Yasuji Ryu MD, Presenter: Nothing to Disclose

Janne Johannes Nappi PhD, Abstract Co-Author: Royalties, Hologic, Inc Royalties, MEDIAN Technologies

Yin Wu, Abstract Co-Author: Nothing to Disclose

Wenli Cai PhD, Abstract Co-Author: Nothing to Disclose

Michael Ethan Zalis MD, Abstract Co-Author: Co-founder, QPID Health Inc Chief Medical Officer, QPID Health Inc Stockholder, QPID Health Inc

Hiroyuki Yoshida PhD, Abstract Co-Author: Patent holder, Hologic, Inc Patent holder, MEDIAN Technologies

Computer-aided detection (CADe) could be useful for effective laxative-free CT colonography (lfCTC) examination; however, its effect on reader performance has not been evaluated on asymptomatic patients. Our purpose was to evaluate the effect of a CADe scheme optimized for lfCTC on human readers’ detection performance for adenomas and carcinomas in a screening population.

A total of 300 cases were randomly sampled from a multi-center lfCTC trial for asymptomatic patients, in which 605 participants had lfCTC examination at 4 medical centers with dietary fecal tagging by non-ionic iodine. The precise spatial locations of lesions confirmed by segmentally unblinded colonoscopy were identified prospectively and retrospectively on the lfCTC images. A fully automated CADe scheme, which had been trained to detect polyps using an independent set of 204 CTC cases with 263 lesions, reviewed the 605 lfCTC cases for significant lesions. The 300 sampled lfCTC cases were subjected to electronic cleansing and used for the observer study, in which the case reading order was designed to distribute the positive cases evenly between quartiles. An expert reader (≥600 cases reading experience) reviewed these cases and recorded all detected lesions ≥6 mm using primary 3D interpretation without and with optimized CADe assistance. Per-patient sensitivities and the areas under the receiver operating curve (AUC) in the detection of adenomas and carcinomas were compared between the two reading modes.

There were 24, 18, and 11 patients with adenomas or carcinomas ≥6 mm, ≥8 mm, and ≥10 mm, respectively. In these size ranges, per-patient sensitivities for unassisted readings were 38%, 44%, and 64%, respectively, whereas those of CADe-assisted readings were 63% (P < 0.05), 72% (P < 0.05), and 91%, respectively. The corresponding AUCs for unassisted readings were .63, .70, and .80, respectively, whereas those of CADe-assisted readings were .76 (P < 0.01), .84 (P < 0.01), and .90 (P < 0.05), respectively.

The use of CADe optimized for lfCTC substantially increased the performance of human readers in the detection of adenomas and carcinomas in a screening population.

The use of CADe can substantially increase the performance of human readers in the detection of adenomas and carcinomas on asymptomatic patients, thus the use of CADe is recommended for colon cancer screening with lfCTC.

Ryu, Y, Nappi, J, Wu, Y, Cai, W, Zalis, M, Yoshida, H, Effect of Computer-aided Detection on Reader Performance in Laxative-free CT Colonography: Observer Performance Study.  Radiological Society of North America 2014 Scientific Assembly and Annual Meeting, - ,Chicago IL. http://archive.rsna.org/2014/14015248.html