Automatic selection of arterial input function on dynamic contrast-enhanced MRI images: comparison of different methods

gadolinium images by deconvolution given the arterial input function, AIF(t), and tissue concentration, C(t),: C(t)=CBF٠[AIF(t)⊗R(t)], where R(t) is the tissue residue function. Often AIF is found by manually inspecting tracer concentration maps which is very time consuming and operator dependent. In our study we compare 5 methods of AIF automatic selection, including a novel one. The performance of the methods are tested on a stenosis data set. To analyze the impact of partial volume presence a simulation study was also performed. MATERIALS AND METHODS Patient: 12 patients with various degree of stenosis were studied. The data are obtained with a gradient echo EPI (TR=1560ms, TE=51ms, 30cm field of view, 5mm slice thickness), the bolus dose is 0.2mmol/Kg of Gd-DTPA, at a rate of 5 ml/s in an antecubital vein. For each patient we obtain twelve slice and every slice is composed by fifty images. Simulation: the simulation follows Murase, et al., [5]. The data set include voxels that represent AIF, gray and white matter, partial volume effect and also some trend that represent false AIF. Gray and white curves are obtained by convolution of AIF and an exponential R(t). Methods: Method A is based on Rempp [1] and uses information regarding the peak value, the moment of maximum concentration and the full width at half maximum. Method B is a K-means clustering based on Ashburner [2]. Method C selects AIF voxels on the basis of the ratio between the peak and the time to peak [3]. Method D assumes that standard deviation (SD) maps of concentration data reliably emphasize vasculature over other tissue classes. By applying filters based on SD maps it is possible to select an adequate number of candidate voxels [4]. Method E selects AIF by manually inspection. Method F is novel and based on a hierarchical clustering applied dichotomously, i.e. at each step one of the two clusters is chosen to be reclusterized on the basis of the peak height or, if the difference is not significant, of the time to peak.