Computer Assisted Radiology and Surgery. H Lemke et al eds. Elsevier Science 1999 (pp139-143) This work has been partially supported by grants FIS 97/270 and Comunidad de Madrid 8.1/49/1998 Increased Sensitivity and Position Accuracy in the Detection of Brain Activation by fMRI using Multiscale Analysis JA Hernández ab , M Desco a , A Santos b , C SantaMarta a , F del Pozo b , P García-Barreno a a Medicina Experimental. Hospital General Universitario “Gregorio Marañón”. Dr. Esquerdo 46. E-28007. Madrid. Spain desco@mce.hggm.es http://www.hggm.es/image b Grupo de Bioingeniería y Telemedicina (Universidad Politécnica de Madrid) E.T.S.I. Telecomunicación. Ciudad Universitaria. E-28040 Madrid (Spain) andres@die.upm.es http://www.gbt.tfo.upm.es Introduction Functional Magnetic Resonance Images (fMRI) are traditionally analyzed by applying statistical tests to each and every voxel in the dataset. This results in maps (Statistical Parametric Maps, SPM) showing the voxels that significantly change between different brain states, usually two (rest-activation). These conventional SPM methods do not exploit the neighborhood information (activation regions usually span over several voxels) to increase the statistical power of the tests. At most, this information is sometimes used in a post-processing step to enhance the shape of the detected activation areas [1][2]. In this paper we present an analysis method that makes full use of that neighborhood information. It processes the images at different scales by using multiresolution decomposition based on a wavelet transform. The statistical tests are then applied in the wavelet domain taking benefit from the possible existing spatial correlation. Another remarkable problem in this type of studies is the lack of a ‘gold standard’ to validate the results, as the exact size and position of activation areas is never known in real patient studies. In this work, sensitivity, specificity and spatial resolution of the multiscale results have been assessed with a realistic computer-simulated phantom that resembles fMRI studies where activation areas are known a priori. Material And Methods Algorithm Functional images have been analyzed through wavelet decomposition up to the sixth level. At each level, the null-hypothesis was tested (z-test) at a given p-value and the inverse transform computed using only the significant coefficients which passed the test. In this way,