A Comparison of Four Versions of a Computer-aided Detection System for Pulmonary Nodules on Chest Radiographs

Abstract

There are few computer-aided detection (CAD) systems that are available for the detection of nodules on chest radiographs. We evaluated the performance of a Food and Drug Administration-approved system and 3 of its subsequent versions to determine their potential to improve readers' accuracy. We tested the performance of 4 generations of CAD software programs, RapidScreen 1.1 and OnGuard 3.0, 4.0, and 5.0 (Riverain Medical), for their ability to detect lung cancer on a sample of 100 patients with and 100 patients without nodules. Each proven nodule (computed tomography scan and/or pathology) was evaluated for its overall difficulty, size, density, shape, contour, and location. The sensitivity and number of false-positive (FP) marks were compared between the different versions; reasons for FP and false-negative marks were analyzed and compared. The newer versions have significantly improved overall sensitivity [62.5% (OnGuard 3.0), 62.5% (4.0), and 64.4% (5.0)] compared with the first version (44.2%). OnGuard 5.0 demonstrated sensitivity of 73.3% for moderately subtle lesions compared with very subtle lesions (20.0%). There was a significant reduction in the average number of FPs per image for each version (3.9 for 1.1, 3.3 for 3.0, 2.6 for 4.0, and 2.0 for 5.0). Rib and vessel crossings were the most common reasons for FPs. The latest version of CAD demonstrates good detection of moderately subtle lesions with a relatively low FP rate.