Combining conditional GAN with VGG perceptual loss for bones CT image reconstruction

International audienceReducing both the radiation dose to patients and the reconstruction time is key for X-ray computed tomography. The imaging of bone microarchitecture at high spatial resolution is all the more challenging as noisy data can severely deteriorate structural details. Deep Learning based algorithms are efficient for post-processing poorquality reconstructions obtained with Filtered BackProjection, though MSE-trained networks hardly capture the structural information relevant for bones. Instead, conditional GANs allow to generate very realistic volumes that correspond to their corrupted FBP. Moreover, perceptual losses are efficient to capture key features for the human eye. In this work we combine both concepts within a new framework called CWGAN-VGG that is designed for the reconstruction of bones at high spatial resolution, with an emphasis put on the preservation of their structural information. We show on simulated low-dose CT bones data that our CWGAN-VGG outperforms state-of-the-art methods that involve GANs and/or perceptual losses in terms of PSNR and other metrics