Room: AAPM ePoster Library
Purpose: MR images generated from CT allow for preserving advantages of both modalities: higher soft-tissue contrast of MR, shorter scanning time and less restrictions for people of CT. In this work, a novel method for generating pseudo MR from CT using super resolution cycle generative adversarial networks (SRCycle GAN) is proposed to assist the diagnosis of soft-tissue sarcoma.
Methods: GAN consists of 2 generator networks and 3 discriminator networks. For the generator, an upsampling layer was used to avoid the checkboard artifacts. Apart from 2 discriminators from original cycle GAN, a super resolution discriminator was added to acquire clearer images. The network was optimized via cycle-consistency loss and gradient loss and mean P distance loss. The GAN loss is shifted from mean squared error (MSE) loss to mean absolute error (MAE) loss. A total of 4216 unaligned images including CT and T1 MRI of 30 patients in the soft-tissue sarcoma dataset from The Cancer Imaging Archive (TCIA) were obtained after preprocessing and were then split into the training dataset (2911 images, 20 patients) and testing dataset (1305 images, 10 patients). For the verification, 4 experienced radiologists were invited to diagnose all patients in the test dataset basing on their CT and real MR, pseudo MR generated by proposed model.
Results: The average accuracy of all radiologists using CT and real MR, pseudo MR generated by proposed network was found to be: 95%, 85%, suggesting that the quality of pseudo MR from SRCycle GAN is very close to that of the real MR.
Conclusion: novel method for generation of pseudo MR from CT has been developed that produces clearer images than the original cycle GAN, thus ready for clinical applications.
IM/TH- Image Analysis (Single Modality or Multi-Modality): Image processing