Automatic Valve Plane Localization in Myocardial Perfusion SPECT/CT by Machine Learning: Anatomic and Clinical...

Area of Science:

• Nuclear Medicine
• Medical Imaging Analysis
• Machine Learning Applications

Background:

• Accurate segmentation of the left ventricle in SPECT MPI is crucial for perfusion quantification.
• Manual adjustment of the mitral valve plane (VP) introduces variability and affects results.
• Developing automated methods for VP placement can enhance SPECT MPI accuracy.

Purpose of the Study:

• To develop and validate a machine learning approach using support vector machines (SVM) for automatic VP placement in SPECT MPI.
• To compare the accuracy of automated VP localization with expert manual placement.
• To assess the impact of automated VP placement on the detection of coronary artery disease.

Main Methods:

• A total of 392 patients undergoing SPECT MPI were studied.
• Support vector machine (SVM) models were trained using expert VP placements.
• Automatic VP localizations were compared to expert placements and validated against invasive coronary angiography.
• Total perfusion deficits and detection of obstructive stenosis were analyzed.

Main Results:

• SVM-based automatic VP localization demonstrated narrower Bland-Altman confidence intervals compared to inter-expert variability for both attenuation-corrected (AC) and non-AC images.
• The SVM method achieved comparable accuracy to experts in detecting obstructive coronary artery stenosis.
• Automated VP placement significantly improved the detection of stenosis compared to unadjusted VP (AUC increased from 0.63-0.65 to 0.79-0.82).

Conclusions:

• Machine learning with SVM enables automatic and accurate VP localization in SPECT MPI.
• This automated approach reduces user dependence and enhances the reliability of SPECT MPI quantification.
• The SVM method shows promise for improving the diagnostic accuracy of SPECT MPI in identifying coronary artery disease.