A method for detecting 3D tooth landmarks based on orthogonal decoupling and asymmetric bi-
branch architecture includes: acquiring raw 3D tooth
point cloud data from a 3D
dental scanning device; extracting high-resolution local point clouds of the
region of interest for a
single tooth and low-resolution global point clouds of the entire
dentition; performing PCA geometric
perception serialization on the low-resolution global
point cloud to generate an ordered
point cloud sequence that conforms to the prior anatomical structure of the
dental arch; constructing a local neighborhood topology structure for the high-resolution local point cloud through graph
convolution to obtain input feature representations for the global and local branches respectively; feeding the feature representations of the global and local branches into a bidirectional state-
space model and a graph convolutional coding network respectively to extract global topological features of the entire
dentition and local geometric features of a
single tooth; feeding the global topological features and local geometric features into an orthogonal gating fusion module, achieving clean
feature fusion by removing redundant components through orthogonal projection, obtaining robust feature representations after fusion, and constructing a geometric heterogeneity hierarchical supervision
system to complete end-to-end model training; and passing the fused feature representations to a prediction head to generate tooth
anatomy category prediction results and 3D... Sub-
millimeter-level precise coordinates of tooth landmarks are used for
clinical decision-making in oral healthcare, such as digital
orthodontics, virtual tooth alignment, and orthognathic
surgery planning.