Compression changes notable in images may not affect diagnosis

May 15, 2008

Changes in images introduced by compression algorithms at levels as low as 8:1 can be observed by readers when they are compared with uncompressed images, according to a research report presented at the Society of Imaging Informatics in Medicine conference Thursday.

Changes in images introduced by compression algorithms at levels as low as 8:1 can be observed by readers when they are compared with uncompressed images, according to a research report presented at the Society of Imaging Informatics in Medicine conference Thursday.

But researchers were unable to say what impact compression has on diagnosis, a question the study did not address.

The study included nearly 15,000 images at three institutions. Readers were asked to flip between compressed and uncompressed images and to say whether they noticed any difference in the two. The readers found differences between 8:1 compressed and uncompressed images 78% of the time. The figures for 12:1 compression were 95% and for 16:1 compression, 99%.

The researchers concluded that even mild compression changes images in ways that are perceptible. But the perceived differences cannot yet be linked to diagnostic performance.

The study used a "flicker" method to compare compressed and uncompressed images. Readers can scroll between images shown in rapid succession, a strategy that creates a sense of motion when the two are different, said Elizabeth Krupinski, Ph.D., one of the principal researchers. That perception of motion may cue readers to the differences.

In one test, however, even introducing a blank image between the compressed and uncompressed images did not change the results, said Krupinski, a professor of radiology at the University of Arizona.

The test looked at thin-slice (0.625 to 1 mm) images from CT scans of the chest, abdomen, and pelvis.

Interestingly, radiologists saw no difference between compressed and uncompressed images 38% of the time, Krupinski said. That figure for Ph.D. readers was 28% and for residents, 29%.

"The trend is for radiologists to be more tolerant of differences or less sensitive," Krupinksi said.

Other researchers in the study were Bradley Erickson, M.D., Ph.D., of the Mayo Clinic in Rochester, MN, and Katherine Andriole, Ph.D., of Brigham and Women's Hospital, Harvard Medical School.