ToothSC-SAM: A Novel Network Model Based on Skip-Connections and SAM for Tooth Segmentation in CBCT Images

This study focuses on the major challenges in addressing the critical limitation of requiring extensive annotated datasets for tooth segmentation from cone-beam computed tomography (CBCT) images, which are essential for clinical patient treatment. Through the modeling process, we found that there are two bottlenecks in the existing deep learning methods: First, the labeling time of a single CBCT case is high, which seriously restricts the clinical application; Second, the number of labeled data influences the performance of the model significantly. Hence, in this paper, we propose a novel two-stage prompt-based network, named ToothSC-SAM, which integrates 3D-SAM with skip-connections for high-resolution tooth segmentation with minimal annotation requirements. The Network first extracts the ROI (Region of Interest) of teeth in the first stage. Then, in the second stage, we use the dots within the ROI as a prompt and send it to a 3D-SAM with skip-connection network for precise tooth segmentation. Finally, we implement tooth labeling and restore CBCT image size through the position provided by the prompt to achieve high-resolution tooth segmentation and labeling. Our method outperforms the SAM approach that directly processes CBCT images by introducing a simple additional prompt step. Moreover, the proposed method performance is approximately 93% of that achieved under the complete supervision baseline, while the annotation cost has been reduced from several hours to just a few minutes. These results highlight the network’s potential to transform dental image analysis by significantly reducing the annotation burden while maintaining clinical-grade accuracy.