Abstract
Efficient 3D cone-beam computed tomography (CBCT) reconstructions typically demand substantial hardware capabilities for both training and execution. This paper introduces a multi-filter and multi-scale U-Net (mU-Net) architecture designed for post-processing CBCT images by incorporating multiple input channels, each processed with distinct frequency filters or reconstruction resolutions. We present a detailed comparison between the mU-Net and a baseline U-Net for low-dose and clinical dose denoising scenarios, reporting key metrics such as Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index Measure (SSIM), and inference time. The mU-Net showed only marginal quantitative differences compared with the baseline U-Net: it performed slightly better in the clinical dose scenario, while the baseline U-Net performed better in the low-dose scenario. The mU-Net required approximately 13% more time per epoch during training and 13.08 s per volume for FDK reconstruction and inference, compared with 10.43 s per volume for the baseline U-Net. In the clinical dose scenario, the mU-Net showed slightly improved visual noise texture and artifact suppression, but these observations were not supported by dedicated texture or observer-study metrics. Notably, the mU-Net and baseline U-Net approaches placed 3rd in the ICASSP 2024 3D-CBCT grand challenge, demonstrating that a properly trained U-Net remains competitive. This study shows that multi-filter and multi-scale inputs can be used for CBCT post-processing under limited computational resources, but the gains over a baseline U-Net were modest.