Diagnostic compression of biomedical volumes

In this work we deal with lossy compression of biomedicai volumes. By force of circumstances, diagnostic compression is bound to a subjective judgment. However, with respect to the algorithms, there is a need to shape the coding methodology so as to highlight beyond compression three important factors: the medical data, the specific usage and the particular end-user. Biomedicai volumes may have very different characteristics which derive from imaging modality, resolution and voxel aspect ratio. Moreover, volumes are usually viewed slice by slice on a lightbox, according to different cutting direction (typically one of the three voxel axes). We will see why and how these aspects impact on the choice of the coding algorithm and on a possible extension of 2D well known algorithms to more efficient 3D versions. Crosscorrelation between reconstruction error and signal is a key aspect to keep into account; we suggest to apply a non uniform quantization to wavelet coefficients in order to reduce slice PSNR variation. Once a good "neutral" coding for a certain volume is obtained, non uniform quantization can also be made space variant in order to reach more objective quality on Volumes of Diagnostic Interest (VoDI), which in turns can determine the diagnostic quality of the entire data set.