Contrast-Free Myocardial Infarction Segmentation with Attention U-Net

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Cardiology

Background:

• Cardiovascular magnetic resonance (CMR) is the gold standard for cardiac assessment but manual segmentation is time-consuming and variable.
• Deep learning (DL) automates segmentation, yet struggles with generalizability in non-contrast cine CMR for myocardial infarction (MI) detection.

Purpose of the Study:

• To develop a DL framework for automated cardiac structure and MI segmentation using contrast-free cine CMR.
• To enhance the clinical applicability of CMR by enabling contrast-independent cardiac assessment.

Main Methods:

• Integrated multiple CNNs for cardiac structure segmentation and an attention-based DL model for MI localization.
• Employed post-processing with stacked autoencoders and active contour modeling for anatomical consistency.
• Evaluated performance using Dice Similarity Coefficient (DSC), Mean Contour Distance (MCD), and Hausdorff Distance (HD).

Main Results:

• Achieved high Dice scores: 0.93 for LV cavity, 0.89 for LV myocardium, 0.91 for RV cavity.
• Demonstrated reliable MI segmentation with a Dice score of 0.80 and high recall.
• Showcased consistently low boundary errors across all segmented cardiac structures.

Conclusions:

• The proposed DL framework enables accurate, contrast-free segmentation of cardiac structures and MI from cine CMR.
• Facilitates broader clinical use, especially for patients with contrast contraindications or in resource-limited settings.
• Supports scalable and reliable cardiac assessment independent of contrast agents.