Automatic Valve Plane Localization in Myocardial Perfusion SPECT/CT by Machine Learning: Anatomic and Clinical Validation
Julián Betancur, Mathieu Rubeaux, Tobias A. Fuchs, Yuka Otaki, Yoav Arnson, Leandro Slipczuk, Dominik C. Benz, Guido Germano, Damini Dey, Chih‐Jen Lin, Daniel S. Berman, Philipp A. Kaufmann, Piotr J. SlomkaJournal of Nuclear Medicine2016 · DOI
Precise definition of the mitral valve plane (VP) during left ventricle segmentation for SPECT myocardial perfusion imaging (MPI) is critical for accurate quantification but often requires time-consuming manual adjustment. This study developed and validated a machine learning approach using support vector machines (SVM) to automate VP placement. Validated against expert consensus, coronary CT angiography (CCTA), and invasive coronary angiography (ICA) in 392 patients, the SVM model demonstrated anatomic precision comparable to or exceeding human experts. Furthermore, the automated approach significantly improved diagnostic accuracy for obstructive stenosis compared to unadjusted software, matching expert-level performance.
Introduction: The Challenge of Manual Segmentation
Myocardial Perfusion Imaging (MPI) using SPECT is a primary tool for detecting coronary artery disease. A critical step in quantifying blood flow deficits is defining the mitral valve plane—the anatomical boundary at the base of the left ventricle. Standard clinical software often fails to locate this plane accurately due to the complex transition between myocardium and valve tissue, necessitating manual "tweaking" by clinicians. This manual intervention introduces subjectivity, inter-operator variability, and slows clinical workflows. This research presents a fully automated SVM-based solution to eliminate this bottleneck.
Methodology: SVM-Based Localization
The researchers framed VP localization as a classification task. The algorithm utilized a two-class SVM with a Radial Basis Function (RBF) kernel, incorporating 22 features including raw intensity values, myocardial mass, and gated data (differences between resting and beating phases).
Machine learning approach for valve plane localization using SVM probability estimates.
As shown in the methodology diagram, the software samples potential VP positions every 1mm along the heart's long axis. The SVM assigns a "correctness probability" to each position, and the point with the highest probability is selected. This probabilistic approach is inherently more robust than traditional coordinate regression in noisy medical imaging environments.
Results and Clinical Validation
The model was validated using a dataset of 392 consecutive patients. Ground truth was established by two expert technologists, with CCTA providing an independent anatomic reference.
Anatomic Precision
In Attenuation-Corrected (AC) images, the SVM demonstrated remarkable consistency. For AC stress images, the SVM showed a bias of only 1mm compared to experts. Notably, the SVM’s 95% confidence intervals (CIs) were narrower than the disagreement between the two human experts (SVM vs. experts: -5 to 7 mm; inter-expert: -8 to 8 mm; $P < 0.01$).
Bland-Altman plots comparing SVM performance to expert consensus and inter-expert variability.
The clinical utility was confirmed by comparing the detection of obstructive stenosis against invasive coronary angiography. The Area Under the Curve (AUC) for detecting ischemia was significantly higher for the SVM (0.82) and experts (0.79–0.81) than for unadjusted software (0.63, $P < 0.01$). There was no statistically significant difference between the SVM and human experts, indicating that the machine learning model achieves expert-level diagnostic power.
Conclusion
Machine learning with SVM provides a robust, automated solution for mitral valve plane localization in SPECT MPI. By matching the accuracy and consistency of experienced clinicians, this technology can eliminate manual contour correction, reduce user dependence, and streamline the quantification of myocardial perfusion in clinical practice.
Summary updated
Abstract
Precise definition of the mitral valve plane (VP) during segmentation of the left ventricle for SPECT myocardial perfusion imaging (MPI) quantification often requires manual adjustment, which affects the quantification of perfusion. We developed a machine learning approach using support vector machines (SVM) for automatic VP placement. <b>Methods:</b> A total of 392 consecutive patients undergoing <sup>99m</sup>Tc-tetrofosmin stress (5 min; mean ± SD, 350 ± 54 MBq) and rest (5 min; 1,024 ± 153 MBq) fast SPECT MPI attenuation corrected (AC) by CT and same-day coronary CT angiography were studied; included in the 392 patients were 48 patients who underwent invasive coronary angiography and had no known coronary artery disease. The left ventricle was segmented with standard clinical software (quantitative perfusion SPECT) by 2 experts, adjusting the VP if needed. Two-class SVM models were computed from the expert placements with 10-fold cross validation to separate the patients used for training and those used for validation. SVM probability estimates were used to compute the best VP position. Automatic VP localizations on AC and non-AC images were compared with expert placement on coronary CT angiography. Stress and rest total perfusion deficits and detection of per-vessel obstructive stenosis by invasive coronary angiography were also compared. <b>Results:</b> Bland-Altman 95% confidence intervals (CIs) for VP localization by SVM and experts for AC stress images (bias, 1; 95% CI, -5 to 7 mm) and AC rest images (bias, 1; 95% CI, -7 to 10 mm) were narrower than interexpert 95% CIs for AC stress images (bias, 0; 95% CI, -8 to 8 mm) and AC rest images (bias, 0; 95% CI, -10 to 10 mm) (<i>P</i> < 0.01). Bland-Altman 95% CIs for VP localization by SVM and experts for non-AC stress images (bias, 1; 95% CI, -4 to 6 mm) and non-AC rest images (bias, 2; 95% CI, -7 to 10 mm) were similar to interexpert 95% CIs for non-AC stress images (bias, 0; 95% CI, -6 to 5 mm) and non-AC rest images (bias, -1; 95% CI, -9 to 7 mm) (<i>P</i> was not significant [NS]). For regional detection of obstructive stenosis, ischemic total perfusion deficit areas under the receiver operating characteristic curve for the 2 experts (AUC, 0.79 [95% CI, 0.7-0.87]; AUC, 0.81 [95% CI, 0.73-0.89]) and the SVM (0.82 [0.74-0.9]) for AC data were the same (<i>P</i> = NS) and were higher than those for the unadjusted VP (0.63 [0.53-0.73]) (<i>P</i> < 0.01). Similarly, for non-AC data, areas under the receiver operating characteristic curve for the experts (AUC, 0.77 [95% CI, 0.69-0.89]; AUC, 0.8 [95% CI, 0.72-0.88]) and the SVM (0.79 [0.71-0.87]) were the same (<i>P</i> = NS) and were higher than those for the unadjusted VP (0.65 [0.56-0.75]) (<i>P</i> < 0.01). <b>Conclusion:</b> Machine learning with SVM allows automatic and accurate VP localization, decreasing user dependence in SPECT MPI quantification.
BibTeX
@article{betancur2016automatic,
author = {Julián Betancur and Mathieu Rubeaux and Tobias A. Fuchs and Yuka Otaki and Yoav Arnson and Leandro Slipczuk and Dominik C. Benz and Guido Germano and Damini Dey and Chih‐Jen Lin and Daniel S. Berman and Philipp A. Kaufmann and Piotr J. Slomka},
title = {{Automatic Valve Plane Localization in Myocardial Perfusion SPECT/CT by Machine Learning: Anatomic and Clinical Validation}},
journal = {Journal of Nuclear Medicine},
year = {2016},
doi = {10.2967/jnumed.116.179911},
}