Automated Determination of the Arterial Input Function for Quantitative MR Perfusion Analysis

Introduction: Parametric maps of CBF and MTT can be derived from dynamic susceptibility changes by curve-fitting the “ first-pass” tissue response then numerically deconvolving it from the AIF[1,2]. Deconvolution yields a “ residue function” which is scaled by CBF. The singular value decomposition deconvolution, however, is sensitive to the shape of the AIF [1,2]. Thus, the accuracy of the input curve may determine the quality of the deconvolution, and the quality of the CBF and MTT maps. AIF is typically determined by manual identification of “arterial” pixels in the images according to their anatomic location. This task is made difficult by the poor spatial resolution and poor blood-totissue contrast of dynamic EPI scans. Although the image data are acquired as low resolution in space due to acquisition constraints, they have high tissue-contrast, temporally. Thus, we discriminate arterial from non-arterial voxels by inspecting their signals in time. We present a two-stage method for automatically deriving an AIF from the dynamic image data based on expected characteristics of arterial concentration curves. Candidate input functions are generated by maximizing three “arterial-likelihood” metrics. Some metrics have been suggested previously[3]. The “best” of the three candidate curves is chosen by a scoring system based on the first four moments of the AIF distributions. After fitting, AIF curves chosen automatically and manually appear comparable and yield comparable CBF results.