Incremental role of resting myocardial computed tomography perfusion for predicting physiologically significant coronary artery disease: A machine learning approach

Abstract

Background. Evaluation of resting myocardial computed tomography perfusion (CTP) by coronary CT angiography (CCTA) might serve as a useful addition for determining coronary artery disease. We aimed to evaluate the incremental benefit of resting CTP over coronary stenosis for predicting ischemia using a computational algorithm trained by machine learning methods.Methods. 252 patients underwent CCTA and invasive fractional flow reserve (FFR). CT stenosis was classified as 0%, 1-30%, 31-49%, 50-70%, and >70% maximal stenosis. Significant ischemia was defined as invasive FFR < 0.80. Resting CTP analysis was performed using a gradient boosting classifier for supervised machine learning.Results. On a per-patient basis, accuracy, sensitivity, specificity, positive predictive, and negative predictive values according to resting CTP when added to CT stenosis (>70%) for predicting ischemia were 68.3%, 52.7%, 84.6%, 78.2%, and 63.0%, respectively. Compared with CT stenosis [area under the receiver operating characteristic curve (AUC): 0.68, 95% confidence interval (CI) 0.62-0.74], the addition of resting CTP appeared to improve discrimination (AUC: 0.75, 95% CI 0.69-0.81, P value .001) and reclassification (net reclassification improvement: 0.52, P value < .001) of ischemia.Conclusions. The addition of resting CTP analysis acquired from machine learning techniques may improve the predictive utility of significant ischemia over coronary stenosis. (J Nucl Cardiol 2018;25:223-33.)

Deep learning analysis of the myocardium in coronary CT angiography for identification of patients with functionally significant coronary artery stenosis

“…The results indicate that the methods utilizing analysis of the blood flow tend to archive higher per-patient accuracy. The results also show that the proposed approach seems to outperform recently proposed method relying on the LV myocardium analysis proposed by Han et al (2017), employing the method presented by Xiong et al (2015). Also, the achieved accuracy equal to the method performing analysis of the blood flow by Min et al (2012).…”

“…The results demonstrate that myocardial information, derived directly from a single CCTA at rest using a trained CAE, has reasonable predictive ability compared with invasive FFR measurements obtained during ICA. As shown in Table 3, the proposed method seems to outperform the results reported by Han et al (2017). Unlike Han et al (2017), that performed patient classification based on engineered LV myocardium features, we have relied on convolutional auto-encoder to learn the discriminative encodings.…”

“…Unlike Han et al (2017), that performed patient classification based on engineered LV myocardium features, we have relied on convolutional auto-encoder to learn the discriminative encodings. As the characteristics of the LV myocardium in a rest CCTA affected by ischemic changes are not (well-)defined and presently there is no consensus regarding their appearance (Spiro et al, 2013;Xiong et al, 2015), it is likely that our unsupervised approach led to higher diagnostic accuracy than the designed features approach presented in Han et al (2017). The learned encodings were able to represent myocardium voxels accurately and densely, as was demonstrated visually by the small mean squared error between the input and the reconstructed examples (Fig.…”

“…In a CCTA image acquired at rest, where pharmacological stimulation is not applied to reveal reversible perfusion defects (stress CCTA) (George et al, 2009), the texture changes are less pronounced than in stress CCTA and are mostly subtle (Han et al, 2017). It would be extremely challenging to manually label myocardial voxels affected by ischemia.…”

“…Contrary to FFR, the grade of stenosis, especially in the intermediate range 30%-70%, is not necessarily related to its functional significance (Tonino et al, 2009). Han et al (2017) employed the method described by Xiong et al (2015) to classify patients with functionally significant stenosis according to the invasively measured FFR.…”