Structure-Aware 3D Tooth Modeling via Prompt-Guided Segmentation and Multi-View Projection

Precise and modular reconstruction of 3D tooth structures is crucial for creating interpretable, adaptable models for digital dental applications. To address the limitations of conventional segmentation approaches under conditions such as missing teeth, misalignment, or incomplete anatomical structures, we propose a process-oriented reconstruction pipeline composed of discrete yet integrated modules. The pipeline begins by decomposing 3D dental meshes into a series of 2D projections, allowing multi-view capture of morphological features. A fine-tuned Segment Anything Model (SAM), enhanced with task-specific bounding box prompts, performs segmentation on each view. T-Rex2, a general object detection module, enables automated prompt generation for high-throughput processing. Segmented 2D components are subsequently reassembled and mapped back onto the original 3D mesh to produce complete and anatomically faithful tooth models. This modular approach enables clear separation of tasks—view projection, segmentation, and reconstruction—enhancing flexibility and robustness. Evaluations on the MICCAI 3DTeethSeg’22 dataset show comparable or superior performance to existing methods, particularly in challenging clinical scenarios. Our method establishes a scalable, interpretable framework for 3D dental modeling, supporting downstream applications in simulation, treatment planning, and morphological analysis.