A Bayesian network for simultaneous keyframe and landmark detection in ultrasonic cine

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Accurate landmark detection in medical imaging is crucial for diagnosis and treatment planning.
• Ultrasound cine analysis requires identifying keyframes and anatomical landmarks, often treated as separate tasks.
• Inter-observer variability in ultrasound poses challenges for both landmark and keyframe detection.

Purpose of the Study:

• To develop a unified approach for simultaneous keyframe and landmark detection in ultrasound cine.
• To address challenges posed by sparse training data and inter-observer variability.
• To improve the accuracy of anatomical structure quantification in cardiac ultrasound.

Main Methods:

• A Bayesian network integrating landmark and keyframe detection using a shared encoder.
• A coarse-to-fine landmark detection architecture with an adaptive Bayesian hypergraph for refinement.
• Bi-directional Gated Recurrent Unit trained with Order Loss for keyframe identification.

Main Results:

• The proposed method outperforms state-of-the-art approaches on multiple echocardiography datasets.
• Achieved mean absolute errors of 2.40 mm (left atrial appendage), 0.83 mm (aortic annulus), and 1.63 mm (left ventricle).
• Demonstrated improved detection accuracy by exploiting temporal and motion information through feature sharing.

Conclusions:

• Simultaneous detection of keyframes and landmarks in ultrasound cine is feasible and effective using the proposed Bayesian network.
• The method shows promise for enhancing quantitative analysis in cardiac ultrasound, particularly with limited data.
• The integrated approach effectively leverages correlations between keyframe and landmark detection tasks.