Having just published DICOM Shrinkinator, our DICOM compression gateway, we have had a few discussions with clients regarding appropriate JPEG compression levels for DICOM studies. As it happens, the Canadian Association of Radiologists published guidelines on this topic in 2008.
The CAR express recommendations as ratios, which makes them difficult to implement given the parameters of a typical compression algorithm. There is also the additional concern about what consistency in image quality results when the ratio is held constant across varying demographics, logos, and other clutter around the actual image data.
Despite these concerns, we post their recommendations for quick reference for where to begin evaluation of compression results.
C.A.R. Guidelines for DICOM Compression
Source: CAR Standards for Irreversible Compression in Digital Diagnostic Imaging within Radiology
There are some surprising aspects of this table, in that MR is apparently amenable to a high degree of lossy compression. This seems counter-intuitive as resonance has both a small voxel size and relies heavily on small differences in contrast. Topic for another post.
[HT] to Innowave Healthcare Private ltd., India
Edit: Paweł, our DICOM guru has also pointed out that color will produce a wildly different ratio from Grayscale for the same perceived quality attribute. Again, this table might be best viewed as a series predictors of acceptable outcomes.