Logo of radaiLink to Publisher's site
PMC full text:
Copyright/LicenseRequest permission to reuse

Figure 3:

Texture translation between radiographs from different manufacturers. A, Texture translation from the Philips DigitalDiagonost (DD) to the Siemens Fluorospot Compact FD (FCFD) with the original (left) and its corresponding fake (right) chest radiograph, shows between-manufacturer changes occurring at high spatial frequencies, with global thoracic structures hardly altered. B, Texture translation from the FCFD to the DD. fDD = fake DD image, fFCFD = fake FCFD image, nDD = native DD image, nFCFD = native FCFD image.

Texture translation between radiographs from different manufacturers. A, Texture translation from the Philips DigitalDiagonost (DD) to the Siemens Fluorospot Compact FD (FCFD) with the original (left) and its corresponding fake (right) chest radiograph, shows between-manufacturer changes occurring at high spatial frequencies, with global thoracic structures hardly altered. B, Texture translation from the FCFD to the DD. fDD = fake DD image, fFCFD = fake FCFD image, nDD = native DD image, nFCFD = native FCFD image.

Images in this article

  • Radiomic feature (RF) reproducibility analysis pipeline. A, A cycle-GAN was first trained to translate textures between manufacturers: Philips DigitalDiagnost (DD) and Siemens Fluorospot Compact FD (FCFD). B, Following texture translation, 92 RFs were extracted from lung parenchyma for each native and fake chest radiograph of an independent testing dataset. C, The intermanufacturer RF variability was compared between pairs of native and translated chest radiographs in this independent dataset computing the concordance correlation coefficient for each RF. fDD = fake DD image, fFCFD = fake FCFD image, GAN = generative adversarial network, nDD = native DD image, nFCFD = native FCFD image.
  • GAN model. A, An nDD is fed into a generator (GDDtoFCFD), which translates its texture to match the FCFD type, producing an fFCFD image, based on the discriminator feedback (DFCFD). B, The inverse translation is similarly performed on the basis of a second generator and discriminator pair (GFCFDtoDD and DDD, respectively). The two discriminators (DFCFD and DDD) are trained to identify native and fake images produced by their corresponding generators (GDDtoFCFD and GFCFDtoDD, respectively), providing quality feedback concerning the counterfeit images to their corresponding generator. The cycle-GAN network architecture is similar to the one used by Zhu et al (15), except for image input and output shape (here, 512 512 pixels). DD = Philips DigitalDiagonost, DDD = DD discriminator, DFCFD = FCFD discriminator, FCFD = Siemens Fluorospot Compact FD, fDD = fake DD image, fFCFD = fake FCFD image, GAN = generative adversarial network, GDDtoFCFD = generator translating DD to FCFD, GFCFDtoDD = generator translating FCFD to DD, nDD = native DD image, nFCFD = native FCFD image.
  • Texture translation between radiographs from different manufacturers. A, Texture translation from the Philips DigitalDiagonost (DD) to the Siemens Fluorospot Compact FD (FCFD) with the original (left) and its corresponding fake (right) chest radiograph, shows between-manufacturer changes occurring at high spatial frequencies, with global thoracic structures hardly altered. B, Texture translation from the FCFD to the DD. fDD = fake DD image, fFCFD = fake FCFD image, nDD = native DD image, nFCFD = native FCFD image.
  • Distribution of the concordance correlation coefficient between the two manufacturers before and after texture translation. fDD = fake Philips DigitalDiagonost image, fFCFD = fake Siemens Fluorospot Compact FD image, nDD = native Philips DigitalDiagonost image, nFCFD = native Siemens Fluorospot Compact FD image.
  • Concordance correlation coefficient (CCC) heatmap for all radiomic features (RFs) before and after texture translation. The heatmaps display the CCC of all RF groups by classes of RF. In each RF class, the left column (framed in orange) compared native RFs (ie, nDDs to nFCFDs). The two other columns (framed in red) compared native to translated RFs (ie, nDDs to fDDs and nFCFDs to fFCFDs). Numerical values of CCC for all RFs are available in Table E1 (supplement). DD = Philips DigitalDiagonost, FCFD = Siemens Fluorospot Compact FD, fDD = fake DD image, fFCFD = fake FCFD image, GAN = generative adversarial network, GLCM = gray-level co-occurrence matrix, GLDM = gray-level dependence matrix, GLRLM = gray-level run length matrix, GLSZM = gray-level size zone matrix, nDD = native DD image, nFCFD = native FCFD image, NGTDM = neighboring gray tone difference matrix.
Click on the image to see a larger version.