(Wednesday, 7/17/2019) 10:30 AM - 11:00 AM
Room: Exhibit Hall | Forum 6
Purpose: To develop a generative adversarial network for image denoising and enhancement on daily low dose CTs (LDCT).
Methods: A generative adversarial network (GAN) based image denoising and enhancement method was proposed to improve daily LDCT (such as cone beam CTs) image quality through style transfer. The daily LDCTs were fed to the generator of the GAN to generate an image called gen_CT. If the input of discriminator is the planning CT, the supervised label is 1, or the label is 0 when the input is gen_CT. The generator and discriminator are trained alternately until optimal results achieved. For comparison purpose, a denoising network was trained as control method. We intentionally added noises to the CT image to obtain a noised_CT, thus a denoising network from noised_CT to CT is trained end to end. The network structure is based on U-Net. The U-Net network contains two stages: the down-sampling stage to encode the features and up-sampling stage to decode it. Different kinds of noises such as gaussian noises and impulse noises were added to make the trained network more robust. We evaluated these networks using a series of 300 CBCTs for ten head-and-neck patients who underwent radiation therapy on a Varian accelerator. For each patient, the planning CT (Siemens) and a series of daily CBCTs were acquired and fed to the above-mentioned networks.
Results: The GAN-based images showed less granulated noises but slightly blurred image. In terms of denoising performance, the GAN-based network is better than the U-Net based de-noising method.
Conclusion: Unsupervised low dose CT denoising and enhancement is a challenging task, we demonstrated the efficacy of the GAN-based image de-noising and enhancement method for improvement of daily CBCTs, which can facilitate clinical adaptive process.
Funding Support, Disclosures, and Conflict of Interest: S. Tan and B. Ye are supported in part by the National Natural Science Foundation of China, under Grant Nos. 61375018 and 61672253.
Keywords
Not Applicable / None Entered.
Taxonomy
IM- CT: Machine learning, computer vision
Contact Email
