Purpose: To investigate the effective dose (E) and convolution kernel's effects on the detection of pulmonary nodules in different artificial intelligence (AI) software systems.
Methods: Simulated nodules of various sizes and densities in the Lungman phantom were CT scanned at different levels of E (3 - 5, 1 - 3, 0.5 - 1, and <0.5 mSv) and were reconstructed with different kernels (B30f, B60f, and B80f). The number of nodules and corresponding volumes in different images were detected by four AI software systems (A, B, C, and D). Sensitivity, false positives (FPs), false negatives (FNs), and relative volume error (RVE) were calculated and compared to the aspects of the E and convolution kernel.
Results: System B had the highest median sensitivity (100 %). The median FPs of systems B (1) and D (1) was lower than A (11.5) and C (5). System D had the smallest RVE (13.12 %). When the E was <0.5 mSv, system D's sensitivity decreased, while the FPs and FNs of systems A and B increased significantly (P < 0.05). When the kernel was changed from B80f to B30f, the FPs of system A decreased, while that of system C increased, and the RVE of systems A, B, and C increased (P < 0.05).
Conclusion: AI software systems B and D have high detection efficiency under normal or low dose conditions and show better stability. However, the detection efficiency of systems A and C would be affected by the E or convolution kernel, but the E would not affect the volume measurement of four systems.
Keywords: Artificial intelligence; Chest CT; Computer-assisted detection; Deep learning; Pulmonary nodule.
Copyright © 2020 Elsevier B.V. All rights reserved.