Image compression and chest radiograph interpretation: image perception comparison between uncompressed chest...

We have assessed the effect of 10:1 lossy (JPEG) compression on six board-certified radiologists' ability to detect three commonly seen abnormalities on chest radiographs. The study radiographs included 150 chest radiographs with one of four diagnoses: normal (n = 101), pulmonary nodule (n = 19), interstitial lung disease (n = 19), and pneumothorax (n = 11). Before compression, these images were printed on laser film and interpreted in a blinded fashion by six radiologists. Following an 8-week interval, the images were reinterpreted on an image display workstation after undergoing 10:1 lossy compression. The results for the compressed images were compared with those of the uncompressed images using receiver operating characteristic (ROC) analyses. For five of six readers, the diagnostic accuracy was higher for the uncompressed images than for the compressed images, but the difference was not significant (P > .1111). Combined readings for the uncompressed images were also more accurate when compared with the compressed images, but this difference was also not significant (P = .1430). The sensitivity, specificity, and accuracy values were 81.5%, 89.2%, and 86.7% for the compressed images, respectively, as compared with 78.9%, 94.5%, and 89.3% for the uncompressed images. There was no correlation between the readers' accuracy and their experience with soft-copy interpretation; the extent of radiographic interpretation experience had no correlation with overall interpretation accuracy. In conclusion, five of six radiologists had a higher diagnostic accuracy when interpreting uncompressed chest radiographs versus the same images modified by 10:1 lossy compression, but this difference was not statistically significant.