Computed tomography diagnosis utilizing compressed image data: an ROC analysis using acute appendicitis as a model

Abstract

Using receiver-operating characteristic (ROC) methodology, the ability to diagnose acute appendicitis with computed tomography (CT) images displayed at varying levels of lossy compression was evaluated. Nine sequential images over the ileocecal region were obtained from 53 consecutive patients with right lower quadrant pain who were clinically suspected to have acute appendicitis. Thirty were proven surgically to have acute appendicitis, alternative diagnoses confirmed in 23. The image sets were subjected to a lossy wavelet-based compression algorithm "Embedded Predictive Wavelet Image Coder" (EPWIC). Compression levels were: none, 8:1, 16:1, and 24:1, resulting in 4 sets of images per patient. Image sets were randomized and evaluated separately by 4 body radiologists on a 1,024 x 768-pixel SVGA color PC monitor in 512 x 512 format. The readers were aware of the clinical suspicion of appendicitis but were unaware of the positive fraction of cases. Individual and combined reader ROC and c2 analyses of sensitivity, specificity, and accuracy were determined. For all readers, sensitivity decreases at 16:1 and 24:1 levels (P <0.01, P <0.001, respectively). Accuracy decreased at 24:1 levels (P <0.01). Specificity was unaffected. By ROC analysis there was statistically significantly decreased area under the curve at 24:1 levels (P <0.02) as compared with uncompressed images. Finite levels of lossy wavelet compression may be applied to CT images without compromising diagnostic performance.