Automatic myocardial segmentation in dynamic contrast enhanced perfusion MRI using Monte Carlo dropout in an...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Cardiac perfusion magnetic resonance imaging (MRI) with dynamic contrast enhancement (DCE) is crucial for identifying myocardial perfusion defects.
• Automatic segmentation of the myocardium is essential for efficient quantification of these perfusion defects.
• Deep convolutional neural networks (CNNs) offer potential for automating this segmentation process.

Purpose of the Study:

• To investigate the utility of uncertainty estimation within deep CNNs for achieving automatic myocardial segmentation.
• To evaluate the performance of an uncertainty-driven automatic segmentation method compared to a semi-automatic approach.

Main Methods:

• A U-Net segmentation model was trained on cardiac cine data and adapted for dynamic perfusion datasets.
• Monte Carlo dropout sampling was employed to generate standard deviation (SD) maps for uncertainty estimation.
• Uncertainty estimates guided the selection of optimal frames for endocardial and epicardial segmentation.

Main Results:

• The proposed method achieved fully automatic myocardial segmentation without manual labeling.
• The mean Dice similarity score for the automatic method was 0.806 (±0.096), comparable to the semi-automatic U-Net's score of 0.808 (±0.084).
• The results indicate high agreement between the automatic and semi-automatic segmentation methods (ICC = 0.61, P < 0.001).

Conclusions:

• An existing U-Net model trained on cine data can be effectively applied to dynamic perfusion data for robust myocardial segmentation.
• Uncertainty estimation facilitates automatic segmentation and can serve as a screening tool, flagging complex cases for expert review.
• This approach enhances the efficiency and reliability of myocardial segmentation in cardiac perfusion MRI analysis.