Computer-aided polyp detection on CT colonography: Comparison of three systems in a high-risk human population

Abstract

Purpose To compare the detection performances of two commercial and one academic computer-aided diagnosis (CAD) systems for polyp detection on CT colonography (CTC) in a high-risk human population and to assess their detection characteristics. Materials and methods This retrospective study had institutional review board approval, but informed consent was waived. Sixty-eight patients who were suspected of having colonic polyps and scheduled for colonoscopic polyp removal were included. After CTC was performed using a 64-row MDCT, two commercial (PEV, CAR) and one academic (Hessian matrix-based) CAD systems were applied to each CTC dataset. Colonoscopy using the segmental unblinded technique was performed as a standard of reference. Per-polyp and per-patient sensitivities were calculated and compared for each CAD system. The mean number of false-positives (FPs) and false-negatives (FNs) was computed and the causes of all FPs and FNs were analyzed. Results A total of 151 polyps in 61 patients were detected (77 polyps <6 mm, 48 6–9.9 mm, 26 ≥ 10 mm). Per-polyp sensitivity for PEV, CAR, and Hessian matrix-based CAD were: 71.6%, 78.4%, and 83.8% for polyps ≥6 mm, and 88.5%, 96.2%, and 96.2% for polyps ≥10 mm. Per-patient sensitivity for polyps ≥6 mm was 80.4%, 89.1%, and 93.5%, and 87%, 95.7%, and 95.7% for polyps ≥10 mm, respectively. Per-polyp and per-patient sensitivities were not significantly different among the three CAD systems regardless of size threshold. Mean number of FPs was 6.9 for PEV, 7.3 for CAR, and 14 for Hessian matrix-based CAD. The most common cause of FPs were feces, followed by extracolonic findings, prominent folds and ileocecal valve, and rectal tube. The distribution of the causes of FPs was significantly different among the three systems. Conclusion Sensitivity of the three CAD systems for polyp detection was comparable regardless of the polyp size threshold; however, the number of FPs was higher in the Hessian matrix-based CAD. In addition, the distribution of the causes of FPs was significantly different among the three systems.