Purpose: Current super-resolution (SR) methods suffer from poor performance on real-world MR images due to the simple translation model (e.g., Gaussian blur) used when preparing low-resolution (LR) images for training. We propose a deep learning-based (DL-based) method that generates more realistic LR images and use them to train our SR network to achieve better performance on real-world LR MR images.
Methods: A two-step workflow was used: (a) dataset preparation via the proposed method. (b) SR network training. In the first step, a Generative Adversarial Network (GAN) was constructed to simulate the downsampling process. The generator of the proposed GAN is comprised of six convolutional layers without non-linear activation layer to simulate the low-pass filtration. The receptive field of the discriminator is 7-by-7. During the adversarial training, the generator output fake LR MR images that mimicked the real LR MR images on patch-level. Matched LR and HR MRI pairs were not needed since the receptive field of discriminator is small. An Enhanced Deep Super-Resolution (EDSR) network was used for SR. A total of 2624 axial HR and LR images from 20 patients scanned with TWIST-VIBE sequence were used for training. Two experiments were conducted using different methods to prepare LR MR images: proposed GAN and conventional Gaussian filtration. Real LR images in this dataset were used to evaluate the SR network.
Results: Severe artifacts were observed with the SR network that was trained with Gaussian blurred LR images. Better resolution enhancement was observed with the SR network that was trained with the data prepared by the proposed adversarial training method. This visual difference was verified by SSIM and PSNR. The SSIM improved by 0.111±0.016 and the PSNR improved by 2.76±0.98 dB.
Conclusion: The proposed method outperformed the Gaussian blur method by demonstrating better resolution enhancement and fewer artifacts on real-world MRI.
Funding Support, Disclosures, and Conflict of Interest: GRF 151021/18M, HMRF 06173276, NIH R01 CA226899