. This provides population data that can be used to quantify inter-scanner variability. This data can be used to determine how robust specific radiomics or other quantitative imaging signatures are. ProtocolComputed tomography scans were acquired on 100 scanners at 35 clinics: 51 GE scanners, 20 Philips scanners, 11 Toshiba scanners, and 1 Philips and Neusoft Medical System scanner. The commonly used chest and head protocols of the local clinic were acquired without changing the protocol parameters. Additionally, a controlled protocol was acquired that was designed to minimize radiomics feature differences between manufacturers. The settings for the controlled protocol were: tube voltage, 120 kV(p); tube current, 200 mA∙s; helical scan type; spiral pitch factor, 1.0; 50-cm display field of view; and image thickness, 3 mm (except for GE scanners, which used an image thickness of 2.5 mm). The convolution kernel was standard for GE; C for Philips; B31f, B31s for Siemens; and FC08 for Toshiba. However, the kernel used for the Toshiba scans switched from FC18 (six scanners) to FC08 (five scanners) halfway through owing to a study by Mackin et al. (Medical Physics, 2018) that found the FC08 kernel to match the GE standard kernel best. K-means clustering showed that the scanners did not cluster by kernel type, thus all Toshiba scanners were included. There were 94 scanners that had a controlled protocol scan that could be used: 48 GE, 18 Philips, 17 Siemens, and 11 Toshiba scanners; 93 scanners that had a local chest protocol scan that could be used: 47 GE, 19 Philips, 17 Siemens, and 10 Toshiba scanners; and 88 scanners that had a local head protocol scan that could be used: 46 GE, 18 Philips, 14 Siemens, and 10 Toshiba scanners. The various reasons that scans could not be used were as follows: the field of view did not encompass all the cartridges, the scan extent did not cover the length of the phantom, and the scan was acquired with variable image thickness. |