SAM-3D-MSF: Parameter-Efficient Adaptation of Segment Anything Model for 3D Tooth CBCT Segmentation

Accurate 3D tooth segmentation from CBCT images is crucial for digital dentistry but faces challenges due to the volumetric nature of the data and limited annotations. While the Segment Anything Model (SAM) excels in 2D segmentation, a significant “dimensionality gap” hinders its direct application to 3D medical images. This paper proposes SAM-3D-MSF, a parameter-efficient framework that adapts SAM for 3D dental CBCT segmentation. Our approach introduces a novel 2D-to-3D adapter that transforms SAM’s pre-trained 2D convolutional kernels into 3D operations, enabling genuine volumetric processing while preserving pre-trained knowledge. Furthermore, a multi-scale feature fusion decoder is designed to enhance the segmentation of fine details and blurred boundaries typical in dental structures. Evaluated on the CTooth+ dataset, SAM-3D-MSF achieves state-of-the-art performance, with a Dice score of 91.04% and an IoU of 86.72%, significantly outperforming comparative methods. Notably, the model maintains high accuracy even with limited training data, demonstrating superior data efficiency. This work provides a feasible paradigm for adapting 2D foundation models to 3D medical image analysis with minimal computational overhead, offering significant potential for clinical applications.