Objectives: To evaluate the performance of a commercially available CAD system for automated detection and measurement of subsolid nodules.
Materials and methods: The CAD system was tested on 50 pure ground-glass and 50 part-solid nodules (median diameter: 17mm) previously found on standard-dose CT scans in 100 different patients. True nodule detection and the total number of CAD marks were evaluated at different sensitivity settings. The influence of nodule and CT acquisition characteristics was analyzed with logistic regression. Software and manually measured diameters were compared with Spearman and Bland-Altman methods.
Results: With sensitivity adjusted for 3-mm nodule detection, 50/100 (50%) subsolid nodules were detected, at the average cost of 17 CAD marks per CT. These figures were respectively 26/100 (26%) and 2 at the 5-mm setting. At the highest sensitivity setting (2-mm nodule detection), the average number of CAD marks per CT was 41 but the nodule detection rate only increased to 54%. Part-solid nodules were better detected than pure ground glass nodules: 36/50 (72%) versus 14/50 (28%) at the 3-mm setting (p<0.0001), with no influence of the solid component size. Except for the type (i.e. part solid or pure ground glass), no other nodule characteristic influenced the detection rate. High-quality segmentation was obtained for 79 nodules, which for automated measurements correlated well with manual measurements (rho=0.90[0.84-0.93]). All part-solid nodules had software-measured attenuation values above -671Hounsfield units (HU).
Conclusion: The detection rate of subsolid nodules by this CAD system was insufficient, but high-quality segmentation was obtained in 79% of cases, allowing automated measurement of size and attenuation.
Keywords: Computer-aided diagnosis; Image interpretation; Lung neoplasm; Solitary pulmonary nodule.
Copyright © 2016 Elsevier Ireland Ltd. All rights reserved.