Steve Halligan MD, Presenter: Nothing to Disclose

Jamshid Dehmeshki, Abstract Co-Author: Nothing to Disclose

Stuart Andrew Taylor MD, Abstract Co-Author: Nothing to Disclose

Hamdan Amin, Abstract Co-Author: Nothing to Disclose

Ye Xujiong, Abstract Co-Author: Nothing to Disclose

Julian Tsang, Abstract Co-Author: Nothing to Disclose

Mary Roddie MD, Abstract Co-Author: Nothing to Disclose

et al, Abstract Co-Author: Nothing to Disclose

To validate a computer-assisted reader (CAR) system for CT colonography, using an external validation paradigm, i.e. by performance analysis of data obtained from a single center uninvolved with the software development.

25 multi-detector CT colonography examinations of patients with colonoscopically validated polyps that had been accumulated at a single center were examined by two expert radiologist observers who used endoscopic and histopathological data to identify polyp coordinates. A CAR system (MedicColon 1.2, Medicsight PLC) that been developed using polyp data obtained from elsewhere, and which had not previously encountered the present data, was then applied to the 25 datasets at sphericity filter settings of 0.75 and 0.50, and used to identify potential polyps. Statistical advice was sought and CAR prompts categorised as true-positive, false-negative, and false-positive by the expert observers with reference to the previously documented polyp coordinates. An estimate of whether false-positive prompts were easily dismissed or required 3D rendering was also made.

25 patients had 57 polyps, median size 6mm (range 1 to 15mm). Per-patient sensitivity for CAR was 96% (24 of 25 patients had at least one polyp correctly detected by computer-assist). CAR correctly detected 43 (75.4%) polyps at a sphericity filter setting 0.75 and 48 (84.2%) polyps at a sphericity filter setting of 0.50: the additional 5 polyps detected all measured 5mm or less. A sphericity setting of 0.75 resulted in a median of 7.5 (range 1 to 17) easily dismissed false-positive prompts per patient and a median of 3 (range 0 to 6) that needed 3-dimensional rendering by the observer before dismissal. These figures rose to 40 (range 17 to 84) and 5 (range 1 to 13) respectively at a sphericity filter setting of 0.5.

We found 96% per-patient sensitivity for CAR using external validation, a more stringent test than either temporal or internal cross-validation (both of which use data obtained from the same source as that used to develop the software). Decreasing sphericity filter values increases sensitivity for small polyps but at the expense of increased false-positive prompts.

J.D.,H.A.,Y.X.,J.T.,M.R.: Authors are employed by Medicsight PLCS.H.,S.A.T.: Authors provide renumerated research and development advice to medicsight PLC

Halligan, S, Dehmeshki, J, Taylor, S, Amin, H, Xujiong, Y, Tsang, J, Roddie, M, et al, , External Validation of Computer-assisted Reading for CT Colonography.  Radiological Society of North America 2005 Scientific Assembly and Annual Meeting, November 27 - December 2, 2005 ,Chicago IL. http://archive.rsna.org/2005/4405655.html