An anthropomorphic phantom study of computer-aided detection performance for polyp detection on CT colonography: a...

The objective of our study was to compare the diagnostic performances of two commercial computer-aided detection (CAD) systems and a CAD system developed in our laboratory, which we refer to as an "academic CAD system," for polyp detection on CT colonography (CTC) and to assess the detection characteristics of the CAD systems.

One hundred three polyps (48 polyps < 6 mm and 55 polyps > or = 6 mm; 45 sessile, 33 flat, and 25 pedunculated polyps) were created. Each CTC data set was analyzed using two commercial CAD systems (Computer Assisted Reader [CAR] and Polyp Enhanced View [PEV]) and one Hessian matrix-based academic CAD system. Per-polyp sensitivities according to polyp size and shape were compared among the three CAD systems. The average number and causes of false-positives (FPs) were analyzed and compared.

Per-polyp sensitivity for all polyps was significantly better for the academic CAD system (83.5%) than for both commercial CAD systems (64.1%) (p < 0.01). However, the difference in per-polyp sensitivity for polyps > or = 6 mm was not significant (p > 0.017). According to morphology, per-polyp sensitivities as determined with the CAR, PEV, and academic CAD systems for flat, sessile, and pedunculated polyps were 51.5%, 57.6%, and 81.8%; 60.0%, 62.2%, and 84.4%; 88.0%, 76.0%, and 84.0%, respectively. The average number of FPs was not significantly different (p > 0.05); however, the distribution of the causes of FPs for the three systems was significantly different (p < 0.001).

For polyps > or = 6 mm, the three CAD systems showed comparable per-polyp sensitivities. Although the number of FPs was not significantly different, the distribution of the causes of FPs for each of the CAD systems was significantly different.