Related Experiment Video
Updated: Mar 29, 2026
14:08
Automated Midline Shift and Intracranial Pressure Estimation based on Brain CT Images
Published on: April 13, 2013
43.8K
Semi-Supervised Vertebra Segmentation and Identification in CT Images
You Fu1, Jiasen Feng2, Hanlin Cheng3
1School of Information Technology, Murdoch University, Murdoch 6150, Australia.
Tomography (Ann Arbor, Mich.)
|March 27, 2026
Summary
This study introduces a semi-supervised deep learning method for automatic vertebra segmentation and identification in spinal CT scans. The approach significantly improves accuracy without needing more labeled data, aiding clinical diagnosis and surgical planning.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Computer-Aided Diagnosis
Background:
- Automatic segmentation and identification of vertebrae in spinal CT are crucial for diagnosing spinal disorders and surgical planning.
- Challenges include high structural similarity between vertebrae and morphological variability, limiting current supervised deep learning methods due to annotation constraints.
- Existing methods struggle with robustness in complex clinical scenarios.
Purpose of the Study:
- To develop a robust semi-supervised deep learning approach for vertebra segmentation and identification.
- To leverage unlabeled data to overcome limitations of fully supervised methods.
- To enhance automated clinical workflows for spinal disorder diagnosis and preoperative planning.
Main Methods:
- A dual-branch 3D U-Net architecture incorporating Mamba modules for long-range dependency modeling along the cranio-caudal axis.
- An identification branch utilizes a 3D convolutional block attention module (3D-CBAM) for improved class discriminability.
- A unified semi-supervised objective based on teacher-student consistency with data augmentation, confidence filtering, class-frequency reweighting, and connected-component analysis for anatomical plausibility.
Main Results:
- The semi-supervised approach improved Dice score from 89.8% to 91.6% and identification accuracy from 92.3% to 97.5% on the VerSe 2019 test set, using VerSe 2020 data as unlabeled training data.
- Achieved relative gains of +1.8% in Dice score and +5.2% in identification accuracy.
- Outperformed competing methods in segmentation accuracy and achieved the highest identification accuracy.
Conclusions:
- The proposed semi-supervised method significantly enhances vertebra segmentation and identification performance without additional annotation costs.
- Offers more robust automated support for clinical diagnosis and preoperative planning.
- Demonstrates the effectiveness of leveraging unlabeled data in medical image analysis.
