fmriquality

This shows you the differences between two versions of the page.

Both sides previous revision Previous revision | |||

fmriquality [2010/10/26 21:23] jochen |
fmriquality [2010/11/26 19:50] (current) jochen |
||
---|---|---|---|

Line 38: | Line 38: | ||

precision), plus some temporary arrays and, if motion correction | precision), plus some temporary arrays and, if motion correction | ||

is selected, with further memory allocation required!</file> | is selected, with further memory allocation required!</file> | ||

+ | |||

+ | ===== Algorithm ===== | ||

+ | The basic algorithm can be unpacked into the following steps: | ||

+ | * read in the data (memory must at least allow one run to be loaded!) | ||

+ | * reserve additional memory for some computation outcomes | ||

+ | * compute mean and standard deviation (over time) images | ||

+ | * detect foreground (brain) and background (black/air voxels) separately | ||

+ | * compute average time courses for foreground and background (global and per-slice) | ||

+ | * estimate the smoothness of the data (one value per volume; over time) | ||

+ | * create a conservative estimate of the average (temporal) standard deviation of the background | ||

+ | * compute a "global signal-to-noise ratio" (GlobalSNR) image (using this global background noise estimate) | ||

+ | * compute a "local signal-to-noise ratio" (LocalSNR) image (using each voxel's individual standard deviation) | ||

+ | * temporally filter timecourses (univariately) and estimate the amount of variance determined by low frequencies | ||

+ | * if requested, perform motion detection/realignment and re-run the computations | ||

+ | |||

+ | Additionally, at the end of steps that produce time courses, an outlier detection is performed which amasses evidence for a given volume being an outlier. | ||

+ | |||

+ | ==== Foreground / Background detection ==== | ||

+ | The detection of the foreground involves the following steps: | ||

+ | * selection of voxels for which the mean value (over time) exceeds the mean value over the entire 4D data slab | ||

+ | * removal of stray voxels (by one step of 3D erosion following by a back-dilation and logical AND with the original selection) | ||

+ | * sub-selection of the "biggest chunk" of cohesive voxels (clustering) | ||

+ | |||

+ | The detection of the background involves the following steps: | ||

+ | * picking the median of over-time mean values for voxels where the mean value does not exceed the mean value over the entire 4D data slab (in other words the opposite of the preliminary foreground mask!) | ||

+ | * selecting voxels for which the mean value is smaller than this median (very conservative background estimate) | ||

+ | * removal of voxels for which more than 5 per cent of values are exactly 0 (depending on the scanner type and sequence parameters as well as field homogeneity corrections, some voxels are, more or less consitently, all-0, and hence not "true background" voxels | ||

+ | * equally, sub-selection of the biggest chunk (to remove stray voxels) | ||

+ | |||

+ | ==== Background noise estimate ==== | ||

+ | The mentioned (conservative) estimate of the background noise is computed by | ||

+ | * sorting the values of the temporal standard deviation in voxels marked as background | ||

+ | * computing the average over the second and third quartile, so as not have voxels which are, for instance, on the fringe of the brain and, due to motion, for some portion of the run contain actual data, pollute the estimate; the idea being that the average noise picked up in air voxels should not be influenced by actual matter, even if only present in part of a run | ||

+ | |||

+ | ==== Outlier detection ==== | ||

+ | The following criteria are being considered for the detection of outliers: | ||

+ | * the estimated smoothness in a given volume is further away than 6 standard deviations from the mean | ||

+ | * the global foreground time course is further away than 5 standard deviations from the mean | ||

+ | * the absolute of the 1st order derivative of the global foreground time course is further away than 5 standard deviations from the mean (detecting stark signal level shifts) | ||

+ | * the Mahalanobis Distance over the foreground slices' time courses is further away than 5 standard deviations from the mean (detecting patterns in time courses across slices, e.g. when partial-volume effects of motion make a volume an outlier, particularly in interleaved sequences!), this is also done with the absolute of the 1st order derivatives of the foreground slices' time courses | ||

+ | * the temporally fitered versions of the global foreground timecourse as well as Mahalanobis Distance over the filtered foreground slices' time courses are further away than 4 standard deviations from their respective mean | ||

===== Usage ===== | ===== Usage ===== |

fmriquality.txt ยท Last modified: 2010/11/26 19:50 by jochen