algorithms.diagnostics.timediff

Module: algorithms.diagnostics.timediff

Time series diagnostics

These started life as tsdiffana.m - see http://imaging.mrc-cbu.cam.ac.uk/imaging/DataDiagnostics

Oliver Josephs (FIL) gave me (MB) the idea of time-point to time-point subtraction as a diagnostic for motion and other sudden image changes.

Functions

nipy.algorithms.diagnostics.timediff.time_slice_diffs(arr, time_axis=-1, slice_axis=None)

Time-point to time-point differences over volumes and slices

We think of the passed array as an image. The image has a “time” dimension given by time_axis and a “slice” dimension, given by slice_axis, and one or more other dimensions. In the case of imaging there will usually be two more dimensions (the dimensions defining the size of an image slice). A single slice in the time dimension we call a “volume”. A single entry in arr is a “voxel”. For example, if time_axis == 0, then v = arr[0] would be the first volume in the series. The volume v above has v.size voxels. If, in addition, slice_axis == 1, then for the volume v (above) s = v[0] would be a “slice”, with s.size voxels. These are obviously terms from neuroimaging.

Parameters

arr : array_like

Array over which to calculate time and slice differences. We’ll call this array an ‘image’ in this doc.

time_axis : int, optional

axis of arr that varies over time. Default is last

slice_axis : None or int, optional

axis of arr that varies over image slice. None gives last non-time axis.

Returns

results : dict

T is the number of time points (arr.shape[time_axis])

S is the number of slices (arr.shape[slice_axis])

v is the shape of a volume (rollimg(arr, time_axis)[0].shape)

d2[t] is the volume of squared differences between voxels at time point t and time point t+1

results has keys:

• ‘volume_mean_diff2’(T-1,) array

  array containing the mean (over voxels in volume) of the squared difference from one time point to the next

• ‘slice_mean_diff2’(T-1, S) array

  giving the mean (over voxels in slice) of the difference from one time point to the next, one value per slice, per timepoint

• ‘volume_means’(T,) array

  mean over voxels for each volume vol[t] for t in 0:T

• ‘slice_diff2_max_vol’v[:] array

  volume, of same shape as input time point volumes, where each slice is is the slice from d2[t] for t in 0:T-1, that has the largest variance across t. Thus each slice in the volume may well result from a different difference time point.

• ‘diff2_mean_vol``v[:] array

  volume with the mean of d2[t] across t for t in 0:T-1.

Raises

ValueError : if time_axis refers to same axis as slice_axis

nipy.algorithms.diagnostics.timediff.time_slice_diffs_image(img, time_axis='t', slice_axis='slice')

Time-point to time-point differences over volumes and slices of image

Parameters

img : Image

The image on which to perform time-point differences

time_axis : str or int, optional

Axis indexing time-points. Default is ‘t’. If time_axis is an integer, gives the index of the input (domain) axis of img. If time_axis is a str, can be an input (domain) name, or an output (range) name, that maps to an input (domain) name.

slice_axis : str or int, optional

Axis indexing MRI slices. If slice_axis is an integer, gives the index of the input (domain) axis of img. If slice_axis is a str, can be an input (domain) name, or an output (range) name, that maps to an input (domain) name.

Returns

results : dict

arr refers to the array as loaded from img

T is the number of time points (img.shape[time_axis])

S is the number of slices (img.shape[slice_axis])

v is the shape of a volume (rollimg(img, time_axis)[0].shape)

d2[t] is the volume of squared differences between voxels at time point t and time point t+1

results has keys:

• ‘volume_mean_diff2’(T-1,) array

  array containing the mean (over voxels in volume) of the squared difference from one time point to the next

• ‘slice_mean_diff2’(T-1, S) array

  giving the mean (over voxels in slice) of the difference from one time point to the next, one value per slice, per timepoint

• ‘volume_means’(T,) array

  mean over voxels for each volume vol[t] for t in 0:T

• ‘slice_diff2_max_vol’v[:] image

  image volume, of same shape as input time point volumes, where each slice is is the slice from d2[t] for t in 0:T-1, that has the largest variance across t. Thus each slice in the volume may well result from a different difference time point.

• ‘diff2_mean_vol``v[:] image

  image volume with the mean of d2[t] across t for t in 0:T-1.