Why the study?
Integrating model uncertainty quantification in machine learning clinical decision support systems can enhance reliability, robustness against domain shifts, and user trust.
Does an uncertainty-informed active learning approach using Monte Carlo dropout improve the classification of carotid ultrasound images for cardiovascular risk stratification?
Population
Carotid ultrasound images from an auxiliary dataset (CUBS) and 87 B-mode ultrasound sequences from ATTIKON hospital
Comparison
Uncertainty rank selection vs pseudo-labeling for certain samples vs pseudo-labeling with variable sample weighting
Design
Model development and active learning study
Key result
An uncertainty-informed active learning approach using pseudo-labeling with variable sample weighting achieved an AUC of 87.28% for cardiovascular risk classification using only 21 annotated samples.
Authors
Loading...
May reduce annotation burden for carotid ultrasound AI; leaves open clinical utility for cardiovascular risk stratification.
Does an uncertainty-informed active learning approach using Monte Carlo dropout improve the classification of carotid ultrasound images for cardiovascular risk stratification?
An uncertainty-informed active learning approach using Monte Carlo dropout can achieve high diagnostic performance for cardiovascular risk stratification in carotid ultrasound with significantly reduced labeling costs.
Ganitidis et al. (2024) studied Cardiovascular disease risk stratification (n=87). Uncertainty-informed active learning using Monte Carlo dropout (pseudo-labeling with variable sample weighting) vs. Other active learning strategies was evaluated on AUC for classifying carotid ultrasound images as high-risk and low-risk for cardiovascular disease. An uncertainty-informed active learning approach using pseudo-labeling with variable sample weighting achieved an AUC of 87.28% for cardiovascular risk classification using only 21 annotated samples.