Deep Learning-Based Segmentation and Uncertainty Assessment for Automated Analysis of Myocardial Perfusion MRI...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiovascular Imaging

Background:

• Accurate segmentation of cardiac MRI is crucial for diagnosing cardiovascular diseases.
• Deep learning models require robust training data and methods to ensure reliable performance in clinical settings.

Purpose of the Study:

• To develop a deep learning framework for segmenting motion-corrected perfusion cardiac MRI.
• To introduce an uncertainty quantification method for detecting out-of-distribution data.
• To enhance the generalization performance of cardiac MRI segmentation models.

Main Methods:

• A patch-level training approach and intensity-based augmentation for cardiac MRI segmentation.
• A novel spatial sliding-window approach for generating image-based uncertainty maps.
• Testing on external MRI data with varied acquisition protocols to assess robustness.

Main Results:

• The proposed method successfully quantifies uncertainty, enabling detection of unsuitable test data.
• The deep learning approach demonstrates robustness to variations in pulse-sequence parameters.
• The spatiotemporal patch-based segmentation outperforms the state-of-the-art 2D approach in generalization.

Conclusions:

• The developed framework enhances the reliability of deep learning in clinical cardiac MRI analysis.
• Uncertainty quantification is a key feature for safe clinical integration of AI in medical imaging.
• The proposed data augmentation and training strategy improve model generalization across different MRI acquisition protocols.