aaalgo
commented
1 year ago Help from NSD google group
https://groups.google.com/g/nsd-users/c/FnHmr4F6B7k
The hemodynamic response is delayed and is broad in time. And also, it is unknown, and might vary across brain voxels and subjects. It is a tricky issue. It is discussed at length in the NSD Nat Neuro paper. Also, we provide "betas" which are BOLD response amplitudes using certain techniques that get around these complications. Perhaps you can read that in depth.
ChatGPT explanation
While the hemodynamic response to neural activity is delayed and broad in time, the estimation of the precise delay of the BOLD signal for each voxel is not typically performed directly on the beta values in fMRI data analysis. Instead, researchers typically use a model-based approach to estimate the hemodynamic response function (HRF) and its delay in a region of interest or across the entire brain.
One common approach is to use a canonical HRF model, which is a mathematical function that describes the typical shape of the hemodynamic response across different brain regions and individuals. The canonical HRF is convolved with the experimental design to generate a model prediction of the expected BOLD response, and the delay of the HRF is adjusted until the model prediction matches the observed BOLD response in the data.
The beta values in fMRI data can then be used to estimate the amplitude of the BOLD response to different experimental conditions or stimuli, while taking into account the estimated delay of the HRF. These beta values can be used to perform statistical tests on the relationship between neural activity and experimental conditions.
It is important to note that the estimation of the HRF delay is a complex and uncertain process, and there can be substantial variability across brain voxels and individuals. The NSD dataset provides betas that are estimated using certain techniques that account for these complications, but the precise estimation of the HRF delay is still an active area of research in the field of fMRI data analysis.
The goal is to establish a mapping between the coco images and the fMRI volumes in the time series. This is how we currently do this.
The following two already have time correspondence, except their lengths do not match.
nsddata_timeseries/ppdata/subj01/func1pt8mm/design/design_session0?_run0?.tsv(Length 225)nsddata_timeseries/ppdata/subj01/func1pt8mm/timeseries/timeseries_session0?_run0?.nii.gz(Length 226)If the i-th row of the design file is not zero, it gives information about a coco image. We assign to it the (i + RESPONSE_DELAY)-th frame in the time series.
RESPONSE_DELAY is currently set to 1.
The goal is to pick out the supposed strongest fMRI signal after the subject viewing the image. The rationale behind setting the value to 1 is:
What is the best way to do this? Should we run some curve fitting model with all 3 frames to calculate some trend signal?