In Review - News from the 2004 Meeting of the RSNA

New model attempts to better define image quality

'Just noticeable differences' standard comes closer to how human visual system actually operates

By: Merlina Trevino

Because it is so closely correlated with human observer perception, a model based on just noticeable differences is better able to predict image quality than traditional quantitative measures, according to researchers at the VA Maryland Health Care System.

Traditionally accepted image quality metrics such as mean square error and peak signal-to-noise ratios (PSNR) are quantitative measures, but humans rely on more subjective measures such as overall image quality and lesion detectibility, said Dr. Bruce Reiner, director of radiology at the Maryland VA.

Researchers have generally relied on receiver operating characteristic studies to quantitatively determine image quality, he said.

The just noticeable differences (JND) values are better predictors of image quality because they can simulate the physiology of the human visual system, said Reiner, who presented study findings on behalf of lead researcher Dr. Khan M. Siddiqui.

To test the model, 11 radiologists examined 80 CT and computed radiography images on 3-megapixel LCD monitors. The images had undergone a range of JPEG compression from lossless up to 60:1.

Normalized reader scores were extremely highly correlated to the JND values for both CT (-0.9) and CR (-0.91). PSNR did not correlate as well to human observations, with values of 0.78 for CT and 0.63 for CR. The differences between PSNR and JND correlation to normalized reader scores were statistically significant, Reiner said.

In an earlier study by the same group, presented at the May 2004 Society for Computer Applications in Radiology meeting, the investigators used the JND system to compare the current 2D compression standard with the developmental 3D JPEG2000 standard.

They examined five thoracic CT data sets from a 16-detector scanner, using 0.75-mm collimation. Slices were reconstructed into thicknesses of 0.75 mm, 1.5 mm, 3 mm, 6 mm, and 10 mm. Compression levels ranged from 4:1 to 64:1 for both the 2D and 3D compression modes. The group measured the outcomes using PSNR and the Sarnoff's JNDmetrix visual discrimination model.

"We were trying to figure out what level of compression one could apply and still have the images remain visually lossless," said Dr. Eliot Siegel, chief of radiology and nuclear medicine at the VA Maryland Health Care System.

The thinnest slices proved to be the least compressible using the 2D compression standard. The effect was especially marked at higher compression levels. The 3D compression method resulted in images of much higher quality.

The JND model simplifies determination of acceptable image compression levels and can help decide whether reduced radiation doses will yield acceptable images, Reiner said.