Unsupervised Myocardial Segmentation for Cardiac BOLD.

Area of Science:

• Cardiovascular Imaging
• Medical Image Analysis
• Biomedical Engineering

Background:

• Cardiac magnetic resonance (CMR) blood-oxygen-level-dependent (BOLD) imaging is crucial for ischemia detection.
• Accurate myocardial segmentation is essential for analyzing spatio-temporal intensity patterns in BOLD CMR.
• Existing supervised segmentation methods struggle with the complex patterns in BOLD data, leading to errors.

Purpose of the Study:

• To develop a fully automated 2-D+time myocardial segmentation framework for CMR BOLD datasets.
• To improve the accuracy of ischemia detection by addressing segmentation challenges in BOLD CMR.
• To create an unsupervised method that outperforms current state-of-the-art techniques.

Main Methods:

• A joint motion and appearance model utilizing dictionary learning to define a suitable subspace.
• Variational pre-processing and spatial regularization with Markov random fields for enhanced performance.
• Unsupervised segmentation framework applied to cardiac phase-resolved BOLD MR and standard CINE MR sequences.

Main Results:

• The proposed unsupervised segmentation method achieved at least 10% higher accuracy (Dice score) on BOLD data compared to supervised methods.
• The technique performed comparably to supervised methods on standard CINE MR data.
• A novel segmental analysis method for BOLD time series demonstrated the preservation of key BOLD patterns.

Conclusions:

• The developed automated segmentation framework offers a superior solution for CMR BOLD data analysis.
• This unsupervised approach effectively overcomes limitations of supervised methods in segmenting BOLD CMR.
• The method enhances the reliability of ischemia detection and analysis of myocardial blood oxygenation patterns.