Unsupervised motion-compensation of multi-slice cardiac perfusion MRI

Area of Science:

• Medical Imaging
• Cardiovascular Imaging
• Machine Learning in Medicine

Background:

• Cardiac perfusion MRI is crucial for diagnosing conditions like myocardial infarction.
• Accurate image registration is essential for quantitative analysis of perfusion data.
• Manual registration is time-consuming and prone to inter-observer variability.

Purpose of the Study:

• To develop and validate a novel, automated method for cardiac perfusion MRI registration.
• To achieve rapid, accurate, and robust image registration without manual intervention.
• To assess the method's performance on clinical data from patients with acute myocardial infarction.

Main Methods:

• Utilized offline computer-intensive analyses of variance and clustering on an annotated training set.
• Employed a slice-coupled active appearance model to model intensity changes during bolus passage.
• Optimized landmark correspondences using the Minimum Description Length (MDL) framework and validated with perfusion-specific prior models.

Main Results:

• Achieved registration in less than a second per frame with no manual interaction.
• Demonstrated high accuracy with a mean point-to-curve distance of 1.25+/-0.36 pixels (left and right ventricle combined) via leave-one-out cross-validation.
• Validated on 2000 clinical quality, short-axis, perfusion MR images from 10 patients.

Conclusions:

• The learning-based method shows significant promise for automating cardiac perfusion investigations.
• The method offers accuracy, robustness, and generalization ability, crucial for clinical applications.
• This automated approach can streamline the analysis of cardiac MRI data, improving diagnostic efficiency.