Vertebrae localization in pathological spine CT via dense classification from sparse annotations.

Area of Science:

• Medical Imaging
• Computer Vision
• Radiology

Background:

• Accurate vertebral localization and identification are critical for spinal imaging analysis in clinical settings.
• Existing automated methods struggle with pathological spine curvatures, limited fields of view, and surgical implant artifacts.
• Parametric models often fail in pathological cases, necessitating more robust approaches.

Purpose of the Study:

• To develop a robust algorithm for automatic localization and identification of vertebrae in spinal CT scans.
• To address the challenges posed by spinal pathologies and image artifacts.
• To improve the accuracy and reliability of automated spinal analysis.

Main Methods:

• A supervised classification forest approach is utilized, avoiding explicit parametric models.
• A semi-automatic labeling strategy transforms sparse centroid annotations into dense probabilistic labels, handling identification uncertainty.
• A discriminative centroid classifier is trained using local and contextual intensity features for robustness.

Main Results:

• The algorithm demonstrates robustness against spinal pathologies like scoliosis and kyphosis, and surgical implants.
• Evaluated on a dataset of 224 spine CT scans and 200 additional CTs, the method shows superior performance.
• Quantitative evaluation confirmed reduced localization errors and improved identification rates compared to state-of-the-art methods.

Conclusions:

• The proposed algorithm offers an efficient and robust solution for vertebral localization and identification in challenging spinal imaging scenarios.
• It significantly outperforms existing methods, particularly in cases with pathological spine conditions.
• This advancement has the potential to enhance clinical diagnosis, surgical planning, and post-operative assessment.