Deep Learning Analysis of Upright-Supine High-Efficiency SPECT Myocardial Perfusion Imaging for Prediction of Obstructive Coronary Artery Disease: A Multicenter Study

DISCUSSION

In this study, we applied DL to automatically combine upright and supine MPI polar maps and to predict obstructive CAD. The performance of DL was compared to clinically established combined perfusion quantification by upright and supine TPD using the rule of the concomitant location of the defect (8,13), and to on-site clinical readers. D-SPECT MPI data were collected in a large multicenter registry, to our knowledge the largest to date with ICA correlations, with the number of samples per vessel disease similar to the number of samples per category in computer vision datasets or new applications in medicine (7,20,21).

In a novel leave-one-center-out cross-validation procedure, equivalent to external validation, we observed that DL from upright and supine polar MPI images outperformed cTPD in the prediction of obstructive disease. The observed gains for CAD prediction were greater than those obtained by attenuation-corrected TPD, or by visual analysis of attenuation-corrected SPECT MPI images (22). These gains by DL derive from an improved processing of the same information used by cTPD with no additional testing, radiation, or cost. It should also be kept in mind that there is an upper limit in diagnostic MPI performance measured versus ICA due to the physiologic constraint of prediction of stenosis from a perfusion defect.

We demonstrate that DL matched the sensitivity/specificity of disease prediction by clinical expert reading, even though on-site reader experts made their diagnosis with all imaging data (including rest scans and gated scans) as well as patient’s clinical information, all unknown by DL and cTPD. Moreover, experts reported in their own laboratory, whereas DL was trained on data excluding patients from the tested laboratory. This validation approach conservatively reflects the performance estimate of the generalization of the DL approach to unseen centers for potential external deployment of such technology.

The proposed DL procedure performs an automatic end-to-end prediction of CAD (from the image pixels to the predicted CAD score) of upright and supine MPI. This demonstrates DL flexibility and extends the previous single-view approach (7). The DL procedure learned image statistics from supine and upright maps and integrated them with sex information to compute a score for obstructive CAD outperforming cTPD prediction. The DL procedure was able to capture complex relationships that were not easily captured with the rule for upright and supine concomitancy as observed on the clinical examples in Figure 6.

We observed AUC improvements for men and for patients undergoing pharmacologic stress, same-day rest/stress, 2-d stress/rest, or stress-only MPI. We also observed per-patient and per-vessel AUC improvements by DL as compared with cTPD, for each participant center. Some intercenter AUC variability was observed, likely due to the differences in patients between centers. For instance, we observed a higher proportion of patients undergoing the exercise stress test in center 1 than the other 2 centers (Supplemental Table 1; supplemental materials are available at http://jnm.snmjournals.org). The variability may also be due to the differences in imaging protocols for obtaining the upright and supine images among the centers.

The proposed DL method can be easily and immediately deployed clinically. The expected increased performance should be reproduced in new upright/supine data from unseen centers as observed here after rigorous external validation. Potentially, the method can be extended for the simultaneous analysis of prone and supine images or for attenuation-corrected and non–attenuation-corrected images. The proposed method is standalone and uses only image data, combining optimally the upright and supine views, which is difficult visually. In comparison to machine learning models combining image and clinical information (23), the presented technique does not require entering clinical information by the physician, which is potentially dependent on the automatic access to the electronic health records. The predicted per-vessel probabilities of obstructive CAD could be integrated with the quantitative software, in the form of polar maps (Fig. 6) to aid the clinician in final reporting. The execution time of such DL models (<1 s), allows for routine application of what can be the first practical deployment of DL technology to nuclear medicine and cardiac imaging and one of the first examples of DL application in clinical practice.

This study has several limitations. First, visual stenosis on ICA was used as the gold standard, which is known to overestimate the prevalence of functionally significant disease when compared with fractional flow reserve. Fractional flow reserve measurements were not available in this population, as these are not commonly performed clinically. The accuracy of stenosis interpretation may also differ between centers. The established endpoint interval of 180 d between SPECT MPI and clinical indicated ICA may be considered long. We did use a 70% diameter stenosis cutoff for the 3 major coronary arteries to indicate functionally significant lesions (24). Second, in this study, we used upright and supine polar maps acquired in D-SPECT cameras because this protocol is specific to this camera. Motion correction was not considered for these scanners because it was not available at any of the participant centers at the time when the images were collected and interpreted; however, the proposed simultaneous processing of 2-position images could potentially mitigate motion artifacts in 1-position images. Third, rest scans and ischemia were not considered; however, in this work, we studied patients without known CAD, for which, traditionally, stress variables are used for detection of obstructive disease. It is possible that externally validated machine learning models, also incorporating clinical information into the scoring of obstructive disease (5), provide further enhancements in AUC performance, especially for the prediction of prognostic outcomes.