problemstatement

The program rapidtide.py looks at fMRI images and determines the delay of a noise signal carried through the brain by the blood.  In healthy normal people, this blood flow varies smoothly from one location to another.  The delays in healthy people have a strong central tendency, and generally fall       
  within a range of -2 to +4 seconds relative to the mode of the distribution.  The delay is calculated by fitting the peak of a crosscorrelation function between the noise in each voxel and a probe signal which is a refined version of the signal moving with the blood.  One common failure mode is that due  
  to noise and autocorrelations in the probe signal, sometimes the fitting routine selects the wrong peak.  The factors that cause this to happen tend to vary slowly in space, so this results in regions that have a smoothly varying delay that is displaced by several seconds from the "correct" delay         
  value.  Rapidtide uses various methods to try to locate and correct these regions, but it is often ineffective.  This is complicated by the fact that vascular pathology can cause the true delay in contiguous regions fed by the same blood vessel to become delayed by several seconds, and we don't want to  
  erroneously shift the delays in those voxels.  Is there a robust approach to solving this sort of problem?