Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: Schematic of GLMdenoise . (A) Inputs and outputs. GLMdenoise takes as input a design matrix (where each column indicates the onsets of a given condition) and fMRI time-series, and returns as output an estimate of the hemodynamic response function (HRF) and BOLD response amplitudes (beta weights). (B) Fitting procedure. The procedure consists of selecting voxels that are unrelated to the experiment (cross-validated R 2 less than 0%), performing principal components analysis (PCA) on the time-series of these voxels to derive noise regressors, and using cross-validation to determine the number of regressors to enter into the model.

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques:

Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: Details of the GLMdenoise fitting procedure . (A) HRF fitting. A canonical HRF representing the response to a brief stimulus (black curve) is convolved with the appropriate square-wave function to predict the response for the condition duration used in a given experiment (red curve). This is the initial seed for the HRF. Iterative linear fitting is then used to estimate the optimal HRF (blue curve). Results are shown for dataset 1 (curves are normalized to peak at one). (B) HRF estimates. Shown are HRF estimates obtained in different datasets. Color scheme same as in (C) . (C) Selecting the number of noise regressors. Voxels passing a minimum threshold are identified (voxels with cross-validated R 2 greater than 0% under any number of noise regressors), and median cross-validated R 2 values are calculated. The minimum number of regressors necessary to achieve within 5% of the maximum performance is selected. The performance curves are generally U-shaped, indicating that noise regressors help but too many noise regressors hurt performance.

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques:

Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: The Denoise Benchmark (DNB) . We designed an architecture that enables automatic evaluation of a candidate denoising method. (A) Cross-validation accuracy. Leave-one-run-out cross-validation is used to quantify the accuracy of the denoising method. In each iteration of this procedure, the denoising method is trained on all runs except one and is asked to predict the task-related signal in the left-out run. Predictions are aggregated across the left-out runs, and the accuracy of the predictions is quantified using coefficient of determination ( R 2 ). (B) Signal-to-noise ratio (SNR). Variability of beta weight estimates across the cross-validation iterations is used to estimate SNR. (C) Candidate denoising methods. Any denoising method that conforms to the prescribed application programming interface (API) can be evaluated in the DNB architecture. Note that the cross-validation used in the DNB is distinct from any internal resampling scheme that might be used by a denoising method (such as the cross-validation used within GLMdenoise).

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques:

Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: GLMdenoise improves accuracy and reliability of BOLD response estimates . Using the DNB, we compared the accuracy and reliability of GLMdenoise to that of an analysis involving no noise regressors (termed Standard GLM). (A) Comparison of R 2 for an example dataset. Each dot indicates cross-validated R 2 values for an individual voxel. (B) Summary of changes in R 2 . Voxels are binned according to the cross-validated R 2 of Standard GLM (bin size 10%). For each bin with at least five voxels, we compute the increase in R 2 provided by GLMdenoise and plot a line indicating the 95% range of results. GLMdenoise provides more accurate BOLD response estimates for nearly all voxels. (C) Comparison of SNR for an example dataset. Format same as (A) , except that only voxels passing a minimum threshold are shown (voxels with cross-validated R 2 greater than 0% for either model). (D) Summary of changes in SNR. Format same as (B) , except that voxels are binned according to SNR (bin size 1). For each bin, we compute the median increase in SNR for each dataset and then the median of these values across datasets. The results are shown as thick black lines (for bins with contributions from at least two datasets). On average, GLMdenoise provides more reliable BOLD response estimates than Standard GLM.

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques:

Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: GLMdenoise outperforms other denoising methods . Using the DNB, we quantified the cross-validation accuracy of a variety of denoising methods on a large number of datasets. (A) Results for individual datasets. For each dataset, we summarize the performance of a method by plotting the median cross-validated R 2 value obtained under that method. Error bars indicate 68% confidence intervals and were obtained via bootstrapping. (B) Overall results. To summarize performance across datasets, we normalize the pattern of results from each dataset such that Standard GLM corresponds to 0 and the best-performing method corresponds to 1. We then compute the mean of this pattern across datasets (error bars indicate standard error of the mean). As an alternative performance summary, we count the number of datasets for which a given method achieves the best or nearly the best performance (specifically, the number of datasets for which the median performance of a method either is the best or provides at least 95% of the performance improvement provided by the best method). The number of datasets (out of 21 total datasets) is indicated in the legend.

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques:

Journal: Frontiers in Neuroscience

Article Title: GLMdenoise: a fast, automated technique for denoising task-based fMRI data

doi: 10.3389/fnins.2013.00247

Figure Lengend Snippet: Example activation maps . As an intuitive way to visualize SNR improvements, we show maps of t -values obtained using Standard GLM and maps obtained using GLMdenoise. Maps have been thresholded at t > 3 and are overlaid on the mean functional volume. (A) Activation map for dataset 3, slice 11, condition 31. The green arrow indicates an activated region that exhibits substantial increases in t -values when using GLMdenoise. The blue arrow indicates a region that exhibits activation under GLMdenoise but not under Standard GLM. (B) Activation map for dataset 7, slice 11, condition 24. Format same as (A) .

Article Snippet: MATLAB code implementing GLMdenoise is available at http://kendrickkay.net/GLMdenoise/ .

Techniques: Activation Assay, Functional Assay