Improved myocardial perfusion PET imaging using artificial neural networks.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Cardiovascular Imaging

Background:

• Myocardial perfusion (MP) PET imaging is crucial for assessing coronary artery disease (CAD).
• Current MP PET imaging methods face challenges in balancing image noise, bias, and contrast, impacting diagnostic accuracy.
• Improving quantification and defect detection in MP PET is essential for patient risk stratification.

Purpose of the Study:

• To develop and evaluate a patch-based artificial neural network (ANN) fusion approach for enhancing MP PET imaging.
• To improve quantitative accuracy and the detection of myocardial perfusion defects.
• To assess the performance of the ANN fusion technique against traditional post-smoothed Maximum Likelihood (ML) reconstruction.

Main Methods:

• A patch-based ANN fusion method was developed, integrating information from ML and post-smoothed ML reconstructions.
• The method was tested on simulated MP PET data (XCAT phantom) with varying noise levels and perfusion defects (normal, non-transmural, transmural).
• Quantitative evaluation included noise-bias-contrast tradeoff analysis and receiver operating characteristic (ROC) analysis using a channelized Hotelling observer (CHO).

Main Results:

• The ANN fusion method demonstrated reduced bias and improved contrast compared to post-smoothed ML reconstruction, with similar noise levels.
• Defect detectability for both non-transmural and transmural perfusion defects was significantly enhanced by the ANN fusion technique.
• Evaluation on patient data confirmed improved noise-mean value tradeoff on the left ventricular (LV) myocardium.

Conclusions:

• The proposed ANN fusion technique effectively enhances MP PET image quality and diagnostic performance.
• This method offers a promising advancement for clinical applications in cardiovascular imaging, particularly for CAD assessment.
• The ANN fusion approach improves the balance between image noise and quantitative accuracy in MP PET.