PROJECT TITLE :

Deep MR Brain Image Super-Resolution Using Spatio-Structural Priors

ABSTRACT:

Accurate diagnosis require high-resolution Magnetic Resonance images (MR). Factors such as hardware and processing limitations limit image resolution in practise. Deep learning algorithms have recently been shown to yield compelling state-of-the-art outcomes in picture enhancement/super-resolution. In order to achieve the desired high-resolution MR image structure and to take advantage of image priors, such as a low-rank structure and a sharpness prior, we suggest a new regularised network (SR). Our contributions then include these priors in an analytically tractable manner, as well as towards a unique prior-guided network architecture that fulfils the super-resolution job. We overcome this problem by pursuing differentiable approximations of the rank for the low rank prior, as the rank is not a differentiable function of the image matrix (and thus the network parameters). When the Laplacian variance is handled by a fixed feedback layer at the network output, we stress its sharpness. Key to this extension is the development of new training data-driven filters for the Laplacian (Fixed Feedback) layer. SNR and image quality measurements can be enhanced significantly by implementing the suggested prior guided network, according to experiments on publically available MR brain image databases and comparisons with current state-of-the-art approaches. Using output images as priors, our method is flexible and can be used with a wide range of current network designs to improve their performance.