The present study showed that the diagnostic accuracy of version 1.1 was better than that of version 1.0 when assessed using the same validation database. The diagnostic accuracy was obviously improved in patients without a history of myocardial infarction or coronary revascularization (that is, without modification by therapeutic intervention).
Computer-aided diagnosis
Visual assessment of myocardial perfusion SPECT for defects and reversibility is the initial step towards an appropriate diagnosis. Computer-assisted quantitation and evaluation play important roles in aiding visual assessment [11], and the most popular method of predicting prognosis has been defect scoring, such as SSS, SRS, and SDS using a 17- or 20-segment model [12, 13]. The amount of infarction and ischemia can also be determined by statistical analysis of the regional count distribution with assistance from normal databases fitted to a study population [14, 15]. In contrast, the ANN determined the probability of abnormalities in candidate regions based on a learning experience similar to that used to train humans, which might be related to integrated information about defect size, location, extent, severity, regional wall motion, sex, and other factors. Therefore, the ANN might mimic the learning processes through which trainees develop the diagnostic ability to become nuclear cardiology experts. The superior diagnostic accuracy of the ANN system over scoring methods has already been established [2, 3].
Gold standard for training
The definition of a true diagnosis was based on the expert reading for both versions 1.0 and 1.1 in the present study. Since the target of the artificial intelligence applied in this study was to achieve diagnostic accuracy comparable to that of human experts, gold standards of coronary stenosis and fractional flow reserve were not applied. A gold standard comprising physicians’ readings had been implemented in a study using the PERFEX system [1]. Although the detection of (for example) anatomical stenosis might be another target of ANN training, stenosis and physiological ischemia might not be identical [16]. Therefore, if experts cannot identify abnormalities on MPI acquired from patients with triple-vessel disease, an ANN would also be unable to do so. However, even when expert interpretation is defined as truth, the improved diagnostic ability of version 1.1 represents progress and support for clinical applications associated with coronary artery disease. Nevertheless, the ability of version 1.1 to accurately diagnose single-vessel disease was improved when coronary stenosis was the gold standard.
Improvement for detecting ischemia
The major improvement in version 1.1 was in its ability to detect stress-induced ischemia in patients without therapeutic modifications resulting from coronary intervention and without myocardial infarction. From our experience with applying ANN version 1.0, we found that small areas or slight degrees of ischemia were overlooked [3]. Therefore, during the development of the new version, we tried to select more candidate regions of abnormalities, and trained the ANN to identify minor degrees of abnormality. That is, the ANN learned to judge minor abnormalities as positive during the present training and development. The ANN was trained using supervised learning; the quality of the content that experts use to teach the ANN is an important part of software development using artificial intelligence.
Although we could not differentiate contribution of each feature in the neural network system, the integrated learning process was effective for improved diagnostic accuracy. Interestingly, intermediate ANN probability values were more often calculated for detecting ischemia by the version 1.1. Due to this change, sensitivity was improved for detecting ischemia while specificity was kept high (or low false-positive rate).
Neural network for clinical practice
The practical method of applying the ANN to clinical practice should be considered. The relationship between ANN probability and defect scores is not linear [3, 4]. Summed stress, rest and difference scores all steeply increased when the ANN probability was > 0.80, which means that the ANN probability could play a unique role in the diagnosis of coronary artery disease. Clinical decisions as to whether or not infarction and ischemia actually exist on MPI are often borderline, and the truth is not always clear. Under such circumstances, expressing perfusion abnormalities as probabilities might be more practical than simply announcing, for example, that ischemia is suspected or cannot be denied. However, diagnostic relevance should be further investigated since such approaches are not common to medical diagnostics. Since estimated areas of ischemia vary widely among physicians, the presence of defects and ischemia suggested by appropriate software packages would help to reduce the inter-observer variability of clinical interpretations [17].
Limitations
One limitation of the present study is that it included only 106 patients who had undergone coronary angiography. Considering that the diagnostic accuracy of version 1.1 has already been established based on 364 patients [4], the present study seems sufficiently valid for comparisons between the two versions. When patients with old myocardial infarction and post-revascularization conditions were included, truth could not be established. However, more precise analyses including follow-up and prognostic investigations might be feasible in future studies that include a sufficient number of patients.