Factors affecting and computation of myocardial perfusion reference images

Many quantitative analysis methods for myocardial perfusion studies require as a central step a comparison with a 'normal' or average density distribution map or reference image. It has been recognized, however, that the normal distribution can be affected by patient attributes, including sex and weight or body habitus, and by acquisition attributes, including the choice of tracer and the position of the patient during imaging. Some authors have proposed separate reference images for the sexes and the tracer. This approach fails if a large number of binary attributes have to be considered, since one would need 2" reference images for each attribute. The problem is compounded when continuous attributes (e.g. age and weight) are included, especially if the approach is to average separate homogeneous groups for each attribute. We propose to create case-specific reference images for the interpretation of myocardial perfusion studies by creating a model based on the influence of each attribute. From a non-homogeneous population of normal cases, or cases presumed to be normal on the basis of the Diamond and Forrester stratification, the effect of patient and study attributes on the density distribution in the stress image and the density differences between rest and stress images were computed. The effects are computed by multi-linear regression, to account for cross-correlation. Significance is assigned on the basis of a partial Fisher test. The data are myocardial perfusion images matched in 3D to a template by an elastic transformation. Even though there was some cross-correlation in the data, we were able to show independent effects of sex, position (prone or supine), age, weight, tracer combination and stress method (exercise, persantine and adenosine). Taken as a whole, the multi-linear regression demonstrated a significant effect in 72% of the pixels within the myocardial volume. In addition, the distribution predicted by the model was equivalent to average images from homogeneous matched groups. In conclusion, our approach makes it possible to produce case-specific reference images without the need for multiple homogeneous large groups to produce averages for each possible patient or study attribute.