Room: Exhibit Hall | Forum 5
Purpose: Magnetic Resonance Imaging (MRI) has been widely used for brain disease diagnosis. However, its employment is limited by the slow image acquirement speed. To allow fast and high-quality medical MRI reconstruction, deep learning (DL) base strategies have been studied recently. Generative adversarial network (GAN) has been shown efficacious in natural image super-resolution, especially in preserving the natural appearance of image. In this work, we implement GAN for the under-sampled brain MRI reconstruction.
Methods: GAN includes a generator network and a discriminator network. The function of the generator is to map latent variables to the distribution of the true high quality data in order to fool the discriminator, while the discriminator D aims to distinguish the true data from the synthesized fake data using the generator. The training of GAN is to choose the network parameters to minimax the loss function. In this work, the input of generator is the inverse Fourier transform of the under-sampled k-space data (denoted as zero-filled image). The network architecture for the generator is ResNet, while it is traditional convolution neural network (CNN) for the discriminator.
Results: The GAN model is evaluated on the IXI brain dataset, where 2400 T1-weighted 2D coronal brain slices are randomly selected from 480 patients and 16 slices from 3 patients for testing. The undersampled ratio is 4 with a Cartesian sampling pattern. A comparison between GAN and ResNet in terms of the image reconstruction quality was also conducted.
Conclusion: We are able to demonstrate the performance of the GAN-based undersample MRI reconstruction. Moreover, the reconstruction time of this DL-based algorithm super-fast, which makes it easily and promising to be integrated into clinical workflow. Future works include to refine the generator network by considering the data-consistency in k-space and/or pretrained networks.