HARU-Net: hybrid attention residual U-Net for edge-preserving denoising in cone-beam computed tomography
Synopsis
This study proposes HARU-Net, a hybrid attention residual U-Net trained on a cadaver dataset of human hemimandibles acquired with a high-resolution cone-beam computed tomography (CBCT) protocol for low-dose CBCT denoising; the architecture embeds a hybrid attention transformer block within each skip connection, a residual hybrid attention transformer group at the bottleneck, and residual learning convolutional blocks, and reports the highest peak signal-to-noise ratio of 37.52 dB, the second-highest SSIM of 0.9557, and the lowest GMSD of 0.1084, while maintaining substantially lower computational complexity than transformer-based methods.
Interpretation
Proposes HARU-Net, integrating three complementary architectural components into a U-Net framework: hybrid attention transformer blocks in skip connections, a residual hybrid attention transformer group at the bottleneck, and residual learning convolutional blocks. Relative to existing CBCT denoising methods, this design combines attention mechanisms with residual learning to selectively emphasize salient anatomical features and strengthen global contextual modeling. Based on training and evaluation on a high-resolution CBCT dataset of human hemimandible cadaver specimens, with reported quantitative metrics.
HARU-Net consistently outperforms state-of-the-art methods, achieving the highest PSNR (37.52 dB), the second-highest SSIM (0.9557), and the lowest GMSD (0.1084). Compared with transformer-based methods, it achieves superior denoising performance while maintaining substantially lower computational complexity. Quantitative metrics come from the reported experiments, but the abstract does not provide sample size, statistical tests, or full comparison details.
The framework provides an effective balance between noise suppression, anatomical structure preservation, and computational efficiency, and is promising for improving image quality and supporting reliable diagnosis in low-dose CBCT imaging. Addresses the challenge of strong spatially varying noise in low-dose CBCT, where classical methods struggle to suppress noise while preserving edges, by offering a deep learning solution. Conclusion is based on the reported experimental results, but clinical deployment effects require further validation.
Perspective
This work targets low-dose CBCT denoising, aiming to suppress noise while preserving fine anatomical structures and maintaining computational efficiency, applicable to dental and maxillofacial imaging scenarios. Its training data are high-resolution CBCT scans of human hemimandible cadaver specimens, so results primarily apply to this imaging setup and similar anatomical regions.
The abstract does not provide sample size, statistical tests, a full list of comparison methods, or specific numerical values for computational complexity, so the robustness and generalizability of the performance advantages require further understanding. Additionally, differences between cadaver specimen data and in vivo clinical scans may affect practical deployment.