Functional Magnetic Resonance Imaging (fMRI) data sets are four dimensional (4D) and very large in size. Compression can enhance system performance in terms of storage and transmission capacities. Two approaches are investigated: adaptive DPCM and integer wavelets. In the DPCM approach, each voxel is coded as a 1D signal in time. Due to the spatial coherence of human anatomy and the similarities in responses of a given substance to stimuli, we classify the voxels by quantizing autoregressive coefficients of the associated time sequences. The resulting 2D classification map is sent as side information. Each voxel time sequence is DPCM coded using a quantized autoregressive model. The prediction residuals are coded by simple Rice coding for high decoder throughput. In the wavelet approach, the 4D fMRI data set is mapped to a 3D data set, with the 3D volume at each time instance being laid out into a 2D plane as a slice mosaic. 3D integer wavelet packets are used for lossless compression of fMRI data. The wavelet coefficients are compressed by 3D context-based adaptive arithmetic coding. An object-oriented compression mode is also introduced in the wavelet codec. An elliptic mask combined with the classification of the background is used to segment the regions of interest from the background. Significantly higher lossless compression of 4D fMRI than JPEG 2000 and JPEG-LS is achieved by both methods. The 2D classification map for compression can also be used for image segmentation in 3D space for analysis and recognition purposes. This segmentation supports object-based random access to very large 4D data volumes. The time sequence of DPCM prediction residuals can be analyzed to yield information on the responses of the imaged anatomy to the stimuli. The proposed wavelet method provides an object-oriented progressive (lossy to lossless) compression of 4D fMRI data set.
Proceedings of Spie--the International Society for Optical Engineering, 2005, p. 190-201