Are Deep Learning Techniques for MR Image Super-Resolution Ready for Deployment?

We address the problem of super-resolving magnetic resonance imaging (MRI) scans. The objective is to examine if the resolution of an MRI scan can be improved by image processing or state-of-the-art deep learning techniques. We consider several interpolation based approaches (bilinear, bicubic, trilinear) as well as deep-learning based approaches such as super-resolution convolutional neural networks, residual networks, and generative adversarial networks.

We also address the issue of faster sampling in the k-space using various subsampling strategies (random, uniform, or Gaussian) and then using super-resolution techniques to enhance the resolution and image quality. We also explore the possibility of improving the image quality and resolution using low-field strength scanners (1.5 T) to match that of 3 T scanners.

The metrics used for performance evaluation are peak signal-to-noise ratio (PSNR) and structural similarity index metric (SSIM). We present validations on standard datasets as well as the CBR dataset of MRI scans. Our experiments show that while the objective performance metrics of machine learning methods are superior to that of classical techniques, the machine learning methods, particularly those relying on generative modeling techniques, are also prone to introducing morphological artifacts.

This is joint work with Soham Nilesh Kolambe, Vidyotha Shetty, and Neelam Sinha.