U-LanD: Uncertainty-Driven Video Landmark Detection.

Area of Science:

• Medical imaging analysis
• Computer vision
• Machine learning

Background:

• Accurate landmark detection in medical videos is crucial for diagnosis.
• Existing methods struggle with noisy and sparse training labels, common in clinical settings.
• Ultrasound videos of the heart often have limited, single-frame annotations.

Purpose of the Study:

• To develop a framework for automatic landmark detection on key video frames.
• To address the challenge of noisy and highly sparse training labels.
• To leverage predictive uncertainty for unsupervised key frame identification.

Main Methods:

• Developed U-LanD, a framework utilizing a deep Bayesian landmark detector.
• Exploited the observation that Bayesian detectors have lower predictive uncertainty on key frames.
• Used this uncertainty as an unsupervised signal to automatically select key frames for landmark detection.
• Tested on ultrasound imaging videos of the heart with sparse, noisy labels.

Main Results:

• U-LanD significantly outperforms state-of-the-art non-Bayesian methods.
• Achieved a 42% absolute improvement in R² score.
• Demonstrated effectiveness on a large dataset of 4,493 patients.
• Imposed minimal overhead on model size.

Conclusions:

• U-LanD offers an effective solution for automatic landmark detection in videos with challenging label conditions.
• The framework successfully identifies key frames using predictive uncertainty, enhancing accuracy.
• The approach shows significant potential for improving analysis in cardiac ultrasound and other medical imaging applications.