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

Abstract

Uncertainty quantification is a vital aspect of explainable artificial intelligence that fosters clinician trust in medical applications and facilitates timely interventions, leading to safer and more reliable outcomes. Although deep learning models have reached clinically acceptable accuracy in anatomical landmark detection, their predictions remain susceptible to contextual noise due to the small size of the target structures, making uncertainty quantification more challenging than in classification and segmentation tasks. This paper presents an end-to-end uncertainty quantification method tailored for heatmap-based anatomical landmark detection models, designed to improve both interpretability and controllability in clinical applications. Leveraging Dempster-Shafer Theory and Subjective Logic Theory, we implement probability assignment and uncertainty quantification through a single forward pass to ensure computational efficiency. We introduce an evidence map that captures the strength of landmark evidence, alongside an uncertainty map that calibrates predicted probabilities within the Subjective Logic framework. The interaction between these two components, facilitated by a cross-attention mechanism, further improves landmark detection accuracy and enhances the effectiveness of uncertainty quantification. Experimental results demonstrate that the proposed method maintains detection accuracy, even in noisy environments, while outperforming state-of-the-art methods in terms of uncertainty quantification and quality control. Furthermore, the model effectively identifies out-of-distribution data solely through calibrated probabilities when encountering inconsistencies in multi-center data and novel data, underscoring its potential for clinical applications. The source code is available at github.com/warmestwind/CalibratedSL