Original Research Calculation of Cerebral Perfusion Parameters Using Regional Arterial Input Functions Identified by Factor Analysis Linda Knutsson, MS, 1 * Elna-Marie Larsson, MD, PhD, 2 Oliver Thilmann, PhD, 1 Freddy Ståhlberg, PhD, 1,2 and Ronnie Wirestam, PhD 1 Purpose: To calculate regional cerebral blood volume (rCBV), regional cerebral blood flow (rCBF), and regional mean transit time (rMTT) accurately, an arterial input function (AIF) is required. In this study we identified a number of AIFs using factor analysis of dynamic studies (FADS), and performed the cerebral perfusion calculation pixel by pixel using the AIF that was located geometrically closest to a certain voxel. Materials and Methods: To verify the robustness of the method, simulated images were generated in which disper- sion or delay was added in some arteries and in the corre- sponding cerebral gray matter (GM), white matter (WM), and ischemic tissue. Thereafter, AIFs were determined us- ing the FADS method and simulations were performed us- ing different signal-to-noise ratios (SNRs). Simulations were also carried out using an AIF from a single pixel that was manually selected. In vivo results were obtained from normal volunteers and patients. Results: The FADS method reduced the underestimation of rCBF due to dispersion or delay that often occurs when only one AIF represents the entire brain. Conclusion: This study indicates that the use of FADS and the nearest-AIF method is preferable to manual selection of one single AIF. Key Words: perfusion; deconvolution; arterial input func- tion; factor analysis of dynamic studies; dynamic suscep- tibility contrast magnetic resonance imaging J. Magn. Reson. Imaging 2006;23:444 – 453. © 2006 Wiley-Liss, Inc. ALTHOUGH DYNAMIC SUSCEPTIBILITY CONTRAST (DSC)-MRI (1) is a promising technique for assessing cerebral perfusion parameters (i.e., regional cerebral blood volume (rCBV), regional cerebral blood flow (rCBF), and regional mean transit time (rMTT)), several problems and pitfalls have been reported. First, it is important to calculate rCBF and rMTT with the use of a reliable deconvolution algorithm, and at present singu- lar value decomposition (SVD) is the most widely used deconvolution tool (2– 4). Another important issue in DSC-MRI is the appropriate registration of the arterial input function (AIF). For example, the standard SVD (sSVD)-based deconvolution (3) shows significant sen- sitivity to any delay and dispersion of the bolus occur- ring between the site of the recorded AIF and the true AIF (i.e., the concentration curve of the artery that ac- tually supplies the tissue of interest), leading to a sig- nificant underestimation of rCBF and overestimation of rMTT (5). Recent studies have presented time-shift in- variant SVD algorithms that take delay effects into con- sideration (6,7); however, it is never easy to compensate for the effects of arterial dispersion when a global AIF is employed. Other problems posed by AIF registration include partial volume effects (8), signal saturation at peak concentration (9), local geometric distortion dur- ing bolus passage, and averaging effects (e.g., false broadening) if more than one pixel is selected for the global AIF. Several authors have recognized the advantages of automatic AIF identification. For example, Rempp et al (10) assigned certain criteria to the full width at half maximum (FWHM), the maximum concentration (MC), and the moment of MC (MMC) of the concentration time curves. Only concentration curves that met the criteria and showed an MC of at least 75% of the highest value observed were selected, and the mean of these concen- tration time curves was used as the AIF. Murase et al (11) used fuzzy c-means (FCM) clustering to determine the AIF automatically. Parameters such as MC (H p ), time of MC (T p ), and FWHM were calculated for each cluster centroid. The cluster centroid that showed the maximum value of H p /[T p  FWHM] was assumed to contain arterial pixels, and the AIF was calculated as the mean value of the concentration time curves from these arterial pixels. Carroll et al (12) proposed an au- tomated algorithm that identified one voxel that opti- mally represented the AIF, using high peak signal 1 Department of Medical Radiation Physics, Lund University Hospital, Lund, Sweden. 2 Department of Diagnostic Radiology, Lund University Hospital, Lund, Sweden. Contract grant sponsor: Swedish Research Council; Contract grant number: 13514; Contract grant sponsors: Crafoord Foundation; T. and E. Segerfalk Foundation; Swedish Society of Medicine. *Address reprint requests to: L.K., Department of Medical Radiation Physics, Lund University Hospital, SE-221 85 Lund, Sweden. E-mail: Linda.Knutsson@med.lu.se Received May 30, 2005; Accepted January 3, 2006. DOI 10.1002/jmri.20535 Published online 7 March 2006 in Wiley InterScience (www.interscience. wiley.com). JOURNAL OF MAGNETIC RESONANCE IMAGING 23:444 – 453 (2006) © 2006 Wiley-Liss, Inc. 444