Interpretable vertebral fracture quantification via anchor-free landmarks localization

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Vertebral compression fractures are key indicators of osteoporosis.
• These fractures are often missed in clinical Computed Tomography (CT) assessments.
• Existing automated methods lack interpretability and struggle with complex cases.

Purpose of the Study:

• To develop a robust, interpretable algorithm for vertebral fracture detection.
• To improve the accuracy and efficiency of diagnosing osteoporosis-related fractures.

Main Methods:

• A two-step algorithm was developed: 3D vertebral column localization followed by 2D individual vertebra detection and fracture quantification.
• Neural networks were trained using a 6-keypoint annotation scheme aligned with clinical standards.
• The algorithm processes 3D CT scans without exclusion criteria.

Main Results:

• Achieved expert-level performance in vertebrae 3D localization (1mm average error) and 2D detection (0.99 precision/recall).
• Demonstrated high accuracy in fracture identification (up to 0.96 ROC AUC at patient level).
• Showcased excellent generalizability on the VerSe dataset (0.95 ROC AUC).

Conclusions:

• The proposed algorithm provides an interpretable and verifiable output for vertebral fracture detection.
• It offers a fast (2 seconds on single GPU) and accurate solution for clinical use.
• The method represents a state-of-the-art advancement in automated osteoporosis fracture assessment.