Sparse representation of dense motion vector fields for lossless compression of 4-D medical CT data

We present a new method for data-adaptive compression of dense vector fields in dynamic medical volume data. Conventional block-based motion compensation used for temporal prediction in video compression cannot conveniently cope with deformable motion typically found in medical image sequences encoded over time. Based on an approximation of physiologic tissue motion between two succeeding slices in time direction computed by optical flow methods, we find the most significant motion vectors with respect to their prediction capability for a second 2-D slice out of the first one. By coding the components of these vectors, we are able to reconstruct a high quality dense motion vector field at the decoder using only minimal side-information. We show that our approach can achieve a smoother prediction than block-based motion compensation for such data, reducing storage demands in spatially predictive lossless compression. We also show that such a predictive approach can yield better compression ratios than JPEG 2000 intra coding.