Comparison of fully automated computer analysis and visual scoring for detection of coronary artery disease from myocardial perfusion SPECT in a large population

Abstract

We compared the performance of fully automated quantification of attenuation-corrected (AC) and noncorrected (NC) myocardial perfusion SPECT (MPS) with the corresponding performance of experienced readers for detection of coronary artery disease (CAD).

Methods: Rest-stress (99m)Tc-sestamibi MPS studies (n = 995; 650 consecutive cases with coronary angiography and 345 with likelihood of CAD < 5%) were obtained by MPS with AC. The total perfusion deficit (TPD) for AC and NC data was compared with the visual summed stress and rest scores of 2 experienced readers. Visual reads were performed in 4 consecutive steps with the following information progressively revealed: NC data, AC + NC data, computer results, and all clinical information.

Results: The diagnostic accuracy of TPD for detection of CAD was similar to both readers (NC: 82% vs. 84%; AC: 86% vs. 85%-87%; P = not significant) with the exception of the second reader when clinical information was used (89%, P < 0.05). The receiver-operating-characteristic area under the curve (ROC AUC) for TPD was significantly better than visual reads for NC (0.91 vs. 0.87 and 0.89, P < 0.01) and AC (0.92 vs. 0.90, P < 0.01), and it was comparable to visual reads incorporating all clinical information. The per-vessel accuracy of TPD was superior to one reader for NC (81% vs. 77%, P < 0.05) and AC (83% vs. 78%, P < 0.05) and equivalent to the second reader (NC, 79%; and AC, 81%). The per-vessel ROC AUC for NC (0.83) and AC (0.84) for TPD was better than that for the first reader (0.78-0.80, P < 0.01) and comparable to that of the second reader (0.82-0.84, P = not significant) for all steps.

Conclusion: For detection of ≥70% stenoses based on angiographic criteria, a fully automated computer analysis of NC and AC MPS data is equivalent for per-patient and can be superior for per-vessel analysis, when compared with expert analysis.

Figure 1

Number of cases with changed diagnosis in each subsequent step for both automated and visual analysis. * Indicates significant difference compared to a prior step (p < 0.05).

Figure 2

Diagnostic performance of automatic analysis versus visual analysis for detection of ≥ 70% coronary artery lesions on per-patient basis (Number of patients with ≥ 70% stenotic lesion on cardiac catheterization = 463), * indicates significant difference compared to a prior step (p < 0.05). The automated analysis was also compared to visual analysis (NC vs. V1 and AC vs. V2–V4). Green color signifies that visual analysis was better than automated, while red color signifies that automated analysis was better than visual analysis (p < 0.05).

Figure 3

The receiver operating characteristic (ROC) curves comparing the automated versus visual reads on per-patient basis 2 readers for detection of ≥ 70% stenosis. * Indicates significant difference compared to a prior step (p < 0.05). The automated analysis was also compared to visual analysis (NC vs. V1 and AC vs. V2–V4), # indicates significant difference (p < 0.01). Red color signifies that automated analysis was better than visual analysis (p < 0.01).

Figure 4

Diagnostic performance of automatic analysis versus visual analysis for detection of ≥70% coronary artery lesions on per-vessel basis in all vessels. * Indicates statistically significant difference compared to a prior step (p < 0.05). The automated analysis was also compared to visual (NC vs. V1 and AC vs. V2–V4). Green color signifies that visual analysis was better than automated analysis; red color signifies that automated analysis was better than visual analysis (p < 0.05).

Figure 5

The receiver operating characteristic (ROC) curves comparing the automated versus visual reads on per-vessel basis for both readers for detection of ≥ 70% stenosis. * Indicates statistically significant difference compared to a prior step (p < 0.01). The automated analysis was also compared to visual (NC vs. V1 and AC vs. V2–V4) and # signifies significant difference (p < 0.01). Red color signifies that automated analysis was better than visual analysis (p < 0.01).