Vertebrae Labeling via End-to-End Integral Regression Localization and Multi-Label Classification Network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Spinal Diagnostics

Background:

• Accurate vertebral identification and localization in CT scans are crucial for spinal diagnosis and treatment.
• Current methods often rely on non-differentiable heatmap-based approaches, hindering end-to-end training.
• Existing techniques face challenges in direct coordinate labeling and preserving spatial information.

Purpose of the Study:

• To develop a robust and accurate automatic vertebral labeling algorithm for CT scans.
• To enable end-to-end differential training for vertebrae coordinate regression.
• To improve both the identification rate and localization accuracy of vertebrae.

Main Methods:

• A novel end-to-end integral regression localization and multi-label classification network is proposed.
• The network captures multi-scale features using residual modules and skip connections.
• An integral regression module combines heatmap advantages with direct coordinate regression for differentiability, enhanced by bidirectional long short-term memory (Bi-LSTM) for contextual learning.

Main Results:

• The proposed method achieved a 91.1% identification rate and a mean localization error of 2.2 mm on a challenging dataset.
• Results significantly outperform existing state-of-the-art methods.
• The algorithm demonstrated good generalization capabilities on a new CT dataset.

Conclusions:

• The developed integral regression localization and multi-label classification network provides an effective solution for automatic vertebral labeling in CT scans.
• The method overcomes the limitations of non-differentiable processes, enabling efficient end-to-end training.
• This approach offers improved accuracy and generalization, advancing pre-processing for spinal clinical applications.