SpineCLUE: Automatic vertebrae identification using contrastive learning and uncertainty estimation.

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Accurate vertebrae identification is vital for diagnosing spinal diseases.
• Current methods struggle with arbitrary fields-of-view in CT scans, often requiring prior knowledge of the number of vertebrae.
• Local CT regions (neck, chest, abdomen) present unique identification challenges.

Purpose of the Study:

• To develop a robust three-stage method for vertebrae identification in 3D CT scans with arbitrary fields-of-view.
• To address limitations of existing spine-level identification methods.
• To improve the accuracy and stability of vertebrae localization, segmentation, and identification.

Main Methods:

• A sequential three-stage approach: vertebrae localization, segmentation, and identification, leveraging anatomical priors.
• Dual-factor density clustering for stable individual vertebrae localization, mitigating issues from abnormal positions.
• Supervised contrastive learning for pretraining the identification network to handle inter-class similarity and intra-class variability.
• Uncertainty estimation and a message fusion module to optimize identification by incorporating global spine information.

Main Results:

• The proposed method achieves state-of-the-art performance on the VerSe20 challenge benchmark.
• Demonstrated improved stability in vertebrae localization compared to existing methods.
• Successfully addressed challenges of inter-class similarity and intra-class variability in vertebrae identification.

Conclusions:

• The three-stage method effectively integrates contextual prior information for accurate vertebrae identification.
• The dual-factor density clustering and supervised contrastive learning contribute to robust performance.
• This approach offers a significant advancement for automated vertebrae identification in diverse CT scan scenarios.