A Three-Stage Semi-Supervised Learning Approach to Spine Image Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate spine segmentation in computed tomography (CT) images is essential for automated analysis.
• Existing datasets often lack labeled fracture data, hindering predictive model development.
• Vertebral similarity presents challenges for precise segmentation.

Purpose of the Study:

• To develop a semi-supervised learning model for spine segmentation using both labeled and unlabeled CT data.
• To reduce the manual annotation workload for medical imaging analysis.
• To create a model capable of processing fracture data without requiring specific labeled fracture datasets.

Main Methods:

• A three-stage, 2.5D semi-supervised learning approach based on the U-Net architecture was employed.
• A cascade framework, mimicking clinical examination, was utilized for precise vertebral segmentation.
• 2D network training supplemented by 2.5D input was strategically used to manage large 3D CT data and GPU constraints.

Main Results:

• Preliminary findings indicate a significant improvement in the model's ability to segment spinal regions.
• The approach demonstrates effectiveness, particularly in environments with limited equipment capabilities.
• The method shows promise for enhancing spine segmentation accuracy and efficiency.

Conclusions:

• The proposed 2.5D semi-supervised learning method offers a viable solution for spine segmentation with limited labeled data.
• This approach effectively addresses challenges posed by data scarcity and computational limitations.
• Further research is needed to fully assess its potential in diverse clinical scenarios, including fracture detection.