Inferência baseada em voxel para fMRI

• Ano de defesa: 2007
• Autor(a) principal: Winkler, Anderson Marcelo
• Tipo de documento: Dissertação
• Tipo de acesso: Acesso aberto
• Idioma: por
• Instituição de defesa: Universidade Tecnológica Federal do Paraná; Curitiba; Brasil; Programa de Pós-Graduação em Engenharia Elétrica e Informática Industrial; UTFPR
• Palavras-chave em Português: Radiologia médica; Ressonância magnética; Engenharia biomédica; Radiology, Medical; Magnetic resonance; Biomedical engineering; CNPQ::ENGENHARIAS::ENGENHARIA BIOMEDICA; Engenharia Elétrica
• Link de acesso: http://repositorio.utfpr.edu.br/jspui/handle/1/26343

Background: Statistical parametric maps are constructed from a massive, univariate, voxel-wise hypothesis testing. Type I errors may happen very often if such a large amount of tests are performed. Though this is a central problem for neuroimaging studies, the best approach is still unclear. Two approaches have emerged as the most suitable for fMRI: the random field theory (RFT) and false discovery rate (FDR). RFT has become the de facto standard method for controlling the family-wise error rate (FWE), despite its complexity and restrictive assumptions. If the researcher is willing to accept some false-positives within the image, methods for controlling the FDR, as the Benjamini and Hochberg (B&H) procedure, can provide more liberal thresholds, with minimal assumptions. This study also features a literature review on recent advances in the field. Objective: Evaluate the performance of RFT and B&H procedures, as well the traditional Bonferroni correction (BON) and no correction (UNC). Method: A real “null” fMRI dataset was acquired at 1.5 T. A temporal high-pass filter was applied, and the brain volumes were randomly permuted, thus avoiding the potential bias due to autocorrelation. Patches of boxcar-like “activation” were added using the canonical haemodynamic response function, which parameters were slightly variable for each “activation” period. The general linear model was applied to both rest and added “activation” datasets, and with and without spatial smoothing. Estimation of the smoothness was based on the residuals of the model fit. For each of these conditions, t-maps were generated and thresholded using the UNC, BON, B&H and RFT procedures. Results: All the evaluated methods resulted in adequate control over error rates, within their theoretical assumptions and limitations. The Bonferroni correction was less conservative than expected. The B&H procedure resulted in variable thresholds, providing better control over FDR for large areas of simulated activity. B&H procedure was also influenced by smoothness. Conservative results were obtained for RFT, but the observed error was close to the nominal level for smoothed maps with a filter of 2.0 voxels of width. Smoothing induced bias in voxel-based measurements.