This is a question from a Medical Physiscist with a reasonable grasp
of basic statistics. All replies will be gratefuly received and should,
as always, be sent directly to me (email address at the bottom) and
not to the list. Thanks.
I work in the field of magnetic resonance imaging and am involved in
modelling the dynamic changes in image signal intensity after
injection of a contrast agent (which works like a dye, causing
the signal intensity of tumours to increase).
We often have a problem with patient motion (breathing, wriggling,
etc.) which can cause large spurious increases or (less often)
decreases in signal intensity. These 'noise spikes' can occur at any
time during imaging (including next to each other) and can have a
profound effect upon the parameter values derived from the model
(via standard non-linear least-squares fitting).
Our data has between 25 and 35 time points and signal intensity
is a continuous, non-integer variable. Random noise is present
in the signal and is typicaly Gaussian in nature with a small
standard deviation relative to true signal changes. A typical
time-series would look like this (the corrupted time-points are
shown as 'o').
TUMOUR SIGNAL INTENSITY
|
| oo
|
|
| xx xx o
| xx xx x
| xx xxxx
| xx xxx x
| x xxxxx
| x o xxxxx
| x
| x
|xxxx
|
--------------------------------------------- TIME
My question is: can anyone recommend a suitable statistical
technique (e.g. filtering or smoothing) that can reliably
identify potential corrupted time-points so that they can be
removed from the NLLS fitting procedure? I'm happy to
sensor/edit the data in this way because the corruption is
arising from a genuine external and systematic influence;
i.e. motion.
At present I'm detecting corrupted time-points just by looking
at the images and the time-course which is too subjective
for my liking.
A possible confounding issue is that the up-slope of the curve
is very steep sometimes and comparable to the rate of change
associated with a positive spike.
-------------------------------------------
David Manton, Ph.D.
YCR Centre for MR Investigations, Hull, UK.
email: [log in to unmask]
-------------------------------------------
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|