projection infilling approach enables 50% reduction in acquired projections, supporting dose reduction in preclinical micro-CT applications.