Uncertainty Estimation for Heatmap-Based Landmark Localization.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer Vision

Background:

• Deep learning advances automatic anatomical landmark localization.
• Quantifying prediction uncertainty is crucial for clinical adoption and error detection.

Purpose of the Study:

• To introduce Quantile Binning, a data-driven method for categorizing predictions by uncertainty with error bounds.
• To enable comparison of different uncertainty measures and improve prediction reliability.

Main Methods:

• Developed Quantile Binning to categorize predictions based on continuous uncertainty measures.
• Evaluated three epistemic uncertainty measures using U-Net and patch-based models.
• Tested on three datasets, including a public Cephalometric dataset.

Main Results:

• Filtering gross mispredictions using Quantile Bins significantly improves the proportion of accurate predictions.
• Quantile Binning effectively handles landmarks with high aleatoric uncertainty due to ambiguity.
• Demonstrated improved prediction accuracy by identifying and excluding uncertain localizations.

Conclusions:

• Quantile Binning is an effective framework for uncertainty quantification in anatomical landmark localization.
• The method enhances the reliability and clinical applicability of deep learning models.
• Provides guidance on selecting and utilizing uncertainty measures for improved medical image analysis.