Uncertainty Quantification and Quality Control for Heatmap-Based Landmark Detection Models.

Area of Science:

• Medical Artificial Intelligence
• Explainable AI
• Computer Vision

Background:

• Deep learning models achieve high accuracy in anatomical landmark detection but struggle with uncertainty quantification due to small target sizes and contextual noise.
• Quantifying uncertainty is crucial for clinician trust and reliable outcomes in medical AI applications.
• Existing methods face challenges in providing robust uncertainty estimates for heatmap-based landmark detection.

Purpose of the Study:

• To develop an end-to-end uncertainty quantification method for heatmap-based anatomical landmark detection.
• To improve the interpretability and controllability of deep learning models in clinical settings.
• To enhance the robustness of landmark detection against noise and identify out-of-distribution data.

Main Methods:

• Leveraged Dempster-Shafer Theory and Subjective Logic Theory for probability assignment and uncertainty quantification in a single forward pass.
• Introduced an evidence map to quantify landmark evidence strength and an uncertainty map for calibrated probability assessment.
• Utilized a cross-attention mechanism to integrate evidence and uncertainty maps, improving detection accuracy and quantification.

Main Results:

• The proposed method maintains high detection accuracy, even in noisy conditions.
• Demonstrated superior performance in uncertainty quantification and quality control compared to state-of-the-art methods.
• Successfully identified out-of-distribution data using calibrated probabilities, showing potential for multi-center and novel data analysis.

Conclusions:

• The developed method offers an efficient and effective solution for uncertainty quantification in anatomical landmark detection.
• The approach enhances model reliability and trustworthiness for clinical deployment of AI in medicine.
• The findings support the use of calibrated probabilities for robust quality control and data validation in medical imaging AI.