Development of a normal database for quantitative analysis and an artificial intelligence application for interpretation of myocardial perfusion scintigraphies

In this thesis, a normal database of myocardial perfusion scintigraphies from patients with completely normal coronary angiographies (n=79) was created and installed in the 4DMSPECT program. 4DMSPECT with the normal database created, ECToolbox and QPS with their supplied databases were compared in a test group consisting of 223 patients with known angiographies. In the comparisons made between the ROC curves created with Summed Stress and Summed Difference Scores obtained from programs, 4DMSpect SSS was the most successful (AUC = 0.768) and there were statistically significant differences between 4DMSpect SSS and 4DMSpect SDS (p=0.031), ECToolbox SSS (p=0.035), ECToolbox SDS (p=0.015), QPS SDS (p=0.001) respectively. In the second part of the thesis an artificial intelligence (AI) application consisting of six artificial neural networks was created. In a test group of 65 patients, this application was tested against human readers and possible collaborations of human readers and AI was researched. When the sensitivity of AI was held at 70% and specificity at 68% with the adjustment of threshold, there were no statisticaly significant differences between the performances of human readers and AI. After the experimental collaboration of human readers and AI, three readers had statistically significant increases in their performances (accuracy of a reader increased from 66% to 77% (p=0.04), specificity of a reader increased from 63% to 80% (p=0.01) and another from 53% to 71% (p=0.03)). When the changes in performances of all readers with and without AI collaboration were evaluated, no statistically significant difference in sensitivities was observed (from 72% to 72% with AI, p=0.855), however, statistically significant difference in specificities was observed (from 64% to 76% with AI, p<0.001) and statistically significant difference in accuracies was observed (from 67% to 75% with AI, p<0.001).