Base-to-apex gradient abnormality detection task performance in myocardial perfusion PET imaging

Objectives Diffuse luminal narrowing in patients with mild angiographic CAD can give rise to graded longitudinal base-to-apex perfusion abnormalities. This can also occur due to diminished or absent flow-dependent dilation of the epicardial coronary arteries in CAD or in the presence of coronary risk factors. We aimed to develop and investigate a novel task-based assessment paradigm for the detection of myocardial perfusion (MP) gradient abnormalities.

Methods Extensive multiple-realization N-13 Ammonia datasets were generated based on realistic K1, k2 and vB values, w/ & w/o base-to-apex gradients in K1. The dynamic acquisition sequence was 10x12sec, 2x30sec, 1x1min, 1x6min, and static perfusion imaging consisted of an 8min frame after 2min delay. Quantitative MP parameter estimation was performed using constrained non-linear optimization as applied to compartmental modeling, w/ & w/o spatial-constraint regularization, and w/ & w/o Butterworth filtering (4th order, cutoff 0.5 cycle cm^-1). Circumferential profiles were used to divide the heart into septal, anterior, lateral and inferior quadrants. Then, for each quadrant, the activity from base to apex was best fit using a 3rd-degree polynomial, the 1st-derivative of which indicated the spatial slope along the long axis for each quadrant. ROC area-under-curve (AUC) analysis was finally performed for gradient detection at each longitudinal location.

Results Base-to-apex gradients as low as 2.5% per cm for a given quadrant could be detected (AUC>0.85). This was especially the case for the septal and lateral quadrants (AUC>0.90) when utilizing Butterworth post-filtering in both static & dynamic MP imaging, for the latter especially if combined with spatial-constraint regularization.

Conclusions In addition to conventional MP defect detection, novel image reconstruction & parametric image generation methods may be assessed in the context of gradient abnormality detection. In particular, our results demonstrate that appropriate post-filtering and spatial regularization can significantly enhance gradient detection task performance.