Deep learning-based segmentation of left ventricular myocardium on dynamic contrast-enhanced MRI: a comprehensive...

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medical Imaging
• Cardiac MRI Analysis

Background:

• Cardiac perfusion MRI is crucial for diagnosing and managing ischemic heart conditions.
• Accurate segmentation of left ventricular (LV) myocardium is essential but challenging due to complex data.
• AI offers a solution for efficient and precise LV myocardium segmentation across dynamic contrast-enhanced MRI (DCE-MRI) timeframes.

Purpose of the Study:

• To develop and evaluate an automated AI-assisted method for LV myocardium segmentation on DCE-MRI data.
• To address the challenges of segmenting multidimensional DCE-MRI data for clinical applications.
• To assess the robustness and accuracy of the proposed automated segmentation technique.

Main Methods:

• Retrospective analysis of DCE-MRI data from 55 subjects using a 1.5 T scanner.
• Identification of the reference frame (post-contrast LV myocardium) using standard deviation.
• Application of iterative image registration (Maxwell's demons algorithm) and a U-Net framework for segmentation.

Main Results:

• The fine-tuned U-Net model (Net_dyn) achieved a mean Dice Similarity Coefficient (DSC) of 0.78 ± 0.03 for LV myocardium segmentation across all timeframes.
• Individual DSC values for Net_dyn ranged from 0.71 to 0.93.
• The average DSC for the reference frame was 0.82 ± 0.06.

Conclusions:

• A fast, fully automated AI method for LV myocardium segmentation on DCE-MRI was developed.
• The method demonstrates robustness and independence from intra-temporal sequence registration accuracy.
• The AI approach can effectively handle timeframes with potential registration errors, improving clinical utility.