Uct s160

Patients whose CT images taken prior to treatment were unavailable were excluded from both data sets. The included CT images were obtained with scanners manufactured by Siemens (Somatom Definition AS+, Biograph64), Philips (Brilliance 40, iCT256, Ingenuity Flex, MX 16-slice), GE Medical System (Bright Speed), and United Imaging (uCT 510, uCT 760, uCT S-160). All images were reconstructed using slice thickness of 0.6, 1, 1.25, 2, 3, or 5 mm.
Two thoracic radiologists (TTW and YY) independently reviewed all scans. Assigned marks were determined by the consensus of the third thoracic radiologists (XWS). The region of interest annotation was performed with 3D slicer (http://www.slicer.org) for each lesion.32 (link) The 3D centre-of-mass location was marked, and a bounding box was constructed to include the whole tumor.
Before training the model, we first normalized the CT images to eliminate radiographic differences between images acquired with different scanners. Afterwards, we generated new data by moving the bounding box marked by the doctors. This data augmentation operation provided new training samples for the model and helped reduce the errors introduced by the doctor’s annotation. Detailed data augmentation methods are described in online supplementary appendix 1.