Abstract
241182
Introduction: Myocardial perfusion (MP) SPECT is an accurate and non-invasive method for diagnosis of coronary artery disease (CAD). Moreover, total perfusion deficit (TPD) can be calculated from MP SPECT based on a normal database and used for obstructive CAD detection. However, TPD cannot reflect the real narrowness of coronary arteries and thus invasive coronary angiography (ICA) is still the gold standard for obstructive CAD diagnosis. This study aims to apply the deep learning technique to predict ICA-based obstructive CAD diagnosis from stress and rest MP SPECT polar plots.
Methods: A total of 220 anonymized patients (58.2% men) underwent standard one-day 99mTc-sestamibi SPECT/CT stress/rest protocol based on a clinical NaI dual-head SPECT scanner were retrospectively recruited in this study. The 64 projections were reconstructed with OS-EM algorithm with CT-based attenuation correction (CTAC) or without AC (NAC), and the reconstructed images were then transformed to polar plots. ICA was performed within 9 months following the MP SPECT scan, and a luminal diameter narrowing ≥70% in the left anterior, left circumflex artery, or right coronary artery was defined as obstructive CAD (1). Thus, 148 of the 220 patients (67.3%) were diagnosed as positive for obstructive CAD with at least one vessel narrowing. A 2D densely connected convolutional networks (DenseNet-121) (2) was implemented using Tensorflow for binary prediction. In addition, a 2D conditional generative adversarial network based AC method (DLAC) (3) was applied to NAC polar plots and used for prediction. We evaluated deep learning (DL) models trained with CTAC, DLAC and NAC data, as well as stress, rest, and combined stress and rest (stress+rest) data. For stress+rest, the stress and rest polar plots were paired with the ICA diagnosis to train the DenseNet as individual cases. In the testing phase, the output probabilities from stress and rest polar plots of the tested patient were averaged and then went through a Softmax activation for final prediction. TPD from polar plots was used as an additional input to improve the prediction of stress+rest data with CTAC (CTAC_TPD). All patients were divided into 150, 26 and 44 for training, validation, and testing. A 5-fold cross-validation was applied to test all 220 patients. The accuracy (ACC) and area under the receiver-operating characteristic curve (AUC) were assessed for evaluation of prediction performance.
Results: For stress, rest and combined stress+rest with CTAC, the ACC were 67.7% (149/220), 63.2% (139/220) and 77.3% (170/220) and their corresponding AUC were 0.533, 0.584, and 0.708. For stress+rest with NAC, DLAC, CTAC and CTAC_TPD, the ACC were 65.0% (143/220), 75.0% (165/220), 77.3% and 81.8% (180/220), while their corresponding AUC values were 0.493, 0.618, 0.708 and 0.778.
Conclusions: Deep learning is promising to predict ICA-based obstructive CAD diagnosis from MP SPECT polar plots. Combined use of stress and rest data, CT- or DL-based AC, and the incorporation of TPD can further enhance the prediction accuracy. The incorporation of other patient clinical information is on-going for further improvement of the prediction.
Funding: This work was supported by the Science and Technology Development Fund of Macau (0016/2023/RIB1).
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.