Coronary artery disease (CAD) is a global epidemic with an increasing impact on healthcare systems [
1]. Significant advances in both diagnosing and treating acute epicardial CAD have improved survival and reduced morbidity during the last decades [
2]. One of the main unresolved issues in diagnosis of chronic CAD represents the definition of clinically relevant ischaemia [
3,
4]. Myocardial perfusion based on cardiovascular magnetic resonance (CMR) imaging provides excellent diagnostic accuracy and prognostic value (summarised in [
4]), and is an established diagnostic method in clinical practice [
1,
5] In clinical practice, perfusion stress CMR is analysed and interpreted based on visually perceptible differences in peaks of contrast signal intensity and contrast kinetics [
6]. Clinical reports usually summarise the extent and transmurality of hypoperfusion, based on the American Heart Association (AHA) 16 segment left ventricle (LV) model [
7], as well as localisation in terms of coronary perfusion territory. Experience reveals that visual analysis of perfusion stress CMR, using the 16 segment model, can be difficult to standardise and to record accurately and reproducibly, as perfusion defects frequently involve several adjacent segments, which are often only partially involved [
8,
9]. Some improvement has been achieved by the subdivision of 16 segments into 32 epi- and endocardial subsegments [
5,
10]. Furthermore, as fully automated analyses, based on voxel-wise quantification, become feasible [
11‐
14], further subdivision may be possible, improving the overall measurement accuracy of regional distribution of myocardial blood flow. However, despite the huge potential, quantitative outputs of voxel-based analyses are reported as an average of all voxel-based measurements, expressed per each transmural segment within the 16 segment model (or one of 32 subsegments, respectively). Consequently, the potential information of voxel-based measurements of spatial differences of myocardial flow is discarded, leading to several obvious problems. Firstly, mixing signals from multiple voxel signals may lead to overestimation of reduced perfusion in segments which are only partially involved. Consequently, the overestimation leads to the underestimation of peak perfusion in normal areas with high inflow of contrast agent and increase in signal intensities. This results in lower effective difference between normal and abnormal perfusion, potentially reducing overall diagnostic performance. Secondly, classifying perfusion defects, in line with presumed coronary artery distributions, may contribute inaccuracies, especially along the border territories in databases with rigid allocation of segments. Meaningful and robust ways of recording and communicating quantification results of myocardial perfusion may be useful to harness the potential of fully automated analyses and to develop reliable diagnostic matrices for artificial intelligence machine learning approaches. We hypothesise that a subdivision of the classical 16 segment model into 32 subsegments (epicardial and endocardial), 48 subsegments (circular division of the 16 segments into 3 segments each) and 96 (sub)-subsegments (dividing the 48 subsegments into epi- and endocardial) would improve the accuracy of myocardial perfusion measurement. Quantitative analysis based on LV segmentation into 96 (sub-)subsegments, each representing approximately 1% of myocardium, may considerably simplify the reporting scheme for the extent of myocardial ischaemia, laying the base for a detailed and robust reporting of voxel-wise analyses for clinical interpretation and databasing. To test this hypothesis, we undertook a proof of concept comparison study of the diagnostic accuracies of myocardial segmentation approaches (transmural 16 and 48 segments, vs. 32 vs. 96 epi- and endocardial subsegments) and based quantitative analysis of stress myocardial perfusion in patients with obstructive CAD and healthy controls.