Monotonic noise suppression used to improve the sensitivity of fMRI activation maps

We have introduced a new method of removing noise from images that identifies significant extrema and forces the pixel intensities between any two extrema to change monotonicly. The method has some similarities to wavelet denoising methods we worked with several years ago but is generally more stable and is effective on images with lower SNR's. In this paper the method of monotonic filtering is used to increase the sensitivity in functional magnetic resonance imaging (fMRI) studies. We have used the increased sensitivity to improve the temporal resolution in fMRI studies by roughly a factor of six. A motor activation study was acquired with single slice 256 x 256 pixel T2* weighted images; six cycles of finger tapping were acquired. Each cycle consisted of five images of rest followed by five images of right hand finger tapping followed by five images of left hand finger tapping. The z-scores were calculated and used as the activation map. The left and right activations were both clearly visible when all six cycles were used in the analysis. However, no definitive activation was seen for any one cycle. When the original 256 x 256 images were averaged down to 64 x 64 pixel images before calculation of the z-scores, the activations were partially identified. When the original images were filtered using the monotonic noise reduction algorithm, the left activation was clearly visible in three of the six cycles and partially visible in two others. The right activation was partially visible in 4 out of 6 cycles. Optimized noise reduction should improve the results significantly. The ability to use a single cycles is very important in fMRI studies because many stimuli are more difficult to maintain over many cycles and because complex processes such as in cognitive or memory activity do not have simple responses.