A segmentation method for oral CBCT image based on Segment Anything Model and semi‐supervised teacher‐student model

• DOI: 10.1002/mp.17854
• Publication Date: 2025-05-08T06:08:56Z
• Authors: Jianhong Gan; Runqing Kang; Xun Deng; Tongli He; Nie Yu; Yuling Gan; Peiyang Wei; Xiangyi Chen; Xiaoli Peng; Zhibin Li
• Keywords: Transfer of learning

Abstract Background Accurate segmentation of oral cone beam computed tomography (CBCT) images is essential for research and clinical diagnosis. However, irregular blurred tooth boundaries in CBCT complicate the labeling tissues, insufficient labeled samples further limit generalization ability models. The Segment Anything Model (SAM) demonstrates strong accuracy across diverse tasks as a vision foundation model. Teacher‐Student (TS) model has proven effective semi‐supervised learning approaches. Purpose To accurately segment various parts CBCT, such enamel, pulp, bone, blood vessels, air, etc., an improved method named SAM‐TS proposed, which combines SAM with TS leverages Low‐Rank Adaptation (LoRA) to fine‐tune on fewer parameters. Methods efficiently utilize numerous unlabeled training models, LoRA strategy SAM. fine‐tuned teacher models collaboratively generate pseudo‐labels images, are filtered utilized train student Then, data augmentation‐based Mean Intersection over Union (MIoU) proposed filter out unreliable or spurious pseudo‐labels. Finally, Exponential Moving Average (EMA) used transfer model's parameters After repeating this process, final optimized obtained. experimental results demonstrate that incorporating into through significantly enhances accuracy. Results Compared baseline algorithm, achieves overall improvement 6.48% MIoU. In task, minimum MIoU maximum increased by at least 10% 27.32%, respectively. bone 7.9% 32.44%, Additionally, segmentation, Hausdorff distance (HD) decreased 5.1 mm, Dice coefficient 2.87%. Conclusion outperforms existing methods, offering more competitive efficient approach image segmentation. This addresses annotation bottleneck opens new avenues applications medical imaging.