Lossless image compression methods for PET imaging

[[abstract]]The authors present a Region-of-Interest-based lossless image compression method which combines segmentation, smooth filling, different quantizers, the lossy coding algorithms and the lossless variable length coding. First, the 16-bit PET imaging is quantized to 8-bits using the uniform quantizer or the non-uniform quantizer of which the quantization error is losslessly encoded. Second, in PET images, there are cross-section regions and background regions and the authors are interested with the cross-section regions only, the lossless coding algorithm is applied for the cross-section regions (or regions of interest). Third, the background regions (unsupported regions) can be filled-in by some values generated by the smooth filling or cosine filling algorithm to reduce the discontinuities on the boundaries thus increase the lossless compression ration. The authors have tried the hierarchical interpolation (HINT) coding, the 2-D linear predictive coding, and the arbitrary-shaped transform (AST) coding of the cross-section and the chain coding of the contour of the circular-shaped cross-section. The prediction error, the chain-code, and the quantization error are coded through the entropy coding (Huffman coding and the Lampel-Ziv coding) to reduce the code bits. In the experiments, the authors find that (1) the AST coding has the best performance than the other two, (2) the uniform quantizer is better than the non-uniform quantizer, (3) Lempel-Ziv code is better than Huffman code, and (4) the 3-D coding is not better than its 2-D counter part simply because the interslice correlation of a multi-slice PET medical data set generally is much lower than the intraslice correlation. The compression ratio of our segmentation-based lossless coding of the PET imaging is more than one-third of the original data.[[fileno]]2030109010040[[department]]電機工程學