Accuracy, uncertainty, and adaptability of automatic myocardial ASL segmentation using deep CNN

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Myocardial perfusion imaging is crucial for diagnosing cardiac conditions.
• Accurate segmentation of myocardial tissue is essential for quantitative analysis.
• Deep learning offers potential for automating complex segmentation tasks.

Purpose of the Study:

• To apply deep convolution neural networks (CNNs) for segmenting myocardial arterial spin labeled perfusion imaging.
• To develop methods for quantifying segmentation uncertainty.
• To adapt CNN models for a desired balance between false-positive and false-negative rates.

Main Methods:

• A Monte Carlo dropout U-Net model was trained and tested on cardiac MRI data.
• Two global uncertainty measures (Dice uncertainty, Monte Carlo dropout uncertainty) were calculated.
• A Tversky loss function was used to adjust the model's sensitivity to false positives versus false negatives.

Main Results:

• The Monte Carlo dropout U-Net achieved high segmentation accuracy (Dice coefficient 0.91 ± 0.04).
• Automated segmentation strongly correlated with manual segmentation for myocardial blood flow (R² = 0.96).
• Uncertainty measures showed good agreement, and model adaptation successfully controlled the false-positive/negative tradeoff.

Conclusions:

• Deep CNNs are feasible for accurate automatic segmentation in myocardial arterial spin labeling.
• Novel methods effectively assess model uncertainty in segmentation tasks.
• The developed approach allows for customizable false-positive/negative tradeoffs in cardiac MRI segmentation.