Central loss guides coordinated Transformer for reliable anatomical landmark detection

Area of Science:

• Computer Vision
• Medical Imaging Analysis
• Machine Learning

Background:

• Heatmap-based anatomical landmark detection faces challenges in accurate heatmap distribution evaluation and exploiting global spatial structure.
• Existing methods struggle with imbalanced datasets (landmarks vs. non-landmarks) and convergence issues in sparse learning.

Purpose of the Study:

• To develop advanced techniques addressing heatmap evaluation and spatial structure exploitation in landmark detection.
• To introduce a position-aware and sample-aware central loss and a Coordinated Transformer (CoorTransformer) for generalized landmark detection.

Main Methods:

• A novel position-aware and sample-aware central loss is proposed to accurately evaluate heatmap distribution and focus on hard-to-detect landmarks.
• A Coordinated Transformer (CoorTransformer) is developed to establish long-range dependencies using landmark coordinate information, enhancing focus on sparse landmarks and global structure.
• A generalized detection model integrating CoorTransformer and central loss is proposed to handle diverse scenarios and exploit landmark relationships.

Main Results:

• The proposed CoorTransformer outperforms state-of-the-art methods in challenging landmark detection tasks.
• The central loss significantly improves model performance, with statistically significant results (p-values < 0.05).
• The integrated model demonstrates effectiveness in various scenarios by incorporating rich structural knowledge.

Conclusions:

• The developed central loss and CoorTransformer effectively address key challenges in heatmap-based anatomical landmark detection.
• These novel methods offer improved accuracy, better handling of data imbalance, and faster convergence.
• The generalized model provides a robust solution for diverse landmark detection applications.