Diagnostic accuracy of an artificial neural network compared with statistical quantitation of myocardial perfusion images: a Japanese multicenter study

Abstract

Purpose: Artificial neural networks (ANN) might help to diagnose coronary artery disease. This study aimed to determine whether the diagnostic accuracy of an ANN-based diagnostic system and conventional quantitation are comparable.

Methods: The ANN was trained to classify potentially abnormal areas as true or false based on the nuclear cardiology expert interpretation of 1001 gated stress/rest ^99mTc-MIBI images at 12 hospitals. The diagnostic accuracy of the ANN was compared with 364 expert interpretations that served as the gold standard of abnormality for the validation study. Conventional summed stress/rest/difference scores (SSS/SRS/SDS) were calculated and compared with receiver operating characteristics (ROC) analysis.

Results: The ANN generated a better area under the ROC curves (AUC) than SSS (0.92 vs. 0.82, p < 0.0001), indicating better identification of stress defects. The ANN also generated a better AUC than SDS (0.90 vs. 0.75, p < 0.0001) for stress-induced ischemia. The AUC for patients with old myocardial infarction based on rest defects was 0.97 (0.91 for SRS, p = 0.0061), and that for patients with and without a history of revascularization based on stress defects was 0.94 and 0.90 (p = 0.0055 and p < 0.0001 vs. SSS, respectively). The SSS/SRS/SDS steeply increased when ANN values (probability of abnormality) were >0.80.

Conclusion: The ANN was diagnostically accurate in various clinical settings, including that of patients with previous myocardial infarction and coronary revascularization. The ANN could help to diagnose coronary artery disease.

Keywords: Artificial intelligence; Computer-aided diagnosis; Coronary artery disease; Diagnostic imaging; Nuclear cardiology.

Fig. 1

Myocardial perfusion study and artificial neural network (ANN) analysis of 70-year-old man after percutaneous coronary intervention to the left circumflex coronary artery. Numbers indicate probability of abnormality. Basal lateral ischemia is evident in short-axis images (upper panel), whereas the ANN system identified abnormality in stress (probability) and subtraction (probability) images with probabilities of 0.96 and 0.91, respectively. Other regions with probability of <0.5 were considered insignificant

Fig. 2

Receiver operating characteristics (ROC) analysis of stress defect (a), rest defect (b) and ischemia (c) using the scoring method (upper panel) and the artificial neural network (ANN; lower panel). All areas under ROC curves (AUC) were higher for the ANN (p < 0.0001)

Fig. 3

Comparison of the scoring method (upper panel) and artificial neural network (ANN; lower panel) based on expert judgments. Positive and negative judgments significantly differed in all comparisons of stress defects (a), rest defects (b) and ischemia (c)

Fig. 4

Relationship between scoring methods and probability of abnormality judged by the ANN. Dotted vertical lines indicate probability of 0.8, and blue lines indicate mean values for probabilities of <0.8 and ≥0.8. Squares and circles denote positive and negative stress defect, respectively by expert interpretations. Red and black marks denote positive and negative ischemia, respectively by expert interpretations