BLM Background Information

The access of the VTC data for latency estimation is performed in one out of three possible modes:

1. VOI-based access and VOI-based estimation.
Given the current VOI definition (visible in the Region-Of-Interest dialog) the VTC time courses are spatially averaged within each separated VOI and one value per VOI is provided for each of the four estimated parameters.

2. VOI-based access and voxel-based estimation.
Given the current VOI definition (visible in the Region-Of-Interest dialog) the VTC time courses are considered voxel-wise only from those voxels which belong to one of the VOIs and one value per voxel is provided for each of the four estimated parameters.

3. VTC Mask-based access and voxel-based estimation.
A mask file (*.msk) can be provided by the user to select the VTC time-courses for voxel-wise latency estimation.

Once specified the protocol condition defining the event-type of interest, the average and single-trial epoch-based responses are isolated from the VTC time-courses. It is possible to specify an interpolation factor (2-3-4) to upsample the VTC data, thus increasing the temporal resolution of the BLM analysis. The total duration of the window of estimation must be specified in milliseconds.

The BOLD latency estimation is based on a piece-wise linear (PWL) fitting (pseudo-trapezoid) with initial and final (optional) baseline and four break-point estimates.

In the single-trial mode an estimate is provided for each repetition of the selected condition. In this case, BLM maps are generated not only for the average trial responses (averaged across all the trials of the single VTCs and, in case of multiple VTCs, further averaged across all VTCs) but also for the average latency of all single trials from all VTCs. In VOI-based VOI-level mode (1) all the signals, the fits and the latencies for each trial are saved in the TXT files and will be available for all kinds of statistical analysis outside Brain Voyager (e.g. Excel, Statistica, SPSS, …).

The results of the BLM analysis is saved as "*.blm" files for all three modes of operation and can be saved in any later session of the main program.

Note 1. The response shape (pseudo-trapezoidal) fitting procedure allows a robust estimation of the BOLD latencies in the presence of noise-related fluctuations in the data and complex shapes of the BOLD response itself. However, depending on the temporal window-of-interest, the original TR and the interpolation factor, the resulting computation time needed for the non-linear fit quite extensive. To avoid unexpected long-lasting calculations, it is suggested to start using the VOI-based approach (mode 1) and to estimate the computation time for high interpolation factors, extended windows, both in single- and average- trial modes. After having an estimate of the computation time for the available number of VOIs as well as of the accuracy of the fits (which may imply a fine tuning of the temporal window of operation and/or the removal of the “return-to-baseline” constraint), the computation time for the number of VTC and masked voxels which would be included in the eventual voxel-based BLM analysis (mode 2 or 3) can be roughly estimated.

Note 2. Sometimes the BOLD responses may exhibit more complex shapes than a simple gamma-like response, especially in those experiments where multiple tasks are coupled in series, following the time-locking external stimulus (e.g. a cue followed by a presentation, followed by a judgment task) . In such cases, it is very important to optimize the temporal window of operation to capture the right latencies and in the most accurate way. This is made possible by making more or less flexible the fitting procedure. The two snapshots below illustrate possible behaviors of the fits in the presence of (sort of) “two-phase” responses, exhibiting two peaks with variable relative weighting. Here we show the same four signals analyzed in two different possible time windows (22s in the upper row and 8s only in the lower row). The upper snapshot shows the results of the fit when using extended temporal windows, while the lower snapshot illustrates the fit when the temporal window is reduced in such a way to model only the “early bump” of the response. Besides the temporal window, the user can decide either to force or not to force an explicit “return-to-baseline” part of the response (see also the detailed operation document). If yes, a signal fit is accepted only when a complete return to baseline of the response is detected (snapshot 1). Depending on the ratio between possible multiple phases of the response, the presence of multiple peaks may produce the effects of slightly altering the onset estimate of the first peak (like in 1) or the program may simply decided to skip the first peak and keep the second when this latter becomes the prevalent effect in size (like in 3 and 4). Instead, if the user decides not to force the “return-to-baseline” while optimally tuning the temporal window of observation around the part of interest of the response, any signal behavior can be addressed properly and the first “main” onset of the response can be detected regardless of the size, as is in the spirit of the BLM analysis to separate as much as possible the pure latency effect from the pure size effect. This way the resulting estimate will be allowed to be highly optimized with respect to the onset latency (snapshot 2) and even the presence of small bumps in the response can be considered properly, if this is the intention. Of course, because the offset of the response is not constrained anymore, the duration latency cannot be used. In conclusion, it is possible to use shorter temporal windows focused on specific temporal “components” of the BOLD responses in change of the full BOLD response duration latency estimation.