Super-Resolution Musculoskeletal MRI Using Deep Learning
In this manuscript, we have demonstrated a method termed ‘DeepResolve’, which can transform low-resolution magnetic resonance images (MRI) into higher-resolution images. In MRI high-resolution images are beneficial in order to better delineate anatomical detail, however, the acquisition of such high-resolution data is time consuming and uncomfortable for patients. To overcome this inefficiency, we trained a convolutional neural network in order to learn features between low and high-resolution representation of the same images in order to teach the network to enhance the quality of arbitrary low-resolution images. Specifically, we showed that DeepResolve was able to outperform (using quantitative image quality metrics and a qualitative radiologist reader study) commonly used interpolation methods for enhancing the through-plane resolution for a variety of downsampling factors.
Chaudhari AS, Fang Z, Kogan F, Wood J, Stevens KJ, Gibbons EK, Lee JH, Gold GE, Hargreaves BA. Super-resolution musculoskeletal MRI using deep learning. Magn Reson Med. 2018 Mar 26. doi: 10.1002/mrm.27178.
In the above image, DeepResolve enhanced the tricubic interpolation image. Compared to the ground-truth, the interpolated image had 3x lower resolution in the left-right direction. The medial collateral ligament (solid arrow), an osteophyte (dashed arrow) and small blood vessels (dotted arrow) had better delineation in the DeepResolve image than the tricubic interpolation image. The DeepResolve images were comparable to the original ground-truth images.
