Dear Ping,

Like suggested by Colin, the most trivial reason would be an accidental flip - trivial but not necessarily easy to trace back. To rule out this aspect start with a new Dicom import and check orientations of the imported data and those of the Dicom files themselves. Check orientations of the functional data and the structural volumes, does the sulcal pattern agree or is one sequence flipped relative to the other? Structural and functional sequences have different scanner settings, maybe the information was lost for only one of those. Maybe you also have trustworthy data for one of the subjects from another appointment or you can still acquire such data? Check the settings at the scanner, maybe the patients' registration was performed incorrectly - the data might require some manual adjustments then. I remember acquiring some animal data, to avoid some dropouts we "tricked" the scanner by switching from supine to prone during their registration (instead of adjusting the encoding direction, don't remember the details but there was something that could not be adjusted via the sequence card) although the animals were scanned in supine position, and we had to reorient the data yourselves afterwards. 

In case you looked at data with other programs, maybe it's just a viewing issue due to their default settings (radiological vs. neurological orientation) - check data within SPM then (its default is neurological orientation = left hemisphere on the left). In older versions of SPM, check the settings for defaults.analyze.flip (see http://atonal.ucdavis.edu/lab_howto/fmri/image_orient_display.htm ).

Alternatively, it might also be the case that the stimulation was flipped. Maybe just the labels were switched. The stimulus projectors also usually permit one to mirror the stimulation & in case someone else has switched those settings it might be difficult to detect for the next users as long as their experiments don't include written instructions or other obvious indicators that would be visible to the experimenter / reported by the subjects. With a mirrored visual stimulation you would indeed observe mirrored activation patterns in visual areas, while motor activations would be lateralised the usual way (in contrast to mirrored data acquisition/reconstruction, in which motor activations would be on the "wrong" side as well). However, activation patterns are a rather bad indicator, as in a particular paradigm it might be the opposite.

> This ipsilateral activation pattern is rare in previous studies.
Well, in general one would expect contralaterally heightend activations, but it might well depend on the exact settings. How much do these studies differ from yours? Is it attending to one hemifield blockwise or does the instruction switch, is it bilateral stimulation, is it a single stimulus or an array, is there a mask, ...?

Best

Helmut