HARU-Net: Hybrid Attention Residual U-Net for Edge-Preserving Denoising in Cone-Beam Computed Tomography

Researchers Khuram Naveed and Ruben Pauwels have introduced HARU-Net (Hybrid Attention Residual U-Net), a novel deep learning architecture specifically designed for edge-preserving denoising in cone-beam computed tomography (CBCT) used in dental and maxillofacial imaging. The model addresses a critical challenge in medical imaging: low-dose CBCT acquisition introduces strong, spatially varying noise that degrades soft-tissue visibility and obscures fine anatomical structures, while existing methods struggle to balance noise suppression with edge preservation. HARU-Net was trained on a cadaver dataset of human hemimandibles acquired using a high-resolution protocol of the 3D Accuitomo 170 CBCT system, overcoming the typical limitation of scarce high-resolution CBCT data for supervised training.

The technical innovation lies in HARU-Net's integration of three complementary components: hybrid attention transformer blocks (HAB) embedded within skip connections to selectively emphasize salient anatomical features, residual hybrid attention transformer groups (RHAG) at the bottleneck to strengthen global contextual modeling, and residual learning convolutional blocks throughout the network for stable feature extraction. This architecture consistently outperforms state-of-the-art methods including SwinIR and Uformer, achieving the highest PSNR (37.52 dB), highest SSIM (0.9557), and lowest GMSD (0.1084) metrics. Crucially, this superior performance comes at significantly lower computational cost than competing methods, making it a practical advancement toward improving diagnostic quality in clinical low-dose CBCT imaging while maintaining computational efficiency.