Orthotvnet: A Self-Supervised Denoising Method on Raw Mesh for Complex Dental Geometry

In recent years, artificial intelligence (AI) has propelled digital innovation in dentistry. Reconstructed crowns restore tooth geometry but often introduce reconstruction noise, which increases the demand for mesh denoising. Current methods are not designed to handle dental mesh with large curvature variation and complex occlusal topology. Moreover, these methods convert mesh into graph and thus lose geometric information, and the high cost of paired high-quality labeled datasets makes supervised approaches impractical. To address these challenges, we propose a self-supervised framework that directly analyzes and denoises raw mesh without graph conversion or paired labels. Our method integrates orthogonal-channel attention, inspired by signal processing principles of orthogonality, to sharpen features and let the network separate anatomical details from noise, analogous to signal separation. Experimental results confirm that our method reduces noise and restores curvature more accurately than existing graph-based methods, demonstrating its effectiveness for dental mesh denoising.