Semantics and instance interactive learning for labeling and segmentation of vertebrae in CT images

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Vertebrae labeling and segmentation in 3D CT images is a complex multi-task problem.
• Existing methods often use separate models, neglecting feature interactions and insufficient semantic information utilization.
• Distinguishing similar adjacent vertebrae and modeling their sequential attributes remain significant challenges.

Purpose of the Study:

• To develop a novel paradigm for synchronous vertebrae labeling and segmentation in CT images.
• To enhance feature interaction between semantic and instance learning for improved accuracy.
• To effectively differentiate adjacent vertebrae and model their sequential relationships.

Main Methods:

• Proposed a Semantics and Instance Interactive Learning (SIIL) paradigm.
• Introduced a Morphological Instance Localization Learning (MILL) module for feature alignment and interaction.
• Devised an Ordinal Contrastive Prototype Learning (OCPL) module for differentiating similar vertebrae and modeling sequential attributes.

Main Results:

• The SIIL paradigm demonstrated significant improvements in vertebrae labeling and segmentation accuracy.
• The MILL and OCPL modules effectively enhanced feature interaction and addressed challenges with similar adjacent vertebrae.
• Experiments on multiple datasets confirmed the superiority of the proposed method over existing approaches.

Conclusions:

• The SIIL paradigm offers a robust solution for synchronous vertebrae labeling and segmentation in 3D CT images.
• The method effectively integrates semantic and instance learning, improving diagnostic accuracy in spinal imaging.
• This work advances automated analysis of spinal CT scans, with potential clinical applications.