Related Experiment Video
Updated: May 12, 2025
09:10
Digital Hybrid Model Preparation for Virtual Planning of Reconstructive Dentoalveolar Surgical Procedures
Published on: August 5, 2021
1.7K
A segmentation method for oral CBCT image based on Segment Anything Model and semi-supervised teacher-student model.
Jianhong Gan1,2,3,4, Runqing Kang1,2,3, Xun Deng1
1College of Software Engineering, Chengdu University of Information Technology, Chengdu, China.
Medical Physics
|May 9, 2025
Summary
The novel SAM-TS method enhances oral CBCT image segmentation by combining the Segment Anything Model (SAM) with a Teacher-Student (TS) approach. This improves accuracy and generalization, addressing data limitations in medical imaging.
Area of Science:
- Medical Imaging
- Computer Vision
- Artificial Intelligence
Background:
- Accurate segmentation of oral cone beam computed tomography (CBCT) images is crucial for clinical diagnosis and research.
- Challenges include irregular tooth boundaries and limited labeled data, hindering segmentation model generalization.
- The Segment Anything Model (SAM) offers strong generalization, while Teacher-Student (TS) models excel in semi-supervised learning.
Purpose of the Study:
- To develop an improved oral CBCT segmentation method, SAM-TS, integrating SAM with TS models.
- To leverage Low-Rank Adaptation (LoRA) for efficient fine-tuning of SAM on limited oral CBCT data.
- To accurately segment diverse oral tissues including enamel, pulp, bone, and blood vessels.
Main Methods:
- Fine-tuning SAM using an improved LoRA strategy for efficient utilization of unlabeled CBCT images.
- Collaborative pseudo-label generation by fine-tuned SAM and teacher models, followed by filtering using a data augmentation-based Mean Intersection over Union (MIoU) method.
- Iterative training of a student model using filtered pseudo-labels and parameter updates via Exponential Moving Average (EMA) for teacher model enhancement.
Main Results:
- SAM-TS achieved a significant overall Mean Intersection over Union (MIoU) improvement of over 6.48% compared to baseline methods.
- Tooth segmentation MIoU increased by at least 10% (minimum) and 27.32% (maximum).
- Bone segmentation MIoU saw minimum and maximum increases of 7.9% and 32.44%, respectively, with improved Hausdorff distance and Dice coefficient for overall segmentation.
Conclusions:
- SAM-TS demonstrates superior performance over existing semi-supervised methods for CBCT image segmentation.
- The approach effectively overcomes the data annotation bottleneck in medical imaging.
- This work presents a competitive and efficient semi-supervised learning strategy for medical image analysis.
