Left ventricle quantification with sample-level confidence estimation via Bayesian neural network.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Cardiac left ventricle (LV) quantification is crucial in clinical practice.
• Deep neural networks have shown promise in LV quantification but lack sample-level confidence.
• Clinicians need reliable indicators to trust automated LV measurements.

Purpose of the Study:

• To introduce uncertainty analysis into deep neural networks for LV quantification.
• To provide sample-level confidence scores for LV quantification results.
• To enhance the clinical utility of automated LV measurement tools.

Main Methods:

• Incorporation of uncertainty analysis theory into a deep neural network for LV quantification.
• Analysis of both Model Uncertainty and Data Uncertainty to derive confidence levels.
• Utilization of an uncertainty-weighted regression loss function for improved quantification.

Main Results:

• The proposed method successfully quantifies the cardiac left ventricle with improved performance.
• The system provides a confidence level for each sample's LV quantification result.
• Experiments on 145 subjects validated the effectiveness of the uncertainty analysis approach.

Conclusions:

• The integration of uncertainty analysis significantly enhances LV quantification by providing sample-specific confidence.
• This approach offers clinicians valuable insights into the reliability of automated measurements.
• The method holds potential for improving clinical decision-making in cardiovascular diagnostics.